Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-STRIX_LEAN.gguf filter=lfs diff=lfs merge=lfs -text

Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-STRIX_LEAN.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56a1eca8639ad1161a236f852063844f09d0c53dc49920291040f02f2e9e805e
+size 18744342464

SHA256SUMS

@@ -0,0 +1,6 @@
+51ad9aa42908c0a765b10ed9fdebe99663fffe3bc0e1a15cbc4e2954effe266a  Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-COHERENT.gguf
+122e62e39ed7141970ca913af5438db86fcae2bfbfd1ad93455f6a149861a4ce  Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-FAST.gguf
+56a1eca8639ad1161a236f852063844f09d0c53dc49920291040f02f2e9e805e  Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-STRIX_LEAN.gguf
+7e5afffc822c64a7b43a1223d0094fb60895a521f6c89b912409d25091bd17f8  Nex-N2.5-mini.imatrix
+9183c7ba8510fb9628edd2265a4cad8c02b3aecec3cbdf65620f47494e75836d  chat_template_enable_thinking.jinja
+a80a82977e4787a838a612316c79a6533c1b098fcac19372cc78929120162a4e  mmproj-Nex-N2.5-mini-BF16.gguf

chat_template_enable_thinking.jinja

@@ -0,0 +1,155 @@
+{%- if reasoning_effort is not defined and enable_thinking is defined %}{%- set reasoning_effort = 'high' if enable_thinking else 'none' %}{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if reasoning_effort is not defined or reasoning_effort is none %}
+        {{- '<think>' }}
+    {%- elif reasoning_effort == 'none' %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- elif reasoning_effort == 'high' %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>' }}
+    {%- endif %}
+{%- endif %}

recipe/logs/N4_kld_q102.log

@@ -0,0 +1,92 @@
+0.00.046.192 I common_init_result: fitting params to device memory ...
+0.00.046.197 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.468.601 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.26.097.875 W llama_context: n_ctx_seq (2048) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.26.161.936 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.26.456.383 I 
+0.26.456.524 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.26.593.623 I kl_divergence: computing over 40 chunks, n_ctx=2048, batch_size=2048, n_seq=1
+0.29.575.306 I kl_divergence: 2.98 seconds per pass - ETA 1.98 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1       5.9944 ±    0.4685       0.05151 ±    0.01742       0.11333 ±    0.00681    10.320 ±  0.638 %    87.390 ±  1.038 %
+   2       6.9832 ±    0.3895       0.04665 ±    0.01189       0.09838 ±    0.00400     8.897 ±  0.415 %    87.048 ±  0.743 %
+   3       7.5417 ±    0.3512       0.07003 ±    0.00990       0.10487 ±    0.00406     9.267 ±  0.393 %    87.195 ±  0.603 %
+   4       7.8954 ±    0.3289       0.08016 ±    0.00897       0.11061 ±    0.00503     9.327 ±  0.343 %    86.901 ±  0.527 %
+   5       7.6999 ±    0.2880       0.07519 ±    0.00798       0.10848 ±    0.00428     9.187 ±  0.313 %    87.097 ±  0.469 %
+   6       6.6265 ±    0.2194       0.07502 ±    0.00740       0.11000 ±    0.00421     9.941 ±  0.317 %    87.455 ±  0.423 %
+   7       6.1509 ±    0.1859       0.06918 ±    0.00711       0.11837 ±    0.00459    10.426 ±  0.309 %    87.530 ±  0.390 %
+   8       6.0395 ±    0.1692       0.06322 ±    0.00648       0.11517 ±    0.00408    10.254 ±  0.281 %    87.476 ±  0.366 %
+   9       6.3422 ±    0.1682       0.05959 ±    0.00609       0.11478 ±    0.00370    10.054 ±  0.261 %    87.216 ±  0.348 %
+  10       6.4559 ±    0.1635       0.05892 ±    0.00570       0.11065 ±    0.00337     9.793 ±  0.242 %    87.243 ±  0.330 %
+  11       6.5040 ±    0.1565       0.05755 ±    0.00536       0.10715 ±    0.00309     9.593 ±  0.227 %    87.283 ±  0.314 %
+  12       6.7574 ±    0.1569       0.05744 ±    0.00504       0.10450 ±    0.00284     9.400 ±  0.214 %    87.211 ±  0.301 %
+  13       6.8020 ±    0.1514       0.05713 ±    0.00478       0.10247 ±    0.00264     9.285 ±  0.203 %    87.232 ±  0.289 %
+  14       6.8553 ±    0.1467       0.05617 ±    0.00457       0.10045 ±    0.00247     9.123 ±  0.192 %    87.236 ±  0.279 %
+  15       6.8957 ±    0.1427       0.05582 ±    0.00438       0.09956 ±    0.00232     9.076 ±  0.184 %    87.247 ±  0.269 %
+  16       7.0571 ±    0.1415       0.05224 ±    0.00422       0.09847 ±    0.00219     8.961 ±  0.175 %    87.243 ±  0.261 %
+  17       7.0917 ±    0.1375       0.05016 ±    0.00405       0.09704 ±    0.00208     8.871 ±  0.169 %    87.281 ±  0.253 %
+  18       7.1772 ±    0.1353       0.04896 ±    0.00392       0.09635 ±    0.00198     8.799 ±  0.162 %    87.243 ±  0.246 %
+  19       7.1236 ±    0.1311       0.04836 ±    0.00378       0.09524 ±    0.00190     8.763 ±  0.159 %    87.303 ±  0.239 %
+  20       6.8769 ±    0.1227       0.05245 ±    0.00378       0.10018 ±    0.00190     9.053 ±  0.155 %    87.175 ±  0.234 %
+  21       6.8931 ±    0.1198       0.05231 ±    0.00368       0.10066 ±    0.00184     9.029 ±  0.150 %    87.120 ±  0.229 %
+  22       6.9236 ±    0.1178       0.05472 ±    0.00363       0.10218 ±    0.00181     9.063 ±  0.147 %    87.168 ±  0.223 %
+  23       6.9807 ±    0.1162       0.05553 ±    0.00354       0.10187 ±    0.00175     9.045 ±  0.143 %    87.131 ±  0.218 %
+  24       6.9769 ±    0.1134       0.05494 ±    0.00347       0.10222 ±    0.00175     9.051 ±  0.143 %    87.093 ±  0.214 %
+  25       7.0111 ±    0.1117       0.05513 ±    0.00339       0.10169 ±    0.00169     8.997 ±  0.139 %    87.015 ±  0.210 %
+  26       6.9763 ±    0.1088       0.05463 ±    0.00332       0.10165 ±    0.00165     9.014 ±  0.137 %    87.055 ±  0.206 %
+  27       7.1528 ±    0.1102       0.05528 ±    0.00326       0.10141 ±    0.00162     8.950 ±  0.134 %    87.079 ±  0.202 %
+  28       7.2365 ±    0.1098       0.05405 ±    0.00318       0.10027 ±    0.00157     8.866 ±  0.131 %    87.118 ±  0.198 %
+  29       7.2391 ±    0.1079       0.05527 ±    0.00313       0.10044 ±    0.00154     8.907 ±  0.129 %    87.077 ±  0.195 %
+  30       7.1873 ±    0.1051       0.05609 ±    0.00307       0.10003 ±    0.00150     8.887 ±  0.126 %    87.061 ±  0.192 %
+  31       7.0778 ±    0.1015       0.05524 ±    0.00300       0.09893 ±    0.00145     8.840 ±  0.123 %    87.154 ±  0.188 %
+  32       6.9728 ±    0.0983       0.05456 ±    0.00300       0.10164 ±    0.00161     8.995 ±  0.124 %    87.121 ±  0.185 %
+  33       6.9098 ±    0.0957       0.05537 ±    0.00295       0.10136 ±    0.00157     8.987 ±  0.122 %    87.153 ±  0.182 %
+  34       6.8891 ±    0.0938       0.05453 ±    0.00288       0.10042 ±    0.00153     8.943 ±  0.120 %    87.154 ±  0.179 %
+  35       6.9008 ±    0.0926       0.05442 ±    0.00283       0.09974 ±    0.00149     8.883 ±  0.118 %    87.214 ±  0.176 %
+  36       6.9190 ±    0.0917       0.05458 ±    0.00279       0.09958 ±    0.00146     8.857 ±  0.116 %    87.197 ±  0.174 %
+  37       6.8168 ±    0.0888       0.05349 ±    0.00273       0.09856 ±    0.00142     8.817 ±  0.113 %    87.242 ±  0.171 %
+  38       6.7442 ±    0.0864       0.05330 ±    0.00269       0.09804 ±    0.00139     8.824 ±  0.112 %    87.277 ±  0.169 %
+  39       6.6583 ±    0.0839       0.05282 ±    0.00265       0.09776 ±    0.00137     8.827 ±  0.110 %    87.247 ±  0.167 %
+  40       6.5617 ±    0.0812       0.05204 ±    0.00261       0.09707 ±    0.00134     8.800 ±  0.108 %    87.287 ±  0.165 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :   6.561727 ±   0.081240
+Mean PPL(base)                :   6.228979 ±   0.075322
+Cor(ln(PPL(Q)), ln(PPL(base))):  97.76%
+Mean ln(PPL(Q)/PPL(base))     :   0.052041 ±   0.002607
+Mean PPL(Q)/PPL(base)         :   1.053419 ±   0.002746
+Mean PPL(Q)-PPL(base)         :   0.332749 ±   0.017588
+
+====== KL divergence statistics ======
+Mean    KLD:   0.097071 ±   0.001342
+Maximum KLD:  12.257548
+99.9%   KLD:   3.289070
+99.0%   KLD:   0.874485
+95.0%   KLD:   0.332573
+90.0%   KLD:   0.209309
+Median  KLD:   0.043524
+10.0%   KLD:   0.000605
+ 5.0%   KLD:   0.000161
+ 1.0%   KLD:  -0.000023
+ 0.1%   KLD:  -0.000269
+Minimum KLD:  -0.000603
+
+====== Token probability statistics ======
+Mean    Δp: -0.366 ± 0.043 %
+Maximum Δp: 99.335%
+99.9%   Δp: 53.261%
+99.0%   Δp: 23.761%
+95.0%   Δp: 10.784%
+90.0%   Δp:  6.407%
+75.0%   Δp:  1.292%
+Median  Δp: -0.004%
+25.0%   Δp: -1.463%
+10.0%   Δp: -6.895%
+ 5.0%   Δp: -12.449%
+ 1.0%   Δp: -30.951%
+ 0.1%   Δp: -71.092%
+Minimum Δp: -99.867%
+RMS Δp    :  8.800 ± 0.108 %
+Same top p: 87.287 ± 0.165 %
+

recipe/logs/N4_kld_q103.log

@@ -0,0 +1,92 @@
+0.00.043.507 I common_init_result: fitting params to device memory ...
+0.00.043.512 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.547.415 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.28.910.009 W llama_context: n_ctx_seq (2048) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.28.974.230 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.29.190.163 I 
+0.29.190.272 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.29.311.893 I kl_divergence: computing over 40 chunks, n_ctx=2048, batch_size=2048, n_seq=1
+0.31.982.391 I kl_divergence: 2.67 seconds per pass - ETA 1.77 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1       6.3324 ±    0.5080       0.10635 ±    0.02072       0.14199 ±    0.00911    10.822 ±  0.626 %    85.728 ±  1.094 %
+   2       7.2445 ±    0.4085       0.08338 ±    0.01320       0.11857 ±    0.00596     9.324 ±  0.416 %    86.510 ±  0.755 %
+   3       7.6437 ±    0.3567       0.08347 ±    0.01046       0.11993 ±    0.00567     9.224 ±  0.356 %    85.956 ±  0.627 %
+   4       7.8662 ±    0.3254       0.07645 ±    0.00919       0.12254 ±    0.00560     9.291 ±  0.320 %    85.997 ±  0.543 %
+   5       7.6438 ±    0.2844       0.06788 ±    0.00816       0.11930 ±    0.00461     9.150 ±  0.277 %    86.041 ±  0.485 %
+   6       6.5928 ±    0.2169       0.06992 ±    0.00760       0.12081 ±    0.00444     9.877 ±  0.294 %    86.298 ±  0.439 %
+   7       6.0938 ±    0.1829       0.05985 ±    0.00717       0.12740 ±    0.00482    10.188 ±  0.281 %    86.468 ±  0.404 %
+   8       6.0037 ±    0.1675       0.05727 ±    0.00661       0.12543 ±    0.00429    10.065 ±  0.258 %    86.522 ±  0.377 %
+   9       6.3212 ±    0.1671       0.05627 ±    0.00620       0.12516 ±    0.00387     9.872 ±  0.238 %    86.282 ±  0.359 %
+  10       6.4505 ±    0.1629       0.05808 ±    0.00580       0.12152 ±    0.00352     9.739 ±  0.221 %    86.354 ±  0.339 %
+  11       6.5075 ±    0.1561       0.05809 ±    0.00543       0.11797 ±    0.00323     9.554 ±  0.208 %    86.404 ±  0.323 %
+  12       6.7603 ±    0.1566       0.05787 ±    0.00509       0.11420 ±    0.00297     9.348 ±  0.196 %    86.355 ±  0.310 %
+  13       6.7920 ±    0.1508       0.05567 ±    0.00486       0.11244 ±    0.00278     9.245 ±  0.186 %    86.397 ±  0.297 %
+  14       6.8545 ±    0.1464       0.05605 ±    0.00465       0.11100 ±    0.00260     9.142 ±  0.177 %    86.385 ±  0.287 %
+  15       6.8911 ±    0.1423       0.05516 ±    0.00446       0.11053 ±    0.00247     9.124 ±  0.169 %    86.354 ±  0.277 %
+  16       7.0572 ±    0.1413       0.05225 ±    0.00431       0.10982 ±    0.00235     9.046 ±  0.163 %    86.254 ±  0.269 %
+  17       7.0949 ±    0.1372       0.05061 ±    0.00415       0.10844 ±    0.00223     8.972 ±  0.157 %    86.263 ±  0.261 %
+  18       7.1844 ±    0.1351       0.04995 ±    0.00401       0.10792 ±    0.00213     8.944 ±  0.152 %    86.168 ±  0.254 %
+  19       7.1415 ±    0.1312       0.05088 ±    0.00389       0.10652 ±    0.00203     8.872 ±  0.147 %    86.191 ±  0.247 %
+  20       6.8778 ±    0.1223       0.05258 ±    0.00387       0.11191 ±    0.00202     9.238 ±  0.145 %    86.144 ±  0.242 %
+  21       6.8965 ±    0.1195       0.05280 ±    0.00379       0.11255 ±    0.00196     9.249 ±  0.143 %    86.110 ±  0.236 %
+  22       6.9128 ±    0.1172       0.05316 ±    0.00370       0.11332 ±    0.00190     9.279 ±  0.140 %    86.150 ±  0.230 %
+  23       6.9751 ±    0.1157       0.05473 ±    0.00362       0.11327 ±    0.00185     9.276 ±  0.138 %    86.124 ±  0.225 %
+  24       6.9742 ±    0.1131       0.05456 ±    0.00353       0.11363 ±    0.00182     9.287 ±  0.137 %    86.091 ±  0.221 %
+  25       7.0117 ±    0.1115       0.05521 ±    0.00346       0.11333 ±    0.00176     9.261 ±  0.133 %    86.088 ±  0.216 %
+  26       6.9819 ±    0.1087       0.05543 ±    0.00339       0.11378 ±    0.00175     9.294 ±  0.131 %    86.127 ±  0.212 %
+  27       7.1532 ±    0.1100       0.05533 ±    0.00332       0.11307 ±    0.00170     9.223 ±  0.128 %    86.148 ±  0.208 %
+  28       7.2405 ±    0.1097       0.05461 ±    0.00324       0.11201 ±    0.00165     9.147 ±  0.126 %    86.193 ±  0.204 %
+  29       7.2447 ±    0.1078       0.05604 ±    0.00319       0.11216 ±    0.00161     9.196 ±  0.125 %    86.133 ±  0.201 %
+  30       7.1827 ±    0.1049       0.05545 ±    0.00314       0.11174 ±    0.00157     9.180 ±  0.122 %    86.103 ±  0.197 %
+  31       7.0787 ±    0.1014       0.05537 ±    0.00308       0.11105 ±    0.00153     9.164 ±  0.119 %    86.201 ±  0.194 %
+  32       6.9666 ±    0.0980       0.05367 ±    0.00307       0.11404 ±    0.00170     9.317 ±  0.121 %    86.150 ±  0.191 %
+  33       6.9006 ±    0.0953       0.05404 ±    0.00303       0.11358 ±    0.00167     9.303 ±  0.119 %    86.167 ±  0.188 %
+  34       6.8792 ±    0.0934       0.05310 ±    0.00297       0.11250 ±    0.00162     9.246 ±  0.117 %    86.180 ±  0.185 %
+  35       6.8869 ±    0.0922       0.05240 ±    0.00291       0.11158 ±    0.00158     9.198 ±  0.116 %    86.242 ±  0.182 %
+  36       6.9043 ±    0.0912       0.05245 ±    0.00287       0.11125 ±    0.00154     9.156 ±  0.113 %    86.239 ±  0.180 %
+  37       6.8080 ±    0.0885       0.05220 ±    0.00282       0.11015 ±    0.00151     9.118 ±  0.111 %    86.304 ±  0.177 %
+  38       6.7280 ±    0.0860       0.05090 ±    0.00278       0.10979 ±    0.00147     9.111 ±  0.109 %    86.343 ±  0.174 %
+  39       6.6466 ±    0.0835       0.05105 ±    0.00274       0.10941 ±    0.00145     9.111 ±  0.107 %    86.370 ±  0.172 %
+  40       6.5498 ±    0.0809       0.05023 ±    0.00270       0.10878 ±    0.00142     9.108 ±  0.106 %    86.405 ±  0.169 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :   6.549829 ±   0.080919
+Mean PPL(base)                :   6.228979 ±   0.075322
+Cor(ln(PPL(Q)), ln(PPL(base))):  97.58%
+Mean ln(PPL(Q)/PPL(base))     :   0.050227 ±   0.002704
+Mean PPL(Q)/PPL(base)         :   1.051509 ±   0.002843
+Mean PPL(Q)-PPL(base)         :   0.320850 ±   0.018076
+
+====== KL divergence statistics ======
+Mean    KLD:   0.108779 ±   0.001420
+Maximum KLD:  12.086925
+99.9%   KLD:   3.352409
+99.0%   KLD:   0.956811
+95.0%   KLD:   0.370801
+90.0%   KLD:   0.235309
+Median  KLD:   0.049730
+10.0%   KLD:   0.000721
+ 5.0%   KLD:   0.000197
+ 1.0%   KLD:  -0.000038
+ 0.1%   KLD:  -0.000343
+Minimum KLD:  -0.000800
+
+====== Token probability statistics ======
+Mean    Δp: -0.515 ± 0.045 %
+Maximum Δp: 98.721%
+99.9%   Δp: 52.916%
+99.0%   Δp: 24.226%
+95.0%   Δp: 11.222%
+90.0%   Δp:  6.613%
+75.0%   Δp:  1.299%
+Median  Δp: -0.007%
+25.0%   Δp: -1.668%
+10.0%   Δp: -7.814%
+ 5.0%   Δp: -13.545%
+ 1.0%   Δp: -32.559%
+ 0.1%   Δp: -70.472%
+Minimum Δp: -99.855%
+RMS Δp    :  9.108 ± 0.106 %
+Same top p: 86.405 ± 0.169 %
+

recipe/logs/N4v_kld_q102.log

@@ -0,0 +1,93 @@
+0.00.044.057 I common_init_result: fitting params to device memory ...
+0.00.044.061 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)
+0.00.458.402 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.01.547.350 W read_raw_unsafe: Falling back to buffered IO due to Bad address
+0.29.871.824 W llama_context: n_ctx_seq (2048) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.29.921.987 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.30.136.871 I 
+0.30.137.012 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.30.272.958 I kl_divergence: computing over 40 chunks, n_ctx=2048, batch_size=2048, n_seq=1
+0.33.203.394 I kl_divergence: 2.93 seconds per pass - ETA 1.95 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1       5.9966 ±    0.4697       0.05187 ±    0.01811       0.11459 ±    0.00733    10.192 ±  0.639 %    86.217 ±  1.078 %
+   2       6.9326 ±    0.3865       0.03938 ±    0.01202       0.09983 ±    0.00427     8.906 ±  0.415 %    86.364 ±  0.759 %
+   3       7.4953 ±    0.3486       0.06386 ±    0.01034       0.10617 ±    0.00431     9.400 ±  0.398 %    86.771 ±  0.612 %
+   4       7.8932 ±    0.3295       0.07988 ±    0.00943       0.11534 ±    0.00606     9.599 ±  0.356 %    86.779 ±  0.530 %
+   5       7.6968 ±    0.2885       0.07479 ±    0.00839       0.11294 ±    0.00512     9.430 ±  0.323 %    86.979 ±  0.471 %
+   6       6.6156 ±    0.2191       0.07337 ±    0.00761       0.11303 ±    0.00469    10.110 ±  0.316 %    87.341 ±  0.424 %
+   7       6.1496 ±    0.1860       0.06895 ±    0.00729       0.12010 ±    0.00470    10.509 ±  0.310 %    87.544 ±  0.390 %
+   8       6.0524 ±    0.1698       0.06536 ±    0.00664       0.11729 ±    0.00419    10.338 ±  0.282 %    87.414 ±  0.367 %
+   9       6.3606 ±    0.1689       0.06248 ±    0.00621       0.11624 ±    0.00378    10.074 ±  0.260 %    87.053 ±  0.350 %
+  10       6.4785 ±    0.1643       0.06242 ±    0.00577       0.11173 ±    0.00343     9.809 ±  0.243 %    87.185 ±  0.330 %
+  11       6.5332 ±    0.1574       0.06204 ±    0.00542       0.10800 ±    0.00313     9.617 ±  0.228 %    87.203 ±  0.315 %
+  12       6.7836 ±    0.1577       0.06132 ±    0.00509       0.10489 ±    0.00288     9.401 ±  0.215 %    87.154 ±  0.302 %
+  13       6.8332 ±    0.1523       0.06171 ±    0.00483       0.10310 ±    0.00268     9.272 ±  0.203 %    87.112 ±  0.291 %
+  14       6.8837 ±    0.1476       0.06029 ±    0.00462       0.10108 ±    0.00251     9.107 ±  0.193 %    87.222 ±  0.279 %
+  15       6.9222 ±    0.1435       0.05966 ±    0.00443       0.10045 ±    0.00236     9.064 ±  0.184 %    87.129 ±  0.270 %
+  16       7.0981 ±    0.1428       0.05802 ±    0.00427       0.09952 ±    0.00223     8.947 ±  0.176 %    87.170 ±  0.261 %
+  17       7.1355 ±    0.1388       0.05632 ±    0.00410       0.09812 ±    0.00211     8.873 ±  0.170 %    87.229 ±  0.253 %
+  18       7.2233 ±    0.1366       0.05535 ±    0.00396       0.09744 ±    0.00201     8.809 ±  0.164 %    87.151 ±  0.247 %
+  19       7.1804 ±    0.1327       0.05631 ±    0.00385       0.09656 ±    0.00193     8.767 ±  0.160 %    87.138 ±  0.240 %
+  20       6.9267 ±    0.1241       0.05967 ±    0.00383       0.10094 ±    0.00190     9.066 ±  0.156 %    87.043 ±  0.235 %
+  21       6.9448 ±    0.1213       0.05978 ±    0.00374       0.10135 ±    0.00184     9.031 ±  0.151 %    87.032 ±  0.229 %
+  22       6.9675 ±    0.1190       0.06104 ±    0.00367       0.10256 ±    0.00183     9.077 ±  0.149 %    87.035 ±  0.224 %
+  23       7.0270 ±    0.1174       0.06214 ±    0.00359       0.10236 ±    0.00177     9.057 ±  0.146 %    86.986 ±  0.219 %
+  24       7.0192 ±    0.1145       0.06099 ±    0.00349       0.10191 ±    0.00171     9.016 ±  0.143 %    86.983 ±  0.215 %
+  25       7.0530 ±    0.1127       0.06108 ±    0.00342       0.10141 ±    0.00166     8.977 ±  0.139 %    86.952 ±  0.211 %
+  26       7.0192 ±    0.1099       0.06075 ±    0.00334       0.10143 ±    0.00162     8.999 ±  0.137 %    86.973 ±  0.206 %
+  27       7.1950 ±    0.1112       0.06116 ±    0.00329       0.10126 ±    0.00160     8.943 ±  0.134 %    86.988 ±  0.202 %
+  28       7.2801 ±    0.1108       0.06006 ±    0.00321       0.10017 ±    0.00155     8.860 ±  0.131 %    86.975 ±  0.199 %
+  29       7.2833 ±    0.1089       0.06136 ±    0.00316       0.10033 ±    0.00152     8.904 ±  0.130 %    86.925 ±  0.196 %
+  30       7.2262 ±    0.1061       0.06148 ±    0.00310       0.09994 ±    0.00148     8.892 ±  0.127 %    86.937 ±  0.192 %
+  31       7.1166 ±    0.1024       0.06072 ±    0.00302       0.09878 ±    0.00144     8.837 ±  0.124 %    87.043 ±  0.189 %
+  32       7.0141 ±    0.0992       0.06047 ±    0.00303       0.10174 ±    0.00161     9.036 ±  0.126 %    87.014 ±  0.186 %
+  33       6.9480 ±    0.0965       0.06088 ±    0.00298       0.10139 ±    0.00157     9.021 ±  0.124 %    87.046 ±  0.183 %
+  34       6.9248 ±    0.0946       0.05971 ±    0.00291       0.10041 ±    0.00152     8.964 ±  0.121 %    87.091 ±  0.180 %
+  35       6.9370 ±    0.0934       0.05965 ±    0.00285       0.09981 ±    0.00149     8.914 ±  0.119 %    87.139 ±  0.177 %
+  36       6.9569 ±    0.0925       0.06004 ±    0.00282       0.09978 ±    0.00146     8.900 ±  0.118 %    87.159 ±  0.174 %
+  37       6.8518 ±    0.0895       0.05860 ±    0.00276       0.09882 ±    0.00142     8.866 ±  0.115 %    87.192 ±  0.172 %
+  38       6.7745 ±    0.0870       0.05779 ±    0.00272       0.09818 ±    0.00139     8.852 ±  0.114 %    87.256 ±  0.169 %
+  39       6.6883 ±    0.0845       0.05732 ±    0.00267       0.09776 ±    0.00137     8.864 ±  0.112 %    87.232 ±  0.167 %
+  40       6.5921 ±    0.0819       0.05666 ±    0.00263       0.09716 ±    0.00134     8.853 ±  0.111 %    87.265 ±  0.165 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :   6.592123 ±   0.081874
+Mean PPL(base)                :   6.228979 ±   0.075322
+Cor(ln(PPL(Q)), ln(PPL(base))):  97.73%
+Mean ln(PPL(Q)/PPL(base))     :   0.056663 ±   0.002630
+Mean PPL(Q)/PPL(base)         :   1.058299 ±   0.002783
+Mean PPL(Q)-PPL(base)         :   0.363144 ±   0.017958
+
+====== KL divergence statistics ======
+Mean    KLD:   0.097165 ±   0.001337
+Maximum KLD:  15.969161
+99.9%   KLD:   3.145117
+99.0%   KLD:   0.858566
+95.0%   KLD:   0.329361
+90.0%   KLD:   0.207554
+Median  KLD:   0.043968
+10.0%   KLD:   0.000631
+ 5.0%   KLD:   0.000159
+ 1.0%   KLD:  -0.000029
+ 0.1%   KLD:  -0.000286
+Minimum KLD:  -0.000564
+
+====== Token probability statistics ======
+Mean    Δp: -0.372 ± 0.044 %
+Maximum Δp: 98.904%
+99.9%   Δp: 55.573%
+99.0%   Δp: 23.490%
+95.0%   Δp: 10.880%
+90.0%   Δp:  6.394%
+75.0%   Δp:  1.299%
+Median  Δp: -0.005%
+25.0%   Δp: -1.458%
+10.0%   Δp: -7.029%
+ 5.0%   Δp: -12.485%
+ 1.0%   Δp: -31.169%
+ 0.1%   Δp: -73.630%
+Minimum Δp: -99.736%
+RMS Δp    :  8.853 ± 0.111 %
+Same top p: 87.265 ± 0.165 %
+

recipe/logs/N6_bench.log

@@ -0,0 +1,46 @@
+[2026-09-16T23:22:44Z] waiting for the quiet-box lock (no HF upload during speed runs)
+[2026-09-17T00:32:03Z] quiet-box lock held
+[2026-09-17T00:32:03Z] bench n-q106-rocm
+{"label": "n-q106-rocm", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "code", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 63.94, "tg_min": 63.56, "tg_max": 63.96, "pp_median": 1158.2, "prompt_n": 7098, "accept": null}
+[2026-09-17T00:33:06Z] bench n-q106-vk
+{"label": "n-q106-vk", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "code", "load_s": 4.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 68.16, "tg_min": 68.1, "tg_max": 68.21, "pp_median": 1006.0, "prompt_n": 7096, "accept": null}
+[2026-09-17T00:33:54Z] bench n-q102-rocm
+{"label": "n-q102-rocm", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "code", "load_s": 24.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 61.74, "tg_min": 61.74, "tg_max": 61.8, "pp_median": 1186.7, "prompt_n": 7100, "accept": null}
+[2026-09-17T00:34:58Z] bench n-q102-vk
+{"label": "n-q102-vk", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "code", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 68.0, "tg_min": 67.88, "tg_max": 68.09, "pp_median": 1001.4, "prompt_n": 7098, "accept": null}
+[2026-09-17T00:36:05Z] bench n-q103-rocm
+{"label": "n-q103-rocm", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "code", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 63.31, "tg_min": 63.14, "tg_max": 63.38, "pp_median": 1156.6, "prompt_n": 7100, "accept": null}
+[2026-09-17T00:37:08Z] bench n-q103-vk
+{"label": "n-q103-vk", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "code", "load_s": 20.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 68.62, "tg_min": 68.6, "tg_max": 68.67, "pp_median": 998.9, "prompt_n": 7098, "accept": null}
+[2026-09-17T00:38:12Z] bench n-q106i-rocm
+{"label": "n-q106i-rocm", "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "code", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 63.29, "tg_min": 63.19, "tg_max": 63.41, "pp_median": 1145.3, "prompt_n": 7100, "accept": null}
+[2026-09-17T00:39:16Z] bench n-q106i-vk
+{"label": "n-q106i-vk", "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "code", "load_s": 20.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 67.77, "tg_min": 67.74, "tg_max": 67.86, "pp_median": 995.0, "prompt_n": 7098, "accept": null}
+[2026-09-17T00:40:20Z] bench n-q102i-rocm
+{"label": "n-q102i-rocm", "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "code", "load_s": 24.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 61.56, "tg_min": 61.53, "tg_max": 61.63, "pp_median": 1183.6, "prompt_n": 7098, "accept": null}
+[2026-09-17T00:41:25Z] bench n-q102i-vk
+{"label": "n-q102i-vk", "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "code", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 67.81, "tg_min": 67.77, "tg_max": 67.85, "pp_median": 998.7, "prompt_n": 7097, "accept": null}
+[2026-09-17T00:42:32Z] bench n-q103i-rocm
+{"label": "n-q103i-rocm", "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "code", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 63.03, "tg_min": 62.55, "tg_max": 63.11, "pp_median": 1154.4, "prompt_n": 7101, "accept": null}
+[2026-09-17T00:43:35Z] bench n-q103i-vk
+{"label": "n-q103i-vk", "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "code", "load_s": 20.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 68.44, "tg_min": 68.42, "tg_max": 68.56, "pp_median": 993.9, "prompt_n": 7098, "accept": null}
+[2026-09-17T00:44:39Z] bench n-q106-rocm-prose
+{"label": "n-q106-rocm-prose", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "ROCm0", "ctx": 65536, "workload": "prose", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 63.06, "tg_min": 62.94, "tg_max": 63.09, "pp_median": 1139.1, "prompt_n": 7454, "accept": null}
+[2026-09-17T00:45:44Z] bench n-q106-vk-prose
+{"label": "n-q106-vk-prose", "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "draft": null, "nmax": null, "strict": false, "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", "dev": "Vulkan0", "ctx": 65536, "workload": "prose", "load_s": 22.0, "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui", "tg_median": 67.34, "tg_min": 67.31, "tg_max": 67.35, "pp_median": 977.7, "prompt_n": 7454, "accept": null}
+[2026-09-17T00:46:53Z] cachegate n-c3-q106
+788d", "cold_sha": "a8545d39788d", "instr": "Write a function that returns every regular expression used in this file. ok ok ok", "warm_instr": "Summarise this file. (warm-up 3) ok ok ok ok"}
+   {"variant": 4, "L": 7039, "warm_prompt_n": 1028, "warm_cache_n": 6011, "cold_prompt_n": 7039, "cold_cache_n": 0, "aligned": true, "warm_draft": [null, null], "cold_draft": [null, null], "identical": true, "first_diff_char": null, "warm_sha": "af97b2dde4e2", "cold_sha": "af97b2dde4e2", "instr": "Write a function that maps each class in this file to its base classes. ok", "warm_instr": "Summarise this file. (warm-up 4) ok ok ok ok"}
+{"label": "n-c3-q106", "gate": "cachegate3-aligned", "n": 5, "L": 7039, "expect_cache_n": 6011, "all_reused": true, "all_aligned": true, "identical": 5, "result": "PASS"}
+[2026-09-17T00:48:34Z] tools n-tools-q106
+alse enum: unit=fahrenheit
+  PASS think=False correct-decline: content='391'
+  PASS think=False multi-turn: final='Tokyo is currently **21°C** with clear skies.'
+  PASS think=False streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  PASS think=False parallel: calls=['lima', 'oslo']
+{"label": "n-tools-q106", "passed": 6, "total": 14, "detail": {"multi-arg|think=True": false, "nested-object|think=True": false, "enum|think=True": false, "correct-decline|think=True": false, "multi-turn|think=True": false, "streaming|think=True": false, "parallel|think=True": false, "multi-arg|think=False": true, "nested-object|think=False": false, "enum|think=False": true, "correct-decline|think=False": true, "multi-turn|think=False": true, "streaming|think=False": true, "parallel|think=False": true}}
+[2026-09-17T00:49:16Z] vision n-vision-q106-faon
+{"label": "n-vision-q106-faon", "fa": "on", "mtp": false, "expected": "red,blue,circle,square", "answer": "The image shows two shapes: a red circle on the left and a blue square on the right.", "hits": ["red", "blue", "circle", "square"], "error": null, "server_died": false, "server_log_errors": [], "result": "PASS"}
+[2026-09-17T00:49:42Z] vision n-vision-q106-faoff
+{"label": "n-vision-q106-faoff", "fa": "off", "mtp": false, "expected": "red,blue,circle,square", "answer": "The image shows two simple shapes:\n\n- A **red circle** on the left.\n- A **blue square** on the right.", "hits": ["red", "blue", "circle", "square"], "error": null, "server_died": false, "server_log_errors": [], "result": "PASS"}
+[2026-09-17T00:50:08Z] NEX_BENCH_DONE

recipe/logs/N6t_tools_roff.log

@@ -0,0 +1,63 @@
+  PASS think=True  multi-arg: args={'city': 'Paris', 'unit': 'celsius'}
+  FAIL think=True  nested-object: exception KeyError('tool_calls')
+  PASS think=True  enum: unit=fahrenheit
+  PASS think=True  correct-decline: content='391'
+  PASS think=True  multi-turn: final='Tokyo is **21°C** and **clear** right now.'
+  PASS think=True  streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  PASS think=True  parallel: calls=['lima', 'oslo']
+  PASS think=False multi-arg: args={'city': 'Paris', 'unit': 'celsius'}
+  PASS think=False nested-object: args={'title': 'Design review', 'when': {'date': '2026-10-02', 'time': '14:00'}, 'attendees': ['ana@x.io', 'bo@x.io']}
+  PASS think=False enum: unit=fahrenheit
+  PASS think=False correct-decline: content='391'
+  PASS think=False multi-turn: final='Tokyo is **21°C** and **clear**.'
+  PASS think=False streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  PASS think=False parallel: calls=['lima', 'oslo']
+{"label": "n-tools-q106-roff", "passed": 13, "total": 14, "detail": {"multi-arg|think=True": true, "nested-object|think=True": false, "enum|think=True": true, "correct-decline|think=True": true, "multi-turn|think=True": true, "streaming|think=True": true, "parallel|think=True": true, "multi-arg|think=False": true, "nested-object|think=False": true, "enum|think=False": true, "correct-decline|think=False": true, "multi-turn|think=False": true, "streaming|think=False": true, "parallel|think=False": true}}
+  PASS think=True  multi-arg: args={'city': 'Paris', 'unit': 'celsius'}
+  PASS think=True  nested-object: args={'title': 'Design review', 'when': {'date': '2026-10-02', 'time': '14:00'}, 'attendees': ['ana@x.io', 'bo@x.io']}
+  PASS think=True  enum: unit=fahrenheit
+  PASS think=True  correct-decline: content='391'
+  PASS think=True  multi-turn: final='Tokyo is **21°C** and **clear**.'
+  PASS think=True  streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  FAIL think=True  parallel: calls=['oslo']
+  PASS think=False multi-arg: args={'city': 'Paris', 'unit': 'celsius'}
+  PASS think=False nested-object: args={'title': 'Design review', 'when': {'date': '2026-10-02', 'time': '14:00'}, 'attendees': ['ana@x.io', 'bo@x.io']}
+  PASS think=False enum: unit=fahrenheit
+  PASS think=False correct-decline: content='391'
+  PASS think=False multi-turn: final='Tokyo’s current weather is **21°C and clear**.'
+  PASS think=False streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  PASS think=False parallel: calls=['lima', 'oslo']
+{"label": "n-tools-q106-roff-r2", "passed": 13, "total": 14, "detail": {"multi-arg|think=True": true, "nested-object|think=True": true, "enum|think=True": true, "correct-decline|think=True": true, "multi-turn|think=True": true, "streaming|think=True": true, "parallel|think=True": false, "multi-arg|think=False": true, "nested-object|think=False": true, "enum|think=False": true, "correct-decline|think=False": true, "multi-turn|think=False": true, "streaming|think=False": true, "parallel|think=False": true}}
+  PASS think=True  multi-arg: args={'city': 'Paris', 'unit': 'celsius'}
+  PASS think=True  nested-object: args={'title': 'Design review', 'when': {'date': '2026-10-02', 'time': '14:00'}, 'attendees': ['ana@x.io', 'bo@x.io']}
+  PASS think=True  enum: unit=fahrenheit
+  PASS think=True  correct-decline: content='391'
+  PASS think=True  multi-turn: final='Tokyo is currently **21°C** with **clear skies**.'
+  PASS think=True  streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  PASS think=True  parallel: calls=['lima', 'oslo']
+  PASS think=False multi-arg: args={'city': 'Paris', 'unit': 'celsius'}
+  PASS think=False nested-object: args={'title': 'Design review', 'when': {'date': '2026-10-02', 'time': '14:00'}, 'attendees': ['ana@x.io', 'bo@x.io']}
+  PASS think=False enum: unit=fahrenheit
+  PASS think=False correct-decline: content='391'
+  PASS think=False multi-turn: final='Tokyo’s weather is **21°C and clear**.'
+  PASS think=False streaming: stream args={'city': 'Rome', 'unit': 'celsius'}
+  PASS think=False parallel: calls=['lima', 'oslo']
+{"label": "n-tools-q106-roff-r3", "passed": 14, "total": 14, "detail": {"multi-arg|think=True": true, "nested-object|think=True": true, "enum|think=True": true, "correct-decline|think=True": true, "multi-turn|think=True": true, "streaming|think=True": true, "parallel|think=True": true, "multi-arg|think=False": true, "nested-object|think=False": true, "enum|think=False": true, "correct-decline|think=False": true, "multi-turn|think=False": true, "streaming|think=False": true, "parallel|think=False": true}}
+{"label": "n-vision-q106-roff-faon", "fa": "on", "mtp": false, "expected": "red,blue,circle,square", "answer": "The image shows two shapes: a red circle on the left and a blue square on the right.", "hits": ["red", "blue", "circle", "square"], "error": null, "server_died": false, "server_log_errors": [], "result": "PASS"}
+probe no-kwargs correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+probe no-kwargs single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+probe no-kwargs multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+probe enable_thinking=false correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+probe enable_thinking=false single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+probe enable_thinking=false multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+probe reasoning_effort=high correct-decline {"content": "We need answer directly. 391.\n</think>\n\n391", "reasoning_len": 0, "tool_calls": [], "leaks": ["</think>"]}
+probe reasoning_effort=high single-word {"content": "We need need output exactly ready.\n</think>\n\nready", "reasoning_len": 0, "tool_calls": [], "leaks": ["</think>"]}
+probe reasoning_effort=high multi-arg {"content": "We need need tool. Current weather Paris celsius.\n</think>\n\n", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": ["</think>"]}
+probe reasoning_effort=medium correct-decline {"content": "\n\n</think>\n\n391", "reasoning_len": 0, "tool_calls": [], "leaks": ["</think>"]}
+probe reasoning_effort=medium single-word {"content": "\n\n</think>\n\nready", "reasoning_len": 0, "tool_calls": [], "leaks": ["</think>"]}
+probe reasoning_effort=medium multi-arg {"content": "\n\n</think>\n\n", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": ["</think>"]}
+probe reasoning_effort=none correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+probe reasoning_effort=none single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+probe reasoning_effort=none multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+NEX_TOOLS_TPL_DONE
+rc=0

recipe/logs/N8b_seats.log

@@ -0,0 +1,8 @@
+[2026-09-17T01:06:02Z] plan: [('max1-nex-fast', 'ROCm0', 'fa on', 262144, '31000M', True), ('max1-nex-fast-imat', 'ROCm0', 'fa on', 262144, '31000M', True)]
+[2026-09-17T01:06:04Z] max1-nex-fast written (ROCm0, -fa on, ctx 262144, MemoryMax 31000M) -> smoke test
+{"unit": "max1-nex-fast", "port": 8097, "load_s": 25, "time": "2026-09-17T01:06:29Z", "direct_reply": "ready", "direct_tg": 61.724584902166534, "default_reply": "", "default_reasoning_len": 5, "default_leak": false, "gateway_model": "nex-n2.5-mini-fast@max1", "gateway_reply": "ready", "result": "FAIL"}
+[2026-09-17T01:06:30Z]   max1-nex-fast stopped (enabled: disabled)
+[2026-09-17T01:06:36Z] max1-nex-fast-imat written (ROCm0, -fa on, ctx 262144, MemoryMax 31000M) -> smoke test
+{"unit": "max1-nex-fast-imat", "port": 8098, "load_s": 25, "time": "2026-09-17T01:07:02Z", "direct_reply": "ready", "direct_tg": 47.69058349428906, "default_reply": "", "default_reasoning_len": 5, "default_leak": false, "gateway_model": "nex-n2.5-mini-fast-imatrix@max1", "gateway_reply": "ready", "result": "FAIL"}
+[2026-09-17T01:07:03Z]   max1-nex-fast-imat stopped (enabled: disabled)
+[2026-09-17T01:07:08Z] NEX_SEATS_DONE fail=1

recipe/logs/N8c_seats.log

@@ -0,0 +1,8 @@
+[2026-09-17T01:19:09Z] plan: [('max1-nex-fast', 'ROCm0', 'fa on', 262144, '31000M', True), ('max1-nex-fast-imat', 'ROCm0', 'fa on', 262144, '31000M', True)]
+[2026-09-17T01:19:10Z] max1-nex-fast written (ROCm0, -fa on, ctx 262144, MemoryMax 31000M) -> smoke test
+{"unit": "max1-nex-fast", "port": 8097, "load_s": 30, "time": "2026-09-17T01:19:40Z", "direct_reply": "ready", "direct_tg": 42.57493188010899, "default_reply": "ready", "default_reasoning_len": 0, "default_leak": false, "thinking_reply": "", "thinking_reasoning_len": 5, "thinking_leak": false, "gateway_model": "nex-n2.5-mini-fast@max1", "gateway_reply": "ready", "result": "FAIL"}
+[2026-09-17T01:19:42Z]   max1-nex-fast stopped (enabled: disabled)
+[2026-09-17T01:19:48Z] max1-nex-fast-imat written (ROCm0, -fa on, ctx 262144, MemoryMax 31000M) -> smoke test
+{"unit": "max1-nex-fast-imat", "port": 8098, "load_s": 25, "time": "2026-09-17T01:20:13Z", "direct_reply": "ready", "direct_tg": 45.33297066956798, "default_reply": "ready", "default_reasoning_len": 0, "default_leak": false, "thinking_reply": "", "thinking_reasoning_len": 5, "thinking_leak": false, "gateway_model": "nex-n2.5-mini-fast-imatrix@max1", "gateway_reply": "ready", "result": "FAIL"}
+[2026-09-17T01:20:15Z]   max1-nex-fast-imat stopped (enabled: disabled)
+[2026-09-17T01:20:20Z] NEX_SEATS_DONE fail=1

recipe/logs/b_n-tools-q106-roff-probe.log

@@ -0,0 +1,285 @@
+0.00.108.543 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.108.553 I device_info:
+0.00.108.699 I   - ROCm0   : AMD Radeon Graphics (131072 MiB, 123524 MiB free)
+0.00.108.869 I   - Vulkan0 : AMD Radeon Graphics (RADV GFX1151) (132096 MiB, 131922 MiB free)
+0.00.108.878 I   - CPU     : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S (127438 MiB, 127438 MiB free)
+0.00.108.975 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.109.043 I srv          init: running without SSL
+0.00.109.069 I srv          init: using 31 threads for HTTP server
+0.00.109.070 I srv          init: the WebUI is disabled
+0.00.109.152 I srv         start: binding port with default address family
+0.00.110.377 I srv          main: loading model
+0.00.110.386 I srv    load_model: loading model '/mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf'
+0.00.167.016 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.22.284.156 W llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.22.593.161 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.22.997.376 I srv    load_model: initializing slots, n_slots = 1
+0.23.238.804 W srv    load_model: speculative decoding will use checkpoints
+0.23.238.818 W common_speculative_init: no implementations specified for speculative decoding
+0.23.238.821 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 65536
+0.23.239.043 I srv    load_model: prompt cache RAM enabled: limit_mib=8192
+0.23.239.047 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
+0.23.239.108 I srv          init: idle slots will be saved to prompt cache upon starting a new task
+0.23.286.447 I init: chat template, example_format: '<|im_start|>system
+You are a helpful assistant<|im_end|>
+<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+Hi there<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+'
+0.23.329.268 I srv          init: init: chat template, thinking = 0
+0.23.329.322 I srv          main: model loaded
+0.23.329.333 I srv          main: server is listening on http://127.0.0.1:18652
+0.23.329.348 I srv  update_slots: all slots are idle
+0.24.570.143 I srv  params_from_: Chat format: peg-native
+0.24.570.532 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.24.570.537 I srv  get_availabl: updating prompt cache
+0.24.570.545 I srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
+0.24.570.551 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 65536 tokens, 8589934592 est)
+0.24.570.554 I srv  get_availabl: prompt cache update took 0.01 ms
+0.24.570.926 I reasoning-budget: activated, budget=2147483647 tokens
+0.24.570.930 I reasoning-budget: deactivated (natural end)
+0.24.570.950 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.25.270.625 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.25.485.037 I slot print_timing: id  0 | task 0 | 
+prompt eval time =     747.50 ms /   427 tokens (    1.75 ms per token,   571.23 tokens per second)
+       eval time =     166.53 ms /     4 tokens (   41.63 ms per token,    24.02 tokens per second)
+      total time =     914.04 ms /   431 tokens
+0.25.485.248 I slot      release: id  0 | task 0 | stop processing: n_tokens = 430, truncated = 0
+0.25.485.267 I srv  update_slots: all slots are idle
+0.25.528.499 I srv  params_from_: Chat format: peg-native
+0.25.530.258 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.942
+0.25.530.747 I reasoning-budget: activated, budget=2147483647 tokens
+0.25.530.752 I reasoning-budget: deactivated (natural end)
+0.25.530.840 I slot launch_slot_: id  0 | task 6 | processing task, is_child = 0
+0.25.530.860 W slot update_slots: id  0 | task 6 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.25.530.863 I slot update_slots: id  0 | task 6 | Checking checkpoint with [422, 422] against 405...
+0.25.530.865 W slot update_slots: id  0 | task 6 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.25.530.870 W slot update_slots: id  0 | task 6 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.26.123.530 I slot create_check: id  0 | task 6 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.26.214.070 I slot print_timing: id  0 | task 6 | 
+prompt eval time =     644.28 ms /   421 tokens (    1.53 ms per token,   653.44 tokens per second)
+       eval time =      38.90 ms /     2 tokens (   19.45 ms per token,    51.42 tokens per second)
+      total time =     683.18 ms /   423 tokens
+0.26.214.297 I slot      release: id  0 | task 6 | stop processing: n_tokens = 422, truncated = 0
+0.26.214.362 I srv  update_slots: all slots are idle
+0.26.263.300 I srv  params_from_: Chat format: peg-native
+0.26.263.827 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.960
+0.26.264.204 I reasoning-budget: activated, budget=2147483647 tokens
+0.26.264.213 I reasoning-budget: deactivated (natural end)
+0.26.264.273 I slot launch_slot_: id  0 | task 10 | processing task, is_child = 0
+0.26.264.297 W slot update_slots: id  0 | task 10 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.26.264.300 I slot update_slots: id  0 | task 10 | Checking checkpoint with [416, 416] against 405...
+0.26.264.302 W slot update_slots: id  0 | task 10 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.26.264.309 W slot update_slots: id  0 | task 10 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.26.862.333 I slot create_check: id  0 | task 10 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.27.692.349 I slot print_timing: id  0 | task 10 | 
+prompt eval time =     646.46 ms /   424 tokens (    1.52 ms per token,   655.88 tokens per second)
+       eval time =     781.59 ms /    39 tokens (   20.04 ms per token,    49.90 tokens per second)
+      total time =    1428.04 ms /   463 tokens
+0.27.692.422 I slot      release: id  0 | task 10 | stop processing: n_tokens = 462, truncated = 0
+0.27.692.446 I srv  update_slots: all slots are idle
+0.27.714.727 I srv  params_from_: Chat format: peg-native
+0.27.715.273 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.951 (> 0.100 thold), f_keep = 0.879
+0.27.715.596 I reasoning-budget: activated, budget=2147483647 tokens
+0.27.715.599 I reasoning-budget: deactivated (natural end)
+0.27.715.649 I slot launch_slot_: id  0 | task 51 | processing task, is_child = 0
+0.27.715.665 W slot update_slots: id  0 | task 51 | n_past = 406, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.27.715.666 I slot update_slots: id  0 | task 51 | Checking checkpoint with [419, 419] against 406...
+0.27.715.667 W slot update_slots: id  0 | task 51 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.27.715.671 W slot update_slots: id  0 | task 51 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.28.253.015 I slot create_check: id  0 | task 51 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.28.373.714 I slot print_timing: id  0 | task 51 | 
+prompt eval time =     579.69 ms /   427 tokens (    1.36 ms per token,   736.60 tokens per second)
+       eval time =      78.34 ms /     4 tokens (   19.59 ms per token,    51.06 tokens per second)
+      total time =     658.04 ms /   431 tokens
+0.28.373.794 I slot      release: id  0 | task 51 | stop processing: n_tokens = 430, truncated = 0
+0.28.373.824 I srv  update_slots: all slots are idle
+0.28.388.433 I srv  params_from_: Chat format: peg-native
+0.28.388.929 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.942
+0.28.389.500 I reasoning-budget: activated, budget=2147483647 tokens
+0.28.389.507 I reasoning-budget: deactivated (natural end)
+0.28.389.598 I slot launch_slot_: id  0 | task 57 | processing task, is_child = 0
+0.28.389.627 W slot update_slots: id  0 | task 57 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.28.389.631 I slot update_slots: id  0 | task 57 | Checking checkpoint with [422, 422] against 405...
+0.28.389.633 W slot update_slots: id  0 | task 57 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.28.389.641 W slot update_slots: id  0 | task 57 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.28.926.082 I slot create_check: id  0 | task 57 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.28.986.227 I slot print_timing: id  0 | task 57 | 
+prompt eval time =     569.23 ms /   421 tokens (    1.35 ms per token,   739.59 tokens per second)
+       eval time =      27.35 ms /     2 tokens (   13.67 ms per token,    73.13 tokens per second)
+      total time =     596.58 ms /   423 tokens
+0.28.986.366 I slot      release: id  0 | task 57 | stop processing: n_tokens = 422, truncated = 0
+0.28.986.404 I srv  update_slots: all slots are idle
+0.29.001.891 I srv  params_from_: Chat format: peg-native
+0.29.002.333 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.960
+0.29.002.794 I reasoning-budget: activated, budget=2147483647 tokens
+0.29.002.796 I reasoning-budget: deactivated (natural end)
+0.29.002.839 I slot launch_slot_: id  0 | task 61 | processing task, is_child = 0
+0.29.002.854 W slot update_slots: id  0 | task 61 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.29.002.855 I slot update_slots: id  0 | task 61 | Checking checkpoint with [416, 416] against 405...
+0.29.002.857 W slot update_slots: id  0 | task 61 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.29.002.861 W slot update_slots: id  0 | task 61 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.29.671.242 I slot create_check: id  0 | task 61 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.30.481.274 I slot print_timing: id  0 | task 61 | 
+prompt eval time =     711.37 ms /   424 tokens (    1.68 ms per token,   596.03 tokens per second)
+       eval time =     767.02 ms /    39 tokens (   19.67 ms per token,    50.85 tokens per second)
+      total time =    1478.38 ms /   463 tokens
+0.30.481.478 I slot      release: id  0 | task 61 | stop processing: n_tokens = 462, truncated = 0
+0.30.481.535 I srv  update_slots: all slots are idle
+0.30.525.398 I srv  params_from_: Chat format: peg-native
+0.30.527.228 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.879
+0.30.527.832 I reasoning-budget: activated, budget=2147483647 tokens
+0.30.527.839 I reasoning-budget: deactivated (natural end)
+0.30.527.933 I slot launch_slot_: id  0 | task 102 | processing task, is_child = 0
+0.30.527.959 W slot update_slots: id  0 | task 102 | n_past = 406, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.30.527.963 I slot update_slots: id  0 | task 102 | Checking checkpoint with [419, 419] against 406...
+0.30.527.965 W slot update_slots: id  0 | task 102 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.30.527.974 W slot update_slots: id  0 | task 102 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.31.124.220 I slot create_check: id  0 | task 102 | created context checkpoint 1 of 32 (pos_min = 420, pos_max = 420, n_tokens = 421, size = 62.813 MiB)
+0.31.601.068 I slot print_timing: id  0 | task 102 | 
+prompt eval time =     634.63 ms /   425 tokens (    1.49 ms per token,   669.68 tokens per second)
+       eval time =     438.43 ms /    17 tokens (   25.79 ms per token,    38.77 tokens per second)
+      total time =    1073.07 ms /   442 tokens
+0.31.601.327 I slot      release: id  0 | task 102 | stop processing: n_tokens = 441, truncated = 0
+0.31.601.391 I srv  update_slots: all slots are idle
+0.31.656.355 I srv  params_from_: Chat format: peg-native
+0.31.658.634 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.967 (> 0.100 thold), f_keep = 0.918
+0.31.659.308 I reasoning-budget: activated, budget=2147483647 tokens
+0.31.659.314 I reasoning-budget: deactivated (natural end)
+0.31.659.409 I slot launch_slot_: id  0 | task 121 | processing task, is_child = 0
+0.31.659.433 W slot update_slots: id  0 | task 121 | n_past = 405, slot.prompt.tokens.size() = 441, seq_id = 0, pos_min = 440, n_swa = 0
+0.31.659.436 I slot update_slots: id  0 | task 121 | Checking checkpoint with [420, 420] against 405...
+0.31.659.438 W slot update_slots: id  0 | task 121 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.31.659.444 W slot update_slots: id  0 | task 121 | erased invalidated context checkpoint (pos_min = 420, pos_max = 420, n_tokens = 421, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.257.720 I slot create_check: id  0 | task 121 | created context checkpoint 1 of 32 (pos_min = 414, pos_max = 414, n_tokens = 415, size = 62.813 MiB)
+0.32.614.361 I slot print_timing: id  0 | task 121 | 
+prompt eval time =     643.10 ms /   419 tokens (    1.53 ms per token,   651.54 tokens per second)
+       eval time =     311.81 ms /    12 tokens (   25.98 ms per token,    38.49 tokens per second)
+      total time =     954.90 ms /   431 tokens
+0.32.614.571 I slot      release: id  0 | task 121 | stop processing: n_tokens = 430, truncated = 0
+0.32.614.632 I srv  update_slots: all slots are idle
+0.32.639.909 I srv  params_from_: Chat format: peg-native
+0.32.640.349 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.960 (> 0.100 thold), f_keep = 0.942
+0.32.640.975 I reasoning-budget: activated, budget=2147483647 tokens
+0.32.640.982 I reasoning-budget: deactivated (natural end)
+0.32.641.069 I slot launch_slot_: id  0 | task 135 | processing task, is_child = 0
+0.32.641.093 W slot update_slots: id  0 | task 135 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.32.641.097 I slot update_slots: id  0 | task 135 | Checking checkpoint with [414, 414] against 405...
+0.32.641.099 W slot update_slots: id  0 | task 135 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.32.641.107 W slot update_slots: id  0 | task 135 | erased invalidated context checkpoint (pos_min = 414, pos_max = 414, n_tokens = 415, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.33.188.406 I slot create_check: id  0 | task 135 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.34.408.355 I slot print_timing: id  0 | task 135 | 
+prompt eval time =     604.54 ms /   422 tokens (    1.43 ms per token,   698.05 tokens per second)
+       eval time =    1162.69 ms /    53 tokens (   21.94 ms per token,    45.58 tokens per second)
+      total time =    1767.23 ms /   475 tokens
+0.34.408.563 I slot      release: id  0 | task 135 | stop processing: n_tokens = 474, truncated = 0
+0.34.408.627 I srv  update_slots: all slots are idle
+0.34.455.891 I srv  params_from_: Chat format: peg-native
+0.34.457.714 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.958 (> 0.100 thold), f_keep = 0.857
+0.34.458.254 I reasoning-budget: activated, budget=2147483647 tokens
+0.34.458.261 I reasoning-budget: deactivated (natural end)
+0.34.458.355 I slot launch_slot_: id  0 | task 190 | processing task, is_child = 0
+0.34.458.379 W slot update_slots: id  0 | task 190 | n_past = 406, slot.prompt.tokens.size() = 474, seq_id = 0, pos_min = 473, n_swa = 0
+0.34.458.382 I slot update_slots: id  0 | task 190 | Checking checkpoint with [417, 417] against 406...
+0.34.458.384 W slot update_slots: id  0 | task 190 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.34.458.391 W slot update_slots: id  0 | task 190 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.35.097.680 I slot create_check: id  0 | task 190 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.35.342.520 I slot print_timing: id  0 | task 190 | 
+prompt eval time =     694.85 ms /   424 tokens (    1.64 ms per token,   610.21 tokens per second)
+       eval time =     189.28 ms /     7 tokens (   27.04 ms per token,    36.98 tokens per second)
+      total time =     884.13 ms /   431 tokens
+0.35.342.606 I slot      release: id  0 | task 190 | stop processing: n_tokens = 430, truncated = 0
+0.35.342.636 I srv  update_slots: all slots are idle
+0.35.396.955 I srv  params_from_: Chat format: peg-native
+0.35.399.122 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.969 (> 0.100 thold), f_keep = 0.942
+0.35.399.482 I reasoning-budget: activated, budget=2147483647 tokens
+0.35.399.485 I reasoning-budget: deactivated (natural end)
+0.35.399.534 I slot launch_slot_: id  0 | task 199 | processing task, is_child = 0
+0.35.399.550 W slot update_slots: id  0 | task 199 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.35.399.552 I slot update_slots: id  0 | task 199 | Checking checkpoint with [419, 419] against 405...
+0.35.399.553 W slot update_slots: id  0 | task 199 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.35.399.556 W slot update_slots: id  0 | task 199 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.35.975.338 I slot create_check: id  0 | task 199 | created context checkpoint 1 of 32 (pos_min = 413, pos_max = 413, n_tokens = 414, size = 62.813 MiB)
+0.36.211.213 I slot print_timing: id  0 | task 199 | 
+prompt eval time =     638.31 ms /   418 tokens (    1.53 ms per token,   654.86 tokens per second)
+       eval time =     173.34 ms /     5 tokens (   34.67 ms per token,    28.84 tokens per second)
+      total time =     811.65 ms /   423 tokens
+0.36.211.296 I slot      release: id  0 | task 199 | stop processing: n_tokens = 422, truncated = 0
+0.36.211.327 I srv  update_slots: all slots are idle
+0.36.257.343 I srv  params_from_: Chat format: peg-native
+0.36.259.204 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.960
+0.36.259.770 I reasoning-budget: activated, budget=2147483647 tokens
+0.36.259.773 I reasoning-budget: deactivated (natural end)
+0.36.259.860 I slot launch_slot_: id  0 | task 206 | processing task, is_child = 0
+0.36.259.881 W slot update_slots: id  0 | task 206 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.36.259.884 I slot update_slots: id  0 | task 206 | Checking checkpoint with [413, 413] against 405...
+0.36.259.886 W slot update_slots: id  0 | task 206 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.36.259.891 W slot update_slots: id  0 | task 206 | erased invalidated context checkpoint (pos_min = 413, pos_max = 413, n_tokens = 414, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.36.853.203 I slot create_check: id  0 | task 206 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.37.881.460 I slot print_timing: id  0 | task 206 | 
+prompt eval time =     638.50 ms /   421 tokens (    1.52 ms per token,   659.35 tokens per second)
+       eval time =     983.06 ms /    42 tokens (   23.41 ms per token,    42.72 tokens per second)
+      total time =    1621.57 ms /   463 tokens
+0.37.881.529 I slot      release: id  0 | task 206 | stop processing: n_tokens = 462, truncated = 0
+0.37.881.557 I srv  update_slots: all slots are idle
+0.37.893.541 I srv  params_from_: Chat format: peg-native
+0.37.894.001 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.951 (> 0.100 thold), f_keep = 0.879
+0.37.894.205 I reasoning-budget: activated, budget=2147483647 tokens
+0.37.894.207 I reasoning-budget: deactivated (natural end)
+0.37.894.246 I slot launch_slot_: id  0 | task 250 | processing task, is_child = 0
+0.37.894.256 W slot update_slots: id  0 | task 250 | n_past = 406, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.37.894.257 I slot update_slots: id  0 | task 250 | Checking checkpoint with [416, 416] against 406...
+0.37.894.258 W slot update_slots: id  0 | task 250 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.37.894.260 W slot update_slots: id  0 | task 250 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.38.426.246 I slot create_check: id  0 | task 250 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.38.626.110 I slot print_timing: id  0 | task 250 | 
+prompt eval time =     594.65 ms /   427 tokens (    1.39 ms per token,   718.07 tokens per second)
+       eval time =     137.18 ms /     4 tokens (   34.29 ms per token,    29.16 tokens per second)
+      total time =     731.83 ms /   431 tokens
+0.38.626.211 I slot      release: id  0 | task 250 | stop processing: n_tokens = 430, truncated = 0
+0.38.626.244 I srv  update_slots: all slots are idle
+0.38.640.241 I srv  params_from_: Chat format: peg-native
+0.38.640.929 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.942
+0.38.641.272 I reasoning-budget: activated, budget=2147483647 tokens
+0.38.641.274 I reasoning-budget: deactivated (natural end)
+0.38.641.328 I slot launch_slot_: id  0 | task 256 | processing task, is_child = 0
+0.38.641.341 W slot update_slots: id  0 | task 256 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.38.641.343 I slot update_slots: id  0 | task 256 | Checking checkpoint with [422, 422] against 405...
+0.38.641.344 W slot update_slots: id  0 | task 256 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.38.641.348 W slot update_slots: id  0 | task 256 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.39.216.267 I slot create_check: id  0 | task 256 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.39.278.250 I slot print_timing: id  0 | task 256 | 
+prompt eval time =     607.81 ms /   421 tokens (    1.44 ms per token,   692.65 tokens per second)
+       eval time =      29.08 ms /     2 tokens (   14.54 ms per token,    68.78 tokens per second)
+      total time =     636.89 ms /   423 tokens
+0.39.278.403 I slot      release: id  0 | task 256 | stop processing: n_tokens = 422, truncated = 0
+0.39.278.445 I srv  update_slots: all slots are idle
+0.39.315.466 I srv  params_from_: Chat format: peg-native
+0.39.315.935 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.960
+0.39.316.462 I reasoning-budget: activated, budget=2147483647 tokens
+0.39.316.465 I reasoning-budget: deactivated (natural end)
+0.39.316.537 I slot launch_slot_: id  0 | task 260 | processing task, is_child = 0
+0.39.316.555 W slot update_slots: id  0 | task 260 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.39.316.557 I slot update_slots: id  0 | task 260 | Checking checkpoint with [416, 416] against 405...
+0.39.316.559 W slot update_slots: id  0 | task 260 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.39.316.565 W slot update_slots: id  0 | task 260 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.39.992.109 I slot create_check: id  0 | task 260 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.40.804.656 I slot print_timing: id  0 | task 260 | 
+prompt eval time =     726.64 ms /   424 tokens (    1.71 ms per token,   583.51 tokens per second)
+       eval time =     761.44 ms /    39 tokens (   19.52 ms per token,    51.22 tokens per second)
+      total time =    1488.09 ms /   463 tokens
+0.40.804.764 I slot      release: id  0 | task 260 | stop processing: n_tokens = 462, truncated = 0
+0.40.804.795 I srv  update_slots: all slots are idle
+0.40.806.202 I srv    operator(): operator(): cleaning up before exit...

recipe/logs/b_n-tools-q106-roff-r2.log

@@ -0,0 +1,301 @@
+0.00.115.039 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.115.043 I device_info:
+0.00.115.119 I   - ROCm0   : AMD Radeon Graphics (131072 MiB, 123866 MiB free)
+0.00.115.243 I   - Vulkan0 : AMD Radeon Graphics (RADV GFX1151) (132096 MiB, 131922 MiB free)
+0.00.115.248 I   - CPU     : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S (127438 MiB, 127438 MiB free)
+0.00.115.311 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.115.325 I srv          init: running without SSL
+0.00.115.360 I srv          init: using 31 threads for HTTP server
+0.00.115.361 I srv          init: the WebUI is disabled
+0.00.115.422 I srv         start: binding port with default address family
+0.00.116.607 I srv          main: loading model
+0.00.116.616 I srv    load_model: loading model '/mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf'
+0.00.167.815 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.22.638.462 W llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.22.931.507 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.23.329.767 I srv    load_model: initializing slots, n_slots = 1
+0.23.568.311 W srv    load_model: speculative decoding will use checkpoints
+0.23.568.330 W common_speculative_init: no implementations specified for speculative decoding
+0.23.568.334 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 65536
+0.23.568.522 I srv    load_model: prompt cache RAM enabled: limit_mib=8192
+0.23.568.527 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
+0.23.568.581 I srv          init: idle slots will be saved to prompt cache upon starting a new task
+0.23.617.611 I init: chat template, example_format: '<|im_start|>system
+You are a helpful assistant<|im_end|>
+<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+Hi there<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+'
+0.23.662.046 I srv          init: init: chat template, thinking = 0
+0.23.662.133 I srv          main: model loaded
+0.23.662.139 I srv          main: server is listening on http://127.0.0.1:18600
+0.23.662.179 I srv  update_slots: all slots are idle
+0.24.822.880 I srv  params_from_: Chat format: peg-native
+0.24.825.218 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.24.825.223 I srv  get_availabl: updating prompt cache
+0.24.825.232 I srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
+0.24.825.240 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 65536 tokens, 8589934592 est)
+0.24.825.243 I srv  get_availabl: prompt cache update took 0.02 ms
+0.24.825.968 I reasoning-budget: activated, budget=2147483647 tokens
+0.24.825.999 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.25.491.127 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.25.806.479 I reasoning-budget: deactivated (natural end)
+0.26.592.455 I slot print_timing: id  0 | task 0 | 
+prompt eval time =     699.89 ms /   422 tokens (    1.66 ms per token,   602.95 tokens per second)
+       eval time =    1066.52 ms /    52 tokens (   20.51 ms per token,    48.76 tokens per second)
+      total time =    1766.41 ms /   474 tokens
+0.26.592.534 I slot      release: id  0 | task 0 | stop processing: n_tokens = 473, truncated = 0
+0.26.592.543 I srv  update_slots: all slots are idle
+0.26.605.042 I srv  params_from_: Chat format: peg-native
+0.26.605.398 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.906 (> 0.100 thold), f_keep = 0.856
+0.26.605.730 I reasoning-budget: activated, budget=2147483647 tokens
+0.26.605.800 I slot launch_slot_: id  0 | task 54 | processing task, is_child = 0
+0.26.605.816 W slot update_slots: id  0 | task 54 | n_past = 405, slot.prompt.tokens.size() = 473, seq_id = 0, pos_min = 472, n_swa = 0
+0.26.605.818 I slot update_slots: id  0 | task 54 | Checking checkpoint with [417, 417] against 405...
+0.26.605.820 W slot update_slots: id  0 | task 54 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.26.605.826 W slot update_slots: id  0 | task 54 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.27.159.769 I slot create_check: id  0 | task 54 | created context checkpoint 1 of 32 (pos_min = 442, pos_max = 442, n_tokens = 443, size = 62.813 MiB)
+0.27.788.657 I reasoning-budget: deactivated (natural end)
+0.29.292.130 I slot print_timing: id  0 | task 54 | n_decoded =    100, tg =  47.73 t/s
+0.29.582.323 I slot print_timing: id  0 | task 54 | 
+prompt eval time =     591.18 ms /   447 tokens (    1.32 ms per token,   756.11 tokens per second)
+       eval time =    2385.31 ms /   115 tokens (   20.74 ms per token,    48.21 tokens per second)
+      total time =    2976.49 ms /   562 tokens
+0.29.582.414 I slot      release: id  0 | task 54 | stop processing: n_tokens = 561, truncated = 0
+0.29.582.457 I srv  update_slots: all slots are idle
+0.29.597.465 I srv  params_from_: Chat format: peg-native
+0.29.597.831 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.953 (> 0.100 thold), f_keep = 0.722
+0.29.598.059 I reasoning-budget: activated, budget=2147483647 tokens
+0.29.598.094 I slot launch_slot_: id  0 | task 171 | processing task, is_child = 0
+0.29.598.104 W slot update_slots: id  0 | task 171 | n_past = 405, slot.prompt.tokens.size() = 561, seq_id = 0, pos_min = 560, n_swa = 0
+0.29.598.104 I slot update_slots: id  0 | task 171 | Checking checkpoint with [442, 442] against 405...
+0.29.598.105 W slot update_slots: id  0 | task 171 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.29.598.107 W slot update_slots: id  0 | task 171 | erased invalidated context checkpoint (pos_min = 442, pos_max = 442, n_tokens = 443, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.30.126.859 I slot create_check: id  0 | task 171 | created context checkpoint 1 of 32 (pos_min = 420, pos_max = 420, n_tokens = 421, size = 62.813 MiB)
+0.30.663.476 I reasoning-budget: deactivated (natural end)
+0.31.461.262 I slot print_timing: id  0 | task 171 | 
+prompt eval time =     582.92 ms /   425 tokens (    1.37 ms per token,   729.09 tokens per second)
+       eval time =    1280.22 ms /    63 tokens (   20.32 ms per token,    49.21 tokens per second)
+      total time =    1863.14 ms /   488 tokens
+0.31.461.343 I slot      release: id  0 | task 171 | stop processing: n_tokens = 487, truncated = 0
+0.31.461.374 I srv  update_slots: all slots are idle
+0.31.475.859 I srv  params_from_: Chat format: peg-native
+0.31.476.263 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.953 (> 0.100 thold), f_keep = 0.832
+0.31.476.455 I reasoning-budget: activated, budget=2147483647 tokens
+0.31.476.489 I slot launch_slot_: id  0 | task 236 | processing task, is_child = 0
+0.31.476.500 W slot update_slots: id  0 | task 236 | n_past = 405, slot.prompt.tokens.size() = 487, seq_id = 0, pos_min = 486, n_swa = 0
+0.31.476.501 I slot update_slots: id  0 | task 236 | Checking checkpoint with [420, 420] against 405...
+0.31.476.502 W slot update_slots: id  0 | task 236 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.31.476.505 W slot update_slots: id  0 | task 236 | erased invalidated context checkpoint (pos_min = 420, pos_max = 420, n_tokens = 421, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.014.645 I slot create_check: id  0 | task 236 | created context checkpoint 1 of 32 (pos_min = 420, pos_max = 420, n_tokens = 421, size = 62.813 MiB)
+0.32.309.451 I reasoning-budget: deactivated (natural end)
+0.32.409.234 I slot print_timing: id  0 | task 236 | 
+prompt eval time =     590.97 ms /   425 tokens (    1.39 ms per token,   719.16 tokens per second)
+       eval time =     341.75 ms /    17 tokens (   20.10 ms per token,    49.74 tokens per second)
+      total time =     932.72 ms /   442 tokens
+0.32.409.323 I slot      release: id  0 | task 236 | stop processing: n_tokens = 441, truncated = 0
+0.32.409.352 I srv  update_slots: all slots are idle
+0.32.432.559 I srv  params_from_: Chat format: peg-native
+0.32.432.999 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.921
+0.32.433.226 I reasoning-budget: activated, budget=2147483647 tokens
+0.32.433.266 I slot launch_slot_: id  0 | task 255 | processing task, is_child = 0
+0.32.433.278 W slot update_slots: id  0 | task 255 | n_past = 406, slot.prompt.tokens.size() = 441, seq_id = 0, pos_min = 440, n_swa = 0
+0.32.433.279 I slot update_slots: id  0 | task 255 | Checking checkpoint with [420, 420] against 406...
+0.32.433.280 W slot update_slots: id  0 | task 255 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.32.433.284 W slot update_slots: id  0 | task 255 | erased invalidated context checkpoint (pos_min = 420, pos_max = 420, n_tokens = 421, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.967.808 I slot create_check: id  0 | task 255 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.33.683.392 I reasoning-budget: deactivated (natural end)
+0.34.620.880 I slot print_timing: id  0 | task 255 | 
+prompt eval time =     589.92 ms /   422 tokens (    1.40 ms per token,   715.35 tokens per second)
+       eval time =    1597.65 ms /    69 tokens (   23.15 ms per token,    43.19 tokens per second)
+      total time =    2187.57 ms /   491 tokens
+0.34.621.091 I slot      release: id  0 | task 255 | stop processing: n_tokens = 490, truncated = 0
+0.34.621.151 I srv  update_slots: all slots are idle
+0.34.685.806 I srv  params_from_: Chat format: peg-native
+0.34.687.624 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.854 (> 0.100 thold), f_keep = 0.859
+0.34.688.188 I reasoning-budget: activated, budget=2147483647 tokens
+0.34.688.287 I slot launch_slot_: id  0 | task 326 | processing task, is_child = 0
+0.34.688.310 W slot update_slots: id  0 | task 326 | n_past = 421, slot.prompt.tokens.size() = 490, seq_id = 0, pos_min = 489, n_swa = 0
+0.34.688.313 I slot update_slots: id  0 | task 326 | Checking checkpoint with [417, 417] against 421...
+0.34.696.358 W slot update_slots: id  0 | task 326 | restored context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_past = 418, size = 62.813 MiB)
+0.34.932.382 I slot create_check: id  0 | task 326 | created context checkpoint 2 of 32 (pos_min = 488, pos_max = 488, n_tokens = 489, size = 62.813 MiB)
+0.35.494.281 I reasoning-budget: deactivated (natural end)
+0.35.771.860 I slot print_timing: id  0 | task 326 | 
+prompt eval time =     289.45 ms /    75 tokens (    3.86 ms per token,   259.11 tokens per second)
+       eval time =     794.07 ms /    36 tokens (   22.06 ms per token,    45.34 tokens per second)
+      total time =    1083.52 ms /   111 tokens
+0.35.772.048 I slot      release: id  0 | task 326 | stop processing: n_tokens = 528, truncated = 0
+0.35.772.102 I srv  update_slots: all slots are idle
+0.35.790.981 I srv  params_from_: Chat format: peg-native
+0.35.791.515 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.972 (> 0.100 thold), f_keep = 0.777
+0.35.792.014 I reasoning-budget: activated, budget=2147483647 tokens
+0.35.792.106 I slot launch_slot_: id  0 | task 364 | processing task, is_child = 0
+0.35.792.127 W slot update_slots: id  0 | task 364 | n_past = 410, slot.prompt.tokens.size() = 528, seq_id = 0, pos_min = 527, n_swa = 0
+0.35.792.130 I slot update_slots: id  0 | task 364 | Checking checkpoint with [488, 488] against 410...
+0.35.792.132 I slot update_slots: id  0 | task 364 | Checking checkpoint with [417, 417] against 410...
+0.35.792.133 W slot update_slots: id  0 | task 364 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.35.792.140 W slot update_slots: id  0 | task 364 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.35.793.559 W slot update_slots: id  0 | task 364 | erased invalidated context checkpoint (pos_min = 488, pos_max = 488, n_tokens = 489, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.36.350.507 I slot create_check: id  0 | task 364 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.36.640.348 I reasoning-budget: deactivated (natural end)
+0.37.440.104 I slot print_timing: id  0 | task 364 | 
+prompt eval time =     591.78 ms /   422 tokens (    1.40 ms per token,   713.10 tokens per second)
+       eval time =    1056.18 ms /    52 tokens (   20.31 ms per token,    49.23 tokens per second)
+      total time =    1647.97 ms /   474 tokens
+0.37.440.166 I slot      release: id  0 | task 364 | stop processing: n_tokens = 473, truncated = 0
+0.37.440.190 I srv  update_slots: all slots are idle
+0.37.456.395 I srv  params_from_: Chat format: peg-native
+0.37.456.822 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.935 (> 0.100 thold), f_keep = 0.856
+0.37.457.040 I reasoning-budget: activated, budget=2147483647 tokens
+0.37.457.081 I slot launch_slot_: id  0 | task 418 | processing task, is_child = 0
+0.37.457.090 W slot update_slots: id  0 | task 418 | n_past = 405, slot.prompt.tokens.size() = 473, seq_id = 0, pos_min = 472, n_swa = 0
+0.37.457.092 I slot update_slots: id  0 | task 418 | Checking checkpoint with [417, 417] against 405...
+0.37.457.092 W slot update_slots: id  0 | task 418 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.37.457.095 W slot update_slots: id  0 | task 418 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.37.989.158 I slot create_check: id  0 | task 418 | created context checkpoint 1 of 32 (pos_min = 428, pos_max = 428, n_tokens = 429, size = 62.813 MiB)
+0.39.035.888 I reasoning-budget: deactivated (natural end)
+0.39.898.949 I slot print_timing: id  0 | task 418 | 
+prompt eval time =     571.43 ms /   433 tokens (    1.32 ms per token,   757.75 tokens per second)
+       eval time =    1870.42 ms /    92 tokens (   20.33 ms per token,    49.19 tokens per second)
+      total time =    2441.84 ms /   525 tokens
+0.39.899.036 I slot      release: id  0 | task 418 | stop processing: n_tokens = 524, truncated = 0
+0.39.899.069 I srv  update_slots: all slots are idle
+0.39.931.009 I srv  params_from_: Chat format: peg-native
+0.39.931.485 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.773
+0.39.932.043 I reasoning-budget: activated, budget=2147483647 tokens
+0.39.932.046 I reasoning-budget: deactivated (natural end)
+0.39.932.130 I slot launch_slot_: id  0 | task 512 | processing task, is_child = 0
+0.39.932.153 W slot update_slots: id  0 | task 512 | n_past = 405, slot.prompt.tokens.size() = 524, seq_id = 0, pos_min = 523, n_swa = 0
+0.39.932.156 I slot update_slots: id  0 | task 512 | Checking checkpoint with [428, 428] against 405...
+0.39.932.157 W slot update_slots: id  0 | task 512 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.39.932.162 W slot update_slots: id  0 | task 512 | erased invalidated context checkpoint (pos_min = 428, pos_max = 428, n_tokens = 429, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.40.516.543 I slot create_check: id  0 | task 512 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.41.428.023 I slot print_timing: id  0 | task 512 | 
+prompt eval time =     630.82 ms /   424 tokens (    1.49 ms per token,   672.15 tokens per second)
+       eval time =     865.02 ms /    39 tokens (   22.18 ms per token,    45.09 tokens per second)
+      total time =    1495.84 ms /   463 tokens
+0.41.428.279 I slot      release: id  0 | task 512 | stop processing: n_tokens = 462, truncated = 0
+0.41.428.346 I srv  update_slots: all slots are idle
+0.41.471.490 I srv  params_from_: Chat format: peg-native
+0.41.472.003 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.902 (> 0.100 thold), f_keep = 0.877
+0.41.472.302 I reasoning-budget: activated, budget=2147483647 tokens
+0.41.472.304 I reasoning-budget: deactivated (natural end)
+0.41.472.355 I slot launch_slot_: id  0 | task 553 | processing task, is_child = 0
+0.41.472.368 W slot update_slots: id  0 | task 553 | n_past = 405, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.41.472.369 I slot update_slots: id  0 | task 553 | Checking checkpoint with [419, 419] against 405...
+0.41.472.370 W slot update_slots: id  0 | task 553 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.41.472.373 W slot update_slots: id  0 | task 553 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.42.065.024 I slot create_check: id  0 | task 553 | created context checkpoint 1 of 32 (pos_min = 444, pos_max = 444, n_tokens = 445, size = 62.813 MiB)
+0.44.104.338 I slot print_timing: id  0 | task 553 | 
+prompt eval time =     633.39 ms /   449 tokens (    1.41 ms per token,   708.88 tokens per second)
+       eval time =    1998.54 ms /    86 tokens (   23.24 ms per token,    43.03 tokens per second)
+      total time =    2631.93 ms /   535 tokens
+0.44.104.526 I slot      release: id  0 | task 553 | stop processing: n_tokens = 534, truncated = 0
+0.44.104.644 I srv  update_slots: all slots are idle
+0.44.130.603 I srv  params_from_: Chat format: peg-native
+0.44.130.995 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.948 (> 0.100 thold), f_keep = 0.758
+0.44.131.303 I reasoning-budget: activated, budget=2147483647 tokens
+0.44.131.305 I reasoning-budget: deactivated (natural end)
+0.44.131.351 I slot launch_slot_: id  0 | task 641 | processing task, is_child = 0
+0.44.131.364 W slot update_slots: id  0 | task 641 | n_past = 405, slot.prompt.tokens.size() = 534, seq_id = 0, pos_min = 533, n_swa = 0
+0.44.131.365 I slot update_slots: id  0 | task 641 | Checking checkpoint with [444, 444] against 405...
+0.44.131.366 W slot update_slots: id  0 | task 641 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.44.131.369 W slot update_slots: id  0 | task 641 | erased invalidated context checkpoint (pos_min = 444, pos_max = 444, n_tokens = 445, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.44.685.484 I slot create_check: id  0 | task 641 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.45.671.889 I slot print_timing: id  0 | task 641 | 
+prompt eval time =     602.32 ms /   427 tokens (    1.41 ms per token,   708.93 tokens per second)
+       eval time =     938.19 ms /    39 tokens (   24.06 ms per token,    41.57 tokens per second)
+      total time =    1540.51 ms /   466 tokens
+0.45.671.971 I slot      release: id  0 | task 641 | stop processing: n_tokens = 465, truncated = 0
+0.45.671.999 I srv  update_slots: all slots are idle
+0.45.686.558 I srv  params_from_: Chat format: peg-native
+0.45.686.878 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.948 (> 0.100 thold), f_keep = 0.871
+0.45.687.125 I reasoning-budget: activated, budget=2147483647 tokens
+0.45.687.126 I reasoning-budget: deactivated (natural end)
+0.45.687.183 I slot launch_slot_: id  0 | task 682 | processing task, is_child = 0
+0.45.687.196 W slot update_slots: id  0 | task 682 | n_past = 405, slot.prompt.tokens.size() = 465, seq_id = 0, pos_min = 464, n_swa = 0
+0.45.687.197 I slot update_slots: id  0 | task 682 | Checking checkpoint with [422, 422] against 405...
+0.45.687.199 W slot update_slots: id  0 | task 682 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.45.687.202 W slot update_slots: id  0 | task 682 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.46.216.440 I slot create_check: id  0 | task 682 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.46.339.458 I slot print_timing: id  0 | task 682 | 
+prompt eval time =     571.41 ms /   427 tokens (    1.34 ms per token,   747.28 tokens per second)
+       eval time =      80.84 ms /     4 tokens (   20.21 ms per token,    49.48 tokens per second)
+      total time =     652.25 ms /   431 tokens
+0.46.339.548 I slot      release: id  0 | task 682 | stop processing: n_tokens = 430, truncated = 0
+0.46.339.578 I srv  update_slots: all slots are idle
+0.46.360.275 I srv  params_from_: Chat format: peg-native
+0.46.360.733 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.958 (> 0.100 thold), f_keep = 0.944
+0.46.361.269 I reasoning-budget: activated, budget=2147483647 tokens
+0.46.361.274 I reasoning-budget: deactivated (natural end)
+0.46.361.357 I slot launch_slot_: id  0 | task 688 | processing task, is_child = 0
+0.46.361.380 W slot update_slots: id  0 | task 688 | n_past = 406, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.46.361.383 I slot update_slots: id  0 | task 688 | Checking checkpoint with [422, 422] against 406...
+0.46.361.385 W slot update_slots: id  0 | task 688 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.46.361.392 W slot update_slots: id  0 | task 688 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.46.930.125 I slot create_check: id  0 | task 688 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.47.764.185 I slot print_timing: id  0 | task 688 | 
+prompt eval time =     611.34 ms /   424 tokens (    1.44 ms per token,   693.56 tokens per second)
+       eval time =     791.43 ms /    40 tokens (   19.79 ms per token,    50.54 tokens per second)
+      total time =    1402.77 ms /   464 tokens
+0.47.764.366 I slot      release: id  0 | task 688 | stop processing: n_tokens = 463, truncated = 0
+0.47.764.400 I srv  update_slots: all slots are idle
+0.47.803.235 I srv  params_from_: Chat format: peg-native
+0.47.803.673 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.935 (> 0.100 thold), f_keep = 1.000
+0.47.804.329 I reasoning-budget: activated, budget=2147483647 tokens
+0.47.804.335 I reasoning-budget: deactivated (natural end)
+0.47.804.467 I slot launch_slot_: id  0 | task 730 | processing task, is_child = 0
+0.48.002.540 I slot create_check: id  0 | task 730 | created context checkpoint 2 of 32 (pos_min = 490, pos_max = 490, n_tokens = 491, size = 62.813 MiB)
+0.48.372.317 I slot print_timing: id  0 | task 730 | 
+prompt eval time =     256.59 ms /    32 tokens (    8.02 ms per token,   124.71 tokens per second)
+       eval time =     311.22 ms /    14 tokens (   22.23 ms per token,    44.98 tokens per second)
+      total time =     567.81 ms /    46 tokens
+0.48.372.410 I slot      release: id  0 | task 730 | stop processing: n_tokens = 508, truncated = 0
+0.48.372.439 I srv  update_slots: all slots are idle
+0.48.412.286 I srv  params_from_: Chat format: peg-native
+0.48.412.805 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.967 (> 0.100 thold), f_keep = 0.807
+0.48.413.102 I reasoning-budget: activated, budget=2147483647 tokens
+0.48.413.104 I reasoning-budget: deactivated (natural end)
+0.48.413.157 I slot launch_slot_: id  0 | task 746 | processing task, is_child = 0
+0.48.413.170 W slot update_slots: id  0 | task 746 | n_past = 410, slot.prompt.tokens.size() = 508, seq_id = 0, pos_min = 507, n_swa = 0
+0.48.413.172 I slot update_slots: id  0 | task 746 | Checking checkpoint with [490, 490] against 410...
+0.48.413.172 I slot update_slots: id  0 | task 746 | Checking checkpoint with [419, 419] against 410...
+0.48.413.173 W slot update_slots: id  0 | task 746 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.48.413.176 W slot update_slots: id  0 | task 746 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.48.414.558 W slot update_slots: id  0 | task 746 | erased invalidated context checkpoint (pos_min = 490, pos_max = 490, n_tokens = 491, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.48.958.571 I slot create_check: id  0 | task 746 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.49.822.447 I slot print_timing: id  0 | task 746 | 
+prompt eval time =     585.25 ms /   424 tokens (    1.38 ms per token,   724.47 tokens per second)
+       eval time =     824.00 ms /    40 tokens (   20.60 ms per token,    48.54 tokens per second)
+      total time =    1409.26 ms /   464 tokens
+0.49.822.516 I slot      release: id  0 | task 746 | stop processing: n_tokens = 463, truncated = 0
+0.49.822.541 I srv  update_slots: all slots are idle
+0.49.837.403 I srv  params_from_: Chat format: peg-native
+0.49.837.865 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.931 (> 0.100 thold), f_keep = 0.875
+0.49.838.376 I reasoning-budget: activated, budget=2147483647 tokens
+0.49.838.380 I reasoning-budget: deactivated (natural end)
+0.49.838.465 I slot launch_slot_: id  0 | task 788 | processing task, is_child = 0
+0.49.838.487 W slot update_slots: id  0 | task 788 | n_past = 405, slot.prompt.tokens.size() = 463, seq_id = 0, pos_min = 462, n_swa = 0
+0.49.838.490 I slot update_slots: id  0 | task 788 | Checking checkpoint with [419, 419] against 405...
+0.49.838.491 W slot update_slots: id  0 | task 788 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.49.838.496 W slot update_slots: id  0 | task 788 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.50.402.186 I slot create_check: id  0 | task 788 | created context checkpoint 1 of 32 (pos_min = 430, pos_max = 430, n_tokens = 431, size = 62.813 MiB)
+0.52.152.175 I slot print_timing: id  0 | task 788 | 
+prompt eval time =     613.67 ms /   435 tokens (    1.41 ms per token,   708.85 tokens per second)
+       eval time =    1699.98 ms /    80 tokens (   21.25 ms per token,    47.06 tokens per second)
+      total time =    2313.65 ms /   515 tokens
+0.52.152.608 I slot      release: id  0 | task 788 | stop processing: n_tokens = 514, truncated = 0
+0.52.152.669 I srv  update_slots: all slots are idle
+0.52.154.330 I srv    operator(): operator(): cleaning up before exit...

recipe/logs/b_n-tools-q106-tpl-probe.log

@@ -0,0 +1,136 @@
+0.00.135.941 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.135.945 I device_info:
+0.00.136.023 I   - ROCm0   : AMD Radeon Graphics (131072 MiB, 122347 MiB free)
+0.00.136.132 I   - Vulkan0 : AMD Radeon Graphics (RADV GFX1151) (132096 MiB, 131922 MiB free)
+0.00.136.137 I   - CPU     : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S (127438 MiB, 127438 MiB free)
+0.00.136.208 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.136.238 I srv          init: running without SSL
+0.00.136.260 I srv          init: using 31 threads for HTTP server
+0.00.136.261 I srv          init: the WebUI is disabled
+0.00.136.327 I srv         start: binding port with default address family
+0.00.137.600 I srv          main: loading model
+0.00.137.603 I srv    load_model: loading model '/mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf'
+0.00.256.612 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.24.306.998 W llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.24.593.444 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.24.922.506 I srv    load_model: initializing slots, n_slots = 1
+0.25.146.190 W srv    load_model: speculative decoding will use checkpoints
+0.25.146.203 W common_speculative_init: no implementations specified for speculative decoding
+0.25.146.206 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 65536
+0.25.146.286 I srv    load_model: prompt cache RAM enabled: limit_mib=8192
+0.25.146.287 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
+0.25.146.331 I srv          init: idle slots will be saved to prompt cache upon starting a new task
+0.25.160.750 I init: chat template, example_format: '<|im_start|>system
+You are a helpful assistant<|im_end|>
+<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+Hi there<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+<think>
+'
+0.25.170.675 I srv          init: init: chat template, thinking = 1
+0.25.170.721 I srv          main: model loaded
+0.25.170.724 I srv          main: server is listening on http://127.0.0.1:18652
+0.25.170.745 I srv  update_slots: all slots are idle
+0.26.899.613 I srv  params_from_: Chat format: peg-native
+0.26.900.013 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.26.900.015 I srv  get_availabl: updating prompt cache
+0.26.900.021 I srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
+0.26.900.027 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 65536 tokens, 8589934592 est)
+0.26.900.029 I srv  get_availabl: prompt cache update took 0.01 ms
+0.26.900.306 I reasoning-budget: activated, budget=2147483647 tokens
+0.26.900.326 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.27.527.394 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 420, pos_max = 420, n_tokens = 421, size = 62.813 MiB)
+0.27.799.897 I reasoning-budget: deactivated (natural end)
+0.27.893.084 I slot print_timing: id  0 | task 0 | 
+prompt eval time =     660.83 ms /   425 tokens (    1.55 ms per token,   643.13 tokens per second)
+       eval time =     331.90 ms /    17 tokens (   19.52 ms per token,    51.22 tokens per second)
+      total time =     992.73 ms /   442 tokens
+0.27.893.160 I slot      release: id  0 | task 0 | stop processing: n_tokens = 441, truncated = 0
+0.27.893.171 I srv  update_slots: all slots are idle
+0.27.910.042 I srv  params_from_: Chat format: peg-native
+0.27.910.501 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.921
+0.27.910.926 I reasoning-budget: activated, budget=2147483647 tokens
+0.27.911.011 I slot launch_slot_: id  0 | task 19 | processing task, is_child = 0
+0.27.911.034 W slot update_slots: id  0 | task 19 | n_past = 406, slot.prompt.tokens.size() = 441, seq_id = 0, pos_min = 440, n_swa = 0
+0.27.911.036 I slot update_slots: id  0 | task 19 | Checking checkpoint with [420, 420] against 406...
+0.27.911.038 W slot update_slots: id  0 | task 19 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.27.911.042 W slot update_slots: id  0 | task 19 | erased invalidated context checkpoint (pos_min = 420, pos_max = 420, n_tokens = 421, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.28.555.134 I slot create_check: id  0 | task 19 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.28.851.175 I reasoning-budget: deactivated (natural end)
+0.29.588.543 I slot print_timing: id  0 | task 19 | 
+prompt eval time =     695.53 ms /   422 tokens (    1.65 ms per token,   606.73 tokens per second)
+       eval time =     981.97 ms /    53 tokens (   18.53 ms per token,    53.97 tokens per second)
+      total time =    1677.50 ms /   475 tokens
+0.29.588.615 I slot      release: id  0 | task 19 | stop processing: n_tokens = 474, truncated = 0
+0.29.588.642 I srv  update_slots: all slots are idle
+0.29.622.104 I srv  params_from_: Chat format: peg-native
+0.29.622.588 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.958 (> 0.100 thold), f_keep = 0.857
+0.29.623.144 I reasoning-budget: activated, budget=2147483647 tokens
+0.29.623.228 I slot launch_slot_: id  0 | task 74 | processing task, is_child = 0
+0.29.623.250 W slot update_slots: id  0 | task 74 | n_past = 406, slot.prompt.tokens.size() = 474, seq_id = 0, pos_min = 473, n_swa = 0
+0.29.623.253 I slot update_slots: id  0 | task 74 | Checking checkpoint with [417, 417] against 406...
+0.29.623.255 W slot update_slots: id  0 | task 74 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.29.623.260 W slot update_slots: id  0 | task 74 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.30.170.371 I slot create_check: id  0 | task 74 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.30.268.328 I reasoning-budget: deactivated (natural end)
+0.30.408.646 I slot print_timing: id  0 | task 74 | 
+prompt eval time =     594.93 ms /   424 tokens (    1.40 ms per token,   712.68 tokens per second)
+       eval time =     190.44 ms /     7 tokens (   27.21 ms per token,    36.76 tokens per second)
+      total time =     785.37 ms /   431 tokens
+0.30.408.832 I slot      release: id  0 | task 74 | stop processing: n_tokens = 430, truncated = 0
+0.30.408.910 I srv  update_slots: all slots are idle
+0.30.423.352 I srv  params_from_: Chat format: peg-native
+0.30.423.958 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.964 (> 0.100 thold), f_keep = 0.944
+0.30.424.428 I reasoning-budget: activated, budget=2147483647 tokens
+0.30.424.483 I slot launch_slot_: id  0 | task 83 | processing task, is_child = 0
+0.30.424.498 W slot update_slots: id  0 | task 83 | n_past = 406, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.30.424.499 I slot update_slots: id  0 | task 83 | Checking checkpoint with [419, 419] against 406...
+0.30.424.500 W slot update_slots: id  0 | task 83 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.30.424.505 W slot update_slots: id  0 | task 83 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.30.909.972 I slot create_check: id  0 | task 83 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.30.972.923 I reasoning-budget: deactivated (natural end)
+0.31.775.290 I slot print_timing: id  0 | task 83 | 
+prompt eval time =     521.51 ms /   421 tokens (    1.24 ms per token,   807.27 tokens per second)
+       eval time =     829.27 ms /    42 tokens (   19.74 ms per token,    50.65 tokens per second)
+      total time =    1350.78 ms /   463 tokens
+0.31.775.369 I slot      release: id  0 | task 83 | stop processing: n_tokens = 462, truncated = 0
+0.31.775.395 I srv  update_slots: all slots are idle
+0.31.814.609 I srv  params_from_: Chat format: peg-native
+0.31.815.041 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.951 (> 0.100 thold), f_keep = 0.879
+0.31.815.337 I reasoning-budget: activated, budget=2147483647 tokens
+0.31.815.385 I slot launch_slot_: id  0 | task 127 | processing task, is_child = 0
+0.31.815.397 W slot update_slots: id  0 | task 127 | n_past = 406, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.31.815.398 I slot update_slots: id  0 | task 127 | Checking checkpoint with [416, 416] against 406...
+0.31.815.399 W slot update_slots: id  0 | task 127 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.31.815.402 W slot update_slots: id  0 | task 127 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.362.167 I slot create_check: id  0 | task 127 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.32.500.485 I slot print_timing: id  0 | task 127 | 
+prompt eval time =     585.39 ms /   427 tokens (    1.37 ms per token,   729.43 tokens per second)
+       eval time =      99.68 ms /     4 tokens (   24.92 ms per token,    40.13 tokens per second)
+      total time =     685.07 ms /   431 tokens
+0.32.500.582 I slot      release: id  0 | task 127 | stop processing: n_tokens = 430, truncated = 0
+0.32.500.610 I srv  update_slots: all slots are idle
+0.32.535.431 I srv  params_from_: Chat format: peg-native
+0.32.535.929 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.958 (> 0.100 thold), f_keep = 0.944
+0.32.536.504 I reasoning-budget: activated, budget=2147483647 tokens
+0.32.536.587 I slot launch_slot_: id  0 | task 133 | processing task, is_child = 0
+0.32.536.611 W slot update_slots: id  0 | task 133 | n_past = 406, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.32.536.614 I slot update_slots: id  0 | task 133 | Checking checkpoint with [422, 422] against 406...
+0.32.536.616 W slot update_slots: id  0 | task 133 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.32.536.622 W slot update_slots: id  0 | task 133 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.33.112.942 I slot create_check: id  0 | task 133 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.34.067.962 I slot print_timing: id  0 | task 133 | 
+prompt eval time =     615.83 ms /   424 tokens (    1.45 ms per token,   688.51 tokens per second)
+       eval time =     915.49 ms /    39 tokens (   23.47 ms per token,    42.60 tokens per second)
+      total time =    1531.32 ms /   463 tokens
+0.34.068.202 I slot      release: id  0 | task 133 | stop processing: n_tokens = 462, truncated = 0
+0.34.068.265 I srv  update_slots: all slots are idle
+0.34.070.064 I srv    operator(): operator(): cleaning up before exit...

recipe/logs/b_n-tools-q106-tpl.log

@@ -0,0 +1,299 @@
+0.00.148.040 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.148.048 I device_info:
+0.00.148.185 I   - ROCm0   : AMD Radeon Graphics (131072 MiB, 122384 MiB free)
+0.00.148.371 I   - Vulkan0 : AMD Radeon Graphics (RADV GFX1151) (132096 MiB, 131922 MiB free)
+0.00.148.380 I   - CPU     : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S (127438 MiB, 127438 MiB free)
+0.00.148.475 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.148.535 I srv          init: running without SSL
+0.00.148.567 I srv          init: using 31 threads for HTTP server
+0.00.148.569 I srv          init: the WebUI is disabled
+0.00.148.676 I srv         start: binding port with default address family
+0.00.149.917 I srv          main: loading model
+0.00.149.927 I srv    load_model: loading model '/mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf'
+0.00.318.164 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.24.651.499 W llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.24.848.592 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.25.156.208 I srv    load_model: initializing slots, n_slots = 1
+0.25.370.167 W srv    load_model: speculative decoding will use checkpoints
+0.25.370.182 W common_speculative_init: no implementations specified for speculative decoding
+0.25.370.188 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 65536
+0.25.370.351 I srv    load_model: prompt cache RAM enabled: limit_mib=8192
+0.25.370.398 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
+0.25.370.453 I srv          init: idle slots will be saved to prompt cache upon starting a new task
+0.25.419.174 I init: chat template, example_format: '<|im_start|>system
+You are a helpful assistant<|im_end|>
+<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+Hi there<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+<think>
+'
+0.25.441.368 I srv          init: init: chat template, thinking = 1
+0.25.441.453 I srv          main: model loaded
+0.25.441.460 I srv          main: server is listening on http://127.0.0.1:18600
+0.25.441.468 I srv  update_slots: all slots are idle
+0.26.789.568 I srv  params_from_: Chat format: peg-native
+0.26.789.991 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.26.789.993 I srv  get_availabl: updating prompt cache
+0.26.790.000 I srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
+0.26.790.006 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 65536 tokens, 8589934592 est)
+0.26.790.010 I srv  get_availabl: prompt cache update took 0.01 ms
+0.26.790.277 I reasoning-budget: activated, budget=2147483647 tokens
+0.26.790.294 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.27.476.977 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.27.881.416 I reasoning-budget: deactivated (natural end)
+0.28.806.104 I slot print_timing: id  0 | task 0 | 
+prompt eval time =     751.68 ms /   422 tokens (    1.78 ms per token,   561.41 tokens per second)
+       eval time =    1264.10 ms /    52 tokens (   24.31 ms per token,    41.14 tokens per second)
+      total time =    2015.78 ms /   474 tokens
+0.28.806.213 I slot      release: id  0 | task 0 | stop processing: n_tokens = 473, truncated = 0
+0.28.806.225 I srv  update_slots: all slots are idle
+0.28.824.839 I srv  params_from_: Chat format: peg-native
+0.28.825.237 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.906 (> 0.100 thold), f_keep = 0.856
+0.28.825.422 I reasoning-budget: activated, budget=2147483647 tokens
+0.28.825.471 I slot launch_slot_: id  0 | task 54 | processing task, is_child = 0
+0.28.825.483 W slot update_slots: id  0 | task 54 | n_past = 405, slot.prompt.tokens.size() = 473, seq_id = 0, pos_min = 472, n_swa = 0
+0.28.825.485 I slot update_slots: id  0 | task 54 | Checking checkpoint with [417, 417] against 405...
+0.28.825.486 W slot update_slots: id  0 | task 54 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.28.825.489 W slot update_slots: id  0 | task 54 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.29.247.541 I slot create_check: id  0 | task 54 | created context checkpoint 1 of 32 (pos_min = 442, pos_max = 442, n_tokens = 443, size = 62.813 MiB)
+0.29.847.933 I reasoning-budget: deactivated (natural end)
+0.31.561.799 I slot print_timing: id  0 | task 54 | n_decoded =    100, tg =  43.85 t/s
+0.31.768.383 I slot print_timing: id  0 | task 54 | 
+prompt eval time =     455.75 ms /   447 tokens (    1.02 ms per token,   980.80 tokens per second)
+       eval time =    2487.14 ms /   110 tokens (   22.61 ms per token,    44.23 tokens per second)
+      total time =    2942.89 ms /   557 tokens
+0.31.768.453 I slot      release: id  0 | task 54 | stop processing: n_tokens = 556, truncated = 0
+0.31.768.488 I srv  update_slots: all slots are idle
+0.31.780.986 I srv  params_from_: Chat format: peg-native
+0.31.781.515 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.953 (> 0.100 thold), f_keep = 0.728
+0.31.781.866 I reasoning-budget: activated, budget=2147483647 tokens
+0.31.781.944 I slot launch_slot_: id  0 | task 166 | processing task, is_child = 0
+0.31.781.959 W slot update_slots: id  0 | task 166 | n_past = 405, slot.prompt.tokens.size() = 556, seq_id = 0, pos_min = 555, n_swa = 0
+0.31.781.960 I slot update_slots: id  0 | task 166 | Checking checkpoint with [442, 442] against 405...
+0.31.781.961 W slot update_slots: id  0 | task 166 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.31.781.965 W slot update_slots: id  0 | task 166 | erased invalidated context checkpoint (pos_min = 442, pos_max = 442, n_tokens = 443, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.284.242 I slot create_check: id  0 | task 166 | created context checkpoint 1 of 32 (pos_min = 420, pos_max = 420, n_tokens = 421, size = 62.813 MiB)
+0.32.618.089 I reasoning-budget: deactivated (natural end)
+0.33.467.501 I slot print_timing: id  0 | task 166 | 
+prompt eval time =     568.15 ms /   425 tokens (    1.34 ms per token,   748.05 tokens per second)
+       eval time =    1117.36 ms /    52 tokens (   21.49 ms per token,    46.54 tokens per second)
+      total time =    1685.50 ms /   477 tokens
+0.33.467.767 I slot      release: id  0 | task 166 | stop processing: n_tokens = 476, truncated = 0
+0.33.467.836 I srv  update_slots: all slots are idle
+0.33.510.830 I srv  params_from_: Chat format: peg-native
+0.33.511.252 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.953 (> 0.100 thold), f_keep = 0.851
+0.33.511.444 I reasoning-budget: activated, budget=2147483647 tokens
+0.33.511.478 I slot launch_slot_: id  0 | task 220 | processing task, is_child = 0
+0.33.511.488 W slot update_slots: id  0 | task 220 | n_past = 405, slot.prompt.tokens.size() = 476, seq_id = 0, pos_min = 475, n_swa = 0
+0.33.511.489 I slot update_slots: id  0 | task 220 | Checking checkpoint with [420, 420] against 405...
+0.33.511.490 W slot update_slots: id  0 | task 220 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.33.511.494 W slot update_slots: id  0 | task 220 | erased invalidated context checkpoint (pos_min = 420, pos_max = 420, n_tokens = 421, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.34.023.181 I slot create_check: id  0 | task 220 | created context checkpoint 1 of 32 (pos_min = 420, pos_max = 420, n_tokens = 421, size = 62.813 MiB)
+0.34.447.908 I reasoning-budget: deactivated (natural end)
+0.34.540.055 I slot print_timing: id  0 | task 220 | 
+prompt eval time =     555.59 ms /   425 tokens (    1.31 ms per token,   764.95 tokens per second)
+       eval time =     472.93 ms /    23 tokens (   20.56 ms per token,    48.63 tokens per second)
+      total time =    1028.52 ms /   448 tokens
+0.34.540.230 I slot      release: id  0 | task 220 | stop processing: n_tokens = 447, truncated = 0
+0.34.540.267 I srv  update_slots: all slots are idle
+0.34.592.769 I srv  params_from_: Chat format: peg-native
+0.34.593.328 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.908
+0.34.593.568 I reasoning-budget: activated, budget=2147483647 tokens
+0.34.593.618 I slot launch_slot_: id  0 | task 245 | processing task, is_child = 0
+0.34.593.629 W slot update_slots: id  0 | task 245 | n_past = 406, slot.prompt.tokens.size() = 447, seq_id = 0, pos_min = 446, n_swa = 0
+0.34.593.631 I slot update_slots: id  0 | task 245 | Checking checkpoint with [420, 420] against 406...
+0.34.593.632 W slot update_slots: id  0 | task 245 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.34.593.635 W slot update_slots: id  0 | task 245 | erased invalidated context checkpoint (pos_min = 420, pos_max = 420, n_tokens = 421, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.35.150.983 I slot create_check: id  0 | task 245 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.35.748.562 I reasoning-budget: deactivated (natural end)
+0.36.564.011 I slot print_timing: id  0 | task 245 | 
+prompt eval time =     621.88 ms /   422 tokens (    1.47 ms per token,   678.58 tokens per second)
+       eval time =    1348.48 ms /    64 tokens (   21.07 ms per token,    47.46 tokens per second)
+      total time =    1970.37 ms /   486 tokens
+0.36.564.097 I slot      release: id  0 | task 245 | stop processing: n_tokens = 485, truncated = 0
+0.36.564.126 I srv  update_slots: all slots are idle
+0.36.618.709 I srv  params_from_: Chat format: peg-native
+0.36.619.282 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.854 (> 0.100 thold), f_keep = 0.868
+0.36.619.528 I reasoning-budget: activated, budget=2147483647 tokens
+0.36.619.573 I slot launch_slot_: id  0 | task 311 | processing task, is_child = 0
+0.36.619.585 W slot update_slots: id  0 | task 311 | n_past = 421, slot.prompt.tokens.size() = 485, seq_id = 0, pos_min = 484, n_swa = 0
+0.36.619.587 I slot update_slots: id  0 | task 311 | Checking checkpoint with [417, 417] against 421...
+0.36.623.567 W slot update_slots: id  0 | task 311 | restored context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_past = 418, size = 62.813 MiB)
+0.36.859.461 I slot create_check: id  0 | task 311 | created context checkpoint 2 of 32 (pos_min = 488, pos_max = 488, n_tokens = 489, size = 62.813 MiB)
+0.37.463.364 I reasoning-budget: deactivated (natural end)
+0.37.731.529 I slot print_timing: id  0 | task 311 | 
+prompt eval time =     287.07 ms /    75 tokens (    3.83 ms per token,   261.26 tokens per second)
+       eval time =     824.84 ms /    39 tokens (   21.15 ms per token,    47.28 tokens per second)
+      total time =    1111.91 ms /   114 tokens
+0.37.731.631 I slot      release: id  0 | task 311 | stop processing: n_tokens = 531, truncated = 0
+0.37.731.658 I srv  update_slots: all slots are idle
+0.37.774.243 I srv  params_from_: Chat format: peg-native
+0.37.774.624 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.972 (> 0.100 thold), f_keep = 0.772
+0.37.774.821 I reasoning-budget: activated, budget=2147483647 tokens
+0.37.774.854 I slot launch_slot_: id  0 | task 352 | processing task, is_child = 0
+0.37.774.864 W slot update_slots: id  0 | task 352 | n_past = 410, slot.prompt.tokens.size() = 531, seq_id = 0, pos_min = 530, n_swa = 0
+0.37.774.865 I slot update_slots: id  0 | task 352 | Checking checkpoint with [488, 488] against 410...
+0.37.774.866 I slot update_slots: id  0 | task 352 | Checking checkpoint with [417, 417] against 410...
+0.37.774.867 W slot update_slots: id  0 | task 352 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.37.774.870 W slot update_slots: id  0 | task 352 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.37.775.796 W slot update_slots: id  0 | task 352 | erased invalidated context checkpoint (pos_min = 488, pos_max = 488, n_tokens = 489, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.38.322.268 I slot create_check: id  0 | task 352 | created context checkpoint 1 of 32 (pos_min = 417, pos_max = 417, n_tokens = 418, size = 62.813 MiB)
+0.38.647.904 I reasoning-budget: deactivated (natural end)
+0.39.470.609 I slot print_timing: id  0 | task 352 | 
+prompt eval time =     598.33 ms /   422 tokens (    1.42 ms per token,   705.29 tokens per second)
+       eval time =    1097.37 ms /    54 tokens (   20.32 ms per token,    49.21 tokens per second)
+      total time =    1695.70 ms /   476 tokens
+0.39.470.815 I slot      release: id  0 | task 352 | stop processing: n_tokens = 475, truncated = 0
+0.39.470.868 I srv  update_slots: all slots are idle
+0.39.526.260 I srv  params_from_: Chat format: peg-native
+0.39.528.818 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.935 (> 0.100 thold), f_keep = 0.853
+0.39.529.421 I reasoning-budget: activated, budget=2147483647 tokens
+0.39.529.515 I slot launch_slot_: id  0 | task 408 | processing task, is_child = 0
+0.39.529.535 W slot update_slots: id  0 | task 408 | n_past = 405, slot.prompt.tokens.size() = 475, seq_id = 0, pos_min = 474, n_swa = 0
+0.39.529.537 I slot update_slots: id  0 | task 408 | Checking checkpoint with [417, 417] against 405...
+0.39.529.539 W slot update_slots: id  0 | task 408 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.39.529.547 W slot update_slots: id  0 | task 408 | erased invalidated context checkpoint (pos_min = 417, pos_max = 417, n_tokens = 418, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.40.127.515 I slot create_check: id  0 | task 408 | created context checkpoint 1 of 32 (pos_min = 428, pos_max = 428, n_tokens = 429, size = 62.813 MiB)
+0.40.699.453 I reasoning-budget: deactivated (natural end)
+0.42.177.715 I slot print_timing: id  0 | task 408 | n_decoded =    100, tg =  49.81 t/s
+0.42.310.602 I slot print_timing: id  0 | task 408 | 
+prompt eval time =     640.38 ms /   433 tokens (    1.48 ms per token,   676.16 tokens per second)
+       eval time =    2140.67 ms /   107 tokens (   20.01 ms per token,    49.98 tokens per second)
+      total time =    2781.05 ms /   540 tokens
+0.42.310.711 I slot      release: id  0 | task 408 | stop processing: n_tokens = 539, truncated = 0
+0.42.310.745 I srv  update_slots: all slots are idle
+0.42.345.777 I srv  params_from_: Chat format: peg-native
+0.42.346.201 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.751
+0.42.346.446 I reasoning-budget: activated, budget=2147483647 tokens
+0.42.346.448 I reasoning-budget: deactivated (natural end)
+0.42.346.491 I slot launch_slot_: id  0 | task 517 | processing task, is_child = 0
+0.42.346.503 W slot update_slots: id  0 | task 517 | n_past = 405, slot.prompt.tokens.size() = 539, seq_id = 0, pos_min = 538, n_swa = 0
+0.42.346.504 I slot update_slots: id  0 | task 517 | Checking checkpoint with [428, 428] against 405...
+0.42.346.506 W slot update_slots: id  0 | task 517 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.42.346.509 W slot update_slots: id  0 | task 517 | erased invalidated context checkpoint (pos_min = 428, pos_max = 428, n_tokens = 429, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.42.884.527 I slot create_check: id  0 | task 517 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.43.790.791 I slot print_timing: id  0 | task 517 | 
+prompt eval time =     576.47 ms /   424 tokens (    1.36 ms per token,   735.51 tokens per second)
+       eval time =     867.80 ms /    39 tokens (   22.25 ms per token,    44.94 tokens per second)
+      total time =    1444.27 ms /   463 tokens
+0.43.790.894 I slot      release: id  0 | task 517 | stop processing: n_tokens = 462, truncated = 0
+0.43.790.929 I srv  update_slots: all slots are idle
+0.43.803.670 I srv  params_from_: Chat format: peg-native
+0.43.804.041 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.902 (> 0.100 thold), f_keep = 0.877
+0.43.804.257 I reasoning-budget: activated, budget=2147483647 tokens
+0.43.804.260 I reasoning-budget: deactivated (natural end)
+0.43.804.297 I slot launch_slot_: id  0 | task 558 | processing task, is_child = 0
+0.43.804.307 W slot update_slots: id  0 | task 558 | n_past = 405, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.43.804.307 I slot update_slots: id  0 | task 558 | Checking checkpoint with [419, 419] against 405...
+0.43.804.308 W slot update_slots: id  0 | task 558 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.43.804.311 W slot update_slots: id  0 | task 558 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.44.348.687 I slot create_check: id  0 | task 558 | created context checkpoint 1 of 32 (pos_min = 444, pos_max = 444, n_tokens = 445, size = 62.813 MiB)
+0.46.245.975 I slot print_timing: id  0 | task 558 | 
+prompt eval time =     596.97 ms /   449 tokens (    1.33 ms per token,   752.13 tokens per second)
+       eval time =    1844.68 ms /    86 tokens (   21.45 ms per token,    46.62 tokens per second)
+      total time =    2441.65 ms /   535 tokens
+0.46.246.065 I slot      release: id  0 | task 558 | stop processing: n_tokens = 534, truncated = 0
+0.46.246.093 I srv  update_slots: all slots are idle
+0.46.272.193 I srv  params_from_: Chat format: peg-native
+0.46.272.663 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.948 (> 0.100 thold), f_keep = 0.758
+0.46.273.230 I reasoning-budget: activated, budget=2147483647 tokens
+0.46.273.238 I reasoning-budget: deactivated (natural end)
+0.46.273.319 I slot launch_slot_: id  0 | task 646 | processing task, is_child = 0
+0.46.273.341 W slot update_slots: id  0 | task 646 | n_past = 405, slot.prompt.tokens.size() = 534, seq_id = 0, pos_min = 533, n_swa = 0
+0.46.273.345 I slot update_slots: id  0 | task 646 | Checking checkpoint with [444, 444] against 405...
+0.46.273.347 W slot update_slots: id  0 | task 646 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.46.273.352 W slot update_slots: id  0 | task 646 | erased invalidated context checkpoint (pos_min = 444, pos_max = 444, n_tokens = 445, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.46.832.222 I slot create_check: id  0 | task 646 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.47.645.063 I slot print_timing: id  0 | task 646 | 
+prompt eval time =     593.12 ms /   427 tokens (    1.39 ms per token,   719.92 tokens per second)
+       eval time =     778.59 ms /    39 tokens (   19.96 ms per token,    50.09 tokens per second)
+      total time =    1371.71 ms /   466 tokens
+0.47.645.151 I slot      release: id  0 | task 646 | stop processing: n_tokens = 465, truncated = 0
+0.47.645.177 I srv  update_slots: all slots are idle
+0.47.664.123 I srv  params_from_: Chat format: peg-native
+0.47.664.583 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.948 (> 0.100 thold), f_keep = 0.871
+0.47.665.092 I reasoning-budget: activated, budget=2147483647 tokens
+0.47.665.095 I reasoning-budget: deactivated (natural end)
+0.47.665.179 I slot launch_slot_: id  0 | task 687 | processing task, is_child = 0
+0.47.665.201 W slot update_slots: id  0 | task 687 | n_past = 405, slot.prompt.tokens.size() = 465, seq_id = 0, pos_min = 464, n_swa = 0
+0.47.665.204 I slot update_slots: id  0 | task 687 | Checking checkpoint with [422, 422] against 405...
+0.47.665.206 W slot update_slots: id  0 | task 687 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.47.665.212 W slot update_slots: id  0 | task 687 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.48.219.459 I slot create_check: id  0 | task 687 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.48.341.963 I slot print_timing: id  0 | task 687 | 
+prompt eval time =     588.36 ms /   427 tokens (    1.38 ms per token,   725.75 tokens per second)
+       eval time =      88.39 ms /     4 tokens (   22.10 ms per token,    45.25 tokens per second)
+      total time =     676.75 ms /   431 tokens
+0.48.342.059 I slot      release: id  0 | task 687 | stop processing: n_tokens = 430, truncated = 0
+0.48.342.091 I srv  update_slots: all slots are idle
+0.48.357.137 I srv  params_from_: Chat format: peg-native
+0.48.357.617 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.958 (> 0.100 thold), f_keep = 0.944
+0.48.358.008 I reasoning-budget: activated, budget=2147483647 tokens
+0.48.358.014 I reasoning-budget: deactivated (natural end)
+0.48.358.075 I slot launch_slot_: id  0 | task 693 | processing task, is_child = 0
+0.48.358.093 W slot update_slots: id  0 | task 693 | n_past = 406, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.48.358.095 I slot update_slots: id  0 | task 693 | Checking checkpoint with [422, 422] against 406...
+0.48.358.096 W slot update_slots: id  0 | task 693 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.48.358.101 W slot update_slots: id  0 | task 693 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.49.035.367 I slot create_check: id  0 | task 693 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.49.869.527 I slot print_timing: id  0 | task 693 | 
+prompt eval time =     708.91 ms /   424 tokens (    1.67 ms per token,   598.10 tokens per second)
+       eval time =     802.48 ms /    40 tokens (   20.06 ms per token,    49.85 tokens per second)
+      total time =    1511.39 ms /   464 tokens
+0.49.869.785 I slot      release: id  0 | task 693 | stop processing: n_tokens = 463, truncated = 0
+0.49.869.856 I srv  update_slots: all slots are idle
+0.49.916.557 I srv  params_from_: Chat format: peg-native
+0.49.916.971 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.935 (> 0.100 thold), f_keep = 1.000
+0.49.917.213 I reasoning-budget: activated, budget=2147483647 tokens
+0.49.917.215 I reasoning-budget: deactivated (natural end)
+0.49.917.257 I slot launch_slot_: id  0 | task 735 | processing task, is_child = 0
+0.50.095.825 I slot create_check: id  0 | task 735 | created context checkpoint 2 of 32 (pos_min = 490, pos_max = 490, n_tokens = 491, size = 62.813 MiB)
+0.50.567.914 I slot print_timing: id  0 | task 735 | 
+prompt eval time =     242.43 ms /    32 tokens (    7.58 ms per token,   132.00 tokens per second)
+       eval time =     408.20 ms /    15 tokens (   27.21 ms per token,    36.75 tokens per second)
+      total time =     650.63 ms /    47 tokens
+0.50.567.999 I slot      release: id  0 | task 735 | stop processing: n_tokens = 509, truncated = 0
+0.50.568.024 I srv  update_slots: all slots are idle
+0.50.583.295 I srv  params_from_: Chat format: peg-native
+0.50.584.124 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.967 (> 0.100 thold), f_keep = 0.806
+0.50.584.432 I reasoning-budget: activated, budget=2147483647 tokens
+0.50.584.436 I reasoning-budget: deactivated (natural end)
+0.50.584.488 I slot launch_slot_: id  0 | task 752 | processing task, is_child = 0
+0.50.584.503 W slot update_slots: id  0 | task 752 | n_past = 410, slot.prompt.tokens.size() = 509, seq_id = 0, pos_min = 508, n_swa = 0
+0.50.584.505 I slot update_slots: id  0 | task 752 | Checking checkpoint with [490, 490] against 410...
+0.50.584.506 I slot update_slots: id  0 | task 752 | Checking checkpoint with [419, 419] against 410...
+0.50.584.507 W slot update_slots: id  0 | task 752 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.50.584.510 W slot update_slots: id  0 | task 752 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.50.585.691 W slot update_slots: id  0 | task 752 | erased invalidated context checkpoint (pos_min = 490, pos_max = 490, n_tokens = 491, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.51.126.714 I slot create_check: id  0 | task 752 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.52.045.755 I slot print_timing: id  0 | task 752 | 
+prompt eval time =     592.34 ms /   424 tokens (    1.40 ms per token,   715.81 tokens per second)
+       eval time =     868.87 ms /    40 tokens (   21.72 ms per token,    46.04 tokens per second)
+      total time =    1461.21 ms /   464 tokens
+0.52.045.943 I slot      release: id  0 | task 752 | stop processing: n_tokens = 463, truncated = 0
+0.52.045.996 I srv  update_slots: all slots are idle
+0.52.104.739 I srv  params_from_: Chat format: peg-native
+0.52.107.431 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.931 (> 0.100 thold), f_keep = 0.875
+0.52.108.068 I reasoning-budget: activated, budget=2147483647 tokens
+0.52.108.073 I reasoning-budget: deactivated (natural end)
+0.52.108.179 I slot launch_slot_: id  0 | task 794 | processing task, is_child = 0
+0.52.108.202 W slot update_slots: id  0 | task 794 | n_past = 405, slot.prompt.tokens.size() = 463, seq_id = 0, pos_min = 462, n_swa = 0
+0.52.108.206 I slot update_slots: id  0 | task 794 | Checking checkpoint with [419, 419] against 405...
+0.52.108.208 W slot update_slots: id  0 | task 794 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.52.108.217 W slot update_slots: id  0 | task 794 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.52.715.448 I slot create_check: id  0 | task 794 | created context checkpoint 1 of 32 (pos_min = 430, pos_max = 430, n_tokens = 431, size = 62.813 MiB)
+0.54.554.592 I slot print_timing: id  0 | task 794 | 
+prompt eval time =     642.08 ms /   435 tokens (    1.48 ms per token,   677.48 tokens per second)
+       eval time =    1804.29 ms /    80 tokens (   22.55 ms per token,    44.34 tokens per second)
+      total time =    2446.37 ms /   515 tokens
+0.54.554.809 I slot      release: id  0 | task 794 | stop processing: n_tokens = 514, truncated = 0
+0.54.554.839 I srv  update_slots: all slots are idle
+0.54.555.755 I srv    operator(): operator(): cleaning up before exit...

recipe/logs/b_n-vision-q106-c1-faon.log

@@ -0,0 +1,71 @@
+0.00.059.279 W Setting 'enable_thinking' via --chat-template-kwargs is deprecated. Use --reasoning on / --reasoning off instead.
+0.00.069.767 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.069.776 I device_info:
+0.00.069.939 I   - ROCm0   : AMD Radeon Graphics (131072 MiB, 123862 MiB free)
+0.00.070.129 I   - Vulkan0 : AMD Radeon Graphics (RADV GFX1151) (132096 MiB, 131922 MiB free)
+0.00.070.138 I   - CPU     : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S (127438 MiB, 127438 MiB free)
+0.00.070.226 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.070.307 I srv          init: running without SSL
+0.00.070.398 I srv          init: using 31 threads for HTTP server
+0.00.070.401 I srv          init: the WebUI is disabled
+0.00.070.610 I srv         start: binding port with default address family
+0.00.071.886 I srv          main: loading model
+0.00.071.903 I srv    load_model: loading model '/mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf'
+0.00.118.721 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.21.623.100 W llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.21.824.309 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.22.126.084 W load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
+0.22.126.088 W load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
+0.22.126.088 W load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842
+
+0.22.393.368 I srv    load_model: loaded multimodal model, '/mnt/models/nex-n2.5-mini/out/mmproj-Nex-N2.5-mini-BF16.gguf'
+0.22.393.383 I srv    load_model: initializing slots, n_slots = 1
+0.22.705.455 W srv    load_model: speculative decoding will use checkpoints
+0.22.705.465 W common_speculative_init: no implementations specified for speculative decoding
+0.22.705.468 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 65536
+0.22.705.551 I srv    load_model: prompt cache RAM enabled: limit_mib=8192
+0.22.705.569 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
+0.22.705.622 I srv          init: idle slots will be saved to prompt cache upon starting a new task
+0.22.759.943 I init: chat template, example_format: '<|im_start|>system
+You are a helpful assistant<|im_end|>
+<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+Hi there<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+'
+0.22.797.944 I srv          init: init: chat template, thinking = 1
+0.22.798.029 I srv          main: model loaded
+0.22.798.044 I srv          main: server is listening on http://127.0.0.1:18600
+0.22.798.080 I srv  update_slots: all slots are idle
+0.23.997.761 I srv  params_from_: Chat format: peg-native
+0.23.998.009 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.23.998.016 I srv  get_availabl: updating prompt cache
+0.23.998.027 I srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
+0.23.998.035 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 65536 tokens, 8589934592 est)
+0.23.998.038 I srv  get_availabl: prompt cache update took 0.02 ms
+0.23.998.225 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.24.059.202 I srv  process_chun: processing image...
+0.24.395.132 W find_slot: non-consecutive token position 4 after 3 for sequence 0 with 196 new tokens
+0.24.395.722 W find_slot: non-consecutive token position 4 after 3 for sequence 0 with 196 new tokens
+0.24.862.757 I srv  process_chun: image processed in 803 ms
+0.24.863.042 W find_slot: non-consecutive token position 34 after 4 for sequence 0 with 17 new tokens
+0.24.863.090 W find_slot: non-consecutive token position 34 after 4 for sequence 0 with 17 new tokens
+0.24.963.440 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 34, pos_max = 34, n_tokens = 217, size = 62.813 MiB)
+0.25.458.909 I slot print_timing: id  0 | task 0 | 
+prompt eval time =    1019.46 ms /   221 tokens (    4.61 ms per token,   216.78 tokens per second)
+       eval time =     441.18 ms /    21 tokens (   21.01 ms per token,    47.60 tokens per second)
+      total time =    1460.64 ms /   242 tokens
+0.25.458.950 I slot      release: id  0 | task 0 | stop processing: n_tokens = 241, truncated = 0
+0.25.458.959 I srv  update_slots: all slots are idle
+0.26.459.931 I srv    operator(): operator(): cleaning up before exit...

recipe/logs/probe_reasoning.log

@@ -0,0 +1,16 @@
+default correct-decline {"content": "We need answer directly. 391.\n</think>\n\n391", "reasoning": "", "tool_calls": [], "leaks": ["</think>"]}
+default multi-arg {"content": "We need need tool. Current weather Paris celsius.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather"], "leaks": ["</think>"]}
+default parallel {"content": "We need need call each city. Parallel okay. Must use commentary tool. Then summarize.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather", "get_weather"], "leaks": ["</think>"]}
+fmt-deepseek correct-decline {"content": "We need answer directly. 391.\n</think>\n\n391", "reasoning": "", "tool_calls": [], "leaks": ["</think>"]}
+fmt-deepseek multi-arg {"content": "We need need tool. Current weather Paris celsius.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather"], "leaks": ["</think>"]}
+fmt-deepseek parallel {"content": "We need need call each city. Parallel okay. Must use commentary tool. Then summarize.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather", "get_weather"], "leaks": ["</think>"]}
+srv-kwargs-high correct-decline {"content": "We need answer directly. 391.\n</think>\n\n391", "reasoning": "", "tool_calls": [], "leaks": ["</think>"]}
+srv-kwargs-high multi-arg {"content": "We need need tool. Current weather Paris celsius.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather"], "leaks": ["</think>"]}
+srv-kwargs-high parallel {"content": "We need need call each city. Parallel okay. Must use commentary tool. Then summarize.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather", "get_weather"], "leaks": ["</think>"]}
+reasoning-on correct-decline {"content": "We need answer directly. 391.\n</think>\n\n391", "reasoning": "", "tool_calls": [], "leaks": ["</think>"]}
+reasoning-on multi-arg {"content": "We need need tool. Current weather Paris celsius.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather"], "leaks": ["</think>"]}
+reasoning-on parallel {"content": "We need need call each city. Parallel okay. Must use commentary tool. Then summarize.\n</think>\n\n", "reasoning": "", "tool_calls": ["get_weather", "get_weather"], "leaks": ["</think>"]}
+tpl-enable-thinking correct-decline {"content": "391", "reasoning": "We need answer directly. 391.\n", "tool_calls": [], "leaks": []}
+tpl-enable-thinking multi-arg {"content": "", "reasoning": "We need need tool. Current weather Paris celsius.\n", "tool_calls": ["get_weather"], "leaks": []}
+tpl-enable-thinking parallel {"content": "", "reasoning": "We need need call each city. Parallel okay. Must use commentary tool. Then summarize.\n", "tool_calls": ["get_weather", "get_weather"], "leaks": []}
+NEX_REASONING_PROBE_DONE /mnt/models/nex-n2.5-mini/results/nex_reasoning_probe.json

recipe/logs/probe_seat_default.log

@@ -0,0 +1,19 @@
+C1 no-kwargs correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C1 no-kwargs single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C1 no-kwargs multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+C1 enable_thinking=true correct-decline {"content": "391", "reasoning_len": 30, "tool_calls": [], "leaks": []}
+C1 enable_thinking=true single-word {"content": "ready", "reasoning_len": 35, "tool_calls": [], "leaks": []}
+C1 enable_thinking=true multi-arg {"content": "", "reasoning_len": 50, "tool_calls": ["get_weather"], "leaks": []}
+C1 enable_thinking=false correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C1 enable_thinking=false single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C1 enable_thinking=false multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+C2 no-kwargs correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C2 no-kwargs single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C2 no-kwargs multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+C2 enable_thinking=true correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C2 enable_thinking=true single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C2 enable_thinking=true multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+C2 enable_thinking=false correct-decline {"content": "391", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C2 enable_thinking=false single-word {"content": "ready", "reasoning_len": 0, "tool_calls": [], "leaks": []}
+C2 enable_thinking=false multi-arg {"content": "", "reasoning_len": 0, "tool_calls": ["get_weather"], "leaks": []}
+NEX_SEAT_DEFAULT_PROBE_DONE

recipe/logs/probe_seat_default_C2.log

@@ -0,0 +1,185 @@
+0.00.244.374 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
+0.00.244.393 I device_info:
+0.00.244.653 I   - ROCm0   : AMD Radeon Graphics (131072 MiB, 123682 MiB free)
+0.00.245.094 I   - Vulkan0 : AMD Radeon Graphics (RADV GFX1151) (132096 MiB, 131922 MiB free)
+0.00.245.113 I   - CPU     : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S (127438 MiB, 127438 MiB free)
+0.00.245.304 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.00.245.404 I srv          init: running without SSL
+0.00.245.477 I srv          init: using 31 threads for HTTP server
+0.00.245.479 I srv          init: the WebUI is disabled
+0.00.245.698 I srv         start: binding port with default address family
+0.00.247.040 I srv          main: loading model
+0.00.247.045 I srv    load_model: loading model '/mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf'
+0.00.310.751 W llama_model_loader: direct I/O is enabled, disabling mmap
+0.22.302.525 W llama_context: n_ctx_seq (16384) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
+0.22.428.373 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
+0.22.720.526 I srv    load_model: initializing slots, n_slots = 1
+0.23.006.696 W srv    load_model: speculative decoding will use checkpoints
+0.23.006.713 W common_speculative_init: no implementations specified for speculative decoding
+0.23.006.717 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 16384
+0.23.006.919 I srv    load_model: prompt cache RAM enabled: limit_mib=8192
+0.23.006.923 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
+0.23.006.964 I srv          init: idle slots will be saved to prompt cache upon starting a new task
+0.23.066.942 I init: chat template, example_format: '<|im_start|>system
+You are a helpful assistant<|im_end|>
+<|im_start|>user
+Hello<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+Hi there<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+<think>'
+0.23.112.988 I srv          init: init: chat template, thinking = 1
+0.23.113.079 I srv          main: model loaded
+0.23.113.093 I srv          main: server is listening on http://127.0.0.1:18653
+0.23.113.102 I srv  update_slots: all slots are idle
+0.24.535.309 I srv  params_from_: Chat format: peg-native
+0.24.535.740 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
+0.24.535.743 I srv  get_availabl: updating prompt cache
+0.24.535.749 I srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
+0.24.535.754 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 16384 tokens, 8589934592 est)
+0.24.535.757 I srv  get_availabl: prompt cache update took 0.01 ms
+0.24.536.006 I reasoning-budget: activated, budget=2147483647 tokens
+0.24.536.022 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
+0.25.228.510 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.25.312.278 I reasoning-budget: deactivated (natural end)
+0.25.454.972 I slot print_timing: id  0 | task 0 | 
+prompt eval time =     749.65 ms /   424 tokens (    1.77 ms per token,   565.60 tokens per second)
+       eval time =     169.27 ms /     7 tokens (   24.18 ms per token,    41.35 tokens per second)
+      total time =     918.92 ms /   431 tokens
+0.25.455.052 I slot      release: id  0 | task 0 | stop processing: n_tokens = 430, truncated = 0
+0.25.455.062 I srv  update_slots: all slots are idle
+0.25.515.651 I srv  params_from_: Chat format: peg-native
+0.25.518.171 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.969 (> 0.100 thold), f_keep = 0.942
+0.25.518.665 I reasoning-budget: activated, budget=2147483647 tokens
+0.25.518.762 I slot launch_slot_: id  0 | task 9 | processing task, is_child = 0
+0.25.518.786 W slot update_slots: id  0 | task 9 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.25.518.790 I slot update_slots: id  0 | task 9 | Checking checkpoint with [419, 419] against 405...
+0.25.518.793 W slot update_slots: id  0 | task 9 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.25.518.809 W slot update_slots: id  0 | task 9 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.26.100.410 I slot create_check: id  0 | task 9 | created context checkpoint 1 of 32 (pos_min = 413, pos_max = 413, n_tokens = 414, size = 62.813 MiB)
+0.26.190.265 I reasoning-budget: deactivated (natural end)
+0.26.318.946 I slot print_timing: id  0 | task 9 | 
+prompt eval time =     636.97 ms /   418 tokens (    1.52 ms per token,   656.23 tokens per second)
+       eval time =     163.18 ms /     5 tokens (   32.64 ms per token,    30.64 tokens per second)
+      total time =     800.14 ms /   423 tokens
+0.26.319.036 I slot      release: id  0 | task 9 | stop processing: n_tokens = 422, truncated = 0
+0.26.319.067 I srv  update_slots: all slots are idle
+0.26.365.926 I srv  params_from_: Chat format: peg-native
+0.26.367.996 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.960
+0.26.368.525 I reasoning-budget: activated, budget=2147483647 tokens
+0.26.368.597 I slot launch_slot_: id  0 | task 16 | processing task, is_child = 0
+0.26.368.617 W slot update_slots: id  0 | task 16 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.26.368.621 I slot update_slots: id  0 | task 16 | Checking checkpoint with [413, 413] against 405...
+0.26.368.623 W slot update_slots: id  0 | task 16 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.26.368.628 W slot update_slots: id  0 | task 16 | erased invalidated context checkpoint (pos_min = 413, pos_max = 413, n_tokens = 414, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.26.955.113 I slot create_check: id  0 | task 16 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.27.041.682 I reasoning-budget: deactivated (natural end)
+0.27.918.891 I slot print_timing: id  0 | task 16 | 
+prompt eval time =     644.55 ms /   421 tokens (    1.53 ms per token,   653.16 tokens per second)
+       eval time =     905.71 ms /    42 tokens (   21.56 ms per token,    46.37 tokens per second)
+      total time =    1550.26 ms /   463 tokens
+0.27.918.973 I slot      release: id  0 | task 16 | stop processing: n_tokens = 462, truncated = 0
+0.27.919.004 I srv  update_slots: all slots are idle
+0.27.975.579 I srv  params_from_: Chat format: peg-native
+0.27.977.357 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.958 (> 0.100 thold), f_keep = 0.879
+0.27.977.785 I reasoning-budget: activated, budget=2147483647 tokens
+0.27.977.885 I slot launch_slot_: id  0 | task 60 | processing task, is_child = 0
+0.27.977.912 W slot update_slots: id  0 | task 60 | n_past = 406, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.27.977.913 I slot update_slots: id  0 | task 60 | Checking checkpoint with [416, 416] against 406...
+0.27.977.915 W slot update_slots: id  0 | task 60 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.27.977.920 W slot update_slots: id  0 | task 60 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.28.575.881 I slot create_check: id  0 | task 60 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.28.639.525 I reasoning-budget: deactivated (natural end)
+0.28.760.322 I slot print_timing: id  0 | task 60 | 
+prompt eval time =     636.60 ms /   424 tokens (    1.50 ms per token,   666.03 tokens per second)
+       eval time =     145.80 ms /     7 tokens (   20.83 ms per token,    48.01 tokens per second)
+      total time =     782.40 ms /   431 tokens
+0.28.760.399 I slot      release: id  0 | task 60 | stop processing: n_tokens = 430, truncated = 0
+0.28.760.426 I srv  update_slots: all slots are idle
+0.28.773.374 I srv  params_from_: Chat format: peg-native
+0.28.773.718 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.969 (> 0.100 thold), f_keep = 0.942
+0.28.773.919 I reasoning-budget: activated, budget=2147483647 tokens
+0.28.773.955 I slot launch_slot_: id  0 | task 69 | processing task, is_child = 0
+0.28.773.965 W slot update_slots: id  0 | task 69 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.28.773.967 I slot update_slots: id  0 | task 69 | Checking checkpoint with [419, 419] against 405...
+0.28.773.968 W slot update_slots: id  0 | task 69 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.28.773.971 W slot update_slots: id  0 | task 69 | erased invalidated context checkpoint (pos_min = 419, pos_max = 419, n_tokens = 420, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.29.251.924 I slot create_check: id  0 | task 69 | created context checkpoint 1 of 32 (pos_min = 413, pos_max = 413, n_tokens = 414, size = 62.813 MiB)
+0.29.313.310 I reasoning-budget: deactivated (natural end)
+0.29.391.023 I slot print_timing: id  0 | task 69 | 
+prompt eval time =     514.69 ms /   418 tokens (    1.23 ms per token,   812.15 tokens per second)
+       eval time =     102.36 ms /     5 tokens (   20.47 ms per token,    48.85 tokens per second)
+      total time =     617.04 ms /   423 tokens
+0.29.391.120 I slot      release: id  0 | task 69 | stop processing: n_tokens = 422, truncated = 0
+0.29.391.150 I srv  update_slots: all slots are idle
+0.29.426.233 I srv  params_from_: Chat format: peg-native
+0.29.426.716 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.960
+0.29.426.958 I reasoning-budget: activated, budget=2147483647 tokens
+0.29.427.000 I slot launch_slot_: id  0 | task 76 | processing task, is_child = 0
+0.29.427.011 W slot update_slots: id  0 | task 76 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.29.427.013 I slot update_slots: id  0 | task 76 | Checking checkpoint with [413, 413] against 405...
+0.29.427.014 W slot update_slots: id  0 | task 76 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.29.427.017 W slot update_slots: id  0 | task 76 | erased invalidated context checkpoint (pos_min = 413, pos_max = 413, n_tokens = 414, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.29.922.653 I slot create_check: id  0 | task 76 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.29.984.421 I reasoning-budget: deactivated (natural end)
+0.30.786.431 I slot print_timing: id  0 | task 76 | 
+prompt eval time =     529.05 ms /   421 tokens (    1.26 ms per token,   795.76 tokens per second)
+       eval time =     830.35 ms /    42 tokens (   19.77 ms per token,    50.58 tokens per second)
+      total time =    1359.40 ms /   463 tokens
+0.30.786.507 I slot      release: id  0 | task 76 | stop processing: n_tokens = 462, truncated = 0
+0.30.786.534 I srv  update_slots: all slots are idle
+0.30.827.357 I srv  params_from_: Chat format: peg-native
+0.30.827.975 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.951 (> 0.100 thold), f_keep = 0.879
+0.30.828.455 I reasoning-budget: activated, budget=2147483647 tokens
+0.30.828.460 I reasoning-budget: deactivated (natural end)
+0.30.828.579 I slot launch_slot_: id  0 | task 120 | processing task, is_child = 0
+0.30.828.602 W slot update_slots: id  0 | task 120 | n_past = 406, slot.prompt.tokens.size() = 462, seq_id = 0, pos_min = 461, n_swa = 0
+0.30.828.605 I slot update_slots: id  0 | task 120 | Checking checkpoint with [416, 416] against 406...
+0.30.828.608 W slot update_slots: id  0 | task 120 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.30.828.615 W slot update_slots: id  0 | task 120 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.31.403.843 I slot create_check: id  0 | task 120 | created context checkpoint 1 of 32 (pos_min = 422, pos_max = 422, n_tokens = 423, size = 62.813 MiB)
+0.31.535.343 I slot print_timing: id  0 | task 120 | 
+prompt eval time =     610.22 ms /   427 tokens (    1.43 ms per token,   699.75 tokens per second)
+       eval time =      96.52 ms /     4 tokens (   24.13 ms per token,    41.44 tokens per second)
+      total time =     706.74 ms /   431 tokens
+0.31.535.416 I slot      release: id  0 | task 120 | stop processing: n_tokens = 430, truncated = 0
+0.31.535.442 I srv  update_slots: all slots are idle
+0.31.579.928 I srv  params_from_: Chat format: peg-native
+0.31.582.191 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.962 (> 0.100 thold), f_keep = 0.942
+0.31.582.782 I reasoning-budget: activated, budget=2147483647 tokens
+0.31.582.789 I reasoning-budget: deactivated (natural end)
+0.31.582.964 I slot launch_slot_: id  0 | task 126 | processing task, is_child = 0
+0.31.582.991 W slot update_slots: id  0 | task 126 | n_past = 405, slot.prompt.tokens.size() = 430, seq_id = 0, pos_min = 429, n_swa = 0
+0.31.582.995 I slot update_slots: id  0 | task 126 | Checking checkpoint with [422, 422] against 405...
+0.31.582.997 W slot update_slots: id  0 | task 126 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.31.583.004 W slot update_slots: id  0 | task 126 | erased invalidated context checkpoint (pos_min = 422, pos_max = 422, n_tokens = 423, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.142.847 I slot create_check: id  0 | task 126 | created context checkpoint 1 of 32 (pos_min = 416, pos_max = 416, n_tokens = 417, size = 62.813 MiB)
+0.32.249.462 I slot print_timing: id  0 | task 126 | 
+prompt eval time =     617.07 ms /   421 tokens (    1.47 ms per token,   682.25 tokens per second)
+       eval time =      49.39 ms /     2 tokens (   24.70 ms per token,    40.49 tokens per second)
+      total time =     666.46 ms /   423 tokens
+0.32.249.559 I slot      release: id  0 | task 126 | stop processing: n_tokens = 422, truncated = 0
+0.32.249.590 I srv  update_slots: all slots are idle
+0.32.262.258 I srv  params_from_: Chat format: peg-native
+0.32.262.613 I slot get_availabl: id  0 | task -1 | selected slot by LCP similarity, sim_best = 0.955 (> 0.100 thold), f_keep = 0.960
+0.32.262.803 I reasoning-budget: activated, budget=2147483647 tokens
+0.32.262.808 I reasoning-budget: deactivated (natural end)
+0.32.262.842 I slot launch_slot_: id  0 | task 130 | processing task, is_child = 0
+0.32.262.852 W slot update_slots: id  0 | task 130 | n_past = 405, slot.prompt.tokens.size() = 422, seq_id = 0, pos_min = 421, n_swa = 0
+0.32.262.852 I slot update_slots: id  0 | task 130 | Checking checkpoint with [416, 416] against 405...
+0.32.262.853 W slot update_slots: id  0 | task 130 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+0.32.262.856 W slot update_slots: id  0 | task 130 | erased invalidated context checkpoint (pos_min = 416, pos_max = 416, n_tokens = 417, n_swa = 0, pos_next = 0, size = 62.813 MiB)
+0.32.810.656 I slot create_check: id  0 | task 130 | created context checkpoint 1 of 32 (pos_min = 419, pos_max = 419, n_tokens = 420, size = 62.813 MiB)
+0.33.748.645 I slot print_timing: id  0 | task 130 | 
+prompt eval time =     596.89 ms /   424 tokens (    1.41 ms per token,   710.35 tokens per second)
+       eval time =     888.89 ms /    39 tokens (   22.79 ms per token,    43.88 tokens per second)
+      total time =    1485.78 ms /   463 tokens
+0.33.748.736 I slot      release: id  0 | task 130 | stop processing: n_tokens = 462, truncated = 0
+0.33.748.767 I srv  update_slots: all slots are idle
+0.33.749.943 I srv    operator(): operator(): cleaning up before exit...

recipe/pipeline/diag_bf16.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+# Which backend computes the Nex BF16 GGUF correctly? d3ca537 ROCm0 (-fa on) gave PPL 139-171 on the first chunks while
+# the 4-bit STRIX_LEAN gave 6.28-7.84 on the same chunks. Same text, same -c/-b; only device / -fa change.
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; B=/opt/llama-rocm/rocmfpx-724/build-hipvk/bin; A=/mnt/models/agnes-3.0-flash
+export LD_LIBRARY_PATH=$B:/opt/rocm-7.2.4/lib HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1
+cd $W; prlimit --pid $$ --core=1:1
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+BF=gguf/Nex-N2.5-mini-BF16.gguf; Q=out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf
+T=$A/calib/wikitext-2-raw/wiki.test.raw
+run(){  # label chunks model args...
+  local lab=$1 ch=$2 m=$3; shift 3
+  timeout 1800 $B/llama-perplexity -m $m -f $T -c 2048 -b 2048 --chunks $ch "$@" > logs/diag_$lab.log 2>&1
+  log "$lab rc=$? $(grep -oE '\[[0-9]+\][0-9.]+' logs/diag_$lab.log | tr '\n' ' ')"
+}
+run bf16_rocm_faoff 4 $BF -ngl 999 -fa off -dev ROCm0
+run bf16_vk_faon    4 $BF -ngl 999 -fa on  -dev Vulkan0
+run q106_vk_faon    4 $Q  -ngl 999 -fa on  -dev Vulkan0
+run bf16_cpu        2 $BF -ngl 0 -t 16
+log "DIAG_BF16_DONE"

recipe/pipeline/diag_tools_run.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+# start the tools diagnostic as soon as the N8 seat smoke tests are finished (so it does not disturb them),
+# capped at 32G; OxCoder phase 2 waits for GPU-free and therefore starts after it.
+cd /mnt/models/nex-n2.5-mini
+until grep -q "nex_seats exit=" logs/N8_unice.log 2>/dev/null; do sleep 5; done
+echo "[$(date -u +%FT%TZ)] seats done -> diag"
+systemd-run --scope --quiet --unit=nex-tools-diag-$(date +%s) -p MemoryMax=32G -p MemorySwapMax=0 python3 nex_tools_diag.py
+echo "[$(date -u +%FT%TZ)] diag exit=$?"

recipe/pipeline/nex_aggregate.py

@@ -0,0 +1,408 @@
+#!/usr/bin/env python3
+"""Collect every measured number for the Nex-N2.5-mini ROCmFP4 cards into results/summary.json.
+Nothing is typed by hand except model facts and protocol constants. Each value is parsed from a log,
+a results file, or the GGUF itself; missing => None."""
+import hashlib, json, os, re, sys
+sys.path.insert(0, "/opt/llama-rocm/rocmfpx-724/gguf-py")
+import gguf  # noqa: E402
+
+W = "/mnt/models/nex-n2.5-mini"
+A = "/mnt/models/agnes-3.0-flash"
+N = "Nex-N2.5-mini"
+OB = "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin"
+FAST = os.environ.get("AGG_FAST") == "1"
+OUT = os.environ.get("AGG_OUT", f"{W}/results/summary.json")
+UB = 1024  # llama-server -ub; cache checkpoints land 4 + n_ubatch tokens before prompt end
+
+
+def rd(p):
+    try:
+        return open(p, errors="replace").read()
+    except FileNotFoundError:
+        return ""
+
+
+def f(rx, s, g=1, cast=float):
+    m = re.search(rx, s)
+    return cast(m.group(g)) if m else None
+
+
+def jl(p):
+    return [json.loads(l) for l in rd(p).splitlines() if l.strip()]
+
+
+def pubname(fn):
+    """Published filename: the Hub cannot parse an underscore-glued quant token."""
+    return fn.replace("Q4_0_ROCMFP4_", "Q4_0-ROCmFP4-")
+
+
+STAGED = {}
+for _k in ("std", "imat"):
+    for _l in rd(f"{W}/hf-upload/{_k}/SHA256SUMS").splitlines():
+        _h, _n = _l.split(maxsplit=1)
+        STAGED[_n.strip()] = _h
+
+
+def sha_file(p):
+    if not os.path.exists(p):
+        return None
+    if pubname(os.path.basename(p)) in STAGED:
+        return STAGED[pubname(os.path.basename(p))]
+    if os.path.basename(p) in STAGED:
+        return STAGED[os.path.basename(p)]
+    if FAST and os.path.getsize(p) > 1 << 30:
+        return None
+    h = hashlib.sha256()
+    with open(p, "rb") as fh:
+        for c in iter(lambda: fh.read(64 << 20), b""):
+            h.update(c)
+    return h.hexdigest()
+
+
+def tensors(p):
+    """name -> (type, bytes); {} if the file cannot be read (e.g. still being written by a running quantize)."""
+    out = {}
+    try:
+        reader = gguf.GGUFReader(p)
+    except (ValueError, OSError) as e:
+        print(f"WARN unreadable GGUF {p}: {e}", file=sys.stderr)
+        return out
+    for t in reader.tensors:
+        try:
+            tn = t.tensor_type.name
+        except Exception:
+            tn = f"TYPE_{int(t.tensor_type)}"
+        out[t.name] = (tn, int(t.n_bytes))
+    return out
+
+
+def rb_line(blob, fn):
+    return next((l for l in blob.splitlines() if len(l.split()) > 1 and l.split()[1] == fn), "")
+
+
+def kld_stats(k):
+    return {
+        "ppl": f(r"Mean PPL\(Q\)\s*:\s*([\d.]+)", k),
+        "ppl_err": f(r"Mean PPL\(Q\)\s*:\s*[\d.]+\s*±\s*([\d.]+)", k),
+        "ppl_ratio": f(r"Mean PPL\(Q\)/PPL\(base\)\s*:\s*([\d.]+)", k),
+        "kld_mean": f(r"Mean\s+KLD:\s*([\d.]+)", k),
+        "kld_err": f(r"Mean\s+KLD:\s*[\d.]+\s*±\s*([\d.]+)", k),
+        "kld_p99": f(r"99\.0%\s+KLD:\s*([\d.]+)", k),
+        "kld_median": f(r"Median\s+KLD:\s*([\d.]+)", k),
+        "same_top_p": f(r"Same top p:\s*([\d.]+)", k),
+        "rms_dp": f(r"RMS Δp\s*:\s*([\d.]+)", k),
+    }
+
+
+def exists_size(p):
+    return os.path.getsize(p) if os.path.exists(p) else None
+
+
+S = {
+    "layout": "no MTP (checkpoint declares mtp_num_hidden_layers=1 but ships no mtp.* tensors)",
+    "model": N,
+    "model_repo_std": f"kingjones777/{N}-ROCmFP4-GGUF",
+    "model_repo_imat": f"kingjones777/{N}-ROCmFP4-imatrix-GGUF",
+    "source": {
+        "repo": "nex-agi/Nex-N2.5-mini",
+        "revision": "87420286149d9cce9bd46cd335ef9bda33c37c1b",
+        "license": "apache-2.0",
+        "gated": False,
+        "params": 35107181936,
+    },
+    "arch": {
+        "name": "qwen3_5_moe",
+        "llama_cpp": "qwen35moe",
+        "layers": 40,
+        "linear_attn_layers": 30,
+        "full_attn_layers": 10,
+        "full_attention_interval": 4,
+        "hidden": 2048,
+        "num_experts": 256,
+        "num_experts_per_tok": 8,
+        "expert_width": 512,
+        "shared_expert_width": 512,
+        "attn_heads": 16,
+        "kv_heads": 2,
+        "head_dim": 256,
+        "linear_key_heads": 16,
+        "linear_value_heads": 32,
+        "linear_head_dim": 128,
+        "vocab": 248320,
+        "max_position_embeddings": 262144,
+        "tie_word_embeddings": False,
+        "vision_layers": 27,
+        "vision_width": 1152,
+        "hf_tensors": 1026,
+        "mtp_num_hidden_layers_declared": 1,
+    },
+}
+
+# ---------- BF16 reference ----------
+# phase 2b: this build's ROCm0 path computes the BF16 GGUF of this MoE wrong, so the reference ran on pure CPU
+b1 = rd(f"{W}/logs/N1c_ppl_bf16_cpu.log")
+k106 = rd(f"{W}/logs/N4_kld_q106.log")
+crb = rd(f"{W}/logs/C_readback.log")
+chunks = f(r"calculating perplexity over (\d+) chunks", b1, cast=int)
+nctx = f(r"n_ctx=(\d+)", b1, cast=int)
+bf_fn = f"{N}-BF16.gguf"
+bf_line = rb_line(crb, bf_fn)
+bf_path = f"{W}/gguf/{bf_fn}"
+bf_ten = tensors(bf_path) if os.path.exists(bf_path) else {}
+S["bf16"] = {
+    "ppl": f(r"Final estimate: PPL = ([\d.]+)", b1),
+    "ppl_err": f(r"Final estimate: PPL = [\d.]+ \+/- ([\d.]+)", b1),
+    "ppl_paired": f(r"Mean PPL\(base\)\s*:\s*([\d.]+)", k106),
+    "ppl_paired_err": f(r"Mean PPL\(base\)\s*:\s*[\d.]+\s*±\s*([\d.]+)", k106),
+    "chunks": chunks,
+    "n_ctx": nctx,
+    "scored_tokens": chunks * (nctx // 2 - 1) if chunks and nctx else None,
+    "size_bytes": exists_size(bf_path),
+    "file": bf_fn,
+    "readback": bf_line.split()[0] if bf_line else None,
+    "arch": f(r"arch=(\S+)", bf_line, cast=str),
+    "ftype": f(r"ftype=(\d+)", bf_line, cast=int),
+    "tensors": f(r"tensors=(\d+)", bf_line, cast=int) or (len(bf_ten) or None),
+    "nextn_tensors": f(r"nextn=(\d+)", bf_line, cast=int),
+    "output_weight": f(r"output\.weight=(\S+)", bf_line, cast=str),
+    "token_embd": f(r"token_embd\.weight=(\S+)", bf_line, cast=str),
+    "mtp_named_tensors": sum(1 for n in bf_ten if n.startswith("mtp.")) if bf_ten else None,
+    "nextn_named_tensors": sum(1 for n in bf_ten if ".nextn." in n) if bf_ten else None,
+}
+
+# ---------- binaries ----------
+qlog = rd(f"{W}/logs/Q1_q106.log")
+S["binary"] = {
+    "repo": "https://github.com/charlie12345/ROCmFPX",
+    "commit": f(r"llama_print_build_info: build = \d+ \(([0-9a-f]+)\)", qlog, cast=str),
+    "dir": OB,
+    "sha256": {x: sha_file(f"{OB}/{x}") for x in
+               ("llama-server", "llama-quantize", "llama-imatrix", "llama-perplexity")},
+}
+
+vp = f"{A}/calib/vision_probe.png"
+if os.path.exists(vp):
+    raw = open(vp, "rb").read()
+    S["vision_probe"] = {
+        "width": int.from_bytes(raw[16:20], "big"),
+        "height": int.from_bytes(raw[20:24], "big"),
+        "bytes": len(raw),
+        "sha256": hashlib.sha256(raw).hexdigest(),
+    }
+
+imat = rd(f"{W}/logs/N2c_imatrix_cpu.log")
+S["imatrix"] = {
+    "chunks": f(r"computing over (\d+) chunks", imat, cast=int),
+    "n_ctx": f(r"n_ctx=(\d+)", imat, cast=int),
+    "calibration": "bartowski calibration_datav3.txt",
+    "file": f"{N}.imatrix",
+    "sha256": (rd(f"{W}/imat/{N}.imatrix.sha256").split() or [None])[0],
+    "size_bytes": exists_size(f"{W}/imat/{N}.imatrix"),
+}
+
+hc_raw = rd(f"{W}/results/hub_check.json").strip()
+if hc_raw:
+    try:
+        S["hub_check"] = json.loads(hc_raw)
+    except json.JSONDecodeError:
+        S["hub_check"] = None
+else:
+    S["hub_check"] = None
+
+
+def type_name(i):
+    try:
+        return gguf.GGMLQuantizationType(int(i)).name
+    except (TypeError, ValueError):
+        return None
+
+
+for _h in (S["hub_check"] or {}).get("header_checks") or []:
+    # ids were read from the other repo's GGUF header; names come from this fork's gguf-py enum
+    _h["output_weight_type"] = type_name(_h.get("output_weight_type_id"))
+    _h["token_embd_type"] = type_name(_h.get("token_embd_type_id"))
+
+REP = jl(f"{W}/results/nex_repeat.jsonl")
+S["repeat"] = next((d for d in REP if d.get("step") == "N5"), None)        # ROCm0
+S["repeat_vk"] = next((d for d in REP if d.get("step") == "N5v"), None)    # Vulkan0
+S["reference"] = next((d for d in jl(f"{W}/results/nex_reference.jsonl") if d.get("step") == "N1c"), None)
+S["reference_device"] = "CPU" if S["reference"] else None
+S["imatrix"]["device"] = "CPU" if imat else None
+
+# ---------- tiers ----------
+TIERS = {
+    "q106":  ("out",      f"{N}-Q4_0_ROCMFP4_STRIX_LEAN.gguf",         "Q1_q106",  "Q_readback.log",  106),
+    "q102":  ("out",      f"{N}-Q4_0_ROCMFP4_COHERENT.gguf",           "Q1_q102",  "Q_readback.log",  102),
+    "q103":  ("out",      f"{N}-Q4_0_ROCMFP4_FAST.gguf",               "Q1_q103",  "Q_readback.log",  103),
+    "q106i": ("out-imat", f"{N}-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "N3_q106i", "N3_readback.log", 106),
+    "q102i": ("out-imat", f"{N}-imatrix-Q4_0_ROCMFP4_COHERENT.gguf",   "N3_q102i", "N3_readback.log", 102),
+    "q103i": ("out-imat", f"{N}-imatrix-Q4_0_ROCMFP4_FAST.gguf",       "N3_q103i", "N3_readback.log", 103),
+}
+RB = {name: rd(f"{W}/logs/{name}") for name in ("Q_readback.log", "N3_readback.log")}
+
+S["tiers"] = {}
+TEN = {}
+for tag, (d, fn, qlog_stem, rblog, ft) in TIERS.items():
+    p = f"{W}/{d}/{fn}"
+    q = rd(f"{W}/logs/{qlog_stem}.log")
+    k = rd(f"{W}/logs/N4_kld_{tag}.log")
+    line = rb_line(RB.get(rblog, ""), fn)
+    ten = tensors(p) if os.path.exists(p) else {}
+    TEN[tag] = ten
+    qs = f(r"quantize time\s*=\s*([\d.]+) ms", q)
+    t = {
+        "file": pubname(fn),
+        "local_file": fn,
+        "dir": d,
+        "ftype": ft,
+        "size_bytes": exists_size(p),
+        "quant_mib": f(r"quant size\s*=\s*([\d.]+) MiB", q),
+        "bpw": f(r"quant size\s*=\s*[\d.]+ MiB \(([\d.]+) BPW\)", q),
+        "quant_seconds": (qs / 1000) if qs is not None else None,
+        "imatrix_entries": f(r"loaded (\d+) importance matrix entries", q, cast=int),
+        "readback": line.split()[0] if line else None,
+        "arch": f(r"arch=(\S+)", line, cast=str),
+        "tensors": f(r"tensors=(\d+)", line, cast=int),
+        "nextn_tensors": f(r"nextn=(\d+)", line, cast=int),
+        "output_weight": f(r"output\.weight=(\S+)", line, cast=str),
+        "token_embd": f(r"token_embd\.weight=(\S+)", line, cast=str),
+        "quality_measured": f(r"Mean\s+KLD:\s*([\d.]+)", k) is not None,
+    }
+    t.update(kld_stats(k) if t["quality_measured"] else {x: None for x in kld_stats("")})
+    kv = rd(f"{W}/logs/N4v_kld_{tag}.log")                      # the same file graded on Vulkan0
+    t["vk"] = kld_stats(kv) if f(r"Mean\s+KLD:\s*([\d.]+)", kv) is not None else None
+    S["tiers"][tag] = t
+
+TWINS = (("q106i", "q106"), ("q102i", "q102"), ("q103i", "q103"))
+for im, st in TWINS:
+    pa = f"{W}/{TIERS[st][0]}/{TIERS[st][1]}"
+    pb = f"{W}/{TIERS[im][0]}/{TIERS[im][1]}"
+    if TEN.get(st) and TEN.get(im):
+        S["tiers"][im]["same_tensor_types_as_standard"] = TEN[st] == TEN[im]
+    else:
+        S["tiers"][im]["same_tensor_types_as_standard"] = None
+    sa, sb = S["tiers"][st]["size_bytes"], S["tiers"][im]["size_bytes"]
+    S["tiers"][im]["file_size_delta_bytes"] = (sb - sa) if None not in (sa, sb) else None
+    ha, hb = (sha_file(pa), sha_file(pb)) if os.path.exists(pa) and os.path.exists(pb) else (None, None)
+    S["tiers"][im]["differs_from_standard"] = (ha != hb) if ha and hb else None
+
+mm_fn = f"mmproj-{N}-BF16.gguf"
+mm_line = rb_line(crb, mm_fn)
+S["aux"] = {mm_fn: exists_size(f"{W}/out/{mm_fn}")}
+S["mmproj"] = {
+    "file": mm_fn,
+    "size_bytes": S["aux"][mm_fn],
+    "readback": mm_line.split()[0] if mm_line else None,
+    "arch": f(r"arch=(\S+)", mm_line, cast=str),
+    "ftype": f(r"ftype=(\d+)", mm_line, cast=int),
+    "tensors": f(r"tensors=(\d+)", mm_line, cast=int),
+}
+
+# ---------- measurements ----------
+S["bench"], S["gates"] = [], []
+for d in jl(f"{W}/results/nex_bench.jsonl"):
+    if "tg_median" in d:
+        row = {k: v for k, v in d.items() if k != "runs"}
+        pns = [r.get("prompt_n") for r in d.get("runs") or [] if r.get("prompt_n") is not None]
+        row["prompt_n_min"], row["prompt_n_max"] = (min(pns), max(pns)) if pns else (None, None)
+        S["bench"].append(row)
+    else:
+        S["gates"].append(d)
+
+PE = re.compile(r"prompt eval time =\s*([\d.]+) ms /\s*(\d+) tokens")
+for g_ in S["gates"]:
+    if g_.get("label") != "n-c3-q106" or not g_.get("rows"):
+        continue
+    ev = [(float(a), int(b)) for a, b in PE.findall(rd(f"{W}/logs/b_{g_['label']}.log"))]
+    per = 3  # warm-up, warm, cold — one server log line each, in order
+    for i, r_ in enumerate(g_["rows"]):
+        blk = ev[per * i: per * i + per]
+        if len(blk) == per and blk[1][1] == r_["warm_prompt_n"] and blk[2][1] == r_["cold_prompt_n"]:
+            r_["warm_prompt_ms"], r_["cold_prompt_ms"] = blk[1][0], blk[2][0]
+        else:
+            r_["warm_prompt_ms"] = r_["cold_prompt_ms"] = None
+
+S["sizing"] = jl(f"{W}/results/nex_sizing.jsonl")
+S["n_ubatch"] = UB
+
+
+def _jobj(p):
+    raw = rd(p).strip()
+    try:
+        return json.loads(raw) if raw else None
+    except json.JSONDecodeError:
+        return None
+
+
+# ---------- chat template fix (tool calls / reasoning) ----------
+TPLF = "chat_template_enable_thinking.jinja"
+S["aux"][TPLF] = exists_size(f"{W}/tpl/{TPLF}")
+shim = {k: _jobj(f"{W}/results/nex_template_shim{k}.json") for k in ("", "_medium", "_c1", "_roff")}
+_roff = shim["_roff"] or {}
+S["template_fix"] = {
+    "file": TPLF,
+    "size_bytes": S["aux"][TPLF],
+    "sha256": _roff.get("sha256"),
+    "source_sha256": _roff.get("source_sha256"),
+    "shim": _roff.get("shim"),
+    "server_flags": ["--chat-template-file", TPLF, "--reasoning", "off"],
+    "probes_roff": _roff.get("probes"),                       # the recommended configuration
+    "probes_high_default_on": (shim[""] or {}).get("probes"),   # same template, thinking on by default
+    "medium_mapping_label": (shim["_medium"] or {}).get("label"),
+}
+_diag = _jobj(f"{W}/results/nex_tools_diag.json") or {}
+_v = _diag.get("variants") or {}
+_on = [r.get("first") or {} for r in (_v.get("gate_on") or {}).values()]
+_nested = [(_v.get("gate_off") or {}).get("nested-object", {}).get("first") or {}] + list(_v.get("nested_off_x3") or [])
+_rp = _jobj(f"{W}/results/nex_reasoning_probe.json") or {}
+S["tools_diag"] = {
+    # stock template, the gate's own sampling: replies with reasoning left in content, of all thinking-on replies
+    "stock_on_replies": len(_on) if _on else None,
+    "stock_on_leaks": sum(1 for r in _on if r.get("leaks_in_content")) if _on else None,
+    "stock_on_reasoning_extracted": sum(1 for r in _on if r.get("reasoning_len")) if _on else None,
+    # stock template, thinking off, the nested-object request repeated: HTTP 500 "does not match peg-native"
+    "nested_off_attempts": len(_nested) if _v else None,
+    "nested_off_http500": sum(1 for r in _nested if r.get("http_error") == 500) if _v else None,
+    "gate_http500_logged": "does not match the expected peg-native format" in rd(f"{W}/logs/b_n-tools-q106.log"),
+    # server-side switches that did NOT fix the stock template (reasoning still in content)
+    "flag_probes": {k: {"leaks": sum(1 for r in (c.get("results") or {}).values() if r.get("leaks")),
+                        "n": len(c.get("results") or {})}
+                    for k, c in (_rp.get("configs") or {}).items()},
+}
+_seats = {}
+for r_ in jl(f"{W}/results/nex_seats.jsonl"):
+    _seats[r_.get("unit")] = r_          # last row per unit = the current unit files
+S["seats"] = _seats
+S["measured"] = (
+    f(r"\[(\d{4}-\d{2}-\d{2})T", rd(f"{W}/logs/N6_bench.log"), cast=str)
+    or f(r"\[(\d{4}-\d{2}-\d{2})T", b1, cast=str)
+    or f(r"\[(\d{4}-\d{2}-\d{2})T", rd(f"{W}/logs/N1_ppl_bf16.log"), cast=str)
+)
+
+_dates = sorted(set(re.findall(r"(\d{4}-\d{2}-\d{2})T\d", rd(f"{W}/logs/N6_bench.log") + rd(f"{W}/logs/N6t_tools_roff.log")
+                              + rd(f"{W}/results/nex_seats.jsonl"))))
+S["measured_range"] = [_dates[0], _dates[-1]] if len(_dates) > 1 else (_dates or None)
+
+
+def _kv_keys(p):
+    try:
+        return set(gguf.GGUFReader(p).fields)
+    except (ValueError, OSError):
+        return None
+
+
+for im, st in TWINS:
+    ka = _kv_keys(f"{W}/{TIERS[st][0]}/{TIERS[st][1]}") if FAST is False else None
+    kb = _kv_keys(f"{W}/{TIERS[im][0]}/{TIERS[im][1]}") if FAST is False else None
+    S["tiers"][im]["header_keys_only_in_imatrix"] = sorted(kb - ka) if ka is not None and kb is not None else None
+    S["tiers"][im]["header_keys_only_in_standard"] = sorted(ka - kb) if ka is not None and kb is not None else None
+
+os.makedirs(os.path.dirname(OUT), exist_ok=True)
+json.dump(S, open(OUT, "w"), indent=2)
+
+print("bench rows", len(S["bench"]), "| gate rows", len(S["gates"]),
+      "| sizing rows", len(S["sizing"]), "| hub_check", S["hub_check"] is not None,
+      "| repeat", (S["repeat"] or {}).get("result"))
+for tg_, v in S["tiers"].items():
+    print(tg_, {k: v[k] for k in ("size_bytes", "bpw", "readback", "tensors", "quality_measured", "kld_mean")})
+print("binary", json.dumps(S["binary"], indent=1))

recipe/pipeline/nex_bench.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+"""Nex-N2.5-mini speed + gates. Every 4-bit file x backend, no draft head (the checkpoint ships no MTP weights),
+same house protocol as Agnes (nex_harness.py). Then gates on STRIX_LEAN: aligned prompt-cache identity, tool calls
+(thinking on + off), vision with -fa on and off. Sequential GPU; results/nex_bench.jsonl."""
+import fcntl, os, subprocess, time
+W = "/mnt/models/nex-n2.5-mini"; N = "Nex-N2.5-mini"; A = "/mnt/models/agnes-3.0-flash"
+J = f"{W}/results/nex_bench.jsonl"
+OB = "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin"
+T = {"q106": f"{W}/out/{N}-Q4_0_ROCMFP4_STRIX_LEAN.gguf", "q102": f"{W}/out/{N}-Q4_0_ROCMFP4_COHERENT.gguf",
+     "q103": f"{W}/out/{N}-Q4_0_ROCMFP4_FAST.gguf",
+     "q106i": f"{W}/out-imat/{N}-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+     "q102i": f"{W}/out-imat/{N}-imatrix-Q4_0_ROCMFP4_COHERENT.gguf",
+     "q103i": f"{W}/out-imat/{N}-imatrix-Q4_0_ROCMFP4_FAST.gguf"}
+ENV = dict(os.environ, AGNES_BIN=OB)
+
+
+def log(m): print(f"[{time.strftime('%FT%TZ', time.gmtime())}] {m}", flush=True)
+
+
+def run(argv, label):
+    log(label)
+    r = subprocess.run(argv, capture_output=True, text=True, env=ENV)
+    print(r.stdout.strip()[-800:], flush=True)
+    if r.returncode != 0:
+        print("  RC", r.returncode, r.stderr.strip()[-600:], flush=True)
+    return r.returncode
+
+
+def harness(cmd, label, model, dev="ROCm0", reps=3, workload="code", extra=()):
+    return run(["python3", f"{W}/nex_harness.py", cmd, "--model", model, "--dev", dev, "--label", label,
+                "--reps", str(reps), "--jsonl", J, "--serverlog", f"{W}/logs/b_{label}.log", "--workload", workload,
+                *extra], f"{cmd} {label}")
+
+
+# One quiet box: publish_v2.sh upload takes the same lock, so an HF upload never overlaps the speed runs.
+QUIET = open("/mnt/models/.quiet-box.lock", "w")
+log("waiting for the quiet-box lock (no HF upload during speed runs)")
+fcntl.flock(QUIET, fcntl.LOCK_EX)
+log("quiet-box lock held")
+for tag in ("q106", "q102", "q103", "q106i", "q102i", "q103i"):
+    for dev, sh in (("ROCm0", "rocm"), ("Vulkan0", "vk")):
+        harness("bench", f"n-{tag}-{sh}", T[tag], dev=dev)
+harness("bench", "n-q106-rocm-prose", T["q106"], workload="prose")
+harness("bench", "n-q106-vk-prose", T["q106"], dev="Vulkan0", workload="prose")
+run(["python3", f"{W}/nex_cachegate.py", "--model", T["q106"], "--label", "n-c3-q106", "--reps", "5",
+     "--jsonl", J, "--serverlog", f"{W}/logs/b_n-c3-q106.log"], "cachegate n-c3-q106")
+harness("tools", "n-tools-q106", T["q106"])
+for fa in ("on", "off"):
+    harness("vision", f"n-vision-q106-fa{fa}", T["q106"],
+            extra=("--fa", fa, "--mmproj", f"{W}/out/mmproj-{N}-BF16.gguf", "--image", f"{A}/calib/vision_probe.png",
+                   "--question", "Describe the two shapes in this image and their colors.",
+                   "--expect", "red,blue,circle,square"))
+log("NEX_BENCH_DONE")

recipe/pipeline/nex_cachegate.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""Chunk-ALIGNED warm/cold identity gate (Nex-N2.5-mini: unpatched server, no draft head).
+Written for the Agnes MTP prompt-cache patch; here it checks the stock server's own checkpoint restore for this
+hybrid (gated-delta + attention) model.
+
+Why a third gate: llama-server splits every prompt so that context checkpoints land (4 + n_ubatch) and 4 tokens
+before its end (tools/server/server-context.cpp `checkpoint_offsets`, upstream PR #20288). A warm request restores a
+checkpoint whose position was fixed by the length of the request that CREATED it. When those lengths differ, the warm
+tail is processed in different chunks than a cold run of the same prompt, so float rounding differs and a greedy
+token can flip - with or without a draft head. cachegate2 mixed prompt lengths and hit exactly that.
+
+Here every prompt in the run is padded to ONE token length L, so warm and cold see identical chunking. What is left
+under test is whether the restored checkpoint (attention KV + recurrent state) is exact.
+Per variant: warm-up A (cache on) -> B warm (cache on, must restore L-(4+ub)) -> B cold (cache off). PASS = every
+warm B reused the cache at the aligned position AND is byte-identical to its cold twin, over all variants."""
+import argparse, hashlib, json, os, sys
+sys.path.insert(0, "/mnt/models/nex-n2.5-mini")
+from nex_harness import Server, post, prompt_8k, THINK_OFF  # noqa: E402  (reads AGNES_BIN at import)
+
+UB = 1024  # agnes_harness.Server passes -ub 1024
+WARM = "Summarise this file."
+INSTR = ["Write a function that lists every tensor name in this file.",
+         "Write a function that counts the model classes registered in this file.",
+         "Write a function that finds the longest method in this file.",
+         "Write a function that returns every regular expression used in this file.",
+         "Write a function that maps each class in this file to its base classes.",
+         "Write a function that extracts all string constants from this file.",
+         "Write a function that reports which imports in this file are unused.",
+         "Write a function that lists every method that raises an exception in this file.",
+         "Write a function that counts the lines of code per class in this file.",
+         "Write a function that finds duplicate method names across classes in this file."]
+KW = THINK_OFF
+
+
+def plen(port, content):
+    p = post(port, "/apply-template", {"messages": [{"role": "user", "content": content}],
+                                       "chat_template_kwargs": KW})["prompt"]
+    return len(post(port, "/tokenize", {"content": p, "add_special": True, "parse_special": True})["tokens"])
+
+
+def pad(port, ctx, instr, L):
+    s = instr
+    n = plen(port, ctx + "\n\n" + s)
+    for filler in (" ok", ".", " x"):
+        while n < L:
+            t = s + filler
+            m = plen(port, ctx + "\n\n" + t)
+            if m > L:
+                break
+            s, n = t, m
+        if n == L:
+            return s
+    raise SystemExit(f"could not pad {instr!r} to {L} (stuck at {n})")
+
+
+def req(port, ctx, instr, n, cache):
+    body = {"messages": [{"role": "user", "content": ctx + "\n\n" + instr}], "max_tokens": n,
+            "temperature": 0, "top_k": 1, "ignore_eos": True, "cache_prompt": cache, "chat_template_kwargs": KW}
+    r = post(port, "/v1/chat/completions", body)
+    tm = r.get("timings", {})
+    return (r["choices"][0]["message"].get("content") or ""), tm.get("prompt_n"), tm.get("cache_n"), \
+        tm.get("draft_n"), tm.get("draft_n_accepted")
+
+
+def first_diff(x, y):
+    if x == y:
+        return None
+    return next((i for i, (p, q) in enumerate(zip(x, y)) if p != q), min(len(x), len(y)))
+
+
+def main(a):
+    s = Server(a, a.port)
+    rows = []
+    try:
+        ctx, _ = prompt_8k("code")
+        instr = INSTR[:a.reps]
+        warm = [f"{WARM} (warm-up {i})" for i in range(len(instr))]
+        # L over the FULL prompt set, whatever --reps is: runs with different --reps then share byte-identical
+        # prompts, so their outputs can be compared across configs (e.g. MTP vs no draft head).
+        allp = INSTR + [f"{WARM} (warm-up {i})" for i in range(len(INSTR))]
+        L = max(plen(a.port, ctx + "\n\n" + t) for t in allp) + 1
+        instr = [pad(a.port, ctx, t, L) for t in instr]
+        warm = [pad(a.port, ctx, t, L) for t in warm]
+        expect_cache = L - (4 + UB)
+        for i, (wa, vb) in enumerate(zip(warm, instr)):
+            req(a.port, ctx, wa, 16, True)
+            tw, pw, cw, dw, aw = req(a.port, ctx, vb, 192, True)
+            tc, pc, cc, dc, ac = req(a.port, ctx, vb, 192, False)
+            rows.append({"variant": i, "L": L, "warm_prompt_n": pw, "warm_cache_n": cw, "cold_prompt_n": pc,
+                         "cold_cache_n": cc, "aligned": (cw == expect_cache and pc == L),
+                         "warm_draft": [aw, dw], "cold_draft": [ac, dc], "identical": tw == tc,
+                         "first_diff_char": first_diff(tw, tc),
+                         "warm_sha": hashlib.sha256(tw.encode()).hexdigest()[:12],
+                         "cold_sha": hashlib.sha256(tc.encode()).hexdigest()[:12], "instr": vb, "warm_instr": wa,
+                         "warm_text": tw, "cold_text": tc})
+            print("  ", json.dumps({k: v for k, v in rows[-1].items() if not k.endswith("_text")}), flush=True)
+    finally:
+        s.stop()
+    reused = all((r["warm_cache_n"] or 0) > 0 for r in rows)
+    aligned = all(r["aligned"] for r in rows)
+    ident = all(r["identical"] for r in rows)
+    res = {"label": a.label, "gate": "cachegate3-aligned", "n": len(rows), "L": rows[0]["L"] if rows else None,
+           "expect_cache_n": expect_cache if rows else None, "all_reused": reused, "all_aligned": aligned,
+           "identical": sum(r["identical"] for r in rows), "rows": rows,
+           "result": "PASS" if (rows and reused and aligned and ident) else "FAIL"}
+    print(json.dumps({k: v for k, v in res.items() if k != "rows"}), flush=True)
+    with open(a.jsonl, "a") as f:
+        f.write(json.dumps(res) + "\n")
+    sys.exit(0 if res["result"] == "PASS" else 1)
+
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--model", required=True); ap.add_argument("--draft")
+    ap.add_argument("--nmax", type=int, default=4); ap.add_argument("--pmin", type=float, default=0.0)
+    ap.add_argument("--dev", default="ROCm0"); ap.add_argument("--ctx", type=int, default=65536)
+    ap.add_argument("--port", type=int, default=18600); ap.add_argument("--reps", type=int, default=10)
+    ap.add_argument("--label", required=True); ap.add_argument("--jsonl", required=True)
+    ap.add_argument("--serverlog", required=True)
+    ap.add_argument("--mtp-infile", action="store_true"); ap.add_argument("--strict", action="store_true")
+    main(ap.parse_args())

recipe/pipeline/nex_cards.sh

@@ -0,0 +1,41 @@
+#!/bin/bash
+# nex_cards.sh cards|upload|verify — Nex-N2.5-mini cards on real data, then the two SEPARATE repos.
+# cards : summary (fast, to satisfy stage's recipe check) -> stage std+imat (hashes) -> summary with staged hashes
+#         -> refresh the staged summary copy -> judge -> render (prints pending judgments)
+# upload: behind the quiet-box lock; each repo through upload_watchdog.sh (40G cap, niced, stall kill + retry).   verify: size + sha256 + recipe + separation.
+set -euo pipefail
+W=/mnt/models/nex-n2.5-mini; cd $W
+export HF_XET_CACHE=/mnt/models/.hf-xet-cache TMPDIR=/mnt/models/.tmp HF_HUB_DISABLE_TELEMETRY=1 PYTHONUNBUFFERED=1
+mkdir -p "$HF_XET_CACHE" "$TMPDIR"
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+case "${1:-}" in
+  cards)
+    grep -q NEX_PHASE2_DONE logs/phase2.log || { log "phase 2 not done -> refuse"; exit 3; }
+    AGG_FAST=1 nice -n 10 python3 nex_aggregate.py > logs/aggregate_fast.log 2>&1
+    for k in std imat; do log "stage $k"; nice -n 10 python3 nex_publish.py stage $k; done
+    nice -n 10 python3 nex_aggregate.py > logs/aggregate.log 2>&1; tail -12 logs/aggregate.log
+    for k in std imat; do cp results/summary.json hf-upload/$k/recipe/results_summary.json; done
+    python3 - <<'PY'
+import json
+d = {}
+for k in ("std", "imat"):
+    d.update(json.load(open(f"hf-upload/staged_sizes_{k}.json")))
+json.dump(d, open("hf-upload/staged_sizes.json", "w"), indent=1)
+print("staged sizes:", len(d))
+PY
+    python3 nex_judge.py results/summary.json results/judgments.json > logs/judge.log 2>&1; head -1 logs/judge.log
+    rm -rf cards
+    python3 nex_render.py results/summary.json cards results/judgments.json \
+      hf-upload/std/SHA256SUMS hf-upload/imat/SHA256SUMS hf-upload/staged_sizes.json
+    ls -la cards; log "NEX_CARDS_DONE" ;;
+  upload)
+    exec 9>/mnt/models/.quiet-box.lock; log "waiting for the quiet-box lock"; flock 9; log "quiet-box lock held"
+    for k in std imat; do
+      bash upload_watchdog.sh nex-upload-$k $W python3 nex_publish.py upload $k || { log "upload $k FAILED"; exit 1; }
+    done
+    log "NEX_UPLOAD_ALL_DONE" ;;
+  verify)
+    rc=0; for k in std imat; do python3 nex_publish.py verify $k || rc=1; done
+    log "NEX_VERIFY rc=$rc"; exit $rc ;;
+  *) echo "usage: $0 cards|upload|verify"; exit 2 ;;
+esac

recipe/pipeline/nex_download.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+# Nex-N2.5-mini: pinned download + byte/sha256 gate. Runs inside a capped scope (2 cores, 16 GiB, idle I/O) so the
+# Agnes measurements still running on this box are not disturbed. HF_HOME stays default (token lives there).
+set -uo pipefail
+R=nex-agi/Nex-N2.5-mini; REV=87420286149d9cce9bd46cd335ef9bda33c37c1b
+W=/mnt/models/nex-n2.5-mini; mkdir -p $W/hf $W/logs
+# 8 parallel files x default range concurrency grew hf to 5.6 GiB RSS and the 6 GiB cap OOM-killed it (20:56Z).
+export HF_XET_CACHE=/mnt/models/.hf-xet-cache TMPDIR=/mnt/models/.tmp HF_HUB_DISABLE_TELEMETRY=1 HF_XET_NUM_CONCURRENT_RANGE_GETS=8
+mkdir -p "$HF_XET_CACHE" "$TMPDIR"
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+t0=$(date +%s)
+log "download $R @ $REV -> $W/hf"
+hf download "$R" --revision "$REV" --local-dir "$W/hf" --max-workers 4
+rc=$?
+log "DONE_RC=$rc elapsed=$(( $(date +%s) - t0 ))s  on disk: $(du -sh --apparent-size $W/hf | cut -f1)"
+if [ $rc -eq 0 ]; then
+  log "verify (size + sha256 vs the tree API lfs.oid at the pinned revision)"
+  python3 $W/verify_download.py "$R" "$REV" "$W/hf" > $W/logs/D2_verify_download.log 2>&1; vrc=$?
+  tail -3 $W/logs/D2_verify_download.log
+  log "VERIFY_RC=$vrc"
+  echo "$vrc" > $W/logs/DOWNLOAD_VERIFY_RC
+fi
+echo "$rc" > $W/logs/DOWNLOAD_RC
+log "NEX_DOWNLOAD_END"

recipe/pipeline/nex_harness.py

@@ -0,0 +1,326 @@
+#!/usr/bin/env python3
+"""Measurement harness for Nex-N2.5-mini on MAX-1 (stdlib only) - derived from the Agnes harness.
+Nex's template has no thinking on/off flag: thinking is `reasoning_effort` = none (off) | high (on) | anything else
+(adaptive). Tool calls use the upstream sampling (temp 0.7, top_p 0.95, top_k 40).
+House protocol (Qwen3.8-27B card): ctx 65536, batch 1, greedy (temp 0, top_k 1), ignore_eos -> exactly
+256 tokens, unique nonce + cache_prompt:false (cached tokens asserted 0), median of reps after 1 warm-up.
+Subcommands:  bench | cachegate | tools | vision   (see argparse)"""
+import argparse, base64, json, os, signal, statistics, subprocess, sys, time, urllib.request, uuid
+
+W = "/mnt/models/nex-n2.5-mini"
+CAL = "/mnt/models/agnes-3.0-flash/calib"      # shared calibration / prose corpora
+THINK_OFF = {"reasoning_effort": "none"}
+THINK_ON = {"reasoning_effort": "high"}
+BIN = os.environ.get("AGNES_BIN", "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin")
+ENV = dict(os.environ, LD_LIBRARY_PATH=f"{BIN}:/opt/rocm-7.2.4/lib", HSA_OVERRIDE_GFX_VERSION="11.5.1",
+           GGML_HIP_ENABLE_UNIFIED_MEMORY="1")
+
+def post(port, path, body, timeout=1800):
+    req = urllib.request.Request(f"http://127.0.0.1:{port}{path}", data=json.dumps(body).encode(),
+                                 headers={"Content-Type": "application/json"})
+    return json.load(urllib.request.urlopen(req, timeout=timeout))
+
+class Server:
+    def __init__(self, a, port, mmproj=None, fa="on"):
+        self.port = port
+        cmd = [f"{BIN}/llama-server", "-m", a.model, "-dev", a.dev, "-ngl", "999", "-fa", fa, "-dio",
+               "--jinja", "-fit", "off", "--parallel", "1", "-c", str(a.ctx), "-b", "2048", "-ub", "1024",
+               "--host", "127.0.0.1", "--port", str(port), "--no-webui"]
+        if a.draft or a.mtp_infile:
+            cmd += ["--spec-type", "draft-mtp"]
+            if a.draft:
+                cmd += ["--model-draft", a.draft, "--spec-draft-ngl", "99", "--spec-draft-device", a.dev]
+            cmd += ["--spec-draft-n-max", str(a.nmax), "--spec-draft-n-min", "0", "--spec-draft-p-min", str(a.pmin)]
+            if a.strict:
+                cmd += ["--spec-mtp-strict-qwen"]
+        if mmproj:
+            cmd += ["--mmproj", mmproj]
+        self.cmd = cmd
+        self.logf = open(a.serverlog, "w")
+        self.t0 = time.time()
+        self.p = subprocess.Popen(cmd, env=ENV, stdout=self.logf, stderr=subprocess.STDOUT, start_new_session=True)
+        while True:
+            if self.p.poll() is not None:
+                raise SystemExit(f"SERVER DIED rc={self.p.returncode} see {a.serverlog}")
+            try:
+                if json.load(urllib.request.urlopen(f"http://127.0.0.1:{port}/health", timeout=3)).get("status") == "ok":
+                    break
+            except Exception:
+                pass
+            if time.time() - self.t0 > 900:
+                self.stop(); raise SystemExit("SERVER LOAD TIMEOUT")
+            time.sleep(2)
+        self.load_s = time.time() - self.t0
+    def stop(self):
+        try: os.killpg(self.p.pid, signal.SIGTERM); self.p.wait(60)
+        except Exception:
+            try: os.killpg(self.p.pid, signal.SIGKILL)
+            except Exception: pass
+        self.logf.close()
+
+def prompt_8k(kind):
+    if kind == "code":
+        src = open("/opt/llama-rocm/rocmfpx-724/convert_hf_to_gguf.py").read()[:30000]
+        return src, "Above is part of a model converter. Write a new, complete Python function that validates a GGUF tensor-name map against a list of HF tensor names and reports unmapped names. Code only."
+    txt = open(f"{CAL}/wikitext-2-raw/wiki.train.raw").read()[:34000]
+    return txt, "Above are encyclopedia excerpts. Write a long, detailed new encyclopedia article in the same style about the history of lighthouses."
+
+def one_request(port, ctx_text, instr, n=256, cache=False, nonce=True):
+    tag = f"[req {uuid.uuid4()}]\n" if nonce else ""
+    body = {"messages": [{"role": "user", "content": tag + ctx_text + "\n\n" + instr}],
+            "max_tokens": n, "temperature": 0, "top_k": 1, "ignore_eos": True, "cache_prompt": cache,
+            "chat_template_kwargs": THINK_OFF}
+    t = time.time(); r = post(port, "/v1/chat/completions", body); el = time.time() - t
+    tm = r.get("timings", {})
+    return {"pred_n": tm.get("predicted_n"), "tg": tm.get("predicted_per_second"),
+            "pp": tm.get("prompt_per_second"), "prompt_n": tm.get("prompt_n"), "cache_n": tm.get("cache_n"),
+            "draft_n": tm.get("draft_n"), "draft_acc": tm.get("draft_n_accepted"), "wall": el,
+            "usage": r.get("usage", {})}
+
+def cmd_bench(a):
+    s = Server(a, a.port)
+    out = {"label": a.label, "model": os.path.basename(a.model),
+           "draft": os.path.basename(a.draft) if a.draft else ("in-file" if a.mtp_infile else None),
+           "nmax": a.nmax if (a.draft or a.mtp_infile) else None, "strict": bool(a.strict), "bin": BIN, "dev": a.dev, "ctx": a.ctx, "workload": a.workload,
+           "load_s": round(s.load_s, 1), "cmd": " ".join(s.cmd), "runs": []}
+    try:
+        ctx_text, instr = prompt_8k(a.workload)
+        one_request(a.port, ctx_text, instr)                      # warm-up, discarded
+        for _ in range(a.reps):
+            r = one_request(a.port, ctx_text, instr)
+            cached = (r["usage"].get("prompt_tokens_details") or {}).get("cached_tokens", 0) or 0
+            if r["pred_n"] != 256: raise SystemExit(f"GATE FAIL: predicted_n={r['pred_n']} != 256")
+            if (r["cache_n"] or 0) != 0 or cached != 0: raise SystemExit(f"GATE FAIL: cache hit cache_n={r['cache_n']} cached={cached}")
+            out["runs"].append(r)
+    finally:
+        s.stop()
+    tg = [r["tg"] for r in out["runs"]]; pp = [r["pp"] for r in out["runs"]]
+    dn = sum(r["draft_n"] or 0 for r in out["runs"]); da = sum(r["draft_acc"] or 0 for r in out["runs"])
+    out.update(tg_median=round(statistics.median(tg), 2), tg_min=round(min(tg), 2), tg_max=round(max(tg), 2),
+               pp_median=round(statistics.median(pp), 1), prompt_n=out["runs"][0]["prompt_n"],
+               accept=(round(da / dn, 3) if dn else None))
+    print(json.dumps({k: v for k, v in out.items() if k != "runs"}))
+    with open(a.jsonl, "a") as f: f.write(json.dumps(out) + "\n")
+
+def cmd_identity(a):
+    """Fixed prompt, no nonce, no cache, greedy: return the exact generated text for cross-config diffing."""
+    s = Server(a, a.port)
+    texts = []
+    try:
+        ctx_text, instr = prompt_8k(a.workload)
+        for _ in range(a.reps):
+            body = {"messages": [{"role": "user", "content": ctx_text + "\n\n" + instr}], "max_tokens": 256,
+                    "temperature": 0, "top_k": 1, "ignore_eos": True, "cache_prompt": False,
+                    "chat_template_kwargs": THINK_OFF}
+            r = post(a.port, "/v1/chat/completions", body)
+            texts.append(r["choices"][0]["message"].get("content") or "")
+    finally:
+        s.stop()
+    import hashlib
+    res = {"label": a.label, "reps": a.reps, "sha256": [hashlib.sha256(t.encode()).hexdigest()[:16] for t in texts],
+           "self_consistent": len(set(texts)) == 1, "text": texts[0]}
+    print(json.dumps({k: v for k, v in res.items() if k != "text"}))
+    with open(a.jsonl, "a") as f: f.write(json.dumps(res) + "\n")
+
+def cmd_cachegate2(a):
+    """Partial-prefix reuse WITH output identity: warm on A, run B warm (partial reuse), run B cold, diff outputs."""
+    import hashlib
+    s = Server(a, a.port)
+    rows = []
+    try:
+        ctx_text, _ = prompt_8k("code")
+        variants = ["Write a function that lists every tensor name in this file.",
+                    "Write a function that counts the model classes registered in this file.",
+                    "Write a function that finds the longest method in this file."]
+        for i, vb in enumerate(variants[:a.reps]):
+            def req(instr, n, cache):
+                body = {"messages": [{"role": "user", "content": ctx_text + "\n\n" + instr}], "max_tokens": n,
+                        "temperature": 0, "top_k": 1, "ignore_eos": True, "cache_prompt": cache,
+                        "chat_template_kwargs": THINK_OFF}
+                r = post(a.port, "/v1/chat/completions", body)
+                tm = r.get("timings", {})
+                return (r["choices"][0]["message"].get("content") or ""), tm.get("prompt_n"), tm.get("cache_n")
+            req(f"Summarise this file. (warm-up {i})", 16, True)
+            tw, pw, cw = req(vb, 192, True)
+            tc, pc, cc = req(vb, 192, False)
+            rows.append({"variant": i, "warm_prompt_n": pw, "warm_cache_n": cw, "cold_prompt_n": pc, "cold_cache_n": cc,
+                         "identical": tw == tc, "warm_sha": hashlib.sha256(tw.encode()).hexdigest()[:12],
+                         "cold_sha": hashlib.sha256(tc.encode()).hexdigest()[:12]})
+            print("  ", json.dumps(rows[-1]), flush=True)
+    finally:
+        s.stop()
+    reused = all((r["warm_cache_n"] or 0) > 0 for r in rows)
+    ident = all(r["identical"] for r in rows)
+    res = {"label": a.label, "rows": rows, "all_reused": reused, "all_identical": ident,
+           "result": "PASS" if (reused and ident) else "FAIL"}
+    print(json.dumps({k: v for k, v in res.items() if k != "rows"}))
+    with open(a.jsonl, "a") as f: f.write(json.dumps(res) + "\n")
+    sys.exit(0 if res["result"] == "PASS" else 1)
+
+def cmd_cachegate(a):
+    """Repeat-prompt test WITH the draft head loaded: turn 2 must reuse turn 1's prefix."""
+    s = Server(a, a.port)
+    try:
+        ctx_text, instr = prompt_8k("code")
+        r1 = one_request(a.port, ctx_text, instr, n=32, cache=True, nonce=False)
+        r2 = one_request(a.port, ctx_text, instr + " Also add type hints.", n=32, cache=True, nonce=False)
+    finally:
+        s.stop()
+    res = {"label": a.label, "turn1_prompt_n": r1["prompt_n"], "turn2_prompt_n": r2["prompt_n"],
+           "turn2_cache_n": r2["cache_n"]}
+    # hybrid recurrent models resume only from context checkpoints (~1024-token spacing), so the test is
+    # "any prefix reuse" -- report the fraction rather than demand near-total reuse.
+    ok = isinstance(r2["cache_n"], int) and r2["cache_n"] > 0
+    res["reuse_fraction"] = round(r2["cache_n"] / r1["prompt_n"], 3) if ok and r1["prompt_n"] else 0.0
+    res["result"] = "PASS" if ok else "FAIL"
+    print(json.dumps(res))
+    with open(a.jsonl, "a") as f: f.write(json.dumps(res) + "\n")
+    sys.exit(0 if ok else 1)
+
+TOOLS = [
+ {"type": "function", "function": {"name": "get_weather", "description": "Current weather for a city",
+  "parameters": {"type": "object", "properties": {"city": {"type": "string"},
+   "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}}, "required": ["city", "unit"]}}},
+ {"type": "function", "function": {"name": "create_event", "description": "Create a calendar event",
+  "parameters": {"type": "object", "properties": {"title": {"type": "string"},
+   "when": {"type": "object", "properties": {"date": {"type": "string"}, "time": {"type": "string"}},
+            "required": ["date", "time"]},
+   "attendees": {"type": "array", "items": {"type": "string"}}}, "required": ["title", "when", "attendees"]}}},
+]
+LEAK = ("<tool_call>", "<function=", "<parameter=", "<think>", "</think>")
+
+def chat(port, msgs, think, stream=False, tools=TOOLS):
+    body = {"messages": msgs, "tools": tools, "tool_choice": "auto", "temperature": 0.7, "top_p": 0.95,
+            "top_k": 40, "max_tokens": 8192,
+            "chat_template_kwargs": (THINK_ON if think else THINK_OFF)}
+    if not stream:
+        return post(port, "/v1/chat/completions", body)["choices"][0]["message"]
+    body["stream"] = True
+    req = urllib.request.Request(f"http://127.0.0.1:{port}/v1/chat/completions", data=json.dumps(body).encode(),
+                                 headers={"Content-Type": "application/json"})
+    calls, content = {}, ""
+    for line in urllib.request.urlopen(req, timeout=1800):
+        line = line.decode().strip()
+        if not line.startswith("data:") or line.endswith("[DONE]"): continue
+        d = json.loads(line[5:])["choices"][0]["delta"]
+        content += d.get("content") or ""
+        for tc in d.get("tool_calls") or []:
+            c = calls.setdefault(tc["index"], {"name": "", "arguments": ""})
+            c["name"] += (tc.get("function") or {}).get("name") or ""
+            c["arguments"] += (tc.get("function") or {}).get("arguments") or ""
+    return {"content": content, "tool_calls": [{"function": v} for _, v in sorted(calls.items())]}
+
+def args_of(m, i=0):
+    return json.loads(m["tool_calls"][i]["function"]["arguments"])
+
+def cmd_tools(a):
+    s = Server(a, a.port)
+    results = {}
+    try:
+        for think in (True, False):
+            def check(name, fn):
+                try: ok, why = fn()
+                except Exception as e: ok, why = False, f"exception {e!r}"[:160]
+                results[f"{name}|think={think}"] = (ok, why)
+                print(f"  {'PASS' if ok else 'FAIL'} think={think!s:5} {name}: {why}", flush=True)
+            def clean(m): return not any(x in (m.get("content") or "") for x in LEAK)
+            def t1():
+                m = chat(a.port, [{"role": "user", "content": "What's the weather in Paris in celsius?"}], think)
+                ag = args_of(m); return (m["tool_calls"][0]["function"]["name"] == "get_weather" and ag.get("city", "").lower().startswith("paris")
+                        and ag.get("unit") == "celsius" and clean(m)), f"args={ag}"
+            def t2():
+                m = chat(a.port, [{"role": "user", "content": "Book 'Design review' on 2026-10-02 at 14:00 with ana@x.io and bo@x.io."}], think)
+                ag = args_of(m); return (isinstance(ag.get("when"), dict) and ag["when"].get("date") == "2026-10-02"
+                        and sorted(ag.get("attendees", [])) == ["ana@x.io", "bo@x.io"] and clean(m)), f"args={ag}"
+            def t3():
+                m = chat(a.port, [{"role": "user", "content": "Weather in Denver, and give it to me in fahrenheit."}], think)
+                return args_of(m).get("unit") == "fahrenheit" and clean(m), f"unit={args_of(m).get('unit')}"
+            def t4():
+                m = chat(a.port, [{"role": "user", "content": "What is 17 times 23? Answer directly."}], think)
+                c = m.get("content") or ""
+                return (not m.get("tool_calls")) and "391" in c and clean(m), f"content={c[:60]!r}"
+            def t5():
+                msgs = [{"role": "user", "content": "What's the weather in Tokyo in celsius?"}]
+                m = chat(a.port, msgs, think)
+                tc = m["tool_calls"][0]
+                msgs += [{"role": "assistant", "content": m.get("content") or "", "tool_calls": [
+                          {"id": "call_1", "type": "function", "function": tc["function"]}]},
+                         {"role": "tool", "tool_call_id": "call_1", "content": json.dumps({"temp_c": 21, "sky": "clear"})}]
+                m2 = chat(a.port, msgs, think)
+                c = m2.get("content") or ""
+                return ("21" in c and not m2.get("tool_calls") and clean(m2)), f"final={c[:70]!r}"
+            def t6():
+                m = chat(a.port, [{"role": "user", "content": "What's the weather in Rome in celsius?"}], think, stream=True)
+                ag = args_of(m); return (m["tool_calls"][0]["function"]["name"] == "get_weather"
+                        and ag.get("city", "").lower().startswith("rome") and clean(m)), f"stream args={ag}"
+            def t7():
+                m = chat(a.port, [{"role": "user", "content": "Get the weather in Oslo AND in Lima, both in celsius. Call the tool for each city."}], think)
+                cities = sorted(args_of(m, i).get("city", "").lower() for i in range(len(m.get("tool_calls") or [])))
+                return (len(cities) == 2 and cities[0].startswith("lima") and cities[1].startswith("oslo") and clean(m)), f"calls={cities}"
+            for nm, fn in (("multi-arg", t1), ("nested-object", t2), ("enum", t3), ("correct-decline", t4),
+                           ("multi-turn", t5), ("streaming", t6), ("parallel", t7)):
+                check(nm, fn)
+    finally:
+        s.stop()
+    n_ok = sum(v[0] for v in results.values())
+    summary = {"label": a.label, "passed": n_ok, "total": len(results),
+               "detail": {k: v[0] for k, v in results.items()}}
+    print(json.dumps(summary))
+    with open(a.jsonl, "a") as f: f.write(json.dumps(summary) + "\n")
+
+def cmd_vision(a):
+    """Image gate. A server that fails to load or dies on the image is a RESULT (FAIL row), not a harness crash."""
+    fa = a.fa or "off"
+    res = {"label": a.label, "fa": fa, "mtp": bool(a.draft or a.mtp_infile), "expected": a.expect,
+           "answer": "", "hits": [], "error": None, "server_died": False, "server_log_errors": []}
+    c = ""
+    try:
+        s = Server(a, a.port, mmproj=a.mmproj, fa=fa)
+    except SystemExit as e:
+        res.update(error=f"server did not start: {e}", server_died=True); s = None
+    if s is not None:
+        try:
+            img = base64.b64encode(open(a.image, "rb").read()).decode()
+            body = {"messages": [{"role": "user", "content": [
+                        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img}"}},
+                        {"type": "text", "text": a.question}]}],
+                    "temperature": 0, "top_k": 1, "max_tokens": 300, "chat_template_kwargs": THINK_OFF}
+            try:
+                r = post(a.port, "/v1/chat/completions", body, timeout=900)
+                c = r["choices"][0]["message"].get("content") or ""
+            except Exception as e:
+                res["error"] = f"{type(e).__name__}: {e}"[:300]
+            time.sleep(1)
+            res["server_died"] = s.p.poll() is not None
+        finally:
+            s.stop()
+    try:
+        res["server_log_errors"] = [l.strip()[-200:] for l in open(a.serverlog, errors="replace")
+                                    if any(k in l for k in ("GGML_ABORT", "abort", "failed to process", " E "))][-5:]
+    except OSError:
+        pass
+    res["answer"] = c[:300]
+    res["hits"] = [w for w in a.expect.split(",") if w.lower() in c.lower()]
+    ok = res["error"] is None and not res["server_died"] and len(res["hits"]) == len(a.expect.split(","))
+    res["result"] = "PASS" if ok else "FAIL"
+    print(json.dumps(res))
+    with open(a.jsonl, "a") as f: f.write(json.dumps(res) + "\n")
+    sys.exit(0 if ok else 1)
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser()
+    ap.add_argument("cmd", choices=["bench", "cachegate", "cachegate2", "tools", "vision", "identity"])
+    ap.add_argument("--model", required=True); ap.add_argument("--draft")
+    ap.add_argument("--nmax", type=int, default=4); ap.add_argument("--pmin", type=float, default=0.0)
+    ap.add_argument("--dev", default="ROCm0"); ap.add_argument("--ctx", type=int, default=65536)
+    ap.add_argument("--port", type=int, default=18600); ap.add_argument("--reps", type=int, default=3)
+    ap.add_argument("--workload", default="code", choices=["code", "prose"])
+    ap.add_argument("--label", default=""); ap.add_argument("--jsonl", default=f"{W}/results/phase_b.jsonl")
+    ap.add_argument("--serverlog", default=f"{W}/logs/server_last.log")
+    ap.add_argument("--mmproj"); ap.add_argument("--image"); ap.add_argument("--question"); ap.add_argument("--expect")
+    ap.add_argument("--mtp-infile", action="store_true"); ap.add_argument("--strict", action="store_true")
+    ap.add_argument("--fa", choices=["on", "off", "auto"], help="vision only; default off")
+    a = ap.parse_args()
+    os.makedirs(os.path.dirname(a.jsonl), exist_ok=True)
+    {"bench": cmd_bench, "cachegate": cmd_cachegate, "tools": cmd_tools, "vision": cmd_vision,
+     "identity": cmd_identity, "cachegate2": cmd_cachegate2}[a.cmd](a)

recipe/pipeline/nex_judge.py

@@ -0,0 +1,584 @@
+#!/usr/bin/env python3
+"""Write judgments.json for the Nex-N2.5-mini cards. Sentence structure is authored; every number is
+computed from summary.json. A judgment whose inputs are missing is NOT emitted (the renderer then
+shows JUDGMENT PENDING).  usage: nex_judge.py <summary.json> <judgments.json>"""
+import json, math, statistics, sys
+
+S = json.load(open(sys.argv[1])); T = S.get("tiers") or {}; J = {}
+MiB = 1024 ** 2
+UB = S.get("n_ubatch") if S.get("n_ubatch") is not None else 1024
+SPEED_EQ = 3.0  # % — tiers closer than this are called speed-equivalent
+TWINS = (("q106i", "q106"), ("q102i", "q102"), ("q103i", "q103"))
+STD_TAGS = ("q106", "q102", "q103")
+IMAT_TAGS = ("q106i", "q102i", "q103i")
+NAMES = {"q106": "STRIX_LEAN", "q102": "COHERENT", "q103": "FAST",
+         "q106i": "STRIX_LEAN", "q102i": "COHERENT", "q103i": "FAST"}
+
+
+def pct(a, b):
+    return (b - a) / a * 100
+
+
+def sig(a, ea, b, eb):
+    return abs(b - a) / math.sqrt(ea * ea + eb * eb)
+
+
+def have(*v):
+    return all(x is not None for x in v)
+
+
+def bench(label):   # last row wins: a re-run supersedes an earlier row with the same label
+    return next((x for x in reversed(S.get("bench") or []) if x["label"] == label), None)
+
+
+def gate(label):
+    return next((x for x in reversed(S.get("gates") or []) if x.get("label") == label), None)
+
+
+def bn(tag, dev):
+    return bench(f"n-{tag}-{dev}")
+
+
+def twin_spread(key):
+    """Largest |gap| between an imatrix file and its standard twin: a measured noise floor."""
+    if not all(T.get(i, {}).get("same_tensor_types_as_standard") is True for i, _ in TWINS):
+        return None
+    vals = []
+    for ti, ts in TWINS:
+        for dev in ("rocm", "vk"):
+            x, y = bn(ti, dev), bn(ts, dev)
+            if not (x and y and x.get(key) is not None and y.get(key) is not None):
+                return None
+            vals.append(abs(pct(y[key], x[key])))
+    return max(vals) if vals else None
+
+
+TG_NOISE, PP_NOISE = twin_spread("tg_median"), twin_spread("pp_median")
+DEC_EQ = max(SPEED_EQ, TG_NOISE) if TG_NOISE is not None else None
+PRE_EQ = max(SPEED_EQ, PP_NOISE) if PP_NOISE is not None else None
+
+
+def rel(g, what, eq):
+    if abs(g) < eq:
+        return f"{what} within {abs(g):.1f} % of"
+    return f"{what} {abs(g):.1f} % {'faster' if g > 0 else 'slower'} than"
+
+
+def cmp_line(a_tag, b_tag, a_nm, b_nm):
+    """-> (text, a_advantage, a_disadvantage) or (None, None, None)."""
+    ra, rb, va, vb = bn(a_tag, "rocm"), bn(b_tag, "rocm"), bn(a_tag, "vk"), bn(b_tag, "vk")
+    if not all((ra, rb, va, vb)) or DEC_EQ is None or PRE_EQ is None:
+        return None, None, None
+    if not have(ra.get("tg_median"), rb.get("tg_median"), va.get("tg_median"), vb.get("tg_median"),
+                ra.get("pp_median"), rb.get("pp_median"),
+                ra.get("tg_min"), ra.get("tg_max"), rb.get("tg_min"), rb.get("tg_max")):
+        return None, None, None
+    g_r, g_v = pct(rb["tg_median"], ra["tg_median"]), pct(vb["tg_median"], va["tg_median"])
+    g_p = pct(rb["pp_median"], ra["pp_median"])
+    txt = (f"`{a_nm}` {rel(g_r, 'decodes', DEC_EQ)} `{b_nm}` on ROCm0 ({ra['tg_median']:.2f} vs "
+           f"{rb['tg_median']:.2f} tok/s; per-run ranges {ra['tg_min']:.2f}–{ra['tg_max']:.2f} and "
+           f"{rb['tg_min']:.2f}–{rb['tg_max']:.2f}) and {rel(g_v, 'decodes', DEC_EQ)[len('decodes '):]} it on Vulkan0 "
+           f"({va['tg_median']:.2f} vs {vb['tg_median']:.2f}), and {rel(g_p, 'prefills', PRE_EQ)} it on ROCm0 "
+           f"({ra['pp_median']:.0f} vs {rb['pp_median']:.0f} tok/s)")
+    adv = g_p >= PRE_EQ or g_r >= DEC_EQ or g_v >= DEC_EQ
+    dis = g_p <= -PRE_EQ or g_r <= -DEC_EQ or g_v <= -DEC_EQ
+    return txt, adv, dis
+
+
+NOISE_NOTE = ("" if DEC_EQ is None or PRE_EQ is None else
+              f"Speed gaps below {DEC_EQ:.1f} % (decode) and {PRE_EQ:.1f} % (prefill) are called a tie: the larger of "
+              f"{SPEED_EQ:.0f} % and the widest gap measured between files that do identical work per token (each imatrix "
+              f"file and its standard twin: decode {TG_NOISE:.1f} %, prefill {PP_NOISE:.1f} %).")
+
+
+def kld_cmp(a, b):
+    """KLD of tag a against tag b -> (pct change b->a, sigma, word). 'lower'/'higher' only at >= 2 sigma."""
+    x, y = T[a], T[b]
+    d = pct(y["kld_mean"], x["kld_mean"])
+    sg = sig(y["kld_mean"], y["kld_err"], x["kld_mean"], x["kld_err"])
+    word = "within noise of" if sg < 2 else ("lower than" if d < 0 else "higher than")
+    return d, sg, word
+
+
+def kld_ready(tags):
+    return all(have(T.get(k, {}).get("kld_mean"), T.get(k, {}).get("kld_err"),
+                    T.get(k, {}).get("size_bytes")) for k in tags)
+
+
+def kld_rank_text(tags):
+    items = sorted(tags, key=lambda t: T[t]["kld_mean"])
+    parts = []
+    for i, tag in enumerate(items):
+        x = T[tag]
+        bit = f"`{NAMES[tag]}` {x['kld_mean']:.4f}"
+        if i > 0:
+            lo = T[items[0]]["kld_mean"]
+            parts.append(f"{bit} ({pct(lo, x['kld_mean']):+.1f} % vs `{NAMES[items[0]]}`)")
+        else:
+            parts.append(bit)
+    return ", ".join(parts)
+
+
+def size_vs(a, b):
+    da, db = T[a]["size_bytes"], T[b]["size_bytes"]
+    if da == db:
+        return f"`{NAMES[a]}` and `{NAMES[b]}` are the same size ({da / MiB:.0f} MiB)"
+    smaller, larger = (a, b) if da < db else (b, a)
+    return (f"`{NAMES[smaller]}` is {(T[larger]['size_bytes'] - T[smaller]['size_bytes']) / MiB:.0f} MiB "
+            f"smaller than `{NAMES[larger]}`")
+
+
+# ---------- quality provenance ----------
+rep, rep_v, ref = S.get("repeat") or {}, S.get("repeat_vk") or {}, S.get("reference") or {}
+if rep.get("result") == "MATCH" and rep.get("rows") is not None and have(ref.get("cpu_chunk1"), ref.get("vulkan0_chunk1")):
+    both = rep_v.get("result") == "MATCH" and rep_v.get("rows") is not None
+    J["quality_provenance"] = (
+        f"Measured directly on these files, against BF16 logits computed **on the CPU** in the same session "
+        f"(first-window perplexity {ref['cpu_chunk1']:.4f}; Vulkan0 gave {ref['vulkan0_chunk1']:.4f} for the same "
+        f"window). Every file was graded on ROCm0 (the columns above) and again on Vulkan0. The STRIX_LEAN grade was run "
+        f"twice {'on each backend' if both else 'on ROCm0'} and every per-chunk row matched: {rep['rows']} of "
+        f"{rep['rows']} on ROCm0" + (f", {rep_v['rows']} of {rep_v['rows']} on Vulkan0." if both else ".")
+        + " Why not the GPU for the reference: see [Known issues](#known-issues-and-limits).")
+
+# ---------- quality by backend ----------
+vk_rows = [(t, T[t]["vk"]) for t in STD_TAGS + IMAT_TAGS if (T.get(t) or {}).get("vk") and have(
+    T[t].get("kld_mean"), T[t].get("kld_err"), T[t]["vk"].get("kld_mean"), T[t]["vk"].get("kld_err"))]
+if len(vk_rows) == 6:
+    parts_b, lower_vk, lower_rocm = [], 0, 0
+    for t, v in vk_rows:
+        a_, b_ = T[t]["kld_mean"], v["kld_mean"]
+        sg_ = sig(a_, T[t]["kld_err"], b_, v["kld_err"])
+        if sg_ >= 2:
+            lower_vk += b_ < a_; lower_rocm += a_ < b_
+        pc_ = f"{pct(a_, b_):+.1f}"
+        pc_ = "0.0" if pc_ in ("+0.0", "-0.0") else pc_
+        parts_b.append(f"{'imatrix ' if t.endswith('i') else ''}{NAMES[t]} {a_:.4f} / {b_:.4f} ({pc_} %, {sg_:.1f}σ)")
+    verdict = ("The two backends agree within noise on every file." if not (lower_vk or lower_rocm) else
+               f"Vulkan0's output is measurably closer to BF16 on {lower_vk} of 6 files and ROCm0's on {lower_rocm}.")
+    J["backend_quality_note"] = (
+        f"**Same files, same reference, graded on each backend** — KLD ROCm0 / Vulkan0: " + "; ".join(parts_b) + f". {verdict}")
+
+
+# ---------- imatrix verdict ----------
+def twin_kld_ok(st, im):
+    keys = ("kld_mean", "kld_err", "kld_median", "kld_p99", "same_top_p")
+    return have(*(T.get(st, {}).get(k) for k in keys), *(T.get(im, {}).get(k) for k in keys))
+
+
+if all(twin_kld_ok(st, im) for im, st in TWINS):
+    bits = []
+    improved, worse = [], []
+    for im, st in TWINS:
+        nm = NAMES[st]
+        a, b = T[st], T[im]
+        dk = pct(a["kld_mean"], b["kld_mean"])
+        sg = sig(a["kld_mean"], a["kld_err"], b["kld_mean"], b["kld_err"])
+        dmed = pct(a["kld_median"], b["kld_median"])
+        dp99 = pct(a["kld_p99"], b["kld_p99"])
+        dtop = b["same_top_p"] - a["same_top_p"]
+        if dk < 0 and sg >= 2:
+            verb = "improves"
+            improved.append(nm)
+        elif dk > 0 and sg >= 2:
+            verb = "is measurably worse than"
+            worse.append(nm)
+        else:
+            verb = "is within noise of"
+        bits.append(
+            f"**{nm}** {verb} the standard file on mean KLD "
+            f"({dk:+.1f} %, {sg:.1f}σ; {a['kld_mean']:.4f} → {b['kld_mean']:.4f}); "
+            f"median {dmed:+.1f} %, 99th-pct {dp99:+.1f} %, top-1 {dtop:+.2f} pp")
+    if len(improved) == 3:
+        head = "**The imatrix measurably improves all three tiers.** "
+    elif improved:
+        head = f"**The imatrix measurably improves {', '.join(improved)}.** "
+    else:
+        head = "**The imatrix does not measurably improve any of the three tiers on this corpus.** "
+    if worse:
+        head += f"**It is measurably worse on {', '.join(worse)}.** "
+    J["imat_verdict"] = head + "; ".join(bits) + "."
+
+
+# ---------- speed_note (standard card) ----------
+def backend_line(tag):
+    r, v = bn(tag, "rocm"), bn(tag, "vk")
+    if not have(r, v) or DEC_EQ is None or PRE_EQ is None:
+        return None
+    if not have(r.get("tg_median"), v.get("tg_median"), r.get("pp_median"), v.get("pp_median"),
+                r.get("tg_min"), r.get("tg_max"), v.get("tg_min"), v.get("tg_max")):
+        return None
+    g_t, g_p = pct(v["tg_median"], r["tg_median"]), pct(v["pp_median"], r["pp_median"])
+    return (f"`{NAMES[tag]}`: ROCm0 {rel(g_t, 'decodes', DEC_EQ)} Vulkan0 "
+            f"({r['tg_median']:.2f} vs {v['tg_median']:.2f} tok/s; ranges "
+            f"{r['tg_min']:.2f}–{r['tg_max']:.2f} / {v['tg_min']:.2f}–{v['tg_max']:.2f}) and "
+            f"{rel(g_p, 'prefills', PRE_EQ)} it ({r['pp_median']:.0f} vs {v['pp_median']:.0f} tok/s)")
+
+
+if DEC_EQ is not None and PRE_EQ is not None:
+    lines = [backend_line(t) for t in STD_TAGS]
+    cr, pr = bench("n-q106-rocm"), bench("n-q106-rocm-prose")
+    cv, pv = bench("n-q106-vk"), bench("n-q106-vk-prose")
+    prose = None
+    if have(cr, pr, cr and cr.get("tg_median"), pr and pr.get("tg_median")):
+        g = pct(cr["tg_median"], pr["tg_median"])
+        prose = (f"STRIX_LEAN workload range on ROCm0: code {cr['tg_median']:.2f} tok/s vs prose "
+                 f"{pr['tg_median']:.2f} ({rel(g, 'prose decodes', DEC_EQ)} code)")
+        if have(cv, pv, cv and cv.get("tg_median"), pv and pv.get("tg_median")):
+            gv = pct(cv["tg_median"], pv["tg_median"])
+            prose += (f"; Vulkan0 code {cv['tg_median']:.2f} vs prose {pv['tg_median']:.2f} "
+                      f"({rel(gv, 'prose decodes', DEC_EQ)} code)")
+        prose += "."
+
+        def _span(wl):
+            v = [(b_.get("prompt_n_min") if b_.get("prompt_n_min") is not None else b_.get("prompt_n"),
+                  b_.get("prompt_n_max") if b_.get("prompt_n_max") is not None else b_.get("prompt_n"))
+                 for b_ in S.get("bench") or [] if b_.get("workload") == wl]
+            v = [x for x in v if None not in x]
+            if not v:
+                return None
+            lo, hi = min(a for a, _ in v), max(b for _, b in v)
+            return f"{lo:,}" if lo == hi else f"{lo:,}–{hi:,}"
+        sc, sp = _span("code"), _span("prose")
+        if sc and sp:
+            prose += (f" Prompt lengths: code {sc} tokens (the first 30,000 characters of `convert_hf_to_gguf.py` plus "
+                      f"an instruction), prose {sp} tokens (the first 34,000 characters of wikitext-2 *train* plus a "
+                      f"writing instruction).")
+    if all(lines) and prose:
+        J["speed_note"] = " ".join(x + "." for x in lines) + " " + prose + (f" {NOISE_NOTE}" if NOISE_NOTE else "")
+
+
+# ---------- cache ----------
+d = gate("n-c3-q106")
+if d and d.get("all_reused") and d.get("all_aligned") and have(d.get("n"), d.get("L"), d.get("expect_cache_n"),
+                                                                d.get("identical")):
+    wm = [r_["warm_prompt_ms"] for r_ in d.get("rows") or [] if r_.get("warm_prompt_ms") is not None]
+    cm = [r_["cold_prompt_ms"] for r_ in d.get("rows") or [] if r_.get("cold_prompt_ms") is not None]
+    t_line = ""
+    if wm and cm:
+        t_line = (f" — median prefill **{statistics.median(wm) / 1000:.1f} s instead of "
+                  f"{statistics.median(cm) / 1000:.1f} s** cold "
+                  f"({statistics.median(cm) / statistics.median(wm):.1f}× faster)")
+    J["cache_note"] = (
+        f"In {d['n']} request pairs sharing a long prefix at one fixed prompt length of {d['L']:,} tokens, every "
+        f"second request resumed from the checkpoint the first one left {4 + UB:,} tokens before its end — "
+        f"**{d['expect_cache_n']:,} tokens reused "
+        f"({d['expect_cache_n'] / d['L'] * 100:.0f} %), {d['L'] - d['expect_cache_n']:,} processed**{t_line}. "
+        f"Each warm reply was byte-identical to a cold run of the same prompt in **{d['identical']}/{d['n']}** "
+        f"exchanges.\n\n"
+        f"llama-server processes the last `n_ubatch` + 4 tokens of every prompt as two batches so it can checkpoint "
+        f"there ({UB:,} + 4 = {4 + UB:,} tokens with the `-ub {UB}` used in these measurements, where the server "
+        f"default is `-ub 512`; [upstream PR #20288](https://github.com/ggml-org/llama.cpp/pull/20288)). A turn that "
+        f"resumes from a checkpoint left by a prompt of a *different* length therefore splits its tail differently "
+        f"from a cold run, and float rounding can flip a greedy token.")
+
+
+# ---------- tools ----------
+def _tools_ok(x):
+    return bool(x) and have(x.get("passed"), x.get("total"), x.get("detail"))
+
+
+tl = gate("n-tools-q106")                                                    # stock template
+FX = [gate(l) for l in ("n-tools-q106-roff", "n-tools-q106-roff-r2", "n-tools-q106-roff-r3")]   # quick-start config
+TD = S.get("tools_diag") or {}
+TF = S.get("template_fix") or {}
+if _tools_ok(tl) and all(_tools_ok(x) for x in FX) and have(
+        TD.get("stock_on_leaks"), TD.get("stock_on_replies"), TD.get("stock_on_reasoning_extracted"),
+        TD.get("nested_off_http500"), TD.get("nested_off_attempts")):
+    misses, on_n, on_ok, off_n, off_ok = [], 0, 0, 0, 0
+    for i, x in enumerate(FX, 1):
+        for k, v in x["detail"].items():
+            name, think = k.split("|think=")
+            if think == "True":
+                on_n += 1; on_ok += bool(v)
+            else:
+                off_n += 1; off_ok += bool(v)
+            if not v:
+                misses.append("`%s` with thinking %s (pass %d)" % (name, "on" if think == "True" else "off", i))
+    fp = TD.get("flag_probes") or {}
+    flag_txt = "; ".join("%s: %d of %d replies still had reasoning in `content`" % (lab, fp[key]["leaks"], fp[key]["n"])
+                         for key, lab in (("fmt-deepseek", "`--reasoning-format deepseek`"),
+                                          ("srv-kwargs-high", "`--chat-template-kwargs` with `reasoning_effort`"),
+                                          ("reasoning-on", "`--reasoning on`")) if key in fp)
+    J["tools_note"] = (
+        "**Stock chat template: %d/%d.** Every thinking-on check failed. Re-run with the raw replies kept, %d of %d "
+        "thinking-on replies carried the reasoning and a `</think>` in `content`, and %d had any "
+        "`reasoning_content`. llama-server builds its reasoning parser by rendering the template with "
+        "`enable_thinking` on and off; this template ignores `enable_thinking` (it switches on `reasoning_effort`), "
+        "so the parser finds no reasoning markers and extracts nothing. Server switches did not help (%s). With "
+        "thinking off, `nested-object` failed on an HTTP 500 — see [Known issues](#known-issues-and-limits).\n\n"
+        "**With the included `%s` and `--reasoning off` (the quick start): %s over three passes of the same suite "
+        "(%d/%d)** — thinking off %d/%d, thinking on %d/%d; the misses were %s. A pass requires a native "
+        "`tool_calls` entry with the right arguments and no think tags in `content`. Each check is a single sample at "
+        "the recommended temperature 0.7." % (
+            tl["passed"], tl["total"], TD["stock_on_leaks"], TD["stock_on_replies"],
+            TD["stock_on_reasoning_extracted"], flag_txt or "not measured", TF.get("file") or "—",
+            ", ".join("%d/%d" % (x["passed"], x["total"]) for x in FX),
+            sum(x["passed"] for x in FX), sum(x["total"] for x in FX), off_ok, off_n, on_ok, on_n,
+            ", ".join(misses) if misses else "none"))
+
+
+# ---------- the template fix: what each request option does (recommended configuration) ----------
+PR = TF.get("probes_roff") or {}
+if PR and all(_tools_ok(x) for x in FX):
+    def _probe(prefix):
+        rows_ = [v for k, v in PR.items() if k.split("|")[0] == prefix]
+        return len(rows_), sum(1 for v in rows_ if v.get("leaks")), sum(1 for v in rows_ if v.get("reasoning_len"))
+    on_ok = sum(bool(v) for x in FX for k, v in x["detail"].items() if k.endswith("|think=True"))
+    on_n = sum(1 for x in FX for k in x["detail"] if k.endswith("|think=True"))
+    off_ok = sum(bool(v) for x in FX for k, v in x["detail"].items() if k.endswith("|think=False"))
+    off_n = sum(1 for x in FX for k in x["detail"] if k.endswith("|think=False"))
+    lines = []
+    for label, prefix, think, dest in (
+            ("no `chat_template_kwargs`", "no-kwargs", "off (the server default with `--reasoning off`)", "—"),
+            ("`\"enable_thinking\": false`", "enable_thinking=false", "off", "—"),
+            ("`\"reasoning_effort\": \"none\"`", "reasoning_effort=none", "off", "—"),
+            ("`\"reasoning_effort\": \"high\"`", "reasoning_effort=high", "on", "**`content`** — do not use"),
+            ("`\"reasoning_effort\": \"medium\"`", "reasoning_effort=medium", "adaptive", "**`content`** — do not use")):
+        n, leaks, _ = _probe(prefix)
+        if n:
+            lines.append("| %s | %s | %s | %d of %d replies with think tags in `content` |" % (label, think, dest, leaks, n))
+    lines.insert(1 if lines else 0,
+                 "| `\"enable_thinking\": true` | on | `reasoning_content` | tool suite with thinking on: %d/%d "
+                 "(a pass requires no think tags in `content`) |" % (on_ok, on_n))
+    J["template_note"] = (
+        "Measured on the standard STRIX_LEAN file with the included template file and `--reasoning off` (greedy "
+        "probes: a direct question, a "
+        "one-word instruction and a tool request, each with the tool schema attached; tool suite: 3 passes):\n\n"
+        "| request | thinking | reasoning ends up in | measured |\n| --- | --- | --- | --- |\n" + "\n".join(lines)
+        + "\n\nSo: switch thinking with `enable_thinking` only. Thinking-off tool checks: %d/%d." % (off_ok, off_n))
+
+
+# ---------- vision ----------
+def vwhy(x):
+    if x.get("server_died"):
+        import re as _re
+        errs = [_re.sub(r"^[0-9.]+ [IWE] (srv +)?", "", e).strip() for e in x.get("server_log_errors") or []]
+        first = next((e for e in errs if "failed" in e.lower() or "error" in e.lower() or "abort" in e.lower()),
+                     None) or (errs[0] if errs else None)
+        return ("the server aborted" + (f" (`{first[:90]}`)" if first else ""))
+    if x.get("error"):
+        return f"the request failed ({x['error'][:80]})"
+    exp = x.get("expected") or ""
+    nexp = len(exp.split(",")) if exp else None
+    hits = x.get("hits") or []
+    if nexp:
+        return f"the reply named {len(hits)} of {nexp} expected terms"
+    return "the reply did not pass"
+
+
+on, off = gate("n-vision-q106-faon"), gate("n-vision-q106-faoff")
+if on and off:
+    def vok(x):
+        return x.get("result") == "PASS"
+    if vok(on) and vok(off):
+        J["vision_note"] = "✅ **Images work with `-fa on` and `-fa off`.**"
+    elif vok(off) and not vok(on):
+        J["vision_note"] = (f"⛔ **With `-fa on`, image requests fail** — {vwhy(on)}. With `-fa off` they work. "
+                            f"**For image input, serve with `-fa off`.**")
+        J["vision_quickstart_warning"] = ("> ⛔ **Images:** with `-fa on`, image requests fail on this build "
+                                          "([measured](#vision)). If you send images, use `-fa off`.")
+    elif vok(on) and not vok(off):
+        J["vision_note"] = (f"⛔ **With `-fa off`, image requests fail** — {vwhy(off)}. With `-fa on` they work. "
+                            f"**For image input, keep `-fa on` (the quick-start default).**")
+    else:
+        J["vision_note"] = (f"⛔ **Images failed in both `-fa` settings.** `-fa on`: {vwhy(on)}; "
+                            f"`-fa off`: {vwhy(off)}.")
+    vfx = gate("n-vision-q106-roff-faon")
+    if vfx and "vision_note" in J:
+        J["vision_note"] += (" The `-fa on` image test was repeated with the included template file and "
+                             f"`--reasoning off` (the quick start): {'✅ passed' if vok(vfx) else '❌ ' + vwhy(vfx)}.")
+
+
+# ---------- memory ----------
+rows_sz = [r for r in (S.get("sizing") or []) if r.get("label") == "strix-lean"]
+if rows_sz:
+    def mem_cell(r, key):
+        if r.get("result") == "LOAD_FAIL":
+            return "did not load"
+        v = r.get(key)
+        return f"{v:.2f} GiB" if v is not None else "—"
+    body = "\n".join(
+        f"| {r['ctx']:,} | {mem_cell(r, 'footprint_loaded_gib')} | {mem_cell(r, 'footprint_after_8k_gib')} |"
+        for r in rows_sz if r.get("ctx") is not None)
+    J["memory_note"] = (
+        "Measured footprint (drop in `MemAvailable`) of STRIX_LEAN with the vision projector, q8_0 KV cache, "
+        "`-cram 512`, one slot, no draft head:\n\n"
+        "| context | after load | after one request (30,000-character code prompt) |\n"
+        "| ---: | ---: | ---: |\n" + body +
+        "\n\nNo row was decode-benchmarked beyond that one request. Nothing beyond these rows was measured.")
+
+
+# ---------- recommendations ----------
+def recommend(lean, coh, fast):
+    """STRIX_LEAN is the flagship tier; the data decides whether COHERENT's quality or FAST's speed is worth taking
+    instead. -> (default tag, markdown) or (None, None) when an input is missing."""
+    lc, lc_adv, lc_dis = cmp_line(lean, coh, "STRIX_LEAN", "COHERENT")
+    fl, fl_adv, fl_dis = cmp_line(fast, lean, "FAST", "STRIX_LEAN")
+    if lc is None or fl is None:
+        return None, None
+    L, C, F = T[lean], T[coh], T[fast]
+    d_c, s_c, w_c = kld_cmp(coh, lean)
+    more = (C["size_bytes"] - L["size_bytes"]) / MiB
+    size_c = f"{abs(more):.0f} MiB {'more' if more > 0 else 'less'}"
+    if w_c == "lower than" and lc_adv and not lc_dis:
+        default = lean
+        head = (f"**Start with `STRIX_LEAN`; take `COHERENT` if quality matters more than speed.** {lc}. `COHERENT`'s "
+                f"KLD is {abs(d_c):.1f} % lower ({s_c:.1f}σ) for {size_c}.")
+    elif w_c == "lower than":
+        default = coh
+        head = (f"**Start with `COHERENT`.** Its KLD is {abs(d_c):.1f} % lower than `STRIX_LEAN`'s ({s_c:.1f}σ) for "
+                f"{size_c}, and the speed comparison does not clearly favour `STRIX_LEAN`: {lc}.")
+    elif w_c == "within noise of" and lc_dis and not lc_adv:
+        default = coh
+        head = (f"**Start with `COHERENT`.** Its KLD is {w_c} `STRIX_LEAN`'s ({d_c:+.1f} %, {s_c:.1f}σ) and it is "
+                f"measurably faster: {lc}. It costs {size_c}.")
+    else:
+        default = lean
+        head = (f"**Start with `STRIX_LEAN`.** `COHERENT`'s KLD is {w_c} it ({d_c:+.1f} %, {s_c:.1f}σ) for "
+                f"{size_c}; {lc}.")
+    d_f, s_f, w_f = kld_cmp(fast, lean)
+    kf = f"its KLD is {w_f} `STRIX_LEAN`'s ({F['kld_mean']:.4f} vs {L['kld_mean']:.4f}, {d_f:+.1f} %, {s_f:.1f}σ)"
+    if fl_adv and not fl_dis:
+        fast_txt = f"**Take `FAST` for speed:** {fl}; {kf}."
+    else:
+        fast_txt = f"`FAST` does not buy a clear speed gain here: {fl}; {kf}."
+    sizes = f"{size_vs(lean, coh)}; {size_vs(lean, fast)}."
+    body = (f"{head}\n\n{fast_txt}\n\nKLD order (lower is closer to BF16): {kld_rank_text((lean, coh, fast))}. "
+            f"{sizes} {NOISE_NOTE}")
+    return default, body
+
+
+def imat_pointer():
+    """Point the standard card at the imatrix repo; name only measurable (>= 2 sigma) twin improvements."""
+    repo = S.get("model_repo_imat")
+    if not repo:
+        return ""
+    link = f"[imatrix build](https://huggingface.co/{repo})"
+    bits = []
+    for im, st in TWINS:
+        if not twin_kld_ok(st, im):
+            return ""
+        d, sg, w = kld_cmp(im, st)
+        if w == "lower than":
+            bits.append(f"`{NAMES[st]}` {T[st]['kld_mean']:.4f} → {T[im]['kld_mean']:.4f} ({d:+.1f} %, {sg:.1f}σ)")
+    if bits:
+        return f"**The {link} is measurably closer to BF16 at the same size:** " + "; ".join(bits) + "."
+    return f"Importance-matrix twins of all three files: **{link}** (not measurably closer to BF16 on this corpus)."
+
+
+if kld_ready(STD_TAGS) and DEC_EQ is not None and PRE_EQ is not None:
+    default, body = recommend("q106", "q102", "q103")
+    if default:
+        ptr = imat_pointer()
+        J["std_recommendation"] = body + (f"\n\n{ptr}" if ptr else "")
+        J["std_default"] = T[default]["file"]
+
+if kld_ready(IMAT_TAGS) and DEC_EQ is not None and PRE_EQ is not None:
+    default_i, body_i = recommend("q106i", "q102i", "q103i")
+    if default_i:
+        cross = ""
+        if kld_ready(("q102",)):
+            d, sg, w = kld_cmp("q106i", "q102")
+            cross = (f"\n\nFor scale: the imatrix `STRIX_LEAN`'s KLD is {w} the *standard* `COHERENT`'s "
+                     f"({T['q106i']['kld_mean']:.4f} vs {T['q102']['kld_mean']:.4f}, {d:+.1f} %, {sg:.1f}σ), "
+                     f"{size_vs('q106i', 'q102')}.")
+        J["imat_recommendation"] = body_i + cross
+        J["imat_default"] = T[default_i]["file"]
+
+parts = []
+for nm, ti, ts in (("STRIX_LEAN", "q106i", "q106"), ("COHERENT", "q102i", "q102"), ("FAST", "q103i", "q103")):
+    for dev, dn in (("rocm", "ROCm0"), ("vk", "Vulkan0")):
+        a_, b_ = bn(ti, dev), bn(ts, dev)
+        if a_ and b_ and have(a_.get("tg_median"), b_.get("tg_median")):
+            parts.append(f"{nm} on {dn} {a_['tg_median']:.2f} vs {b_['tg_median']:.2f} tok/s")
+same = [T.get(i, {}).get("same_tensor_types_as_standard") for i, _ in TWINS]
+_hk = [T.get(i, {}).get("header_keys_only_in_imatrix") for i, _ in TWINS]
+_hs = [T.get(i, {}).get("header_keys_only_in_standard") for i, _ in TWINS]
+HDR = ""
+if all(x is not None for x in _hk + _hs) and not any(_hs) and len({tuple(x) for x in _hk}) == 1 and _hk[0]:
+    HDR = ", all in the GGUF header, which carries %d extra keys in each imatrix file: %s" % (
+        len(_hk[0]), ", ".join("`%s`" % k for k in _hk[0]))
+deltas = [T.get(i, {}).get("file_size_delta_bytes") for i, _ in TWINS]
+if len(parts) == 6 and all(x is True for x in same) and all(d is not None for d in deltas):
+    J["imat_speed_note"] = (
+        f"The imatrix changes scale values, not tensor types or sizes: tensor by tensor, each imatrix file has the "
+        f"same names, types and byte sizes as its standard twin, so the work per token is the same (the files differ "
+        f"in size by {deltas[0]} / {deltas[1]} / {deltas[2]} bytes{HDR})." + " "
+        f"Measured speed still differs — imatrix vs standard: " + "; ".join(parts) + "."
+        + (f" The largest of these decode gaps is {TG_NOISE:.1f} %; both cards call decode gaps below {DEC_EQ:.1f} % "
+           f"a tie." if TG_NOISE is not None and DEC_EQ is not None else ""))
+
+
+# ---------- known issues (model facts + protocol; always emit) ----------
+ki = [
+    "- **No MTP head.** The checkpoint's `config.json` declares `mtp_num_hidden_layers: 1`, but the weights contain "
+    "**no** `mtp.*` tensors (1,026 tensors total). There is no multi-token-prediction head and no speculative decoding "
+    "on these files — do not pass `--spec-type draft-mtp`.",
+    "- **The stock chat template ignores `enable_thinking`, and llama-server cannot separate its reasoning** "
+    "(see [Tool calling](#tool-calling)). Serve with the included `chat_template_enable_thinking.jinja` and "
+    "`--reasoning off`, and switch thinking per request with `enable_thinking` — see "
+    "[Reasoning controls](#reasoning-controls).",
+    "- **Earlier assistant turns are re-rendered with their reasoning**, so multi-turn contexts grow faster than with "
+    "templates that drop it.",
+    "- **`llama-server`'s host-RAM prompt cache defaults to 8 GiB** (`-cram 8192`). On a shared box, set `-cram` explicitly.",
+    "- Measured on Linux only (Ryzen AI Max+ 395, ROCm 7.2.4, unpatched `d3ca537`).",
+]
+_pn_vals = [b_.get("prompt_n_max") if b_.get("prompt_n_max") is not None else b_.get("prompt_n")
+            for b_ in (S.get("bench") or [])]
+_pn = max((v for v in _pn_vals if v is not None), default=None)   # longest prompt any timed decode followed
+_sz = {r_.get("ctx"): r_ for r_ in (S.get("sizing") or []) if r_.get("label") == "strix-lean"}
+_big = max(_sz) if _sz else None
+_ctx_txt = ("" if _big is None else
+            f" (a {_big:,}-token context was loaded in the memory test, not benchmarked)"
+            if _sz[_big].get("result") != "LOAD_FAIL" else "")
+for _c, _r in sorted(_sz.items()):
+    if _r.get("result") == "LOAD_FAIL":
+        ki.append(f"- **A {_c:,}-token context did not load** in the memory test (STRIX_LEAN + vision projector, "
+                  f"q8_0 KV cache, one slot) on this box.")
+ki.append(f"- **Not measured:** decode beyond a {_pn:,}-token prompt{_ctx_txt}, " if _pn is not None else
+          "- **Not measured:** decode at long context, ")
+ki[-1] += ("long-context quality, video input, concurrency above 1, and task-level accuracy. Perplexity/KLD measure "
+           "next-token fidelity to BF16 on prose, not reasoning or code correctness.")
+if have(ref.get("hip_rocm0_chunk1"), ref.get("cpu_chunk1"), ref.get("hip_rocm0_final")) and have(S.get("bf16", {}).get("ppl")):
+    ki.insert(0, f"- **Do not run the BF16 GGUF of this model on ROCm0 with this build.** `d3ca537` computes it wrong on "
+                 f"that path: wikitext-2 perplexity {ref['hip_rocm0_final']:.1f} on ROCm0 vs {S['bf16']['ppl']:.2f} on the CPU "
+                 f"(first window {ref['hip_rocm0_chunk1']:.1f} vs {ref['cpu_chunk1']:.2f}; also wrong with `-fa off`). "
+                 f"The 4-bit files are not affected — their ROCm0 grades are in the table — and the BF16 file is not "
+                 f"published here; the quality reference was computed on the CPU instead.")
+if have(TD.get("nested_off_http500"), TD.get("nested_off_attempts")):
+    ki.insert(1, "- **llama-server rejects a tool call whose required arguments are not in the order the schema lists "
+                 "them** — HTTP 500, *The model produced output that does not match the expected peg-native format*: "
+                 "this build's parser for the XML tool-call format expects required arguments in definition order. "
+                 "On the standard STRIX_LEAN file with the stock template and thinking off, the `nested-object` "
+                 "request (three required arguments) "
+                 "hit it in the tool suite%s and in %d of %d repeats; the model had written a well-formed call with "
+                 "the arguments reordered. Be ready to retry on this error." % (
+                     "" if TD.get("gate_http500_logged") else " (not confirmed in the suite's server log)",
+                     TD["nested_off_http500"], TD["nested_off_attempts"]))
+_PR = (S.get("template_fix") or {}).get("probes_roff") or {}
+_lk = [v for k, v in _PR.items() if k.split("|")[0] in ("reasoning_effort=high", "reasoning_effort=medium")]
+if _lk:
+    ki.insert(2, "- **With the included template, do not set `reasoning_effort` to `high` or `medium`** — the reasoning "
+                 "goes back into `content` (%d of %d probe replies). `enable_thinking: true` is the way to turn "
+                 "thinking on." % (sum(1 for v in _lk if v.get("leaks")), len(_lk)))
+_seat = [r_ for r_ in (S.get("seats") or {}).values() if r_.get("thinking_reasoning_len") is not None]
+_short = [r_ for r_ in _seat if not r_.get("thinking_reply") and r_.get("thinking_reasoning_len")]
+if _short:
+    ki.insert(3, "- **With thinking on, a very short answer can stay inside the think block.** Served as in the quick "
+                 "start, `Reply with the single word: ready` (no tools, `enable_thinking: true`, temperature 0) came "
+                 "back with the word in `reasoning_content` and an empty `content` on %d of %d FAST files tested.%s" % (
+                     len(_short), len(_seat),
+                     " With thinking off the same request returned `ready` in `content`."
+                     if all((r_.get("default_reply") or "").strip().lower().startswith("ready") for r_ in _seat) else ""))
+J["std_known_issues"] = "\n".join(ki)
+J["imat_known_issues"] = "\n".join([
+    "- Calibration text is general-purpose English/code (bartowski `calibration_datav3.txt`); an imatrix built from "
+    "your own domain can do better on that domain.",
+    "- Graded on wikitext-2 *test*, a different corpus from the calibration text. Improvements on reasoning/code "
+    "tasks were not measured.",
+] + ki)
+
+json.dump(J, open(sys.argv[2], "w"), indent=2)
+print("judgments written:", sorted(J))
+for k, v in J.items():
+    print(f"\n[{k}]\n{v}")

recipe/pipeline/nex_phase1.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+# Nex-N2.5-mini phase 1 (CPU): verified download -> BF16 GGUF + vision projector -> the three standard 4-bit tiers
+# (King: STRIX_LEAN + COHERENT + FAST; no Q8/Q6) with head protection, read back by exact tensor name.
+# Runs in the capped scope `nex-conv` (Agnes memory sizing waits for nex-* scopes). Waits until the Agnes BF16
+# re-grade (60 GiB on the GPU) is finished so the two large memory users never overlap.
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; T=/opt/llama-rocm/rocmfpx-724; B=$T/build-hipvk/bin; N=Nex-N2.5-mini
+A=/mnt/models/agnes-3.0-flash
+export LD_LIBRARY_PATH=$B:/opt/rocm-7.2.4/lib TMPDIR=/mnt/models/.tmp PYTHONUNBUFFERED=1
+cd $W; mkdir -p gguf out logs
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+until [ -f logs/DOWNLOAD_RC ]; do sleep 20; done
+if [ "$(cat logs/DOWNLOAD_RC)" != 0 ] || [ "$(cat logs/DOWNLOAD_VERIFY_RC 2>/dev/null)" != 0 ]; then
+  log "download or verify failed -> stop"; log "NEX_PHASE1_FAILED"; exit 1
+fi
+log "download verified; waiting for the Agnes BF16 re-grade to leave the GPU"
+until grep -q "R3 grade" $A/logs/regrade.log 2>/dev/null; do sleep 20; done
+
+log "C1 convert BF16 (the checkpoint has no mtp.* tensors, so no MTP block is emitted)"
+python3 $T/convert_hf_to_gguf.py hf --outtype bf16 --model-name "$N" --outfile gguf/$N-BF16.gguf > logs/C1_convert.log 2>&1
+rc=$?; log "C1 exit=$rc"; [ $rc -eq 0 ] || { tail -30 logs/C1_convert.log; log "NEX_PHASE1_FAILED"; exit 2; }
+log "C2 convert vision projector"
+python3 $T/convert_hf_to_gguf.py hf --outtype bf16 --mmproj --model-name "$N" --outfile out/mmproj-$N-BF16.gguf > logs/C2_mmproj.log 2>&1
+rc=$?; log "C2 exit=$rc"; [ $rc -eq 0 ] || { tail -30 logs/C2_mmproj.log; log "NEX_PHASE1_FAILED"; exit 3; }
+python3 readback.py - - gguf/$N-BF16.gguf | tee logs/C_readback.log
+python3 readback.py - - out/mmproj-$N-BF16.gguf | tee -a logs/C_readback.log
+
+BF=gguf/$N-BF16.gguf; Q=$B/llama-quantize
+log "Q1 standard tiers"
+$Q --output-tensor-type q6_K                             $BF out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf Q4_0_ROCMFP4_STRIX_LEAN 16 > logs/Q1_q106.log 2>&1; log "  q106 exit=$?"
+$Q --output-tensor-type q6_K --token-embedding-type q6_K $BF out/$N-Q4_0_ROCMFP4_COHERENT.gguf   Q4_0_ROCMFP4_COHERENT   16 > logs/Q1_q102.log 2>&1; log "  q102 exit=$?"
+$Q --output-tensor-type q6_K                             $BF out/$N-Q4_0_ROCMFP4_FAST.gguf       Q4_0_ROCMFP4_FAST       16 > logs/Q1_q103.log 2>&1; log "  q103 exit=$?"
+python3 readback.py Q6_K Q5_K out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf | tee logs/Q_readback.log
+python3 readback.py Q6_K Q6_K out/$N-Q4_0_ROCMFP4_COHERENT.gguf   | tee -a logs/Q_readback.log
+python3 readback.py Q6_K -    out/$N-Q4_0_ROCMFP4_FAST.gguf       | tee -a logs/Q_readback.log
+for l in Q1_q106 Q1_q102 Q1_q103; do printf "%-8s " $l; grep -oE "quant size\s*=\s*[0-9.]+ MiB \([0-9.]+ BPW\)" logs/$l.log; done | tee logs/Q_sizes.log
+log "NEX_PHASE1_DONE"

recipe/pipeline/nex_phase2.sh

@@ -0,0 +1,74 @@
+#!/bin/bash
+# Nex-N2.5-mini phase 2 (GPU, box iced): BF16 reference + imatrix -> imatrix tiers (CPU, alongside the standard grades)
+# -> KLD grade of all six 4-bit files + a repeat run -> speed on both backends + gates -> memory sizing -> un-ice.
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; A=/mnt/models/agnes-3.0-flash; B=/opt/llama-rocm/rocmfpx-724/build-hipvk/bin; N=Nex-N2.5-mini
+export LD_LIBRARY_PATH=$B:/opt/rocm-7.2.4/lib HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1
+cd $W; mkdir -p kld imat out-imat results
+ulimit -c 1   # a crashing server must not pipe a multi-GiB core into apport on the full root disk
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+table(){
+  python3 - "$1" <<'EOF'
+import sys
+out = []
+for l in open(sys.argv[1], errors="replace"):
+    t = l.split()
+    if len(t) > 2 and t[0].isdigit() and l.rstrip().endswith("%") and l.count("%") == 2:
+        out.append(" ".join(t))
+print("\n".join(out))
+EOF
+}
+until grep -q -E "NEX_PHASE1_(DONE|FAILED)" logs/phase1.log 2>/dev/null; do sleep 30; done
+grep -q NEX_PHASE1_DONE logs/phase1.log || { log "phase 1 failed -> stop (box stays iced)"; log "NEX_PHASE2_FAILED"; exit 1; }
+until grep -q AGNES_FINISH_DONE $A/logs/agnes_finish.log 2>/dev/null; do sleep 30; done
+log "box handed over from Agnes"
+PPL="$B/llama-perplexity -c 2048 -b 2048 -ngl 999 -fa on -dio -dev ROCm0"
+BF=gguf/$N-BF16.gguf; Q=$B/llama-quantize
+
+log "N1 BF16 perplexity + base logits (wikitext-2 test, 40 chunks)"
+$PPL -m $BF -f $A/calib/wikitext-2-raw/wiki.test.raw --chunks 40 --kl-divergence-base kld/bf16.kld > logs/N1_ppl_bf16.log 2>&1
+log "N1 exit=$? $(grep -E 'Final estimate' logs/N1_ppl_bf16.log)"
+
+log "N2 imatrix (BF16, bartowski calibration_datav3, c=512)"
+$B/llama-imatrix -m $BF -f $A/calib/calibration_datav3.txt -o imat/$N.imatrix -ngl 999 -c 512 -b 512 -fa on -dio -dev ROCm0 \
+  > logs/N2_imatrix.log 2>&1
+rc=$?; log "N2 exit=$rc"
+[ $rc -eq 0 ] && [ -s imat/$N.imatrix ] || { log "imatrix failed -> stop (box stays iced)"; log "NEX_PHASE2_FAILED"; exit 2; }
+sha256sum imat/$N.imatrix > imat/$N.imatrix.sha256
+
+log "N3 imatrix tiers on CPU (background) while N4a grades the standard tiers on the GPU"
+(
+  systemd-run --scope --quiet -p MemoryMax=24G -p MemorySwapMax=0 nice -n 5 bash -c "
+    $Q --imatrix imat/$N.imatrix --output-tensor-type q6_K $BF out-imat/$N-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf Q4_0_ROCMFP4_STRIX_LEAN 16 > logs/N3_q106i.log 2>&1
+    $Q --imatrix imat/$N.imatrix --output-tensor-type q6_K --token-embedding-type q6_K $BF out-imat/$N-imatrix-Q4_0_ROCMFP4_COHERENT.gguf Q4_0_ROCMFP4_COHERENT 16 > logs/N3_q102i.log 2>&1
+    $Q --imatrix imat/$N.imatrix --output-tensor-type q6_K $BF out-imat/$N-imatrix-Q4_0_ROCMFP4_FAST.gguf Q4_0_ROCMFP4_FAST 16 > logs/N3_q103i.log 2>&1"
+  python3 readback.py Q6_K Q5_K out-imat/$N-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf >  logs/N3_readback.log
+  python3 readback.py Q6_K Q6_K out-imat/$N-imatrix-Q4_0_ROCMFP4_COHERENT.gguf   >> logs/N3_readback.log
+  python3 readback.py Q6_K -    out-imat/$N-imatrix-Q4_0_ROCMFP4_FAST.gguf       >> logs/N3_readback.log
+  echo N3_DONE >> logs/N3_readback.log
+) &
+n3=$!
+grade(){ $PPL -m "$2" --kl-divergence-base kld/bf16.kld --kl-divergence > logs/N4_kld_$1.log 2>&1
+         log "  $1 exit=$? $(grep -E 'Mean +KLD' logs/N4_kld_$1.log | tr -s ' ')"; }
+grade q106 out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf
+grade q102 out/$N-Q4_0_ROCMFP4_COHERENT.gguf
+grade q103 out/$N-Q4_0_ROCMFP4_FAST.gguf
+wait $n3; log "N3 done: $(grep -c ^PASS logs/N3_readback.log)/3 read-back PASS"; cat logs/N3_readback.log
+for l in N3_q106i N3_q102i N3_q103i; do printf "%-9s " $l; grep -oE "quant size\s*=\s*[0-9.]+ MiB \([0-9.]+ BPW\)" logs/$l.log; done | tee -a logs/Q_sizes.log
+log "N4b grade the imatrix tiers"
+grade q106i out-imat/$N-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf
+grade q102i out-imat/$N-imatrix-Q4_0_ROCMFP4_COHERENT.gguf
+grade q103i out-imat/$N-imatrix-Q4_0_ROCMFP4_FAST.gguf
+
+log "N5 repeat the STRIX_LEAN grade; every per-chunk row must match"
+$PPL -m out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf --kl-divergence-base kld/bf16.kld --kl-divergence > logs/N5_kld_q106_repeat.log 2>&1
+a=$(table logs/N5_kld_q106_repeat.log); b=$(table logs/N4_kld_q106.log); n=$(printf '%s\n' "$a" | grep -c .)
+echo "{\"step\":\"N5\",\"rows\":$n,\"result\":\"$( [ "$n" -ge 40 ] && [ "$a" = "$b" ] && echo MATCH || echo MISMATCH )\"}" | tee results/nex_repeat.jsonl
+
+log "N6 speed + gates"
+python3 nex_bench.py > logs/N6_bench.log 2>&1; log "N6 exit=$?"
+log "N7 memory sizing"
+bash nex_sizing.sh > logs/N7_sizing.log 2>&1; log "N7 exit=$?"
+log "N8 un-ice"
+bash $A/unice.sh > logs/N8_unice.log 2>&1; log "N8 exit=$? $(tail -1 logs/N8_unice.log)"
+log "NEX_PHASE2_DONE"

recipe/pipeline/nex_phase2b.sh

@@ -0,0 +1,87 @@
+#!/bin/bash
+# Nex-N2.5-mini phase 2b. Phase 2 stopped at 22:24Z: this build's ROCm0 path computes the BF16 MoE GGUF wrong
+# (wikitext PPL 106 vs 7.8 for its own 4-bit quant; -fa off and "-ngl 0" with the default op offload are wrong too).
+# Controls, chunk 1: pure CPU (-dev none --no-op-offload) 5.6964, Vulkan0 5.6953.
+# => BF16 reference logits and the imatrix are computed on PURE CPU; every 4-bit file is graded on BOTH GPU backends
+#    against that reference; then speed + gates -> memory sizing -> Nex FAST seats -> un-ice (as phase 2).
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; A=/mnt/models/agnes-3.0-flash; B=/opt/llama-rocm/rocmfpx-724/build-hipvk/bin; N=Nex-N2.5-mini
+export LD_LIBRARY_PATH=$B:/opt/rocm-7.2.4/lib HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1
+cd $W; mkdir -p kld imat out-imat results
+prlimit --pid $$ --core=1:1   # RLIMIT_CORE of exactly 1 BYTE (bash `ulimit -c 1` is 1 KiB and does not stop apport)
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+table(){
+  python3 - "$1" <<'EOF'
+import sys
+out = []
+for l in open(sys.argv[1], errors="replace"):
+    t = l.split()
+    if len(t) > 2 and t[0].isdigit() and l.rstrip().endswith("%") and l.count("%") == 2:
+        out.append(" ".join(t))
+print("\n".join(out))
+EOF
+}
+chunk1(){ grep -oE '\[1\][0-9.]+' "$1" | head -1 | cut -d']' -f2; }
+CPU="-dev none -ngl 0 --no-op-offload -t 16"
+PPL="$B/llama-perplexity -c 2048 -b 2048"
+BF=gguf/$N-BF16.gguf; Q=$B/llama-quantize; TXT=$A/calib/wikitext-2-raw/wiki.test.raw
+
+log "N1c BF16 perplexity + base logits on pure CPU (wikitext-2 test, 40 chunks)"
+$PPL -m $BF -f $TXT --chunks 40 --kl-divergence-base kld/bf16.kld $CPU > logs/N1c_ppl_bf16_cpu.log 2>&1
+log "N1c exit=$? $(grep -E 'Final estimate' logs/N1c_ppl_bf16_cpu.log)"
+a=$(chunk1 logs/N1c_ppl_bf16_cpu.log); b=$(chunk1 logs/diag_bf16_vk_faon.log)
+if ! python3 -c "import sys; a, b = float('${a:-nan}'), float('${b:-nan}'); sys.exit(0 if abs(a - b) / b < 0.005 else 1)"; then
+  log "GATE FAIL: CPU chunk-1 PPL ${a:-missing} vs Vulkan0 control ${b:-missing} -> stop (box stays iced)"; log "NEX_PHASE2B_FAILED"; exit 1
+fi
+echo "{\"step\":\"N1c\",\"cpu_chunk1\":$a,\"vulkan0_chunk1\":$b,\"hip_rocm0_chunk1\":$(chunk1 logs/N1_ppl_bf16.log),\"hip_rocm0_faoff_chunk1\":$(chunk1 logs/diag_bf16_rocm_faoff.log),\"hip_rocm0_final\":$(grep -oE 'Final estimate: PPL = [0-9.]+' logs/N1_ppl_bf16.log | grep -oE '[0-9.]+$'),\"result\":\"PASS\"}" | tee results/nex_reference.jsonl
+log "N1c gate PASS (CPU $a vs Vulkan0 $b)"
+
+log "N2c imatrix on pure CPU (BF16, bartowski calibration_datav3, c=512)"
+$B/llama-imatrix -m $BF -f $A/calib/calibration_datav3.txt -o imat/$N.imatrix -c 512 -b 512 $CPU > logs/N2c_imatrix_cpu.log 2>&1
+rc=$?; log "N2c exit=$rc"
+[ $rc -eq 0 ] && [ -s imat/$N.imatrix ] || { log "imatrix failed -> stop (box stays iced)"; log "NEX_PHASE2B_FAILED"; exit 2; }
+sha256sum imat/$N.imatrix > imat/$N.imatrix.sha256
+
+log "N3 imatrix tiers on CPU (background) while the standard tiers are graded on the GPU"
+(
+  systemd-run --scope --quiet -p MemoryMax=24G -p MemorySwapMax=0 nice -n 5 bash -c "
+    $Q --imatrix imat/$N.imatrix --output-tensor-type q6_K $BF out-imat/$N-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf Q4_0_ROCMFP4_STRIX_LEAN 16 > logs/N3_q106i.log 2>&1
+    $Q --imatrix imat/$N.imatrix --output-tensor-type q6_K --token-embedding-type q6_K $BF out-imat/$N-imatrix-Q4_0_ROCMFP4_COHERENT.gguf Q4_0_ROCMFP4_COHERENT 16 > logs/N3_q102i.log 2>&1
+    $Q --imatrix imat/$N.imatrix --output-tensor-type q6_K $BF out-imat/$N-imatrix-Q4_0_ROCMFP4_FAST.gguf Q4_0_ROCMFP4_FAST 16 > logs/N3_q103i.log 2>&1"
+  python3 readback.py Q6_K Q5_K out-imat/$N-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf >  logs/N3_readback.log
+  python3 readback.py Q6_K Q6_K out-imat/$N-imatrix-Q4_0_ROCMFP4_COHERENT.gguf   >> logs/N3_readback.log
+  python3 readback.py Q6_K -    out-imat/$N-imatrix-Q4_0_ROCMFP4_FAST.gguf       >> logs/N3_readback.log
+  echo N3_DONE >> logs/N3_readback.log
+) &
+n3=$!
+grade(){  # tag file device logprefix
+  $PPL -m "$2" --kl-divergence-base kld/bf16.kld --kl-divergence -ngl 999 -fa on -dio -dev $3 > logs/$4_kld_$1.log 2>&1
+  log "  $4 $1 exit=$? $(grep -E 'Mean +KLD' logs/$4_kld_$1.log | tr -s ' ')"
+}
+for t in "q106 out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf" "q102 out/$N-Q4_0_ROCMFP4_COHERENT.gguf" "q103 out/$N-Q4_0_ROCMFP4_FAST.gguf"; do
+  set -- $t; grade $1 $2 ROCm0 N4; grade $1 $2 Vulkan0 N4v
+done
+wait $n3; log "N3 done: $(grep -c ^PASS logs/N3_readback.log)/3 read-back PASS"; cat logs/N3_readback.log
+for l in N3_q106i N3_q102i N3_q103i; do printf "%-9s " $l; grep -oE "quant size\s*=\s*[0-9.]+ MiB \([0-9.]+ BPW\)" logs/$l.log; done | tee -a logs/Q_sizes.log
+log "N4 grade the imatrix tiers (both backends)"
+for t in "q106i out-imat/$N-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf" "q102i out-imat/$N-imatrix-Q4_0_ROCMFP4_COHERENT.gguf" "q103i out-imat/$N-imatrix-Q4_0_ROCMFP4_FAST.gguf"; do
+  set -- $t; grade $1 $2 ROCm0 N4; grade $1 $2 Vulkan0 N4v
+done
+
+log "N5 repeat the STRIX_LEAN grade on both backends; every per-chunk row must match"
+: > results/nex_repeat.jsonl
+for pair in "ROCm0 N4 N5" "Vulkan0 N4v N5v"; do
+  set -- $pair
+  $PPL -m out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf --kl-divergence-base kld/bf16.kld --kl-divergence -ngl 999 -fa on -dio -dev $1 \
+    > logs/$3_kld_q106_repeat.log 2>&1
+  x=$(table logs/$3_kld_q106_repeat.log); y=$(table logs/$2_kld_q106.log); n=$(printf '%s\n' "$x" | grep -c .)
+  echo "{\"step\":\"$3\",\"device\":\"$1\",\"rows\":$n,\"result\":\"$( [ "$n" -ge 40 ] && [ "$x" = "$y" ] && echo MATCH || echo MISMATCH )\"}" | tee -a results/nex_repeat.jsonl
+done
+
+log "N6 speed + gates"
+python3 nex_bench.py > logs/N6_bench.log 2>&1; log "N6 exit=$?"
+log "N7 memory sizing"
+bash nex_sizing.sh > logs/N7_sizing.log 2>&1; log "N7 exit=$?"
+log "N8 Nex FAST seats + un-ice"
+bash $A/unice.sh > logs/N8_unice.log 2>&1; log "N8 exit=$? $(tail -1 logs/N8_unice.log)"
+log "NEX_PHASE2_DONE"

recipe/pipeline/nex_publish.py

@@ -0,0 +1,171 @@
+#!/usr/bin/env python3
+"""Stage (hardlinks), upload, and byte-verify the two Nex-N2.5-mini repos.
+usage: nex_publish.py stage|upload|verify|readme <std|imat> [README path]
+Upload through upload_watchdog.sh (40G cap, stall kill + retry) - see nex_cards.sh upload
+Env: default HF_HOME (token); HF_XET_CACHE / TMPDIR pinned to /mnt/models by the caller."""
+import hashlib, json, os, shutil, sys, urllib.request
+
+W = "/mnt/models/nex-n2.5-mini"; N = "Nex-N2.5-mini"; U = "kingjones777"
+REPOS = {"std": f"{U}/{N}-ROCmFP4-GGUF", "imat": f"{U}/{N}-ROCmFP4-imatrix-GGUF"}
+
+
+def pubname(fn):
+    """Published filename: the quant token must be hyphen-delimited or the Hub cannot parse the variant
+    (`Q4_0_ROCMFP4_STRIX_LEAN` reads as the garbage label `Q4_0_ROCMFP`)."""
+    return fn.replace("Q4_0_ROCMFP4_", "Q4_0-ROCmFP4-")
+
+
+MMPROJ = {f"mmproj-{N}-BF16.gguf": (f"{W}/out", f"mmproj-{N}-BF16.gguf")}
+# the stock chat template plus one line that maps enable_thinking onto reasoning_effort (served with
+# --chat-template-file; see the card's Reasoning controls) - shipped in both repos
+TPL = {"chat_template_enable_thinking.jinja": (f"{W}/tpl", "chat_template_enable_thinking.jinja")}
+TIERS = ("Q4_0_ROCMFP4_STRIX_LEAN", "Q4_0_ROCMFP4_COHERENT", "Q4_0_ROCMFP4_FAST")
+FILES = {
+    "std": {**{pubname(f"{N}-{t}.gguf"): (f"{W}/out", f"{N}-{t}.gguf") for t in TIERS}, **MMPROJ, **TPL},
+    "imat": {**{pubname(f"{N}-imatrix-{t}.gguf"): (f"{W}/out-imat", f"{N}-imatrix-{t}.gguf") for t in TIERS},
+             f"{N}.imatrix": (f"{W}/imat", f"{N}.imatrix"), **MMPROJ, **TPL},
+}
+P = "recipe/pipeline/"
+SCRIPTS = ("nex_download.sh", "verify_download.py", "nex_phase1.sh", "nex_phase2.sh", "diag_bf16.sh", "nex_phase2b.sh",
+           "nex_harness.py", "nex_bench.py", "nex_sizing.sh",
+           "nex_cachegate.py", "readback.py", "nex_aggregate.py", "nex_judge.py", "nex_render.py", "nex_publish.py",
+           "nex_cards.sh", "upload_watchdog.sh", "nex_seats.sh",
+           # tool-call / reasoning diagnosis (stock template 6/14) and the template fix
+           "diag_tools_run.sh", "nex_tools_diag.py", "nex_reasoning_probe.py", "nex_tools_tpl.py",
+           "nex_seat_default_probe.py", "run_tools_c1.sh", "run_tools_roff.sh", "nex_refresh_cards.sh", "nex_publish_finish.sh")
+LOGS = (
+    "D2_verify_download.log", "phase1.log", "phase2.log", "Q_sizes.log", "N6_bench.log", "N7_sizing.log",
+    "b_n-vision-q106-faon.log", "b_n-vision-q106-faoff.log",
+    "C1_convert.log", "C2_mmproj.log", "C_readback.log",
+    "Q1_q106.log", "Q1_q102.log", "Q1_q103.log", "Q_readback.log",
+    "N1c_ppl_bf16_cpu.log", "N2c_imatrix_cpu.log",
+    # the stopped first attempt on ROCm0 and the backend diagnosis that followed (the Known issues evidence)
+    "N1_ppl_bf16.log", "diag_ppl_q106_rocm_c4.log", "diag_bf16.log", "diag_bf16_rocm_faoff.log", "diag_bf16_vk_faon.log",
+    "diag_q106_vk_faon.log", "diag_bf16_cpu.log", "diag_bf16_purecpu_c1.log",
+    "N3_q106i.log", "N3_q102i.log", "N3_q103i.log", "N3_readback.log",
+    *(f"{p}_kld_{t}.log" for p in ("N4", "N4v") for t in ("q106", "q102", "q103", "q106i", "q102i", "q103i")),
+    "N5_kld_q106_repeat.log", "N5v_kld_q106_repeat.log", "b_n-c3-q106.log",
+    "b_n-tools-q106.log", "diag_tools.log", "diag_tools_server.log",
+    "probe_reasoning.log", *(f"probe_reasoning_{c}.log" for c in
+                             ("default", "fmt-deepseek", "srv-kwargs-high", "reasoning-on", "tpl-enable-thinking")),
+    "N6t_tools_tpl.log", "b_n-tools-q106-tpl.log", "b_n-tools-q106-tpl-probe.log",
+    "N6t_tools_tpl_medium.log", "b_n-tools-q106-tpl-medium.log", "b_n-tools-q106-tpl-medium-probe.log",
+    "probe_seat_default.log", "probe_seat_default_C1.log", "probe_seat_default_C2.log",
+    "N6t_tools_c1.log", "b_n-tools-q106-c1.log", "b_n-tools-q106-c1-probe.log", "b_n-vision-q106-c1-faon.log",
+    "N6t_tools_roff.log", "b_n-tools-q106-roff.log", "b_n-tools-q106-roff-r2.log", "b_n-tools-q106-roff-r3.log",
+    "b_n-tools-q106-roff-probe.log", "b_n-vision-q106-roff-faon.log",
+    "N8_unice.log", "N8a_seats.log", "N8b_seats.log", "N8c_seats.log", "N8d_seats.log",
+)
+RAW = ("nex_repeat.jsonl", "nex_reference.jsonl", "nex_bench.jsonl", "nex_sizing.jsonl", "hub_check.json",
+       "nex_tools_diag.json", "nex_reasoning_probe.json", "nex_seat_default_probe.json", "nex_template_shim.json",
+       "nex_template_shim_medium.json", "nex_template_shim_c1.json", "nex_template_shim_roff.json",
+       "nex_seats.jsonl", "nex_seats_plan.json")
+TEMPLATES = ("chat_template_enable_thinking_medium.jinja", "chat_template_enable_thinking_v2.jinja")  # tested, not used
+RECIPE = {
+    "recipe/results_summary.json": "results/summary.json",
+    **{P + x: x for x in SCRIPTS},
+    **{f"recipe/raw/{x}": f"results/{x}" for x in RAW},
+    **{f"recipe/logs/{x}": f"logs/{x}" for x in LOGS},
+    **{f"recipe/templates/{x}": f"tpl/{x}" for x in TEMPLATES},
+    "recipe/templates/chat_template_stock.jinja": "hf/chat_template.jinja",
+}
+
+
+def separation_errors(kind, names):
+    """The imatrix build is its OWN repo (King): no imatrix weights in the standard repo, no standard weights in the
+    imatrix repo. The vision projector is the only model file both carry."""
+    errs = []
+    for fn in names:
+        if not fn.endswith(".gguf") or fn.startswith("mmproj-"):
+            continue
+        is_imat = "-imatrix-" in fn
+        if kind == "std" and is_imat:
+            errs.append(f"imatrix model file in the standard repo: {fn}")
+        if kind == "imat" and not is_imat:
+            errs.append(f"standard model file in the imatrix repo: {fn}")
+    if kind == "imat" and f"{N}.imatrix" not in names:
+        errs.append("imatrix repo is missing the .imatrix file")
+    return errs
+
+
+assert REPOS["std"] != REPOS["imat"]
+assert set(FILES["std"]) & set(FILES["imat"]) == set(MMPROJ) | set(TPL), "only the projector + template may be shared"
+for _k in FILES:
+    assert not separation_errors(_k, list(FILES[_k])), separation_errors(_k, list(FILES[_k]))
+
+
+def git_blob_sha1(p):
+    """What the Hub reports as `oid` for a file stored in plain git (no LFS/Xet pointer)."""
+    h = hashlib.sha1(b"blob %d\0" % os.path.getsize(p))
+    with open(p, "rb") as fh:
+        for b in iter(lambda: fh.read(64 << 20), b""):
+            h.update(b)
+    return h.hexdigest()
+
+
+def sha(p):
+    h = hashlib.sha256()
+    with open(p, "rb") as fh:
+        for b in iter(lambda: fh.read(64 << 20), b""):
+            h.update(b)
+    return h.hexdigest()
+
+
+kind = sys.argv[2]; stage = f"{W}/hf-upload/{kind}"; repo = REPOS[kind]
+if sys.argv[1] == "stage":
+    missing = [f"{d}/{src}" for fn, (d, src) in FILES[kind].items() if not os.path.exists(f"{d}/{src}")] + \
+              [src for src in RECIPE.values() if not os.path.exists(f"{W}/{src}")]
+    if missing:
+        raise SystemExit(f"STAGE ABORT - missing: {missing}")
+    shutil.rmtree(stage, ignore_errors=True)
+    for fn, (d, src) in FILES[kind].items():
+        os.makedirs(stage, exist_ok=True); os.link(f"{d}/{src}", f"{stage}/{fn}")
+    for dst, src in RECIPE.items():
+        os.makedirs(os.path.dirname(f"{stage}/{dst}"), exist_ok=True); shutil.copy2(f"{W}/{src}", f"{stage}/{dst}")
+    sums = {fn: sha(f"{stage}/{fn}") for fn in sorted(FILES[kind])}
+    with open(f"{stage}/SHA256SUMS", "w") as fh:
+        for fn, h in sums.items():
+            fh.write(f"{h}  {fn}\n")
+    sizes = {x: os.path.getsize(f"{stage}/{x}") for x in sums}
+    json.dump(sizes, open(f"{stage}/../staged_sizes_{kind}.json", "w"))
+    print(json.dumps({"stage": stage, "files": len(sums), "bytes": sum(sizes.values()), "recipe_files": len(RECIPE)}))
+elif sys.argv[1] == "upload":
+    from huggingface_hub import HfApi
+    api = HfApi()
+    print("whoami:", api.whoami()["name"], flush=True)
+    api.create_repo(repo, repo_type="model", private=False, exist_ok=True)
+    api.upload_large_folder(repo_id=repo, repo_type="model", folder_path=stage,
+                            ignore_patterns=["README.md", ".cache/**"], num_workers=2)
+    print("UPLOAD_DONE", repo, flush=True)
+elif sys.argv[1] == "verify":
+    tree = json.load(urllib.request.urlopen(f"https://huggingface.co/api/models/{repo}/tree/main?recursive=true", timeout=60))
+    remote = {t["path"]: t for t in tree if t.get("type") == "file"}
+    local = dict(l.split()[::-1] for l in open(f"{stage}/SHA256SUMS").read().splitlines())
+    bad = []
+    for fn, h in local.items():
+        r = remote.get(fn)
+        if not r:
+            bad.append((fn, "MISSING remote")); continue
+        if r["size"] != os.path.getsize(f"{stage}/{fn}"):
+            bad.append((fn, f"size {r['size']}")); continue
+        lfs = r.get("lfs")
+        if lfs:
+            if lfs.get("oid") != h:
+                bad.append((fn, f"sha {lfs.get('oid')} != {h[:12]}"))
+        elif r.get("oid") != git_blob_sha1(f"{stage}/{fn}"):
+            bad.append((fn, f"git oid {r.get('oid')} does not match the staged bytes"))
+    for fn in list(RECIPE) + ["SHA256SUMS"]:
+        if fn not in remote:
+            bad.append((fn, "MISSING remote"))
+    stale = sorted(x for x in remote if x not in local and x not in RECIPE and x not in ("SHA256SUMS", "README.md", ".gitattributes"))
+    bad += [(x, "SEPARATION") for x in separation_errors(kind, list(remote))]
+    print(json.dumps({"repo": repo, "checked": len(local), "bad": bad, "unexpected_remote_files": stale,
+                      "result": "PASS" if not bad and not stale else "FAIL"}))
+    sys.exit(0 if not bad and not stale else 1)
+elif sys.argv[1] == "readme":
+    from huggingface_hub import HfApi
+    HfApi().upload_file(path_or_fileobj=sys.argv[3], path_in_repo="README.md", repo_id=repo, repo_type="model",
+                        commit_message="Model card: measured results, reproduction, known issues")
+    print("README_DONE", repo)
+else:
+    raise SystemExit(f"usage: {sys.argv[0]} stage|upload|verify|readme <std|imat> [README path]")

recipe/pipeline/nex_publish_finish.sh

@@ -0,0 +1,61 @@
+#!/bin/bash
+# Publish the two Nex repos (Jay, 2026-09-17). Order:
+#  1. wait until the OxCoder speed + memory runs are finished (its phase-2 log reaches "O7 exit"), so this upload never
+#     holds the quiet-box lock they need;
+#  2. upload std, then imat, through upload_watchdog.sh (40G cap, stall kill + retry) under the quiet-box lock;
+#  3. verify both repos: model files by size + sha256, every recipe file by content (git sha1 / LFS sha256);
+#  4. push the READMEs only once cards/.factcheck_ok exists (created by hand after the lane fact-checks are reviewed).
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; O=/mnt/models/oxcoder-9b; cd $W
+export HF_XET_CACHE=/mnt/models/.hf-xet-cache TMPDIR=/mnt/models/.tmp HF_HUB_DISABLE_TELEMETRY=1 PYTHONUNBUFFERED=1
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+log "waiting for the OxCoder speed + memory runs to finish (phase-2 log reaches O7)"
+until grep -q -E "O7 exit=|OXCODER_PHASE2_(DONE|FAILED)" $O/logs/phase2.log 2>/dev/null; do sleep 30; done
+exec 9>/mnt/models/.quiet-box.lock; log "waiting for the quiet-box lock"; flock 9; log "quiet-box lock held"
+ok=1
+for k in std imat; do
+  bash upload_watchdog.sh nex-upload-$k $W python3 nex_publish.py upload $k || { log "upload $k FAILED"; ok=0; }
+done
+exec 9>&-; log "quiet-box lock released"
+for k in std imat; do
+  python3 nex_publish.py verify $k > logs/verify_$k.log 2>&1 || ok=0
+  log "verify $k: $(tail -c 300 logs/verify_$k.log)"
+done
+python3 - > logs/verify_recipe_content.log 2>&1 <<'PY' || ok=0
+import hashlib, json, os, sys, urllib.request
+W = "/mnt/models/nex-n2.5-mini"
+REPOS = {"std": "kingjones777/Nex-N2.5-mini-ROCmFP4-GGUF", "imat": "kingjones777/Nex-N2.5-mini-ROCmFP4-imatrix-GGUF"}
+bad, n = [], 0
+for k, repo in REPOS.items():
+    tree = json.load(urllib.request.urlopen(
+        "https://huggingface.co/api/models/%s/tree/main?recursive=true" % repo, timeout=60))
+    remote = {t["path"]: t for t in tree if t.get("type") == "file"}
+    root = "%s/hf-upload/%s" % (W, k)
+    for dp, _, fns in os.walk(root + "/recipe"):
+        for fn in fns:
+            p = os.path.join(dp, fn)
+            rel = os.path.relpath(p, root)
+            r = remote.get(rel)
+            n += 1
+            if r is None:
+                bad.append((k, rel, "missing remote"))
+                continue
+            data = open(p, "rb").read()
+            if r.get("lfs"):
+                same = r["lfs"].get("oid") == hashlib.sha256(data).hexdigest()
+            else:
+                same = r.get("oid") == hashlib.sha1(b"blob %d\0" % len(data) + data).hexdigest()
+            if not same:
+                bad.append((k, rel, "content differs"))
+print(json.dumps({"recipe_files_checked": n, "bad": bad, "result": "PASS" if not bad else "FAIL"}))
+sys.exit(1 if bad else 0)
+PY
+log "recipe content: $(tail -c 300 logs/verify_recipe_content.log)"
+if [ $ok -ne 1 ]; then log "NEX_PUBLISH_FAILED (files)"; exit 1; fi
+log "files verified; waiting for cards/.factcheck_ok before the READMEs"
+until [ -f cards/.factcheck_ok ]; do sleep 30; done
+for k in std imat; do
+  python3 nex_publish.py readme $k cards/README_$k.md > logs/readme_$k.log 2>&1 || ok=0
+  log "readme $k: $(tail -1 logs/readme_$k.log)"
+done
+[ $ok -eq 1 ] && log "NEX_PUBLISH_DONE" || log "NEX_PUBLISH_FAILED (readme)"

recipe/pipeline/nex_reasoning_probe.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+"""Find a llama-server configuration that separates Nex's reasoning from `content` when thinking is on.
+Diagnostic follow-up to nex_tools_diag.py (reasoning_content empty, `...</think>` left in content for
+reasoning_effort=high and for the template's default). Each config: start the unpatched server with extra flags
+(and, for the last one, a template copy in which enable_thinking drives reasoning_effort), send the same three
+thinking-on requests, record content / reasoning_content / tool calls. Diagnostic only -> results/nex_reasoning_probe.json."""
+import json, os, re, sys, time
+from types import SimpleNamespace
+
+os.environ["AGNES_BIN"] = "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin"
+sys.path.insert(0, "/mnt/models/nex-n2.5-mini")
+import nex_harness as H  # noqa: E402
+
+W = H.W
+PORT = 18651
+OUT = f"{W}/results/nex_reasoning_probe.json"
+TPL_SRC = f"{W}/hf/chat_template.jinja"
+TPL_PATCHED = f"{W}/results/nex_template_enable_thinking.jinja"
+
+# template copy: if the caller passes enable_thinking (and no reasoning_effort), map it onto reasoning_effort
+src = open(TPL_SRC).read()
+shim = ("{%- if reasoning_effort is not defined and enable_thinking is defined %}"
+        "{%- set reasoning_effort = 'high' if enable_thinking else 'none' %}{%- endif %}\n")
+open(TPL_PATCHED, "w").write(shim + src)
+
+CONFIGS = [
+    ("default", [], {"reasoning_effort": "high"}),
+    ("fmt-deepseek", ["--reasoning-format", "deepseek"], {"reasoning_effort": "high"}),
+    ("srv-kwargs-high", ["--chat-template-kwargs", json.dumps({"reasoning_effort": "high"})], {}),
+    ("reasoning-on", ["--reasoning", "on"], {"reasoning_effort": "high"}),
+    ("tpl-enable-thinking", ["--chat-template-file", TPL_PATCHED], {"enable_thinking": True}),
+]
+REQS = [
+    ("correct-decline", "What is 17 times 23? Answer directly."),
+    ("multi-arg", "What's the weather in Paris in celsius?"),
+    ("parallel", "Get the weather in Oslo AND in Lima, both in celsius. Call the tool for each city."),
+]
+
+
+def patched_popen_cmd(a, port, extra):
+    """H.Server builds and launches its argv in __init__; append the extra flags through a temporary Popen hook."""
+    import subprocess
+    orig = subprocess.Popen
+
+    def popen(cmd, *args, **kw):
+        if cmd and str(cmd[0]).endswith("llama-server"):
+            cmd = list(cmd) + list(extra)
+        return orig(cmd, *args, **kw)
+    subprocess.Popen = popen
+    try:
+        return H.Server(a, port)
+    finally:
+        subprocess.Popen = orig
+
+
+report = {"template_shim": shim, "configs": {}}
+for name, extra, kwargs in CONFIGS:
+    a = SimpleNamespace(model=f"{W}/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", dev="ROCm0", ctx=16384,
+                        draft=None, mtp_infile=False, nmax=4, pmin=0.0, strict=False,
+                        serverlog=f"{W}/logs/probe_reasoning_{name}.log")
+    rec = {"extra": extra, "kwargs": kwargs, "results": {}}
+    try:
+        s = patched_popen_cmd(a, PORT, extra)
+    except SystemExit as e:
+        rec["error"] = str(e)
+        report["configs"][name] = rec
+        print(name, "SERVER FAILED", e, flush=True)
+        continue
+    try:
+        for rn, prompt in REQS:
+            body = {"messages": [{"role": "user", "content": prompt}], "tools": H.TOOLS, "tool_choice": "auto",
+                    "temperature": 0, "top_k": 1, "max_tokens": 2048, "chat_template_kwargs": kwargs}
+            try:
+                m = H.post(PORT, "/v1/chat/completions", body)["choices"][0]["message"]
+                c = m.get("content") or ""
+                rec["results"][rn] = {"content": c[:300], "reasoning": (m.get("reasoning_content") or "")[:300],
+                                      "tool_calls": [(t.get("function") or {}).get("name") for t in m.get("tool_calls") or []],
+                                      "leaks": [x for x in H.LEAK if x in c]}
+            except Exception as e:  # noqa: BLE001
+                rec["results"][rn] = {"error": repr(e)[:300]}
+            print(name, rn, json.dumps(rec["results"][rn])[:300], flush=True)
+        log = open(a.serverlog, errors="replace").read()
+        rec["chat_format_lines"] = sorted(set(re.findall(r"Chat format: \S+", log)))
+        rec["thinking_line"] = next((l.strip()[-80:] for l in log.splitlines() if "thinking =" in l), None)
+    finally:
+        s.stop()
+    report["configs"][name] = rec
+    time.sleep(3)
+json.dump(report, open(OUT, "w"), indent=1)
+print("NEX_REASONING_PROBE_DONE", OUT)

recipe/pipeline/nex_refresh_cards.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+# Re-render the Nex cards after script or text fixes without re-staging the model files: summary (reuses the staged
+# sha256), refresh every recipe copy in both stages with a FRESH mtime (plain cp; `cp -p` can hide a change from
+# upload_large_folder's cache), judge, render. Light work, niced; no HF traffic.
+set -euo pipefail
+W=/mnt/models/nex-n2.5-mini; cd $W
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+[ -s hf-upload/std/SHA256SUMS ] && [ -s hf-upload/imat/SHA256SUMS ] || { log "not staged yet -> run nex_cards.sh cards"; exit 3; }
+nice -n 19 ionice -c3 python3 nex_aggregate.py > logs/aggregate_refresh.log 2>&1; tail -3 logs/aggregate_refresh.log
+python3 - <<'PY'
+import os, shutil, sys
+sys.argv = ["nex_publish.py", "noop", "std"]
+src = open("nex_publish.py").read()
+ns = {}
+exec(compile(src[:src.index("kind = sys.argv[2]")], "nex_publish_head", "exec"), ns)
+W = ns["W"]
+changed = 0
+for kind in ("std", "imat"):
+    stage = f"{W}/hf-upload/{kind}"
+    for dst, s in ns["RECIPE"].items():
+        a, b = f"{W}/{s}", f"{stage}/{dst}"
+        if not os.path.exists(b) or open(a, "rb").read() != open(b, "rb").read():
+            os.makedirs(os.path.dirname(b), exist_ok=True)
+            shutil.copyfile(a, b)          # fresh mtime on purpose
+            changed += 1
+            print("refreshed", kind, dst)
+print("recipe files refreshed:", changed)
+PY
+python3 nex_judge.py results/summary.json results/judgments.json > logs/judge.log 2>&1; head -1 logs/judge.log | cut -c1-200
+rm -rf cards
+python3 nex_render.py results/summary.json cards results/judgments.json \
+  hf-upload/std/SHA256SUMS hf-upload/imat/SHA256SUMS hf-upload/staged_sizes.json
+log "NEX_CARDS_REFRESHED"

recipe/pipeline/nex_render.py

@@ -0,0 +1,752 @@
+#!/usr/bin/env python3
+"""Render both HF cards from summary.json + judgments.json. Every number comes from the summary;
+absent => '—'. Judgment prose lives in nex_judge.py and only ever references computed values.
+usage: nex_render.py <summary.json> <outdir> <judgments.json> <SHA256SUMS_std> <SHA256SUMS_imat> <staged_sizes.json>"""
+import json, math, os, sys
+
+S = json.load(open(sys.argv[1])); OUT = sys.argv[2]; os.makedirs(OUT, exist_ok=True)
+JUDGE = json.load(open(sys.argv[3]))
+SUMS = {}
+for sf in sys.argv[4:6]:
+    if os.path.exists(sf):
+        for line in open(sf):
+            h, fn = line.split(maxsplit=1); SUMS[fn.strip()] = h
+SIZES = json.load(open(sys.argv[6])) if len(sys.argv) > 6 and os.path.exists(sys.argv[6]) else {}
+N = S["model"]; STD, IMAT = S["model_repo_std"], S["model_repo_imat"]
+T = S.get("tiers") or {}; B = S.get("binary") or {}; SRC = S.get("source") or {}; AR = S.get("arch") or {}
+GiB, MiB = 1024 ** 3, 1024 ** 2
+UB = S.get("n_ubatch") if S.get("n_ubatch") is not None else 1024
+CTX, GEN, REPS = 65536, 256, 3
+STD_TAGS, IMAT_TAGS = ("q106", "q102", "q103"), ("q106i", "q102i", "q103i")
+NAMES = {"q106": "STRIX_LEAN", "q102": "COHERENT", "q103": "FAST",
+         "q106i": "STRIX_LEAN", "q102i": "COHERENT", "q103i": "FAST"}
+
+
+def g(v, fmt="{:.2f}"):
+    return "—" if v is None else fmt.format(v)
+
+
+def gib(b):
+    return g(None if b is None else b / GiB, "{:.2f} GiB")
+
+
+def pm(v, e, fmt="{:.4f}"):
+    return "—" if v is None else (fmt.format(v) + ("" if e is None else " ± " + fmt.format(e)))
+
+
+def bench(label):
+    return next((x for x in S.get("bench") or [] if x["label"] == label), None)
+
+
+def tg(label):
+    x = bench(label)
+    return None if x is None else x.get("tg_median")
+
+
+def pp(label):
+    x = bench(label)
+    return None if x is None else x.get("pp_median")
+
+
+def gate(label):
+    return next((x for x in S.get("gates") or [] if x.get("label") == label), None)
+
+
+def J(k):
+    return JUDGE.get(k, f"**[JUDGMENT PENDING: {k}]**")
+
+
+def Jopt(k):
+    return JUDGE.get(k, "")
+
+
+def speed_label(tag, dev):
+    return f"n-{tag}-{dev}"
+
+
+def prompt_range(workload="code"):
+    """Prompt tokens processed by the timed requests of one workload (each carries a unique nonce)."""
+    lo, hi = [], []
+    for b in S.get("bench") or []:
+        if b.get("workload") != workload:
+            continue
+        a_ = b.get("prompt_n_min") if b.get("prompt_n_min") is not None else b.get("prompt_n")
+        z_ = b.get("prompt_n_max") if b.get("prompt_n_max") is not None else b.get("prompt_n")
+        if a_ is not None and z_ is not None:
+            lo.append(a_); hi.append(z_)
+    if not lo:
+        return "—"
+    return f"{min(lo):,}" if min(lo) == max(hi) else f"{min(lo):,}–{max(hi):,}"
+
+
+def tier(tag):
+    return T.get(tag) or {}
+
+
+def tier_row(tag):
+    t = tier(tag); ratio = t.get("ppl_ratio")
+    r = [f"`{t.get('file') or '—'}`", g(t.get("ftype"), "{}"), gib(t.get("size_bytes")), g(t.get("bpw")),
+         pm(t.get("kld_mean"), t.get("kld_err")), g(t.get("same_top_p"), "{:.2f} %"),
+         pm(t.get("ppl"), t.get("ppl_err")) + ("" if ratio is None else f" (×{ratio:.4f})"),
+         g(tg(speed_label(tag, "rocm"))), g(tg(speed_label(tag, "vk"))),
+         g(pp(speed_label(tag, "rocm")), "{:.0f}")]
+    return "| " + " | ".join(r) + " |"
+
+
+PN_TXT = prompt_range("code")
+PROTOCOL = (
+    f"Ryzen AI Max+ 395 (MAX-1), ROCm 7.2.4, unpatched `llama-server` at `d3ca537` (see [Quick start](#quick-start)), "
+    f"`-c {CTX}`, one request at a time (`--parallel 1`), greedy (`temp 0`, `top_k 1`), `ignore_eos` so every arm "
+    f"generates exactly {GEN} tokens "
+    f"after a code prompt of {PN_TXT} tokens (the first 30,000 characters of `convert_hf_to_gguf.py` plus an "
+    f"instruction), a unique nonce per request and `cache_prompt: false` (`cache_n = 0` asserted on "
+    f"every timed request), 1 warm-up then the median of {REPS}. Decode numbers are the server's own "
+    f"`predicted_per_second`. Box iced: no other model loaded.")
+TABLE_HEAD = ("| File | ftype | Size | BPW | KLD vs BF16 ↓² | Same top-1 ↑ | PPL (× BF16) | TG ROCm0 | TG Vulkan0 | PP ROCm0 |\n"
+              "| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |")
+TG_NOTE = f"TG = decode tokens/s after the {PN_TXT}-token code prompt, no draft head. PP = prefill tokens/s on ROCm0."
+
+
+def quality_blurb():
+    b = S.get("bf16") or {}
+    ch, nc, st = b.get("chunks"), b.get("n_ctx"), b.get("scored_tokens")
+    if ch is not None and nc is not None and st is not None:
+        scored = (f"{ch} chunks × {nc // 2 - 1:,} scored tokens each — the second half of every window, less its first token — = {st:,}")
+    else:
+        scored = "— chunks × — scored tokens"
+    return (f"Quality is graded against the **BF16 GGUF** (reference logits computed on the CPU) on a **held-out** corpus (wikitext-2 *test*, `-c 2048`, "
+            f"{scored}), never on the imatrix calibration text. **KLD** is the per-token KL divergence of each "
+            "quant's next-token distribution from BF16's on the same tokens — far more sensitive than perplexity.")
+
+
+def bf16_row():
+    b = S.get("bf16") or {}
+    return (f"| *BF16 reference* | {g(b.get('ftype'), '{}')} | {gib(b.get('size_bytes'))}³ | 16 | 0 | 100 % | "
+            f"{pm(b.get('ppl_paired'), b.get('ppl_paired_err'))}¹ | — | — | — |")
+
+
+def footnotes(where="below"):
+    b = S.get("bf16") or {}
+    return (f"{TG_NOTE}\n"
+            f"¹ The BF16 PPL shown is the paired base every \"×\" ratio is computed against (averaged over the same scored tokens "
+            f"in the KL-divergence runs). The standalone BF16 run's own summary line reads {pm(b.get('ppl'), b.get('ppl_err'))}.\n"
+            f"² Quality columns: see *Where the quality numbers come from* {where}.\n"
+            f"³ BF16 conversion of the checkpoint; not published.")
+
+
+YAML = """---
+license: apache-2.0
+base_model: nex-agi/Nex-N2.5-mini
+base_model_relation: quantized
+pipeline_tag: image-text-to-text
+library_name: gguf
+tags:
+  - gguf
+  - llama.cpp
+  - rocm
+  - amd
+  - rocmfp4
+  - rocmfpx
+  - strix-halo
+  - amd-strix-halo
+  - gfx1151
+  - ryzen-ai-max
+  - ryzen-ai-max-395
+  - radeon-8060s
+  - moe
+  - reasoning
+  - multimodal
+  - vision
+  - nex
+  - qwen3.5
+  - quantized{extra}
+---
+"""
+
+
+def cmake_block():
+    commit = B.get("commit") or "d3ca537"
+    return f"""```bash
+git clone https://github.com/charlie12345/ROCmFPX.git && cd ROCmFPX
+git checkout {commit}
+HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \\
+cmake -S . -B build -DCMAKE_BUILD_TYPE=Release \\
+  -DGGML_HIP=ON -DGGML_VULKAN=ON -DGPU_TARGETS=gfx1151 \\
+  -DGGML_HIP_GRAPHS=ON -DGGML_HIP_NO_VMM=ON -DLLAMA_CURL=OFF
+cmake --build build --target llama-server -j
+```"""
+
+
+def serve_block(model_file):
+    env = ("env LD_LIBRARY_PATH=$PWD/build/bin:/opt/rocm/lib HSA_OVERRIDE_GFX_VERSION=11.5.1 "
+           "GGML_HIP_ENABLE_UNIFIED_MEMORY=1 \\\n")
+    mm = f"  --mmproj ~/models/nex/mmproj-{N}-BF16.gguf \\\n"
+    tpl = ("  --chat-template-file ~/models/nex/chat_template_enable_thinking.jinja --reasoning off \\\n")
+    tail = f"  -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -dev ROCm0 \\\n  -c {CTX} --host 127.0.0.1 --port 8080"
+    head = f"build/bin/llama-server \\\n  -m ~/models/nex/{model_file} \\\n"
+    cmd = f"```bash\n{env}{head}{mm}{tpl}{tail}\n```"
+    w = Jopt("vision_quickstart_warning")
+    return cmd + ("\n\n" + w if w else "")
+
+
+def curl_block():
+    return """```bash
+curl http://127.0.0.1:8080/v1/chat/completions \\
+  -H 'Content-Type: application/json' \\
+  -d '{
+    "messages": [{"role": "user", "content": "Hello"}],
+    "temperature": 0.7,
+    "top_p": 0.95,
+    "top_k": 40,
+    "chat_template_kwargs": {"enable_thinking": true}
+  }'
+```"""
+
+
+def quick_start(model_file, repo):
+    return f"""**1. Download**
+
+```bash
+hf download {repo} --local-dir ~/models/nex
+```
+
+**2. Build `llama-server`** — ROCmFPX at the measured commit (ROCm and Vulkan
+prerequisites: the project's [build guide](https://github.com/charlie12345/ROCmFPX/blob/{B.get('commit') or 'd3ca537'}/docs/build.md)). No patch.
+
+{cmake_block()}
+
+(The CMake options of the measured build are listed in [Reproduction](#reproduction).)
+
+`d3ca537` is also in the history of the official [ROCmFPX/ROCmFPX](https://github.com/ROCmFPX/ROCmFPX) repository.
+
+**3. Serve**
+
+{serve_block(model_file)}
+
+(`LD_LIBRARY_PATH` avoids a soname clash on machines that also have a Vulkan-only llama.cpp build.) The exact measured
+argv is in [Reproduction](#reproduction).
+
+**4. Call** — upstream sampling. Thinking is off unless the request sets `"enable_thinking": true` (as here; drop that
+line for a direct answer):
+
+{curl_block()}
+
+| Flag | Why |
+| --- | --- |
+| `--chat-template-file …/chat_template_enable_thinking.jinja` | The model's own template plus one line (see [Reasoning controls](#reasoning-controls)). Without it llama-server leaves the reasoning in `content` and thinking-on tool calls fail ([measured](#tool-calling)). |
+| `--reasoning off` | Thinking stays off unless a request passes `"enable_thinking": true`. |
+| `--jinja` | Already on by default in this build; keep it on — the reasoning controls (`chat_template_kwargs`) and tool calling rely on the Jinja chat template. |
+| `-fit off` | Autofit reads `MemAvailable` on integrated GPUs and can silently shrink context or push tensors to CPU. |
+| `-cram <MiB>` | Not set above (default 8 GiB of host RAM for saved prompts). Set it on a shared box — see [Known issues](#known-issues-and-limits). |
+| `--mmproj` | Loads the {AR.get('vision_layers', '—')}-layer vision tower. Drop the flag for text-only. |
+
+Requires a llama.cpp build with ROCmFP4 / ROCmFPX tensor-type support; stock llama.cpp rejects these tensor types."""
+
+
+def reasoning_block():
+    tf = S.get("template_fix") or {}
+    shim = (tf.get("shim") or "").rstrip("\n")
+    shim_md = ("```jinja\n" + shim + "\n```") if shim else "**[JUDGMENT PENDING: template shim]**"
+    return f"""## Reasoning controls
+
+The model's own chat template switches thinking with `chat_template_kwargs.reasoning_effort` and ignores
+`enable_thinking`:
+
+| `reasoning_effort` | What the stock template emits |
+| --- | --- |
+| `"none"` | empty `<think>\\n\\n</think>` (no thinking) |
+| `"high"` | opens `<think>\\n` (always think) |
+| `"medium"`, unset, or anything else | opens `<think>` and lets the model decide (adaptive; upstream default is `"medium"`) |
+
+llama-server decides how to split reasoning from the answer by rendering the template with `enable_thinking` on and
+off. With this template both renders are the same, so it never extracts the reasoning ([measured](#tool-calling)).
+`chat_template_enable_thinking.jinja` in this repo is the model's `chat_template.jinja` (sha256
+`{tf.get('source_sha256') or '—'}`) with one line added at the top (file sha256 `{tf.get('sha256') or '—'}`):
+
+{shim_md}
+
+Serve it with `--chat-template-file` and `--reasoning off`.
+
+{J('template_note')}
+
+Upstream serving (SGLang) uses `--reasoning-parser qwen3 --tool-call-parser qwen3_coder`. Recommended sampling:
+temperature 0.7, top_p 0.95, top_k 40.
+
+Earlier assistant turns are re-rendered **with** their reasoning (contexts grow faster than with templates that drop
+it). With thinking on and a small `max_tokens`, the whole budget can go to reasoning and `content` comes back empty —
+raise `max_tokens` before concluding the model is broken.
+
+Tool calls use the XML-style `<tool_call><function=…><parameter=…>` format, which llama.cpp parses natively
+through the Jinja chat template (on by default)."""
+
+
+def speed_table():
+    rows = ["| File | Backend | Workload | Decode tok/s (min–max) | Prefill tok/s |",
+            "| --- | --- | --- | ---: | ---: |"]
+    for tag in STD_TAGS + IMAT_TAGS:
+        t = tier(tag)
+        fn = t.get("file") or "—"
+        for dev, dn in (("rocm", "ROCm0"), ("vk", "Vulkan0")):
+            b = bench(speed_label(tag, dev))
+            if b is None or b.get("tg_median") is None:
+                rows.append(f"| `{fn}` | {dn} | code | — | — |")
+            else:
+                rows.append(f"| `{fn}` | {dn} | code | {b['tg_median']:.2f} ({g(b.get('tg_min'))}–{g(b.get('tg_max'))}) | "
+                            f"{g(b.get('pp_median'), '{:.0f}')} |")
+    for lab, dn in (("n-q106-rocm-prose", "ROCm0"), ("n-q106-vk-prose", "Vulkan0")):
+        b = bench(lab)
+        fn = tier("q106").get("file") or "—"
+        if b is None or b.get("tg_median") is None:
+            rows.append(f"| `{fn}` | {dn} | prose | — | — |")
+        else:
+            rows.append(f"| `{fn}` | {dn} | prose | {b['tg_median']:.2f} ({g(b.get('tg_min'))}–{g(b.get('tg_max'))}) | "
+                        f"{g(b.get('pp_median'), '{:.0f}')} |")
+    return "\n".join(rows)
+
+
+def cache_table():
+    x = gate("n-c3-q106")
+    rows = ["| server | second-request prompt tokens reused | processed | warm reply = cold reply |",
+            "| --- | ---: | ---: | :---: |"]
+    if not x:
+        rows.append("| d3ca537, unpatched | — | — | — |")
+        return "\n".join(rows)
+    n, L, ident = x.get("n"), x.get("L"), x.get("identical")
+    got = sorted({r_.get("warm_cache_n") for r_ in x.get("rows") or []}, key=lambda v: (v is None, v))
+    if n is None or L is None or not got or None in got:
+        rows.append("| d3ca537, unpatched | — | — | — |")
+        return "\n".join(rows)
+    if len(got) == 1:
+        reused, proc = f"**{got[0]:,}** of {L:,} (all {n} pairs)", f"{L - got[0]:,}"
+    else:
+        reused, proc = f"{got[0]:,}–{got[-1]:,} of {L:,} (varies across {n} pairs)", f"{L - got[-1]:,}–{L - got[0]:,}"
+    rows.append(f"| d3ca537, unpatched | {reused} | {proc} | {ident if ident is not None else '—'}/{n} |")
+    return "\n".join(rows)
+
+
+def tools_block():
+    t = gate("n-tools-q106")
+    fx = [gate(l) for l in ("n-tools-q106-roff", "n-tools-q106-roff-r2", "n-tools-q106-roff-r3")]
+    if not t and not any(fx):
+        return "_Not measured._"
+    names = ["multi-arg", "nested-object", "enum", "correct-decline", "multi-turn", "streaming", "parallel"]
+    rows = ["| check | quick start, thinking ON | quick start, thinking OFF | stock template, thinking ON | "
+            "stock template, thinking OFF |",
+            "| --- | :---: | :---: | :---: | :---: |"]
+    detail = (t or {}).get("detail") or {}
+    mk = lambda x: "—" if x is None else ("✅" if x else "❌")
+
+    def count(n, think):
+        vals = [((x or {}).get("detail") or {}).get(f"{n}|think={think}") for x in fx]
+        if any(v is None for v in vals):
+            return "—"
+        return f"{sum(bool(v) for v in vals)}/{len(vals)}"
+    for n in names:
+        rows.append(f"| {n} | {count(n, True)} | {count(n, False)} | {mk(detail.get(f'{n}|think=True'))} | "
+                    f"{mk(detail.get(f'{n}|think=False'))} |")
+    fn = os.path.basename(tier("q106").get("file") or "—")
+    tot = (f"**{sum(x['passed'] for x in fx)}/{sum(x['total'] for x in fx)}** over three passes with the quick-start "
+           f"configuration, **{t['passed']}/{t['total']}** with the stock template"
+           if all(fx) and t and t.get("passed") is not None else "Tool-calling suite")
+    return (f"{tot}, run on `{fn}`. Quick start = the included template file + `--reasoning off`, thinking switched "
+            f"with `enable_thinking`; stock = the model's own template, thinking switched with `reasoning_effort` "
+            f"(`high` / `none`). A check passes only with a native `tool_calls` entry carrying the right arguments "
+            f"and no raw XML or think tags left in `content`.\n\n" + "\n".join(rows))
+
+
+def vision_block():
+    on, off = gate("n-vision-q106-faon"), gate("n-vision-q106-faoff")
+    if not on and not off:
+        return "_Not measured._"
+    vp = S.get("vision_probe") or {}
+    exp_row = next((x for x in (on, off) if x and x.get("expected")), None)
+    exp_txt = ", ".join(f"`{e.strip()}`" for e in exp_row["expected"].split(",")) if exp_row else "—"
+    def cell(x):
+        if not x:
+            return "—"
+        exp = x.get("expected") or ""
+        nexp = len(exp.split(",")) if exp else None
+        hits = x.get("hits") or []
+        n = f"{len(hits)}/{nexp} terms" if nexp is not None else "—"
+        if x.get("result") == "PASS":
+            return f"✅ {n}"
+        return "❌ " + ("server stopped" if x.get("server_died") else ("request failed" if x.get("error") else n))
+    rows = ["| | `-fa on` | `-fa off` |", "| --- | :---: | :---: |",
+            f"| STRIX_LEAN + projector | {cell(on)} | {cell(off)} |"]
+    txt = (f"Probe: a synthetic {vp.get('width', '—')}×{vp.get('height', '—')} image with a red circle and a blue square "
+           f"(a model that ignores the image cannot name both), sent to `{tier('q106').get('file') or '—'}` with "
+           f"`--mmproj`, temperature 0. Pass = the reply names every expected term ({exp_txt}).\n\n"
+           + "\n".join(rows) + "\n\n" + J("vision_note"))
+    ans = next((x for x in (on, off) if x and x.get("result") == "PASS" and x.get("answer")), None)
+    if ans:
+        a_ = (ans.get("answer") or "").strip()
+        which = "`-fa on`" if ans is on else "`-fa off`"
+        cut = a_[:300]
+        txt += f"\n\nReply ({which}):\n\n> {cut}" + (" …" if len(a_) >= 300 else "")
+    return txt
+
+
+def files_table(names):
+    rows = ["| File | Size | sha256 |", "| --- | ---: | --- |"]
+    for fn, size in names:
+        size = SIZES.get(fn, size)
+        sz = ("—" if size is None else gib(size) if size >= GiB // 10 else
+              f"{size / MiB:.1f} MiB" if size >= MiB else f"{size / 1024:.1f} KiB")
+        rows.append(f"| `{fn}` | {sz} | `{SUMS.get(fn, '—')}` |")
+    return "\n".join(rows)
+
+
+def receipts_table(tags, imat=False):
+    head = ("| File | `output.weight` | `token_embd.weight` | tensors |"
+            + (" imatrix entries | bytes differ from standard |" if imat else ""))
+    sep = "| --- | --- | --- | ---: |" + (" ---: | :---: |" if imat else "")
+    rows = [head, sep]
+    for k in tags:
+        t = tier(k)
+        r = (f"| `{t.get('file') or '—'}` | {t.get('output_weight') or '—'} | {t.get('token_embd') or '—'} | "
+             f"{g(t.get('tensors'), '{}')} |")
+        if imat:
+            dfs = t.get("differs_from_standard")
+            r += f" {t.get('imatrix_entries') if t.get('imatrix_entries') is not None else '—'} | "
+            r += f"{'yes' if dfs else ('no' if dfs is False else '—')} |"
+        rows.append(r)
+    return "\n".join(rows)
+
+
+def repro(model_file, label):
+    b = bench(label)
+    sha = B.get("sha256") or {}
+    cmd = (b.get("cmd") if b else None) or "—"
+    return f"""```
+server   : {B.get('repo') or '—'} @ {B.get('commit') or '—'}
+           unpatched; build dir {os.path.dirname(B['dir']) if B.get('dir') else '—'}, Release, Unix Makefiles, GGML_HIP=ON GGML_VULKAN=ON
+           GGML_HIP_GRAPHS=ON GGML_HIP_NO_VMM=ON GGML_NATIVE=ON AMDGPU_TARGETS=gfx1151 LLAMA_CURL=OFF
+           CMAKE_HIP_COMPILER=/opt/rocm-7.2.4/lib/llvm/bin/clang
+           sha256 llama-quantize   {sha.get('llama-quantize') or '—'}
+           sha256 llama-imatrix    {sha.get('llama-imatrix') or '—'}
+           sha256 llama-perplexity {sha.get('llama-perplexity') or '—'}
+           sha256 llama-server     {sha.get('llama-server') or '—'}
+source   : {SRC.get('repo') or '—'}  revision {SRC.get('revision') or '—'}
+model    : {model_file} (the argv below; every file was measured the same way)
+argv     : {cmd}
+template : the quick-start tool-suite and image rows add --chat-template-file chat_template_enable_thinking.jinja
+           --reasoning off to this argv (recipe/pipeline/run_tools_roff.sh -> nex_tools_tpl.py; their server logs
+           read "chat template, thinking = 0"); the speed rows use the stock template
+env      : LD_LIBRARY_PATH=<build>/bin:/opt/rocm-7.2.4/lib
+           HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1
+box      : aimax — AMD Ryzen AI Max+ 395 / Radeon 8060S (gfx1151), 124 GiB, GTT 131072 MiB,
+           kernel 6.17.6-061706-generic, ROCm 7.2.4
+protocol : {PN_TXT}-token code prompt, {GEN} generated tokens, temp 0 / top_k 1, ignore_eos, cache_prompt false,
+           1 warm-up + median of {REPS}, no co-resident models (box iced)
+measured : {' to '.join(S.get('measured_range') or []) or S.get('measured') or '—'}, by the pipeline in recipe/ (every raw number in recipe/results_summary.json and recipe/raw/)
+```"""
+
+
+def methodology_std():
+    return f"""```bash
+# 1. convert: text model and the vision projector (the checkpoint has no mtp.* tensors)
+python convert_hf_to_gguf.py hf --outtype bf16 --model-name {N} --outfile {N}-BF16.gguf
+python convert_hf_to_gguf.py hf --outtype bf16 --mmproj --model-name {N} --outfile mmproj-{N}-BF16.gguf
+
+# 2. quantize from BF16 only; the LM head is forced up on every tier and read back by exact tensor name
+llama-quantize --output-tensor-type q6_K                            {N}-BF16.gguf OUT Q4_0_ROCMFP4_STRIX_LEAN 16
+llama-quantize --output-tensor-type q6_K --token-embedding-type q6_K {N}-BF16.gguf OUT Q4_0_ROCMFP4_COHERENT 16
+llama-quantize --output-tensor-type q6_K                            {N}-BF16.gguf OUT Q4_0_ROCMFP4_FAST 16
+
+# 3. BF16 reference logits on the CPU only (this build's ROCm0 path computes the BF16 MoE wrong — Known issues)
+llama-perplexity -m {N}-BF16.gguf -f wikitext-2-raw/wiki.test.raw -c 2048 -b 2048 --chunks 40 --kl-divergence-base bf16.kld \\
+  -dev none -ngl 0 --no-op-offload -t 16
+# 4. grade each shipped file against those logits, on each GPU backend
+llama-perplexity -m OUT --kl-divergence-base bf16.kld --kl-divergence -c 2048 -b 2048 -ngl 999 -fa on -dio -dev ROCm0
+llama-perplexity -m OUT --kl-divergence-base bf16.kld --kl-divergence -c 2048 -b 2048 -ngl 999 -fa on -dio -dev Vulkan0
+```
+
+Receipts (the built file is the receipt — exact tensor names, never a substring match; `recipe/logs/`):
+
+{receipts_table(STD_TAGS)}"""
+
+
+def first_rocm_bullet():
+    hc = S.get("hub_check")
+    others = [h for h in (hc or {}).get("header_checks") or [] if h.get("output_weight_type")]
+    if others:
+        return "".join(
+            f"- Another public ROCmFP4 build of this model exists — [{h['repo']}](https://huggingface.co/{h['repo']}): "
+            f"its `{h['file']}` stores `output.weight` as `{h['output_weight_type']}`"
+            f"{' and carries no imatrix metadata' if h.get('imatrix_keys') == [] else ''}. Every tier here keeps "
+            f"`output.weight` at `Q6_K`, and the imatrix builds are a separate repo.\n" for h in others)
+    if isinstance(hc, dict) and hc.get("rocm_builds_found") == 0:
+        return ("- **First ROCmFP4 build of this model** — no ROCm or Strix Halo build of Nex-N2.5-mini was on the Hub "
+                "at publication.\n")
+    return ""
+
+
+def intro_arch():
+    p = SRC.get("params")
+    ptxt = f"{p:,} parameters (BF16)" if p is not None else "— parameters"
+    return (f"{AR.get('layers', '—')}-layer Qwen3.5 MoE ({AR.get('linear_attn_layers', '—')} Gated DeltaNet linear-attention + "
+            f"{AR.get('full_attn_layers', '—')} full-attention layers), {AR.get('num_experts', '—')} routed experts / "
+            f"{AR.get('num_experts_per_tok', '—')} active, {AR.get('max_position_embeddings', '—'):,}-token context"
+            if AR.get("max_position_embeddings") is not None else
+            f"{AR.get('layers', '—')}-layer Qwen3.5 MoE, {ptxt}")
+
+
+def std_card():
+    L = tier("q106")
+    bf = S.get("bf16") or {}
+    nextn = bf.get("nextn_tensors")
+    no_mtp = ("the checkpoint ships no `mtp.*` weights" +
+              (f" (the converted BF16 GGUF reads back {nextn} `nextn` tensors)" if nextn is not None else ""))
+    return YAML.format(extra="") + f"""
+# Nex-N2.5-mini — ROCmFP4 for AMD Strix Halo (gfx1151)
+
+ROCmFP4 / ROCmFPX quantizations of **[nex-agi/Nex-N2.5-mini](https://huggingface.co/nex-agi/Nex-N2.5-mini)** —
+{g(SRC.get('params'), '{:,}')} parameters (BF16), {intro_arch()}, text + image — built and measured on an AMD Ryzen AI
+Max+ 395 (Radeon 8060S, `gfx1151`). Upstream publishes no GGUF.
+
+{first_rocm_bullet()}- **Vision projector included.**
+- **No MTP head.** `mtp_num_hidden_layers: 1` is declared in `config.json`, but {no_mtp}. There is no speculative
+  decoding on these files.
+- Importance-matrix builds of the same three 4-bit tiers: **[{IMAT}](https://huggingface.co/{IMAT})**.
+
+## Which file should I use?
+
+{PROTOCOL}
+
+{TABLE_HEAD}
+{chr(10).join(tier_row(k) for k in STD_TAGS)}
+{bf16_row()}
+
+{footnotes()}
+
+{J('std_recommendation')}
+
+{quality_blurb()}
+
+**Where the quality numbers come from.** {J('quality_provenance')}
+
+{J('backend_quality_note')}
+
+## Quick start
+
+{quick_start(JUDGE.get('std_default') or L.get('file') or '—', STD)}
+
+{reasoning_block()}
+
+## Speed
+
+{speed_table()}
+
+{J('speed_note')}
+
+## Prompt caching
+
+Measured: pairs of requests that share a long code prefix and differ only in the closing instruction. The second
+request of each pair runs warm (`cache_prompt: true`, resuming from what the first one left) and then cold
+(`cache_prompt: false`), and the two replies are compared byte for byte. Every prompt is padded to one token length so
+warm and cold see identical chunking.
+
+{cache_table()}
+
+{J('cache_note')}
+
+## Tool calling
+
+The template emits the XML-style `<tool_call><function=…><parameter=…>` format, which llama.cpp parses natively
+through the Jinja chat template (on by default). Suite run through `llama-server`, at the checkpoint's recommended
+sampling (temperature 0.7, top-p 0.95, top-k 40):
+
+{tools_block()}
+
+{J('tools_note')}
+
+## Vision
+
+`mmproj-{N}-BF16.gguf` is the {AR.get('vision_layers', '—')}-layer vision tower (width {AR.get('vision_width', '—')}),
+loaded with `--mmproj`. Its attention follows the server's `-fa` setting, so both settings were checked.
+
+{vision_block()}
+
+## Memory
+
+{J('memory_note')}
+
+## Quantization methodology
+
+{methodology_std()}
+
+`tie_word_embeddings` is false, so the output head is a real tensor and `--output-tensor-type q6_K` does real work.
+All three tiers pin `output.weight` to `q6_K`; COHERENT also pins `token_embd.weight` to `q6_K`, while STRIX_LEAN and
+FAST keep their tier's own embedding type (shown in the receipts).
+
+## Reproduction
+
+{repro(L.get('file') or '—', 'n-q106-rocm')}
+
+## Files
+
+{files_table([(tier(k).get('file'), tier(k).get('size_bytes')) for k in STD_TAGS] + list((S.get('aux') or {}).items()))}
+
+`SHA256SUMS` covers every model file and the chat template file. `recipe/` holds the measurement pipeline (`recipe/pipeline/`), raw per-run
+results (`recipe/raw/`), build and receipt logs (`recipe/logs/`), and `results_summary.json` with every measured value
+on this card. Architecture facts (layer counts, vocabulary, vision depth) come from the checkpoint's `config.json` at
+revision `{SRC.get('revision') or '—'}`.
+
+## Known issues and limits
+
+{J('std_known_issues')}
+
+## License and attribution
+
+Apache-2.0, inherited from the base model. Weights and architecture: **Nex-AGI**
+([nex-agi/Nex-N2.5-mini](https://huggingface.co/nex-agi/Nex-N2.5-mini)). ROCmFP4 / ROCmFPX quantization format and
+runtime: the ROCmFPX project. Quantization and measurements: kingjones777.
+"""
+
+
+def imat_effect_table():
+    rows = ["| Tier | Build | Size | KLD vs BF16 ↓ | Same top-1 ↑ | PPL (× BF16) | 99th-pct KLD |",
+            "| --- | --- | ---: | ---: | ---: | ---: | ---: |"]
+    def d(a, b, k, ek):
+        x, y = tier(a).get(k), tier(b).get(k)
+        ex, ey = (tier(a).get(ek), tier(b).get(ek)) if ek else (None, None)
+        if None in (x, y):
+            return "—"
+        s = f"{(y - x) / x * 100:+.1f} %"
+        if ex is not None and ey is not None:
+            s += f" ({abs(y - x) / math.sqrt(ex * ex + ey * ey):.1f}σ)"
+        return s
+    for base, imat in (("q106", "q106i"), ("q102", "q102i"), ("q103", "q103i")):
+        for tag, lab in ((base, "standard"), (imat, "**imatrix**")):
+            x = tier(tag)
+            ratio = x.get("ppl_ratio")
+            rows.append(f"| {NAMES[tag]} | {lab} | {gib(x.get('size_bytes'))} | {pm(x.get('kld_mean'), x.get('kld_err'))} | "
+                        f"{g(x.get('same_top_p'), '{:.2f} %')} | {pm(x.get('ppl'), x.get('ppl_err'))} "
+                        f"({'—' if ratio is None else '×%.4f' % ratio}) | {g(x.get('kld_p99'), '{:.4f}')} |")
+        dpp = ("—" if None in (tier(base).get("same_top_p"), tier(imat).get("same_top_p"))
+               else f"{tier(imat)['same_top_p'] - tier(base)['same_top_p']:+.2f} pp")
+        rows.append(f"| | *Δ imatrix* | | {d(base, imat, 'kld_mean', 'kld_err')} | {dpp} | "
+                    f"{d(base, imat, 'ppl', 'ppl_err')} | {d(base, imat, 'kld_p99', None)} |")
+    return "\n".join(rows)
+
+
+def imat_card():
+    im = S.get("imatrix") or {}
+    Li = tier("q106i")
+    entries = [tier(k).get("imatrix_entries") for k in IMAT_TAGS]
+    ent = next((e for e in entries if e is not None), None)
+    ent_txt = g(ent, "{:,}") if (ent is None or len(set(e for e in entries if e is not None)) <= 1) else \
+        " / ".join(g(e, "{:,}") for e in entries)
+    return YAML.format(extra="\n  - imatrix") + f"""
+# Nex-N2.5-mini — ROCmFP4 **imatrix** for AMD Strix Halo (gfx1151)
+
+Importance-matrix-calibrated ROCmFP4 quantizations of
+**[nex-agi/Nex-N2.5-mini](https://huggingface.co/nex-agi/Nex-N2.5-mini)** ({g(SRC.get('params'), '{:,}')} parameters,
+{intro_arch()}, text + image). Companion to the standard build **[{STD}](https://huggingface.co/{STD})** — the same
+three 4-bit tiers, same vision projector, same unpatched `d3ca537` server; the only difference is how each 4-bit
+block's scale was chosen. There is no MTP head on either repo.
+
+## What the imatrix changes
+
+ROCmFP4 has an importance-weighted quantizer path: with `--imatrix`, each block's scale is chosen by an exhaustive
+search that minimises error **weighted by how strongly the calibration activations use each weight**, instead of the
+unweighted default. It changes **which** scales are picked at the **same** bit width and tensor types — so it moves
+quality, not size, and per-token compute is identical.
+
+| | |
+| --- | --- |
+| calibration text | {im.get('calibration') or '—'} (the widely used community calibration set) |
+| computed on | BF16 GGUF, {g(im.get('chunks'), '{}')} chunks × {g(im.get('n_ctx'), '{}')} tokens, {im.get('device') or '—'} |
+| entries loaded | {ent_txt} (from the N3 quantize logs) |
+| file | `{im.get('file') or '—'}` (GGUF format), sha256 `{im.get('sha256') or '—'}` |
+
+## Measured effect
+
+{quality_blurb()} The calibration text and the grading text are different corpora.
+
+{imat_effect_table()}
+
+σ = difference divided by the two runs' combined standard error. The two runs score the **same** tokens, so this is
+conservative (paired noise is smaller).
+
+{J('imat_verdict')}
+
+**Where the quality numbers come from.** {J('quality_provenance')}
+
+{J('backend_quality_note')}
+
+## Which file should I use?
+
+{J('imat_recommendation')}
+
+{TABLE_HEAD}
+{chr(10).join(tier_row(k) for k in IMAT_TAGS)}
+{bf16_row()}
+
+{footnotes("above")}
+
+{J('imat_speed_note')}
+
+Full speed tables (both backends, prose vs code), prompt-cache, tool-calling and vision results are on
+the [standard card](https://huggingface.co/{STD}).
+
+## Quick start
+
+{quick_start(JUDGE.get('imat_default') or Li.get('file') or '—', IMAT)}
+
+{reasoning_block()}
+
+## Quantization methodology
+
+```bash
+llama-imatrix -m {N}-BF16.gguf -f calibration_datav3.txt -o {N}.imatrix \\
+  -c 512 -b 512 -dev none -ngl 0 --no-op-offload -t 16
+llama-quantize --imatrix {N}.imatrix --output-tensor-type q6_K \\
+  {N}-BF16.gguf {N}-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf Q4_0_ROCMFP4_STRIX_LEAN 16
+llama-quantize --imatrix {N}.imatrix --output-tensor-type q6_K --token-embedding-type q6_K \\
+  {N}-BF16.gguf {N}-imatrix-Q4_0_ROCMFP4_COHERENT.gguf Q4_0_ROCMFP4_COHERENT 16
+llama-quantize --imatrix {N}.imatrix --output-tensor-type q6_K \\
+  {N}-BF16.gguf {N}-imatrix-Q4_0_ROCMFP4_FAST.gguf Q4_0_ROCMFP4_FAST 16
+```
+
+Receipts that the weighted path was actually taken, and that each shipped file differs from its standard twin:
+
+{receipts_table(IMAT_TAGS, imat=True)}
+
+## Reproduction
+
+{repro(Li.get('file') or '—', 'n-q106i-rocm')}
+
+## Files
+
+{files_table([(tier(k).get('file'), tier(k).get('size_bytes')) for k in IMAT_TAGS] + [(im.get('file'), im.get('size_bytes'))] + list((S.get('aux') or {}).items()))}
+
+## Known issues and limits
+
+{J('imat_known_issues')}
+
+## License and attribution
+
+Apache-2.0, inherited from the base model. Weights and architecture: **Nex-AGI**
+([nex-agi/Nex-N2.5-mini](https://huggingface.co/nex-agi/Nex-N2.5-mini)). Calibration text: bartowski's
+`calibration_datav3`. ROCmFP4 / ROCmFPX: the ROCmFPX project. Imatrix, quantization and measurements: kingjones777.
+"""
+
+
+open(os.path.join(OUT, "README_std.md"), "w").write(std_card())
+def fix_anchors(md, other_repo):
+    """Links to sections that exist only on the other card point there instead of to a missing anchor."""
+    import re
+    slugs = {re.sub(r"[^a-z0-9 -]", "", h.strip().lower()).replace(" ", "-")
+             for h in re.findall(r"^#{1,6} (.+)$", md, flags=re.M)}
+    return re.sub(r"\]\(#([a-z0-9-]+)\)",
+                  lambda m: m.group(0) if m.group(1) in slugs else "](https://huggingface.co/%s#%s)" % (
+                      other_repo, m.group(1)), md)
+
+
+open(os.path.join(OUT, "README_imat.md"), "w").write(fix_anchors(imat_card(), STD))
+cards = open(os.path.join(OUT, "README_std.md")).read() + open(os.path.join(OUT, "README_imat.md")).read()
+pend = sorted(set(x.split("JUDGMENT PENDING: ")[1].split("]")[0] for x in cards.split("**[")[1:] if "JUDGMENT PENDING" in x))
+dash_cells = cards.count("| — |")
+print("rendered | bench rows =", len(S.get("bench") or []), "| gates =", len(S.get("gates") or []),
+      "| pending judgments:", pend, "| '—' cells:", dash_cells)

recipe/pipeline/nex_seat_default_probe.py

@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+"""Which server-side default makes Nex behave well for clients that pass no chat_template_kwargs (the gateway case)?
+Follow-up to nex_tools_tpl.py: with the `high` shim, a no-kwargs one-word request came back entirely in
+reasoning_content (the model never closed its forced-open think block). Two candidate defaults, greedy probes:
+  C1  high shim + `--chat-template-kwargs {"enable_thinking": false}`      (thinking off unless a client asks)
+  C2  v2 shim (explicit enable_thinking=false wins; otherwise enable_thinking maps to high unless reasoning_effort is
+      set) + `--chat-template-kwargs {"reasoning_effort": "medium"}`      (upstream's adaptive default)
+Diagnostic only -> results/nex_seat_default_probe.json."""
+import json, os, subprocess, sys, time
+from types import SimpleNamespace
+
+os.environ["AGNES_BIN"] = "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin"
+sys.path.insert(0, "/mnt/models/nex-n2.5-mini")
+import nex_harness as H  # noqa: E402
+
+W = H.W
+SRC = open(f"{W}/hf/chat_template.jinja", "rb").read()
+SHIM_V2 = ("{%- if enable_thinking is defined and not enable_thinking %}{%- set reasoning_effort = 'none' %}"
+           "{%- elif reasoning_effort is not defined and enable_thinking is defined %}"
+           "{%- set reasoning_effort = 'high' %}{%- endif %}\n")
+TPL_V2 = f"{W}/tpl/chat_template_enable_thinking_v2.jinja"
+open(TPL_V2, "wb").write(SHIM_V2.encode() + SRC)
+CONFIGS = [
+    ("C1", [f"--chat-template-file", f"{W}/tpl/chat_template_enable_thinking.jinja",
+            "--chat-template-kwargs", json.dumps({"enable_thinking": False})]),
+    ("C2", ["--chat-template-file", TPL_V2, "--chat-template-kwargs", json.dumps({"reasoning_effort": "medium"})]),
+]
+PROMPTS = [("correct-decline", "What is 17 times 23? Answer directly."),
+           ("single-word", "Reply with the single word: ready"),
+           ("multi-arg", "What's the weather in Paris in celsius?")]
+KW = [("no-kwargs", None), ("enable_thinking=true", {"enable_thinking": True}),
+      ("enable_thinking=false", {"enable_thinking": False})]
+_orig = subprocess.Popen
+report = {"shim_v2": SHIM_V2, "configs": {}}
+for name, extra in CONFIGS:
+    def _popen(cmd, *a, _extra=extra, **k):
+        if cmd and str(cmd[0]).endswith("llama-server"):
+            cmd = list(cmd) + list(_extra)
+        return _orig(cmd, *a, **k)
+    subprocess.Popen = _popen
+    a = SimpleNamespace(model=f"{W}/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", dev="ROCm0", ctx=16384,
+                        draft=None, mtp_infile=False, nmax=4, pmin=0.0, strict=False,
+                        serverlog=f"{W}/logs/probe_seat_default_{name}.log")
+    try:
+        s = H.Server(a, 18653)
+    finally:
+        subprocess.Popen = _orig
+    rec = {"extra": extra, "results": {}}
+    try:
+        for kn, kw in KW:
+            for pn, prompt in PROMPTS:
+                body = {"messages": [{"role": "user", "content": prompt}], "tools": H.TOOLS, "tool_choice": "auto",
+                        "temperature": 0, "top_k": 1, "max_tokens": 2048}
+                if kw is not None:
+                    body["chat_template_kwargs"] = kw
+                try:
+                    m = H.post(18653, "/v1/chat/completions", body)["choices"][0]["message"]
+                    c = m.get("content") or ""
+                    r = {"content": c[:120], "reasoning_len": len(m.get("reasoning_content") or ""),
+                         "tool_calls": [(t.get("function") or {}).get("name") for t in m.get("tool_calls") or []],
+                         "leaks": [x for x in H.LEAK if x in c]}
+                except Exception as e:  # noqa: BLE001
+                    r = {"error": repr(e)[:300]}
+                rec["results"][f"{kn}|{pn}"] = r
+                print(name, kn, pn, json.dumps(r)[:220], flush=True)
+    finally:
+        s.stop()
+    report["configs"][name] = rec
+    time.sleep(3)
+json.dump(report, open(f"{W}/results/nex_seat_default_probe.json", "w"), indent=1)
+print("NEX_SEAT_DEFAULT_PROBE_DONE")

recipe/pipeline/nex_seats.sh

@@ -0,0 +1,177 @@
+#!/bin/bash
+# King 2026-09-16: "leave the fast of the models on this machine for swap i want to test them" -> the two Nex-N2.5-mini
+# FAST files (standard + imatrix) become ON-DEMAND seats: DISABLED systemd units King starts by hand (llama-swap is
+# masked on MAX-1). Every setting comes from this build's measurements (backend, -fa, context, memory cap); a missing
+# measurement stops the script instead of guessing. Each seat is smoke-tested once (start -> /health -> one real
+# completion -> stop) while the box is still iced, and through the gateway when a route already exists.
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; B=/opt/llama-rocm/rocmfpx-724/build-hipvk/bin; N=Nex-N2.5-mini
+# the stock template switches thinking only via reasoning_effort, which llama-server's parser analysis cannot see
+# (reasoning leaks into content); this copy maps enable_thinking onto it, and --reasoning off makes thinking opt-in
+# per request (enable_thinking: true). Measured: tool suite 40/42 over 3 passes vs 6/14 stock (nex_tools_tpl.py).
+TPL=$W/tpl/chat_template_enable_thinking.jinja
+GW=http://192.168.0.87:4000; GWKEY=sk-homecloud-master
+cd $W
+prlimit --pid $$ --core=1:1   # 1 BYTE (bash `ulimit -c 1` is 1 KiB and does not stop apport)
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+
+python3 - > results/nex_seats_plan.json <<'PY' || { log "PLAN FAILED -> no seats"; exit 2; }
+import json, math, sys
+W = "/mnt/models/nex-n2.5-mini"; N = "Nex-N2.5-mini"
+rows = [json.loads(l) for l in open(f"{W}/results/nex_bench.jsonl") if l.strip()]
+bench = {r["label"]: r for r in rows if "tg_median" in r}
+gate = {r["label"]: r for r in rows if "tg_median" not in r}
+siz = [json.loads(l) for l in open(f"{W}/results/nex_sizing.jsonl") if l.strip()]
+EQ = 0.03  # a backend must win decode by more than this AND not lose prefill by more than this to replace ROCm0
+
+def backend(tag):
+    r, v = bench.get(f"n-{tag}-rocm"), bench.get(f"n-{tag}-vk")
+    if not (r and v) or None in (r.get("tg_median"), v.get("tg_median"), r.get("pp_median"), v.get("pp_median")):
+        sys.exit(f"missing bench rows for {tag}")
+    vk = v["tg_median"] > r["tg_median"] * (1 + EQ) and v["pp_median"] >= r["pp_median"] * (1 - EQ)
+    return ("Vulkan0" if vk else "ROCm0"), {"rocm": [r["tg_median"], r["pp_median"]], "vk": [v["tg_median"], v["pp_median"]]}
+
+on, off = gate.get("n-vision-q106-faon"), gate.get("n-vision-q106-faoff")
+if not (on and off):
+    sys.exit("missing vision gates")
+fa, mmproj = ("on", True) if on.get("result") == "PASS" else (("off", True) if off.get("result") == "PASS" else ("on", False))
+
+ok = [s for s in siz if s.get("label") == "strix-lean" and s.get("footprint_after_8k_gib") is not None]
+if not ok:
+    sys.exit("missing sizing rows")
+best = max(ok, key=lambda s: s["ctx"])          # largest context that loaded and served a request
+cap = int(math.ceil((best["footprint_after_8k_gib"] * 1.15 + 2) * 1024 / 1000.0) * 1000)
+
+seats = []
+for unit, port, tag, path, label in (
+        ("max1-nex-fast", 8097, "q103", f"{W}/out/{N}-Q4_0_ROCMFP4_FAST.gguf", "FAST"),
+        ("max1-nex-fast-imat", 8098, "q103i", f"{W}/out-imat/{N}-imatrix-Q4_0_ROCMFP4_FAST.gguf", "FAST imatrix")):
+    dev, speeds = backend(tag)
+    seats.append({"unit": unit, "port": port, "tag": tag, "model": path, "label": label, "dev": dev, "speeds": speeds,
+                  "fa": fa, "mmproj": f"{W}/out/mmproj-{N}-BF16.gguf" if mmproj else None, "ctx": best["ctx"],
+                  "memory_max_mib": cap, "sized_from": best,
+                  "gateway_model": "nex-n2.5-mini-fast@max1" if tag == "q103" else "nex-n2.5-mini-fast-imatrix@max1"})
+print(json.dumps({"seats": seats, "vision": {"faon": on.get("result"), "faoff": off.get("result")}}, indent=1))
+PY
+log "plan: $(python3 -c "import json; d=json.load(open('results/nex_seats_plan.json')); print([(s['unit'], s['dev'], 'fa '+s['fa'], s['ctx'], str(s['memory_max_mib'])+'M', bool(s['mmproj'])) for s in d['seats']])")"
+
+n=$(python3 -c "import json; print(len(json.load(open('results/nex_seats_plan.json'))['seats']))")
+fail=0
+for i in $(seq 0 $((n - 1))); do
+  eval "$(python3 - $i <<'PY'
+import json, shlex, sys
+s = json.load(open("/mnt/models/nex-n2.5-mini/results/nex_seats_plan.json"))["seats"][int(sys.argv[1])]
+for k in ("unit", "port", "model", "label", "dev", "fa", "ctx", "memory_max_mib", "gateway_model"):
+    print(f"{k.upper()}={shlex.quote(str(s[k]))}")
+print(f"MMPROJ={shlex.quote(s['mmproj'] or '')}")
+PY
+)"
+  [ -s "$MODEL" ] || { log "$UNIT: model file missing ($MODEL)"; fail=1; continue; }
+  [ -s "$TPL" ] || { log "$UNIT: template file missing ($TPL)"; fail=1; continue; }
+  if ss -tln | grep -q ":$PORT "; then log "$UNIT: port $PORT already in use -> skip"; fail=1; continue; fi
+  MM=""; [ -n "$MMPROJ" ] && MM="--mmproj $MMPROJ"
+  tmp=$(mktemp /tmp/$UNIT.XXXX.service)
+  cat > $tmp <<EOF
+[Unit]
+Description=MAX-1 on-demand test seat: $N $LABEL (:$PORT) - disabled; start by hand, stop another seat first if memory is tight
+After=network.target
+
+[Service]
+User=root
+SupplementaryGroups=render video
+Environment=HSA_OVERRIDE_GFX_VERSION=11.5.1
+Environment=GGML_HIP_ENABLE_UNIFIED_MEMORY=1
+Environment=LD_LIBRARY_PATH=$B:/opt/rocm-7.2.4/lib
+ExecStart=$B/llama-server -m $MODEL $MM -dev $DEV -ngl 999 -fa $FA --no-mmap -fit off -c $CTX -ctk q8_0 -ctv q8_0 -np 1 -b 2048 -ub 1024 -t 16 --poll 100 --jinja --chat-template-file $TPL --reasoning off -cram 512 --host 0.0.0.0 --port $PORT
+Restart=on-failure
+RestartSec=15
+MemoryMax=${MEMORY_MAX_MIB}M
+LimitMEMLOCK=infinity
+# RLIMIT_CORE of exactly 1 byte: a crashing test seat must not pipe a 12 GiB core into apport on the 98 %-full root disk
+LimitCORE=1
+
+[Install]
+WantedBy=multi-user.target
+EOF
+  cp $tmp /tmp/$UNIT.service
+  systemd-analyze verify /tmp/$UNIT.service > /tmp/$UNIT.verify 2>&1; vrc=$?
+  if [ $vrc -ne 0 ] || grep -qiE "error|invalid|unknown|ignoring" /tmp/$UNIT.verify; then
+    log "$UNIT: unit did not verify (rc=$vrc):"; cat /tmp/$UNIT.verify; rm -f $tmp; fail=1; continue
+  fi
+  [ -f /etc/systemd/system/$UNIT.service ] && cp /etc/systemd/system/$UNIT.service /etc/systemd/system/.$UNIT.service.bak-$(date +%s)
+  install -m 0644 $tmp /etc/systemd/system/$UNIT.service; rm -f $tmp /tmp/$UNIT.service /tmp/$UNIT.verify
+  systemctl daemon-reload
+  systemctl disable $UNIT.service >/dev/null 2>&1   # on-demand: never part of boot
+  log "$UNIT written ($DEV, -fa $FA, ctx $CTX, MemoryMax ${MEMORY_MAX_MIB}M) -> smoke test"
+  t0=$(date +%s); systemctl start $UNIT.service
+  ok=0
+  for k in $(seq 1 180); do
+    curl -s -m 3 http://127.0.0.1:$PORT/health | grep -q '"ok"' && { ok=1; break; }
+    systemctl is-active --quiet $UNIT.service || { log "  $UNIT not active"; break; }
+    sleep 5
+  done
+  load_s=$(( $(date +%s) - t0 ))
+  if [ $ok -ne 1 ]; then
+    log "  FAIL $UNIT never healthy"; journalctl -u $UNIT.service -n 20 --no-pager | tail -20
+    systemctl stop $UNIT.service; fail=1; continue
+  fi
+  python3 - "$UNIT" "$PORT" "$GATEWAY_MODEL" "$load_s" "$GW" "$GWKEY" <<'PY' | tee -a results/nex_seats.jsonl
+import json, sys, time, urllib.request
+unit, port, gwm, load_s, gw, key = sys.argv[1], int(sys.argv[2]), sys.argv[3], int(sys.argv[4]), sys.argv[5], sys.argv[6]
+body = {"messages": [{"role": "user", "content": "Reply with the single word: ready"}], "max_tokens": 64,
+        "temperature": 0, "chat_template_kwargs": {"enable_thinking": False}}
+def call(url, b, hdr):
+    req = urllib.request.Request(url, data=json.dumps(b).encode(), headers={"Content-Type": "application/json", **hdr})
+    return json.load(urllib.request.urlopen(req, timeout=600))
+row = {"unit": unit, "port": port, "load_s": load_s, "time": time.strftime("%FT%TZ", time.gmtime())}
+try:
+    r = call(f"http://127.0.0.1:{port}/v1/chat/completions", body, {})
+    row["direct_reply"] = (r["choices"][0]["message"].get("content") or "")[:60]
+    row["direct_tg"] = r.get("timings", {}).get("predicted_per_second")
+except Exception as e:
+    row["direct_error"] = str(e)[:200]
+try:   # default request (no kwargs): thinking off by default; nothing of a think block may reach content
+    r = call(f"http://127.0.0.1:{port}/v1/chat/completions",
+             {"messages": body["messages"], "max_tokens": 1024, "temperature": 0}, {})
+    m = r["choices"][0]["message"]
+    row["default_reply"] = (m.get("content") or "")[:60]
+    row["default_reasoning_len"] = len(m.get("reasoning_content") or "")
+    row["default_leak"] = any(x in (m.get("content") or "") for x in ("<think>", "</think>"))
+except Exception as e:
+    row["default_error"] = str(e)[:200]
+try:   # thinking requested: nothing of the think block may reach content. With thinking forced on and no tools, a
+       # one-word answer can stay inside the think block (answer in reasoning_content, content empty) - model behaviour
+    r = call(f"http://127.0.0.1:{port}/v1/chat/completions",
+             {"messages": body["messages"], "max_tokens": 1024, "temperature": 0,
+              "chat_template_kwargs": {"enable_thinking": True}}, {})
+    m = r["choices"][0]["message"]
+    row["thinking_reply"] = (m.get("content") or "")[:60]
+    row["thinking_reasoning_len"] = len(m.get("reasoning_content") or "")
+    row["thinking_leak"] = any(x in (m.get("content") or "") for x in ("<think>", "</think>"))
+except Exception as e:
+    row["thinking_error"] = str(e)[:200]
+try:
+    names = [m["id"] for m in json.load(urllib.request.urlopen(urllib.request.Request(
+        f"{gw}/v1/models", headers={"Authorization": f"Bearer {key}"}), timeout=30))["data"]]
+    if gwm in names:
+        r = call(f"{gw}/v1/chat/completions", {**body, "model": gwm}, {"Authorization": f"Bearer {key}"})
+        row["gateway_model"] = gwm
+        row["gateway_reply"] = (r["choices"][0]["message"].get("content") or "")[:60]
+    else:
+        row["gateway_model"] = None   # route not added yet
+except Exception as e:
+    row["gateway_error"] = str(e)[:200]
+row["result"] = "PASS" if row.get("direct_reply") and not row.get("gateway_error") and \
+    row.get("default_reply") and row.get("default_leak") is False and \
+    (row.get("thinking_reply") or row.get("thinking_reasoning_len")) and row.get("thinking_leak") is False and \
+    (row.get("gateway_model") is None or row.get("gateway_reply")) else "FAIL"
+print(json.dumps(row))
+PY
+  grep -q "\"unit\": \"$UNIT\".*\"result\": \"PASS\"" <(tail -1 results/nex_seats.jsonl) || fail=1
+  systemctl stop $UNIT.service
+  for k in $(seq 1 30); do systemctl is-active --quiet $UNIT.service || break; sleep 2; done
+  log "  $UNIT stopped (enabled: $(systemctl is-enabled $UNIT.service 2>&1))"
+  sleep 5
+done
+log "NEX_SEATS_DONE fail=$fail"
+exit $fail

recipe/pipeline/nex_sizing.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+# Nex-N2.5-mini memory footprint (drop in MemAvailable): STRIX_LEAN + vision projector, q8_0 KV, one slot, no draft head.
+set -uo pipefail
+W=/mnt/models/nex-n2.5-mini; B=/opt/llama-rocm/rocmfpx-724/build-hipvk/bin; N=Nex-N2.5-mini; cd $W
+export LD_LIBRARY_PATH=$B:/opt/rocm-7.2.4/lib HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1
+prlimit --pid $$ --core=1:1   # 1 BYTE (bash `ulimit -c 1` is 1 KiB and does not stop apport)
+# One quiet box: publish_v2.sh upload takes the same lock, so an HF upload never skews MemAvailable here.
+exec 9>/mnt/models/.quiet-box.lock
+echo "[$(date -u +%FT%TZ)] waiting for the quiet-box lock"; flock 9; echo "[$(date -u +%FT%TZ)] quiet-box lock held"
+avail(){ awk '/MemAvailable/{printf "%.2f", $2/1048576}' /proc/meminfo; }
+sync; sleep 5
+measure(){  # $1 label  $2 ctx  $3 model
+  local a0=$(avail)
+  $B/llama-server -m "$3" -dev ROCm0 -ngl 999 -fa on --no-mmap -fit off -c $2 -ctk q8_0 -ctv q8_0 -np 1 -b 2048 -ub 1024 -t 16 \
+     --jinja --mmproj out/mmproj-$N-BF16.gguf -cram 512 --host 127.0.0.1 --port 18720 9>&- > logs/N7_sizing_$1_c$2.log 2>&1 &
+  local pid=$! ok=0
+  for i in $(seq 1 120); do curl -s -m 2 http://127.0.0.1:18720/health | grep -q ok && { ok=1; break; }; kill -0 $pid 2>/dev/null || break; sleep 3; done
+  if [ $ok -ne 1 ]; then echo "{\"label\":\"$1\",\"ctx\":$2,\"result\":\"LOAD_FAIL\"}" | tee -a results/nex_sizing.jsonl; kill $pid 2>/dev/null; sleep 5; return; fi
+  local a1=$(avail)
+  python3 - <<EOF
+import json, urllib.request
+txt = open("/opt/llama-rocm/rocmfpx-724/convert_hf_to_gguf.py").read()[:30000]
+b = {"messages": [{"role": "user", "content": txt + "\n\nSummarise this file in one sentence."}], "max_tokens": 64,
+     "temperature": 0, "chat_template_kwargs": {"reasoning_effort": "none"}}
+urllib.request.urlopen(urllib.request.Request("http://127.0.0.1:18720/v1/chat/completions", data=json.dumps(b).encode(),
+                       headers={"Content-Type": "application/json"}), timeout=600).read()
+EOF
+  local a2=$(avail)
+  echo "{\"label\":\"$1\",\"ctx\":$2,\"avail_before\":$a0,\"footprint_loaded_gib\":$(python3 -c "print(round($a0-$a1,2))"),\"footprint_after_8k_gib\":$(python3 -c "print(round($a0-$a2,2))")}" | tee -a results/nex_sizing.jsonl
+  kill $pid; wait $pid 2>/dev/null; sleep 8
+}
+for C in 65536 262144; do measure strix-lean $C out/$N-Q4_0_ROCMFP4_STRIX_LEAN.gguf; done
+echo NEX_SIZING_DONE

recipe/pipeline/nex_tools_diag.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""Diagnose the Nex tool-call gate (n-tools-q106: 6/14 - every thinking-on check failed, and nested-object failed
+with thinking off on an HTTP 500 "does not match the expected peg-native format"). nex_bench.py keeps only the last
+800 characters of the harness output, so the per-check reasons were lost.
+
+Same binary, file, server flags and tool schema as the gate, but every raw reply is kept (content, reasoning_content,
+tool_calls, HTTP error body). Variants, one server session:
+  gate_on       reasoning_effort=high, gate sampling (temp 0.7, top_p 0.95, top_k 40)   - the failing half
+  gate_off      reasoning_effort=none, gate sampling                                    - the passing half
+  on_greedy     reasoning_effort=high, temp 0 / top_k 1                                 - is it sampling?
+  adaptive      no reasoning_effort (the template's adaptive mode), gate sampling
+  nested_off_x3 the nested-object request three more times with thinking off           - is the 500 repeatable?
+Diagnostic only (not a card measurement) -> results/nex_tools_diag.json."""
+import json, os, sys, time, urllib.error, urllib.request
+from types import SimpleNamespace
+
+os.environ["AGNES_BIN"] = "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin"   # the gate ran on the unpatched server
+sys.path.insert(0, "/mnt/models/nex-n2.5-mini")
+import nex_harness as H  # noqa: E402
+
+W = H.W
+PORT = 18650
+OUT = f"{W}/results/nex_tools_diag.json"
+a = SimpleNamespace(model=f"{W}/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf", dev="ROCm0", ctx=65536,
+                    draft=None, mtp_infile=False, nmax=4, pmin=0.0, strict=False,
+                    serverlog=f"{W}/logs/diag_tools_server.log")
+PROMPTS = {
+    "multi-arg": "What's the weather in Paris in celsius?",
+    "nested-object": "Book 'Design review' on 2026-10-02 at 14:00 with ana@x.io and bo@x.io.",
+    "enum": "Weather in Denver, and give it to me in fahrenheit.",
+    "correct-decline": "What is 17 times 23? Answer directly.",
+    "multi-turn": "What's the weather in Tokyo in celsius?",
+    "streaming": "What's the weather in Rome in celsius?",
+    "parallel": "Get the weather in Oslo AND in Lima, both in celsius. Call the tool for each city.",
+}
+GATE = dict(temperature=0.7, top_p=0.95, top_k=40)
+GREEDY = dict(temperature=0, top_k=1)
+
+
+def call(msgs, kwargs, sampling, stream=False):
+    body = {"messages": msgs, "tools": H.TOOLS, "tool_choice": "auto", "max_tokens": 8192, **sampling}
+    if kwargs is not None:
+        body["chat_template_kwargs"] = kwargs
+    t0 = time.time()
+    try:
+        if not stream:
+            r = H.post(PORT, "/v1/chat/completions", body)
+            ch = r["choices"][0]
+            return {"message": ch["message"], "finish": ch.get("finish_reason"), "s": round(time.time() - t0, 2)}
+        body["stream"] = True
+        req = urllib.request.Request(f"http://127.0.0.1:{PORT}/v1/chat/completions", data=json.dumps(body).encode(),
+                                     headers={"Content-Type": "application/json"})
+        calls, content, reasoning, finish = {}, "", "", None
+        for line in urllib.request.urlopen(req, timeout=1800):
+            line = line.decode().strip()
+            if not line.startswith("data:") or line.endswith("[DONE]"):
+                continue
+            ch = json.loads(line[5:])["choices"][0]
+            d = ch.get("delta") or {}
+            finish = ch.get("finish_reason") or finish
+            content += d.get("content") or ""
+            reasoning += d.get("reasoning_content") or ""
+            for tc in d.get("tool_calls") or []:
+                c = calls.setdefault(tc["index"], {"name": "", "arguments": ""})
+                c["name"] += (tc.get("function") or {}).get("name") or ""
+                c["arguments"] += (tc.get("function") or {}).get("arguments") or ""
+        msg = {"content": content, "reasoning_content": reasoning,
+               "tool_calls": [{"function": v} for _, v in sorted(calls.items())]}
+        return {"message": msg, "finish": finish, "s": round(time.time() - t0, 2)}
+    except urllib.error.HTTPError as e:
+        return {"http_error": e.code, "body": e.read().decode(errors="replace")[:600], "s": round(time.time() - t0, 2)}
+    except Exception as e:  # noqa: BLE001 - a diagnostic records every failure mode
+        return {"exception": repr(e)[:300], "s": round(time.time() - t0, 2)}
+
+
+def summarize(res):
+    if "message" not in res:
+        return res
+    m = res["message"]
+    c = m.get("content") or ""
+    return {"finish": res.get("finish"), "s": res["s"],
+            "tool_calls": [(t.get("function") or {}).get("name") for t in (m.get("tool_calls") or [])],
+            "args": [(t.get("function") or {}).get("arguments") for t in (m.get("tool_calls") or [])],
+            "leaks_in_content": [x for x in H.LEAK if x in c],
+            "content": c[:600], "reasoning_len": len(m.get("reasoning_content") or ""),
+            "reasoning_head": (m.get("reasoning_content") or "")[:300]}
+
+
+def run_variant(name, kwargs, sampling, stream_ok=True):
+    out = {}
+    for k, p in PROMPTS.items():
+        msgs = [{"role": "user", "content": p}]
+        r = call(msgs, kwargs, sampling, stream=(k == "streaming" and stream_ok))
+        rec = {"first": summarize(r)}
+        if k == "multi-turn" and "message" in r and r["message"].get("tool_calls"):
+            tc = r["message"]["tool_calls"][0]
+            msgs += [{"role": "assistant", "content": r["message"].get("content") or "",
+                      "tool_calls": [{"id": "call_1", "type": "function", "function": tc["function"]}]},
+                     {"role": "tool", "tool_call_id": "call_1", "content": json.dumps({"temp_c": 21, "sky": "clear"})}]
+            rec["second"] = summarize(call(msgs, kwargs, sampling))
+        out[k] = rec
+        print(name, k, json.dumps(rec)[:400], flush=True)
+    return out
+
+
+s = H.Server(a, PORT)
+report = {"model": os.path.basename(a.model), "bin": H.BIN, "cmd": " ".join(s.cmd), "variants": {}}
+try:
+    report["variants"]["gate_on"] = run_variant("gate_on", H.THINK_ON, GATE)
+    report["variants"]["gate_off"] = run_variant("gate_off", H.THINK_OFF, GATE)
+    report["variants"]["on_greedy"] = run_variant("on_greedy", H.THINK_ON, GREEDY, stream_ok=False)
+    report["variants"]["adaptive"] = run_variant("adaptive", None, GATE, stream_ok=False)
+    report["variants"]["nested_off_x3"] = [
+        summarize(call([{"role": "user", "content": PROMPTS["nested-object"]}], H.THINK_OFF, GATE)) for _ in range(3)]
+    print("nested_off_x3", json.dumps(report["variants"]["nested_off_x3"])[:600], flush=True)
+finally:
+    s.stop()
+json.dump(report, open(OUT, "w"), indent=1)
+print("NEX_TOOLS_DIAG_DONE", OUT)

recipe/pipeline/nex_tools_tpl.py

@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+"""Tool-call gate on the Nex STRIX_LEAN file with a one-line template shim.
+
+Why: the stock template switches thinking with `reasoning_effort` only. llama-server's template analysis probes
+`enable_thinking`, finds no difference, and builds a parser that never extracts reasoning: with thinking on,
+`reasoning_content` stays empty and the reasoning plus `</think>` land in `content` (n-tools-q106: 0/7 with thinking
+on; results/nex_tools_diag.json, results/nex_reasoning_probe.json). `--reasoning-format deepseek`, server-level
+`--chat-template-kwargs` and `--reasoning on` do not change that. Prepending one line that maps `enable_thinking`
+onto `reasoning_effort` does.
+
+This run: the same suite, sampling, file and server as n-tools-q106, plus `--chat-template-file` with the shim, and
+thinking switched with `enable_thinking` true / false -> label n-tools-q106-tpl in results/nex_bench.jsonl (full
+per-check output in logs/N6t_tools_tpl.log). Then probes of the shim's other paths (no kwargs; an explicit
+reasoning_effort) -> results/nex_template_shim.json."""
+import argparse, hashlib, json, os, subprocess, sys
+from types import SimpleNamespace
+
+os.environ["AGNES_BIN"] = "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin"   # same unpatched server as n-tools-q106
+sys.path.insert(0, "/mnt/models/nex-n2.5-mini")
+import nex_harness as H  # noqa: E402
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--on", default="high", choices=["high", "medium"], help="reasoning_effort that enable_thinking=true maps to")
+ap.add_argument("--label", default="n-tools-q106-tpl")
+ap.add_argument("--tpl", default="chat_template_enable_thinking.jinja")
+ap.add_argument("--out", default="nex_template_shim.json")
+ap.add_argument("--extra", default="[]", help="JSON list of extra llama-server flags")
+ap.add_argument("--vision", action="store_true", help="also run the -fa on image gate in this configuration")
+ap.add_argument("--repeats", type=int, default=1, help="tool-suite passes (labels <label>, <label>-r2, ...)")
+args = ap.parse_args()
+W = H.W
+TPL = f"{W}/tpl/{args.tpl}"
+SHIM = ("{%- if reasoning_effort is not defined and enable_thinking is defined %}"
+        "{%- set reasoning_effort = '" + args.on + "' if enable_thinking else 'none' %}{%- endif %}\n")
+os.makedirs(os.path.dirname(TPL), exist_ok=True)
+src = open(f"{W}/hf/chat_template.jinja", "rb").read()
+open(TPL, "wb").write(SHIM.encode() + src)
+
+H.THINK_ON, H.THINK_OFF = {"enable_thinking": True}, {"enable_thinking": False}
+_orig_popen = subprocess.Popen
+
+
+def _popen(cmd, *pargs, **kw):
+    if cmd and str(cmd[0]).endswith("llama-server"):
+        cmd = list(cmd) + ["--chat-template-file", TPL] + json.loads(args.extra)
+    return _orig_popen(cmd, *pargs, **kw)
+
+
+subprocess.Popen = _popen
+MODEL = f"{W}/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf"
+a = SimpleNamespace(model=MODEL, dev="ROCm0", ctx=65536, draft=None, mtp_infile=False, nmax=4, pmin=0.0, strict=False,
+                    port=18600, label=args.label, jsonl=f"{W}/results/nex_bench.jsonl",
+                    serverlog=f"{W}/logs/b_{args.label}.log")
+H.cmd_tools(a)
+for r in range(2, args.repeats + 1):
+    H.cmd_tools(SimpleNamespace(**{**vars(a), "label": f"{args.label}-r{r}",
+                                   "serverlog": f"{W}/logs/b_{args.label}-r{r}.log"}))
+if args.vision:
+    A = "/mnt/models/agnes-3.0-flash"
+    v = SimpleNamespace(**{**vars(a), "label": args.label.replace("tools", "vision") + "-faon",
+                           "serverlog": f"{W}/logs/b_{args.label.replace('tools', 'vision')}-faon.log",
+                           "fa": "on", "mmproj": f"{W}/out/mmproj-Nex-N2.5-mini-BF16.gguf",
+                           "image": f"{A}/calib/vision_probe.png",
+                           "question": "Describe the two shapes in this image and their colors.",
+                           "expect": "red,blue,circle,square"})
+    try:
+        H.cmd_vision(v)
+    except SystemExit:
+        pass   # cmd_vision exits 1 on FAIL after writing its row; the probes below still run
+
+# the shim's other paths, greedy: no kwargs at all, and an explicit reasoning_effort (the shim then steps aside)
+p = SimpleNamespace(**{**vars(a), "label": "", "serverlog": f"{W}/logs/b_{args.label}-probe.log"})
+s = H.Server(p, 18652)
+probes = {}
+try:
+    for name, kw in (("no-kwargs", None), ("enable_thinking=false", {"enable_thinking": False}),
+                     ("reasoning_effort=high", {"reasoning_effort": "high"}),
+                     ("reasoning_effort=medium", {"reasoning_effort": "medium"}),
+                     ("reasoning_effort=none", {"reasoning_effort": "none"})):
+        for rn, prompt in (("correct-decline", "What is 17 times 23? Answer directly."),
+                           ("single-word", "Reply with the single word: ready"),
+                           ("multi-arg", "What's the weather in Paris in celsius?")):
+            body = {"messages": [{"role": "user", "content": prompt}], "tools": H.TOOLS, "tool_choice": "auto",
+                    "temperature": 0, "top_k": 1, "max_tokens": 2048}
+            if kw is not None:
+                body["chat_template_kwargs"] = kw
+            try:
+                m = H.post(18652, "/v1/chat/completions", body)["choices"][0]["message"]
+                c = m.get("content") or ""
+                probes[f"{name}|{rn}"] = {
+                    "content": c[:200], "reasoning_len": len(m.get("reasoning_content") or ""),
+                    "tool_calls": [(t.get("function") or {}).get("name") for t in m.get("tool_calls") or []],
+                    "leaks": [x for x in H.LEAK if x in c]}
+            except Exception as e:  # noqa: BLE001 - a probe records every failure mode
+                probes[f"{name}|{rn}"] = {"error": repr(e)[:300]}
+            print("probe", name, rn, json.dumps(probes[f"{name}|{rn}"])[:260], flush=True)
+finally:
+    s.stop()
+json.dump({"label": args.label, "enable_thinking_true_maps_to": args.on,
+           "template_file": os.path.basename(TPL), "template_path": TPL,
+           "sha256": hashlib.sha256(open(TPL, "rb").read()).hexdigest(),
+           "source_sha256": hashlib.sha256(src).hexdigest(), "shim": SHIM, "probes": probes},
+          open(f"{W}/results/{args.out}", "w"), indent=1)
+print("NEX_TOOLS_TPL_DONE")

recipe/pipeline/readback.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python3
+"""Exact-name read-back gate: the built file is the receipt, not the ftype name.
+usage: readback.py <expect_output> <expect_embd> <file.gguf>   (use '-' to skip a check)"""
+import sys
+sys.path.insert(0, "/opt/llama-rocm/rocmfpx-724/gguf-py")
+import gguf
+exp_out, exp_emb, path = sys.argv[1:4]
+r = gguf.GGUFReader(path)
+def tname(t):
+    try: return t.tensor_type.name
+    except Exception: return f"TYPE_{int(t.tensor_type)}"
+ty = {t.name: tname(t) for t in r.tensors}
+f = r.fields.get("general.file_type")
+ftype = int(f.parts[f.data[0]][0]) if f else None
+arch = bytes(r.fields["general.architecture"].parts[-1]).decode()
+nextn = sum(".nextn." in k for k in ty)
+out, emb = ty.get("output.weight"), ty.get("token_embd.weight")
+ok = (exp_out == "-" or out == exp_out) and (exp_emb == "-" or emb == exp_emb)
+print(f"{'PASS' if ok else 'FAIL'} {path.split('/')[-1]} arch={arch} ftype={ftype} tensors={len(ty)} nextn={nextn} "
+      f"output.weight={out} token_embd.weight={emb}")
+sys.exit(0 if ok else 1)

recipe/pipeline/run_tools_c1.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+# C1 = high shim template + server default enable_thinking=false: full tool suite + -fa on image gate + probes
+cd /mnt/models/nex-n2.5-mini
+systemd-run --scope --quiet --unit=nex-tools-c1-$(date +%s) -p MemoryMax=32G -p MemorySwapMax=0 \
+  python3 nex_tools_tpl.py --on high --label n-tools-q106-c1 --tpl chat_template_enable_thinking.jinja \
+  --extra '["--chat-template-kwargs", "{\"enable_thinking\": false}"]' --vision --out nex_template_shim_c1.json
+echo "rc=$?"

recipe/pipeline/run_tools_roff.sh

@@ -0,0 +1,8 @@
+#!/bin/bash
+# Recommended Nex serving: high shim template + --reasoning off (default thinking off; enable_thinking=true per request
+# turns it on). Three tool-suite passes + -fa on image gate + probes.
+cd /mnt/models/nex-n2.5-mini
+systemd-run --scope --quiet --unit=nex-tools-roff-$(date +%s) -p MemoryMax=32G -p MemorySwapMax=0 \
+  python3 nex_tools_tpl.py --on high --label n-tools-q106-roff --tpl chat_template_enable_thinking.jinja \
+  --extra '["--reasoning", "off"]' --repeats 3 --vision --out nex_template_shim_roff.json
+echo "rc=$?"

recipe/pipeline/upload_watchdog.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+# upload_watchdog.sh <name> <workdir> <command...>
+# Runs a Hugging Face upload command in a 40G memory scope (niced) and kills + retries it (3 attempts) when its progress
+# output — upload_large_folder's minute-by-minute "Files:" status and the xet progress bars — has not changed for
+# 20 minutes. Reason: on 2026-09-16 the Agnes imatrix upload hung silently for 51 min (both HTTPS sockets in
+# CLOSE-WAIT, 0 B/s, no error line) while holding the quiet-box lock. upload_large_folder resumes from the folder's
+# .cache, so a retry only sends what is missing. Exit 0 once an attempt prints UPLOAD_DONE.
+# The caller holds the quiet-box lock on fd 9; the upload process itself does not inherit it.
+set -uo pipefail
+name=$1; dir=$2; shift 2
+cd "$dir" || exit 2
+mkdir -p logs
+log(){ echo "[$(date -u +%FT%TZ)] $*"; }
+for try in 1 2 3; do
+  L=logs/upload_$name.r$try.log; U=$name-r$try-$(date +%s)
+  log "$name attempt $try -> $dir/$L (scope $U, MemoryMax=40G)"
+  systemd-run --scope --quiet --unit="$U" -p MemoryMax=40G -p MemorySwapMax=0 nice -n 10 "$@" > "$L" 2>&1 9>&- &
+  pid=$!; last=""; same=0
+  while kill -0 $pid 2>/dev/null; do
+    sleep 60
+    cur=$(tr '\r' '\n' < "$L" | grep -E '^Files:|B / ' | tail -3 | md5sum)
+    if [ "$cur" = "$last" ]; then same=$((same + 1)); else same=0; last=$cur; fi
+    if [ $same -ge 20 ]; then
+      log "  $name attempt $try: no progress for 20 min -> kill ($(tr '\r' '\n' < "$L" | grep -E '^Files:' | tail -1))"
+      systemctl kill --signal=KILL "$U"; same=0
+    fi
+  done
+  wait $pid; rc=$?
+  if [ $rc -eq 0 ] && grep -q UPLOAD_DONE "$L"; then log "  $name attempt $try done"; exit 0; fi
+  log "  $name attempt $try failed rc=$rc: $(tr '\r' '\n' < "$L" | grep -v '^\s*$' | tail -1 | cut -c1-200)"
+  sleep 30
+done
+exit 1

recipe/pipeline/verify_download.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+"""Byte + sha256 gate for a pinned HF snapshot. Compares against the tree API's lfs.oid
+(the real content sha256) -- never the plain etag, which is a Xet/CAS id.
+usage: verify_download.py <repo> <revision> <local_dir>"""
+import hashlib, json, os, sys, urllib.request
+from concurrent.futures import ThreadPoolExecutor
+repo, rev, d = sys.argv[1:4]
+tree = json.load(urllib.request.urlopen(
+    f"https://huggingface.co/api/models/{repo}/tree/{rev}?recursive=true", timeout=60))
+files = [t for t in tree if t.get("type") == "file"]
+def sha(p):
+    h = hashlib.sha256()
+    with open(p, "rb") as f:
+        for b in iter(lambda: f.read(64 << 20), b""): h.update(b)
+    return h.hexdigest()
+def one(t):
+    p = os.path.join(d, t["path"])
+    if not os.path.exists(p): return (t["path"], "MISSING")
+    if os.path.getsize(p) != t["size"]: return (t["path"], f"SIZE {os.path.getsize(p)} != {t['size']}")
+    lfs = t.get("lfs")
+    if lfs:
+        got = sha(p)
+        return (t["path"], "ok-sha" if got == lfs["oid"] else f"SHA {got[:12]} != {lfs['oid'][:12]}")
+    return (t["path"], "ok-size")
+with ThreadPoolExecutor(6) as ex: res = list(ex.map(one, files))
+bad = [r for r in res if not r[1].startswith("ok")]
+print(f"files={len(res)} sha-verified={sum(r[1]=='ok-sha' for r in res)} size-only={sum(r[1]=='ok-size' for r in res)} bad={len(bad)}")
+for r in bad: print("  BAD", r)
+print("RESULT:", "PASS" if not bad else "FAIL")
+sys.exit(0 if not bad else 1)

recipe/results_summary.json

@@ -0,0 +1,1721 @@
+{
+  "layout": "no MTP (checkpoint declares mtp_num_hidden_layers=1 but ships no mtp.* tensors)",
+  "model": "Nex-N2.5-mini",
+  "model_repo_std": "kingjones777/Nex-N2.5-mini-ROCmFP4-GGUF",
+  "model_repo_imat": "kingjones777/Nex-N2.5-mini-ROCmFP4-imatrix-GGUF",
+  "source": {
+    "repo": "nex-agi/Nex-N2.5-mini",
+    "revision": "87420286149d9cce9bd46cd335ef9bda33c37c1b",
+    "license": "apache-2.0",
+    "gated": false,
+    "params": 35107181936
+  },
+  "arch": {
+    "name": "qwen3_5_moe",
+    "llama_cpp": "qwen35moe",
+    "layers": 40,
+    "linear_attn_layers": 30,
+    "full_attn_layers": 10,
+    "full_attention_interval": 4,
+    "hidden": 2048,
+    "num_experts": 256,
+    "num_experts_per_tok": 8,
+    "expert_width": 512,
+    "shared_expert_width": 512,
+    "attn_heads": 16,
+    "kv_heads": 2,
+    "head_dim": 256,
+    "linear_key_heads": 16,
+    "linear_value_heads": 32,
+    "linear_head_dim": 128,
+    "vocab": 248320,
+    "max_position_embeddings": 262144,
+    "tie_word_embeddings": false,
+    "vision_layers": 27,
+    "vision_width": 1152,
+    "hf_tensors": 1026,
+    "mtp_num_hidden_layers_declared": 1
+  },
+  "bf16": {
+    "ppl": 6.2303,
+    "ppl_err": 0.07538,
+    "ppl_paired": 6.228979,
+    "ppl_paired_err": 0.075322,
+    "chunks": 40,
+    "n_ctx": 2048,
+    "scored_tokens": 40920,
+    "size_bytes": 69376636928,
+    "file": "Nex-N2.5-mini-BF16.gguf",
+    "readback": "PASS",
+    "arch": "qwen35moe",
+    "ftype": 32,
+    "tensors": 733,
+    "nextn_tensors": 0,
+    "output_weight": "BF16",
+    "token_embd": "BF16",
+    "mtp_named_tensors": 0,
+    "nextn_named_tensors": 0
+  },
+  "binary": {
+    "repo": "https://github.com/charlie12345/ROCmFPX",
+    "commit": "d3ca537",
+    "dir": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+    "sha256": {
+      "llama-server": "d077b98f3f04df4b482a5826c8922e39cda3aa8d2e2b2fbe1870f36d245cebe7",
+      "llama-quantize": "1559189968a30781f837e4d25ca3cdf26c8eb0351023675406535eddcbe0db13",
+      "llama-imatrix": "bfdce5c21a01bdc8043d69335b74a7b68b8335ae989991a1e10e21d5faeaeb8e",
+      "llama-perplexity": "c2175925bb0c61df89ff787845d8e03a35403f684be52719ac2987b8ca636a53"
+    }
+  },
+  "vision_probe": {
+    "width": 448,
+    "height": 448,
+    "bytes": 1945,
+    "sha256": "53924b4715db9757d5b70614dd0e560b8f7de686ff3df9792721d157e2594f5e"
+  },
+  "imatrix": {
+    "chunks": 129,
+    "n_ctx": 512,
+    "calibration": "bartowski calibration_datav3.txt",
+    "file": "Nex-N2.5-mini.imatrix",
+    "sha256": "7e5afffc822c64a7b43a1223d0094fb60895a521f6c89b912409d25091bd17f8",
+    "size_bytes": 192223936,
+    "device": "CPU"
+  },
+  "hub_check": {
+    "checked_at_utc": "2026-09-16T21:27:05Z",
+    "queries": [
+      "Nex-N2.5-mini",
+      "Nex-N2.5",
+      "nex-n2.5-mini gguf",
+      "Nex N2.5 mini"
+    ],
+    "repos": [
+      {
+        "repo": "6block/Nex-N2.5-Pro-GGUF",
+        "gguf_files": 58,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-Pro-IQ2_XS-00001-of-00019.gguf",
+          "Nex-N2.5-Pro-IQ2_XS-00002-of-00019.gguf",
+          "Nex-N2.5-Pro-IQ2_XS-00003-of-00019.gguf"
+        ]
+      },
+      {
+        "repo": "Arahide/Nex-N2.5-mini-INT4-W4A16",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "DevQuasar/nex-agi.Nex-N2.5-Max-GGUF",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "DevQuasar/nex-agi.Nex-N2.5-Pro-GGUF",
+        "gguf_files": 66,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Q2_K/nex-agi.Nex-N2.5-Pro.f16.gguf.Q2_K.gguf-00001-of-00011.gguf",
+          "Q2_K/nex-agi.Nex-N2.5-Pro.f16.gguf.Q2_K.gguf-00002-of-00011.gguf",
+          "Q2_K/nex-agi.Nex-N2.5-Pro.f16.gguf.Q2_K.gguf-00003-of-00011.gguf"
+        ]
+      },
+      {
+        "repo": "DevQuasar/nex-agi.Nex-N2.5-mini-GGUF",
+        "gguf_files": 7,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Q2_K/nex-agi.Nex-N2.5-mini.f16.gguf.Q2_K.gguf",
+          "Q3_K_M/nex-agi.Nex-N2.5-mini.f16.gguf.Q3_K_M.gguf",
+          "Q4_K_M/nex-agi.Nex-N2.5-mini.f16.gguf.Q4_K_M.gguf"
+        ]
+      },
+      {
+        "repo": "IsValorum/Nex-N2.5-mini-APEX-I-MiniPlus-GGUF",
+        "gguf_files": 2,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini.APEX-I-MiniPlus.gguf",
+          "mmproj-nex-agi_Nex-N2.5-mini-Q8_0.gguf"
+        ]
+      },
+      {
+        "repo": "Kagandi/Nex-N2.5-mini-mlx-4Bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "MrFuzzihead/Nex-N2.5-mini-APEX-GGUF",
+        "gguf_files": 12,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-APEX-Balanced.gguf",
+          "Nex-N2.5-mini-APEX-Compact.gguf",
+          "Nex-N2.5-mini-APEX-I-Balanced.gguf"
+        ]
+      },
+      {
+        "repo": "NandoG-AI/Nex-N2.5-mini-GGUF",
+        "gguf_files": 1,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-Q4_K_M.gguf"
+        ]
+      },
+      {
+        "repo": "ProCreations/Nex-N2.5-mini-NVFP4",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "SAIFIINDUSTRIES/Nex-N2.5-mini",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Tech2wild/Nex-N2.5-mini-Int4-Int8Mix",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-4bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-6bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-8bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-oQ2",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-oQ3",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-oQ4",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-oQ6",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "Vontra/Nex-N2.5-mini-MLX-oQ8",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "abenzerps/Nex-N2.5-mini-GGUF",
+        "gguf_files": 18,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-IQ1_M.gguf",
+          "Nex-N2.5-mini-IQ1_S.gguf",
+          "Nex-N2.5-mini-IQ2_M.gguf"
+        ]
+      },
+      {
+        "repo": "abenzerps/Nex-N2.5-mini-MLX-4bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "abenzerps/Nex-N2.5-mini-MLX-6bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "abenzerps/Nex-N2.5-mini-MLX-8bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "auryn-macmillan/Nex-N2.5-mini-exl3-5bpw",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "bartowski/Nex-N2.5-Pro-GGUF",
+        "gguf_files": 144,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-Pro-IQ1_M/Nex-N2.5-Pro-IQ1_M-00001-of-00003.gguf",
+          "Nex-N2.5-Pro-IQ1_M/Nex-N2.5-Pro-IQ1_M-00002-of-00003.gguf",
+          "Nex-N2.5-Pro-IQ1_M/Nex-N2.5-Pro-IQ1_M-00003-of-00003.gguf"
+        ]
+      },
+      {
+        "repo": "bartowski/nex-agi_Nex-N2.5-mini-GGUF",
+        "gguf_files": 28,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "mmproj-nex-agi_Nex-N2.5-mini-bf16.gguf",
+          "mmproj-nex-agi_Nex-N2.5-mini-f16.gguf",
+          "nex-agi_Nex-N2.5-mini-IQ2_M.gguf"
+        ]
+      },
+      {
+        "repo": "cbert33/Nex-N2.5-mini-FP8-Calibrated",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "ghazni101/Nex-N2.5-mini-MQ4R",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "julianmb/Nex-N2.5-mini-ROCmFP4-GGUF",
+        "gguf_files": 2,
+        "rocm_like_files": [
+          "Nex-N2.5-mini-ROCmFP4-STRIX_LEAN.gguf"
+        ],
+        "gguf_sample": [
+          "Nex-N2.5-mini-ROCmFP4-STRIX_LEAN.gguf",
+          "mmproj-Nex-N2.5-mini.gguf"
+        ]
+      },
+      {
+        "repo": "karmx/Nex-N2.5-mini-Mixed-Q2Q3-128K-GGUF",
+        "gguf_files": 2,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-mixed-Q2Q3-vision-trial.gguf",
+          "mmproj-Nex-N2.5-mini-f16.gguf"
+        ]
+      },
+      {
+        "repo": "liskasYR/Nex-N2.5-Max",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "mlx-community/Nex-N2.5-mini-OptiQ-4bit",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "mlx-community/Nex-N2.5-mini-oQ4",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "mradermacher/Nex-N2.5-mini-GGUF",
+        "gguf_files": 13,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini.IQ4_XS.gguf",
+          "Nex-N2.5-mini.Q2_K.gguf",
+          "Nex-N2.5-mini.Q3_K_L.gguf"
+        ]
+      },
+      {
+        "repo": "mradermacher/Nex-N2.5-mini-i1-GGUF",
+        "gguf_files": 24,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini.i1-IQ1_M.gguf",
+          "Nex-N2.5-mini.i1-IQ1_S.gguf",
+          "Nex-N2.5-mini.i1-IQ2_M.gguf"
+        ]
+      },
+      {
+        "repo": "nex-agi/Nex-N2.5-Max",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "nex-agi/Nex-N2.5-mini",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "ngquocvinh/Nex-N2.5-mini-GGUF",
+        "gguf_files": 16,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-IQ1_M.gguf",
+          "Nex-N2.5-mini-IQ2_XS.gguf",
+          "Nex-N2.5-mini-IQ3_M.gguf"
+        ]
+      },
+      {
+        "repo": "orcarouter/Nex-N2.5-mini-Uncensored",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "orcarouter/Nex-N2.5-mini-Uncensored-FP8",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "orcarouter/Nex-N2.5-mini-Uncensored-GGUF",
+        "gguf_files": 6,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-Uncensored-IQ4_XS.gguf",
+          "Nex-N2.5-mini-Uncensored-Q2_K.gguf",
+          "Nex-N2.5-mini-Uncensored-Q3_K_M.gguf"
+        ]
+      },
+      {
+        "repo": "orcarouter/Nex-N2.5-mini-Uncensored-MLX",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "orcarouter/Nex-N2.5-mini-Uncensored-NVFP4",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "primitive-ai/Nex-N2.5-mini-FP8",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "primitive-ai/Nex-N2.5-mini-NVFP4",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "primitive-ai/Nex-N2.5-mini-mixed-NVFP4-FP8",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "qtum/Nex-N2.5-Pro-GGUF",
+        "gguf_files": 58,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-Pro-IQ2_XS-00001-of-00019.gguf",
+          "Nex-N2.5-Pro-IQ2_XS-00002-of-00019.gguf",
+          "Nex-N2.5-Pro-IQ2_XS-00003-of-00019.gguf"
+        ]
+      },
+      {
+        "repo": "quant-mind/Nex-N2.5-mini-W4A16-AutoRound",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "sakamakismile/Nex-N2.5-Max-GGUF",
+        "gguf_files": 39,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Q3_K_M/Nex-N2.5-Max-Q3_K_M-00001-of-00018.gguf",
+          "Q3_K_M/Nex-N2.5-Max-Q3_K_M-00002-of-00018.gguf",
+          "Q3_K_M/Nex-N2.5-Max-Q3_K_M-00003-of-00018.gguf"
+        ]
+      },
+      {
+        "repo": "sigmanih/nex-agi-Nex-N2.5-mini-GGUF-Q4_K_S",
+        "gguf_files": 1,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "nex-agi--Nex-N2.5-mini.Q4_K_S.gguf"
+        ]
+      },
+      {
+        "repo": "sojufx/Nex-N2.5-mini-NVFP4",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "suzu89/Nex-N2.5-mini-oQ5",
+        "gguf_files": 0,
+        "rocm_like_files": [],
+        "gguf_sample": []
+      },
+      {
+        "repo": "vikrant1123/Nex-N2.5-mini-Uncensored-APEX-GGUF",
+        "gguf_files": 5,
+        "rocm_like_files": [],
+        "gguf_sample": [
+          "Nex-N2.5-mini-Uncensored-APEX-I-Balanced.gguf",
+          "Nex-N2.5-mini-Uncensored-APEX-I-Compact.gguf",
+          "Nex-N2.5-mini-Uncensored-APEX-I-Quality.gguf"
+        ]
+      }
+    ],
+    "gguf_repos": 19,
+    "rocm_builds_found": 1,
+    "header_checks": [
+      {
+        "repo": "julianmb/Nex-N2.5-mini-ROCmFP4-GGUF",
+        "file": "Nex-N2.5-mini-ROCmFP4-STRIX_LEAN.gguf",
+        "file_type": 106,
+        "tensors": 733,
+        "block_count": 40,
+        "output_weight_type_id": 101,
+        "token_embd_type_id": 13,
+        "imatrix_keys": [],
+        "read_via": "HTTP range read of the GGUF header",
+        "output_weight_type": "Q4_0_ROCMFP4_FAST",
+        "token_embd_type": "Q5_K"
+      }
+    ]
+  },
+  "repeat": {
+    "step": "N5",
+    "device": "ROCm0",
+    "rows": 40,
+    "result": "MATCH"
+  },
+  "repeat_vk": {
+    "step": "N5v",
+    "device": "Vulkan0",
+    "rows": 40,
+    "result": "MATCH"
+  },
+  "reference": {
+    "step": "N1c",
+    "cpu_chunk1": 5.6964,
+    "vulkan0_chunk1": 5.6953,
+    "hip_rocm0_chunk1": 139.1466,
+    "hip_rocm0_faoff_chunk1": 136.6879,
+    "hip_rocm0_final": 105.9103,
+    "result": "PASS"
+  },
+  "reference_device": "CPU",
+  "tiers": {
+    "q106": {
+      "file": "Nex-N2.5-mini-Q4_0-ROCmFP4-STRIX_LEAN.gguf",
+      "local_file": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "dir": "out",
+      "ftype": 106,
+      "size_bytes": 18744342208,
+      "quant_mib": 17865.52,
+      "bpw": 4.32,
+      "quant_seconds": 394.87116,
+      "imatrix_entries": null,
+      "readback": "PASS",
+      "arch": "qwen35moe",
+      "tensors": 733,
+      "nextn_tensors": 0,
+      "output_weight": "Q6_K",
+      "token_embd": "Q5_K",
+      "quality_measured": true,
+      "ppl": 6.473964,
+      "ppl_err": 0.079799,
+      "ppl_ratio": 1.03933,
+      "kld_mean": 0.104436,
+      "kld_err": 0.001413,
+      "kld_p99": 0.926687,
+      "kld_median": 0.047702,
+      "same_top_p": 86.659,
+      "rms_dp": 9.028,
+      "vk": {
+        "ppl": 6.474466,
+        "ppl_err": 0.079947,
+        "ppl_ratio": 1.03941,
+        "kld_mean": 0.104408,
+        "kld_err": 0.001395,
+        "kld_p99": 0.899358,
+        "kld_median": 0.048009,
+        "same_top_p": 86.794,
+        "rms_dp": 9.03
+      }
+    },
+    "q102": {
+      "file": "Nex-N2.5-mini-Q4_0-ROCmFP4-COHERENT.gguf",
+      "local_file": "Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf",
+      "dir": "out",
+      "ftype": 102,
+      "size_bytes": 19846166208,
+      "quant_mib": 18916.3,
+      "bpw": 4.58,
+      "quant_seconds": 443.71327,
+      "imatrix_entries": null,
+      "readback": "PASS",
+      "arch": "qwen35moe",
+      "tensors": 733,
+      "nextn_tensors": 0,
+      "output_weight": "Q6_K",
+      "token_embd": "Q6_K",
+      "quality_measured": true,
+      "ppl": 6.561727,
+      "ppl_err": 0.08124,
+      "ppl_ratio": 1.053419,
+      "kld_mean": 0.097071,
+      "kld_err": 0.001342,
+      "kld_p99": 0.874485,
+      "kld_median": 0.043524,
+      "same_top_p": 87.287,
+      "rms_dp": 8.8,
+      "vk": {
+        "ppl": 6.592123,
+        "ppl_err": 0.081874,
+        "ppl_ratio": 1.058299,
+        "kld_mean": 0.097165,
+        "kld_err": 0.001337,
+        "kld_p99": 0.858566,
+        "kld_median": 0.043968,
+        "same_top_p": 87.265,
+        "rms_dp": 8.853
+      }
+    },
+    "q103": {
+      "file": "Nex-N2.5-mini-Q4_0-ROCmFP4-FAST.gguf",
+      "local_file": "Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf",
+      "dir": "out",
+      "ftype": 103,
+      "size_bytes": 18648495808,
+      "quant_mib": 17774.11,
+      "bpw": 4.3,
+      "quant_seconds": 392.98796999999996,
+      "imatrix_entries": null,
+      "readback": "PASS",
+      "arch": "qwen35moe",
+      "tensors": 733,
+      "nextn_tensors": 0,
+      "output_weight": "Q6_K",
+      "token_embd": "Q4_0_ROCMFP4_FAST",
+      "quality_measured": true,
+      "ppl": 6.549829,
+      "ppl_err": 0.080919,
+      "ppl_ratio": 1.051509,
+      "kld_mean": 0.108779,
+      "kld_err": 0.00142,
+      "kld_p99": 0.956811,
+      "kld_median": 0.04973,
+      "same_top_p": 86.405,
+      "rms_dp": 9.108,
+      "vk": {
+        "ppl": 6.551305,
+        "ppl_err": 0.081055,
+        "ppl_ratio": 1.051746,
+        "kld_mean": 0.108756,
+        "kld_err": 0.001422,
+        "kld_p99": 0.930024,
+        "kld_median": 0.050197,
+        "same_top_p": 86.356,
+        "rms_dp": 9.101
+      }
+    },
+    "q106i": {
+      "file": "Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-STRIX_LEAN.gguf",
+      "local_file": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "dir": "out-imat",
+      "ftype": 106,
+      "size_bytes": 18744342464,
+      "quant_mib": 17865.52,
+      "bpw": 4.32,
+      "quant_seconds": 306.08412,
+      "imatrix_entries": 510,
+      "readback": "PASS",
+      "arch": "qwen35moe",
+      "tensors": 733,
+      "nextn_tensors": 0,
+      "output_weight": "Q6_K",
+      "token_embd": "Q5_K",
+      "quality_measured": true,
+      "ppl": 6.353616,
+      "ppl_err": 0.077001,
+      "ppl_ratio": 1.020009,
+      "kld_mean": 0.085158,
+      "kld_err": 0.001309,
+      "kld_p99": 0.737186,
+      "kld_median": 0.038436,
+      "same_top_p": 87.849,
+      "rms_dp": 8.281,
+      "vk": {
+        "ppl": 6.364168,
+        "ppl_err": 0.077334,
+        "ppl_ratio": 1.021703,
+        "kld_mean": 0.083563,
+        "kld_err": 0.001206,
+        "kld_p99": 0.713036,
+        "kld_median": 0.038,
+        "same_top_p": 87.903,
+        "rms_dp": 8.201
+      },
+      "same_tensor_types_as_standard": true,
+      "file_size_delta_bytes": 256,
+      "differs_from_standard": true,
+      "header_keys_only_in_imatrix": [
+        "quantize.imatrix.chunks_count",
+        "quantize.imatrix.dataset",
+        "quantize.imatrix.entries_count",
+        "quantize.imatrix.file"
+      ],
+      "header_keys_only_in_standard": []
+    },
+    "q102i": {
+      "file": "Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-COHERENT.gguf",
+      "local_file": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf",
+      "dir": "out-imat",
+      "ftype": 102,
+      "size_bytes": 19846166464,
+      "quant_mib": 18916.3,
+      "bpw": 4.58,
+      "quant_seconds": 409.08034000000004,
+      "imatrix_entries": 510,
+      "readback": "PASS",
+      "arch": "qwen35moe",
+      "tensors": 733,
+      "nextn_tensors": 0,
+      "output_weight": "Q6_K",
+      "token_embd": "Q6_K",
+      "quality_measured": true,
+      "ppl": 6.36005,
+      "ppl_err": 0.077045,
+      "ppl_ratio": 1.021042,
+      "kld_mean": 0.076947,
+      "kld_err": 0.001181,
+      "kld_p99": 0.685225,
+      "kld_median": 0.034241,
+      "same_top_p": 88.463,
+      "rms_dp": 7.909,
+      "vk": {
+        "ppl": 6.36321,
+        "ppl_err": 0.077202,
+        "ppl_ratio": 1.021549,
+        "kld_mean": 0.076844,
+        "kld_err": 0.001332,
+        "kld_p99": 0.669352,
+        "kld_median": 0.033844,
+        "same_top_p": 88.556,
+        "rms_dp": 7.783
+      },
+      "same_tensor_types_as_standard": true,
+      "file_size_delta_bytes": 256,
+      "differs_from_standard": true,
+      "header_keys_only_in_imatrix": [
+        "quantize.imatrix.chunks_count",
+        "quantize.imatrix.dataset",
+        "quantize.imatrix.entries_count",
+        "quantize.imatrix.file"
+      ],
+      "header_keys_only_in_standard": []
+    },
+    "q103i": {
+      "file": "Nex-N2.5-mini-imatrix-Q4_0-ROCmFP4-FAST.gguf",
+      "local_file": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf",
+      "dir": "out-imat",
+      "ftype": 103,
+      "size_bytes": 18648496064,
+      "quant_mib": 17774.11,
+      "bpw": 4.3,
+      "quant_seconds": 328.50694,
+      "imatrix_entries": 510,
+      "readback": "PASS",
+      "arch": "qwen35moe",
+      "tensors": 733,
+      "nextn_tensors": 0,
+      "output_weight": "Q6_K",
+      "token_embd": "Q4_0_ROCMFP4_FAST",
+      "quality_measured": true,
+      "ppl": 6.368075,
+      "ppl_err": 0.077317,
+      "ppl_ratio": 1.02233,
+      "kld_mean": 0.088974,
+      "kld_err": 0.001318,
+      "kld_p99": 0.784672,
+      "kld_median": 0.039738,
+      "same_top_p": 87.454,
+      "rms_dp": 8.37,
+      "vk": {
+        "ppl": 6.392929,
+        "ppl_err": 0.077851,
+        "ppl_ratio": 1.026321,
+        "kld_mean": 0.08915,
+        "kld_err": 0.001266,
+        "kld_p99": 0.784911,
+        "kld_median": 0.039583,
+        "same_top_p": 87.424,
+        "rms_dp": 8.509
+      },
+      "same_tensor_types_as_standard": true,
+      "file_size_delta_bytes": 256,
+      "differs_from_standard": true,
+      "header_keys_only_in_imatrix": [
+        "quantize.imatrix.chunks_count",
+        "quantize.imatrix.dataset",
+        "quantize.imatrix.entries_count",
+        "quantize.imatrix.file"
+      ],
+      "header_keys_only_in_standard": []
+    }
+  },
+  "aux": {
+    "mmproj-Nex-N2.5-mini-BF16.gguf": 902821920,
+    "chat_template_enable_thinking.jinja": 7895
+  },
+  "mmproj": {
+    "file": "mmproj-Nex-N2.5-mini-BF16.gguf",
+    "size_bytes": 902821920,
+    "readback": "PASS",
+    "arch": "clip",
+    "ftype": 32,
+    "tensors": 334
+  },
+  "bench": [
+    {
+      "label": "n-q106-rocm",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 63.94,
+      "tg_min": 63.56,
+      "tg_max": 63.96,
+      "pp_median": 1158.2,
+      "prompt_n": 7098,
+      "accept": null,
+      "prompt_n_min": 7098,
+      "prompt_n_max": 7100
+    },
+    {
+      "label": "n-q106-vk",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 4.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 68.16,
+      "tg_min": 68.1,
+      "tg_max": 68.21,
+      "pp_median": 1006.0,
+      "prompt_n": 7096,
+      "accept": null,
+      "prompt_n_min": 7096,
+      "prompt_n_max": 7098
+    },
+    {
+      "label": "n-q102-rocm",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 24.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 61.74,
+      "tg_min": 61.74,
+      "tg_max": 61.8,
+      "pp_median": 1186.7,
+      "prompt_n": 7100,
+      "accept": null,
+      "prompt_n_min": 7096,
+      "prompt_n_max": 7100
+    },
+    {
+      "label": "n-q102-vk",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_COHERENT.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 68.0,
+      "tg_min": 67.88,
+      "tg_max": 68.09,
+      "pp_median": 1001.4,
+      "prompt_n": 7098,
+      "accept": null,
+      "prompt_n_min": 7098,
+      "prompt_n_max": 7101
+    },
+    {
+      "label": "n-q103-rocm",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 63.31,
+      "tg_min": 63.14,
+      "tg_max": 63.38,
+      "pp_median": 1156.6,
+      "prompt_n": 7100,
+      "accept": null,
+      "prompt_n_min": 7094,
+      "prompt_n_max": 7100
+    },
+    {
+      "label": "n-q103-vk",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 20.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_FAST.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 68.62,
+      "tg_min": 68.6,
+      "tg_max": 68.67,
+      "pp_median": 998.9,
+      "prompt_n": 7098,
+      "accept": null,
+      "prompt_n_min": 7098,
+      "prompt_n_max": 7101
+    },
+    {
+      "label": "n-q106i-rocm",
+      "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 63.29,
+      "tg_min": 63.19,
+      "tg_max": 63.41,
+      "pp_median": 1145.3,
+      "prompt_n": 7100,
+      "accept": null,
+      "prompt_n_min": 7097,
+      "prompt_n_max": 7100
+    },
+    {
+      "label": "n-q106i-vk",
+      "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 20.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 67.77,
+      "tg_min": 67.74,
+      "tg_max": 67.86,
+      "pp_median": 995.0,
+      "prompt_n": 7098,
+      "accept": null,
+      "prompt_n_min": 7098,
+      "prompt_n_max": 7100
+    },
+    {
+      "label": "n-q102i-rocm",
+      "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 24.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 61.56,
+      "tg_min": 61.53,
+      "tg_max": 61.63,
+      "pp_median": 1183.6,
+      "prompt_n": 7098,
+      "accept": null,
+      "prompt_n_min": 7096,
+      "prompt_n_max": 7100
+    },
+    {
+      "label": "n-q102i-vk",
+      "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_COHERENT.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 67.81,
+      "tg_min": 67.77,
+      "tg_max": 67.85,
+      "pp_median": 998.7,
+      "prompt_n": 7097,
+      "accept": null,
+      "prompt_n_min": 7096,
+      "prompt_n_max": 7097
+    },
+    {
+      "label": "n-q103i-rocm",
+      "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 63.03,
+      "tg_min": 62.55,
+      "tg_max": 63.11,
+      "pp_median": 1154.4,
+      "prompt_n": 7101,
+      "accept": null,
+      "prompt_n_min": 7099,
+      "prompt_n_max": 7101
+    },
+    {
+      "label": "n-q103i-vk",
+      "model": "Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "code",
+      "load_s": 20.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out-imat/Nex-N2.5-mini-imatrix-Q4_0_ROCMFP4_FAST.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 68.44,
+      "tg_min": 68.42,
+      "tg_max": 68.56,
+      "pp_median": 993.9,
+      "prompt_n": 7098,
+      "accept": null,
+      "prompt_n_min": 7098,
+      "prompt_n_max": 7102
+    },
+    {
+      "label": "n-q106-rocm-prose",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "ROCm0",
+      "ctx": 65536,
+      "workload": "prose",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev ROCm0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 63.06,
+      "tg_min": 62.94,
+      "tg_max": 63.09,
+      "pp_median": 1139.1,
+      "prompt_n": 7454,
+      "accept": null,
+      "prompt_n_min": 7453,
+      "prompt_n_max": 7455
+    },
+    {
+      "label": "n-q106-vk-prose",
+      "model": "Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf",
+      "draft": null,
+      "nmax": null,
+      "strict": false,
+      "bin": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin",
+      "dev": "Vulkan0",
+      "ctx": 65536,
+      "workload": "prose",
+      "load_s": 22.0,
+      "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/nex-n2.5-mini/out/Nex-N2.5-mini-Q4_0_ROCMFP4_STRIX_LEAN.gguf -dev Vulkan0 -ngl 999 -fa on -dio --jinja -fit off --parallel 1 -c 65536 -b 2048 -ub 1024 --host 127.0.0.1 --port 18600 --no-webui",
+      "tg_median": 67.34,
+      "tg_min": 67.31,
+      "tg_max": 67.35,
+      "pp_median": 977.7,
+      "prompt_n": 7454,
+      "accept": null,
+      "prompt_n_min": 7454,
+      "prompt_n_max": 7457
+    }
+  ],
+  "gates": [
+    {
+      "label": "n-c3-q106",
+      "gate": "cachegate3-aligned",
+      "n": 5,
+      "L": 7039,
+      "expect_cache_n": 6011,
+      "all_reused": true,
+      "all_aligned": true,
+      "identical": 5,
+      "rows": [
+        {
+          "variant": 0,
+          "L": 7039,
+          "warm_prompt_n": 1028,
+          "warm_cache_n": 6011,
+          "cold_prompt_n": 7039,
+          "cold_cache_n": 0,
+          "aligned": true,
+          "warm_draft": [
+            null,
+            null
+          ],
+          "cold_draft": [
+            null,
+            null
+          ],
+          "identical": true,
+          "first_diff_char": null,
+          "warm_sha": "013b7b4662ba",
+          "cold_sha": "013b7b4662ba",
+          "instr": "Write a function that lists every tensor name in this file. ok ok ok ok",
+          "warm_instr": "Summarise this file. (warm-up 0) ok ok ok ok",
+          "warm_text": "```python\ndef list_tensor_names(model: ModelBase) -> list[str]:\n    \"\"\"Return every tensor name indexed by the model, in model-tensor order.\"\"\"\n    return list(model.model_tensors.keys())\n```\n\nExample usage:\n\n```python\nmodel = SomeModelClass(\n    dir_model=Path(\"...\"),\n    ftype=gguf.LlamaFileType.GUESSED,\n    fname_out=Path(\"model.gguf\"),\n)\n\nfor name in list_tensor_names(model):\n    print(name)\n```\n\nIf you want to include tensors that are only produced later by `modify_tensors()` or quantization dequantization, use this version instead:\n\n```python\ndef list_all_tensor_names(model: ModelBase) -> list[str]:\n    \"\"\"Return tensor names from indexing plus tensors produced by tensor transforms.\"\"\"\n    names = set(model.model_tensors.keys())\n\n    for name, gen in model.model",
+          "cold_text": "```python\ndef list_tensor_names(model: ModelBase) -> list[str]:\n    \"\"\"Return every tensor name indexed by the model, in model-tensor order.\"\"\"\n    return list(model.model_tensors.keys())\n```\n\nExample usage:\n\n```python\nmodel = SomeModelClass(\n    dir_model=Path(\"...\"),\n    ftype=gguf.LlamaFileType.GUESSED,\n    fname_out=Path(\"model.gguf\"),\n)\n\nfor name in list_tensor_names(model):\n    print(name)\n```\n\nIf you want to include tensors that are only produced later by `modify_tensors()` or quantization dequantization, use this version instead:\n\n```python\ndef list_all_tensor_names(model: ModelBase) -> list[str]:\n    \"\"\"Return tensor names from indexing plus tensors produced by tensor transforms.\"\"\"\n    names = set(model.model_tensors.keys())\n\n    for name, gen in model.model",
+          "warm_prompt_ms": 999.89,
+          "cold_prompt_ms": 6158.66
+        },
+        {
+          "variant": 1,
+          "L": 7039,
+          "warm_prompt_n": 1028,
+          "warm_cache_n": 6011,
+          "cold_prompt_n": 7039,
+          "cold_cache_n": 0,
+          "aligned": true,
+          "warm_draft": [
+            null,
+            null
+          ],
+          "cold_draft": [
+            null,
+            null
+          ],
+          "identical": true,
+          "first_diff_char": null,
+          "warm_sha": "76498f09c9c8",
+          "cold_sha": "76498f09c9c8",
+          "instr": "Write a function that counts the model classes registered in this file. ok ok ok",
+          "warm_instr": "Summarise this file. (warm-up 1) ok ok ok ok",
+          "warm_text": "```python\ndef count_model_classes(path: str | os.PathLike[str]) -> int:\n    \"\"\"Count classes that inherit from ModelBase in the given Python file.\"\"\"\n    with open(path, \"r\", encoding=\"utf-8\") as f:\n        tree = ast.parse(f.read(), filename=path)\n\n    count = 0\n\n    for node in tree.body:\n        if not isinstance(node, ast.ClassDef):\n            continue\n\n        for base in node.bases:\n            if isinstance(base, ast.Name) and base.id == \"ModelBase\":\n                count += 1\n                break\n\n    return count\n```\n\nExample usage:\n\n```python\nif __name__ == \"__main__\":\n    print(count_model_classes(\"your_file.py\"))\n```\n\nThis counts only direct subclasses of `ModelBase`, not indirect subclasses. If you want to count indirect subclasses too, use this version:",
+          "cold_text": "```python\ndef count_model_classes(path: str | os.PathLike[str]) -> int:\n    \"\"\"Count classes that inherit from ModelBase in the given Python file.\"\"\"\n    with open(path, \"r\", encoding=\"utf-8\") as f:\n        tree = ast.parse(f.read(), filename=path)\n\n    count = 0\n\n    for node in tree.body:\n        if not isinstance(node, ast.ClassDef):\n            continue\n\n        for base in node.bases:\n            if isinstance(base, ast.Name) and base.id == \"ModelBase\":\n                count += 1\n                break\n\n    return count\n```\n\nExample usage:\n\n```python\nif __name__ == \"__main__\":\n    print(count_model_classes(\"your_file.py\"))\n```\n\nThis counts only direct subclasses of `ModelBase`, not indirect subclasses. If you want to count indirect subclasses too, use this version:",
+          "warm_prompt_ms": 1002.97,
+          "cold_prompt_ms": 6165.72
+        },
+        {
+          "variant": 2,
+          "L": 7039,
+          "warm_prompt_n": 1028,
+          "warm_cache_n": 6011,
+          "cold_prompt_n": 7039,
+          "cold_cache_n": 0,
+          "aligned": true,
+          "warm_draft": [
+            null,
+            null
+          ],
+          "cold_draft": [
+            null,
+            null
+          ],
+          "identical": true,
+          "first_diff_char": null,
+          "warm_sha": "16577469a2ea",
+          "cold_sha": "16577469a2ea",
+          "instr": "Write a function that finds the longest method in this file. ok ok ok ok",
+          "warm_instr": "Summarise this file. (warm-up 2) ok ok ok ok",
+          "warm_text": "```python\nimport ast\nfrom pathlib import Path\n\n\ndef find_longest_method(path: str | Path) -> tuple[str, int, ast.FunctionDef | ast.AsyncFunctionDef]:\n    \"\"\"\n    Return the longest method in a Python file.\n\n    Length is measured by the number of AST nodes in the method body.\n    \"\"\"\n    source_path = Path(path)\n    tree = ast.parse(source_path.read_text(encoding=\"utf-8\"), filename=str(source_path))\n\n    methods = [\n        node\n        for node in ast.walk(tree)\n        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))\n    ]\n\n    if not methods:\n        raise ValueError(f\"No methods found in {source_path}\")\n\n    longest = max(\n        methods,\n        key=lambda node: sum(1 for _ in ast.walk(node.body)),\n    )\n\n    return (\n        longest.name",
+          "cold_text": "```python\nimport ast\nfrom pathlib import Path\n\n\ndef find_longest_method(path: str | Path) -> tuple[str, int, ast.FunctionDef | ast.AsyncFunctionDef]:\n    \"\"\"\n    Return the longest method in a Python file.\n\n    Length is measured by the number of AST nodes in the method body.\n    \"\"\"\n    source_path = Path(path)\n    tree = ast.parse(source_path.read_text(encoding=\"utf-8\"), filename=str(source_path))\n\n    methods = [\n        node\n        for node in ast.walk(tree)\n        if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))\n    ]\n\n    if not methods:\n        raise ValueError(f\"No methods found in {source_path}\")\n\n    longest = max(\n        methods,\n        key=lambda node: sum(1 for _ in ast.walk(node.body)),\n    )\n\n    return (\n        longest.name",
+          "warm_prompt_ms": 1003.48,
+          "cold_prompt_ms": 6169.04
+        },
+        {
+          "variant": 3,
+          "L": 7039,
+          "warm_prompt_n": 1028,
+          "warm_cache_n": 6011,
+          "cold_prompt_n": 7039,
+          "cold_cache_n": 0,
+          "aligned": true,
+          "warm_draft": [
+            null,
+            null
+          ],
+          "cold_draft": [
+            null,
+            null
+          ],
+          "identical": true,
+          "first_diff_char": null,
+          "warm_sha": "a8545d39788d",
+          "cold_sha": "a8545d39788d",
+          "instr": "Write a function that returns every regular expression used in this file. ok ok ok",
+          "warm_instr": "Summarise this file. (warm-up 3) ok ok ok ok",
+          "warm_text": "```python\nimport re\n\ndef extract_regular_expressions(source: str) -> list[str]:\n    \"\"\"\n    Return every regular expression string literal used in the given Python source.\n\n    This is a lexical scanner, not a full AST parser, so it handles normal string\n    literals but does not evaluate them.\n    \"\"\"\n    expressions: list[str] = []\n    i = 0\n    n = len(source)\n\n    while i < n:\n        if source[i] not in \"'\\\"\":\n            i += 1\n            continue\n\n        quote = source[i]\n        i += 1\n        chars: list[str] = []\n\n        while i < n:\n            ch = source[i]\n\n            if ch == \"\\\\\":\n                if i + 1 >= n:\n                    break\n                chars.append(ch)\n                chars.append(source[i + 1])\n                i",
+          "cold_text": "```python\nimport re\n\ndef extract_regular_expressions(source: str) -> list[str]:\n    \"\"\"\n    Return every regular expression string literal used in the given Python source.\n\n    This is a lexical scanner, not a full AST parser, so it handles normal string\n    literals but does not evaluate them.\n    \"\"\"\n    expressions: list[str] = []\n    i = 0\n    n = len(source)\n\n    while i < n:\n        if source[i] not in \"'\\\"\":\n            i += 1\n            continue\n\n        quote = source[i]\n        i += 1\n        chars: list[str] = []\n\n        while i < n:\n            ch = source[i]\n\n            if ch == \"\\\\\":\n                if i + 1 >= n:\n                    break\n                chars.append(ch)\n                chars.append(source[i + 1])\n                i",
+          "warm_prompt_ms": 1003.44,
+          "cold_prompt_ms": 6170.8
+        },
+        {
+          "variant": 4,
+          "L": 7039,
+          "warm_prompt_n": 1028,
+          "warm_cache_n": 6011,
+          "cold_prompt_n": 7039,
+          "cold_cache_n": 0,
+          "aligned": true,
+          "warm_draft": [
+            null,
+            null
+          ],
+          "cold_draft": [
+            null,
+            null
+          ],
+          "identical": true,
+          "first_diff_char": null,
+          "warm_sha": "af97b2dde4e2",
+          "cold_sha": "af97b2dde4e2",
+          "instr": "Write a function that maps each class in this file to its base classes. ok",
+          "warm_instr": "Summarise this file. (warm-up 4) ok ok ok ok",
+          "warm_text": "```python\ndef map_classes_to_base_classes(source: str) -> dict[str, list[str]]:\n    \"\"\"\n    Return a mapping of class names to their direct base class names.\n\n    Example:\n        class A(B): pass\n        class C(A): pass\n    returns:\n        {\"A\": [\"B\"], \"C\": [\"A\"]}\n    \"\"\"\n    import ast\n\n    tree = ast.parse(source)\n    result: dict[str, list[str]] = {}\n\n    for node in tree.body:\n        if isinstance(node, ast.ClassDef):\n            bases = [ast.unparse(base) for base in node.bases]\n            result[node.name] = bases\n\n    return result\n```\n\nUsage:\n\n```python\nsource = open(\"file.py\", \"r\", encoding=\"utf-8\").read()\nprint(map_classes_to_base_classes(source))\n```\n\nFor the provided file",
+          "cold_text": "```python\ndef map_classes_to_base_classes(source: str) -> dict[str, list[str]]:\n    \"\"\"\n    Return a mapping of class names to their direct base class names.\n\n    Example:\n        class A(B): pass\n        class C(A): pass\n    returns:\n        {\"A\": [\"B\"], \"C\": [\"A\"]}\n    \"\"\"\n    import ast\n\n    tree = ast.parse(source)\n    result: dict[str, list[str]] = {}\n\n    for node in tree.body:\n        if isinstance(node, ast.ClassDef):\n            bases = [ast.unparse(base) for base in node.bases]\n            result[node.name] = bases\n\n    return result\n```\n\nUsage:\n\n```python\nsource = open(\"file.py\", \"r\", encoding=\"utf-8\").read()\nprint(map_classes_to_base_classes(source))\n```\n\nFor the provided file",
+          "warm_prompt_ms": 1005.61,
+          "cold_prompt_ms": 6177.63
+        }
+      ],
+      "result": "PASS"
+    },
+    {
+      "label": "n-tools-q106",
+      "passed": 6,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": false,
+        "nested-object|think=True": false,
+        "enum|think=True": false,
+        "correct-decline|think=True": false,
+        "multi-turn|think=True": false,
+        "streaming|think=True": false,
+        "parallel|think=True": false,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": false,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-vision-q106-faon",
+      "fa": "on",
+      "mtp": false,
+      "expected": "red,blue,circle,square",
+      "answer": "The image shows two shapes: a red circle on the left and a blue square on the right.",
+      "hits": [
+        "red",
+        "blue",
+        "circle",
+        "square"
+      ],
+      "error": null,
+      "server_died": false,
+      "server_log_errors": [],
+      "result": "PASS"
+    },
+    {
+      "label": "n-vision-q106-faoff",
+      "fa": "off",
+      "mtp": false,
+      "expected": "red,blue,circle,square",
+      "answer": "The image shows two simple shapes:\n\n- A **red circle** on the left.\n- A **blue square** on the right.",
+      "hits": [
+        "red",
+        "blue",
+        "circle",
+        "square"
+      ],
+      "error": null,
+      "server_died": false,
+      "server_log_errors": [],
+      "result": "PASS"
+    },
+    {
+      "label": "n-tools-q106-tpl",
+      "passed": 14,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": true,
+        "nested-object|think=True": true,
+        "enum|think=True": true,
+        "correct-decline|think=True": true,
+        "multi-turn|think=True": true,
+        "streaming|think=True": true,
+        "parallel|think=True": true,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": true,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-tools-q106-tpl-medium",
+      "passed": 7,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": false,
+        "nested-object|think=True": false,
+        "enum|think=True": false,
+        "correct-decline|think=True": false,
+        "multi-turn|think=True": false,
+        "streaming|think=True": false,
+        "parallel|think=True": false,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": true,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-tools-q106-c1",
+      "passed": 13,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": true,
+        "nested-object|think=True": false,
+        "enum|think=True": true,
+        "correct-decline|think=True": true,
+        "multi-turn|think=True": true,
+        "streaming|think=True": true,
+        "parallel|think=True": true,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": true,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-vision-q106-c1-faon",
+      "fa": "on",
+      "mtp": false,
+      "expected": "red,blue,circle,square",
+      "answer": "The image shows two shapes: a red circle on the left and a blue square on the right.",
+      "hits": [
+        "red",
+        "blue",
+        "circle",
+        "square"
+      ],
+      "error": null,
+      "server_died": false,
+      "server_log_errors": [],
+      "result": "PASS"
+    },
+    {
+      "label": "n-tools-q106-roff",
+      "passed": 13,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": true,
+        "nested-object|think=True": false,
+        "enum|think=True": true,
+        "correct-decline|think=True": true,
+        "multi-turn|think=True": true,
+        "streaming|think=True": true,
+        "parallel|think=True": true,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": true,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-tools-q106-roff-r2",
+      "passed": 13,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": true,
+        "nested-object|think=True": true,
+        "enum|think=True": true,
+        "correct-decline|think=True": true,
+        "multi-turn|think=True": true,
+        "streaming|think=True": true,
+        "parallel|think=True": false,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": true,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-tools-q106-roff-r3",
+      "passed": 14,
+      "total": 14,
+      "detail": {
+        "multi-arg|think=True": true,
+        "nested-object|think=True": true,
+        "enum|think=True": true,
+        "correct-decline|think=True": true,
+        "multi-turn|think=True": true,
+        "streaming|think=True": true,
+        "parallel|think=True": true,
+        "multi-arg|think=False": true,
+        "nested-object|think=False": true,
+        "enum|think=False": true,
+        "correct-decline|think=False": true,
+        "multi-turn|think=False": true,
+        "streaming|think=False": true,
+        "parallel|think=False": true
+      }
+    },
+    {
+      "label": "n-vision-q106-roff-faon",
+      "fa": "on",
+      "mtp": false,
+      "expected": "red,blue,circle,square",
+      "answer": "The image shows two shapes: a red circle on the left and a blue square on the right.",
+      "hits": [
+        "red",
+        "blue",
+        "circle",
+        "square"
+      ],
+      "error": null,
+      "server_died": false,
+      "server_log_errors": [],
+      "result": "PASS"
+    }
+  ],
+  "sizing": [
+    {
+      "label": "strix-lean",
+      "ctx": 65536,
+      "avail_before": 122.34,
+      "footprint_loaded_gib": 21.11,
+      "footprint_after_8k_gib": 21.29
+    },
+    {
+      "label": "strix-lean",
+      "ctx": 262144,
+      "avail_before": 122.15,
+      "footprint_loaded_gib": 24.36,
+      "footprint_after_8k_gib": 24.52
+    }
+  ],
+  "n_ubatch": 1024,
+  "template_fix": {
+    "file": "chat_template_enable_thinking.jinja",
+    "size_bytes": 7895,
+    "sha256": "9183c7ba8510fb9628edd2265a4cad8c02b3aecec3cbdf65620f47494e75836d",
+    "source_sha256": "f1753536417ee87cded4bc5017354eb2123c079d20c634403a3d808b2ba3cc5b",
+    "shim": "{%- if reasoning_effort is not defined and enable_thinking is defined %}{%- set reasoning_effort = 'high' if enable_thinking else 'none' %}{%- endif %}\n",
+    "server_flags": [
+      "--chat-template-file",
+      "chat_template_enable_thinking.jinja",
+      "--reasoning",
+      "off"
+    ],
+    "probes_roff": {
+      "no-kwargs|correct-decline": {
+        "content": "391",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "no-kwargs|single-word": {
+        "content": "ready",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "no-kwargs|multi-arg": {
+        "content": "",
+        "reasoning_len": 0,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": []
+      },
+      "enable_thinking=false|correct-decline": {
+        "content": "391",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "enable_thinking=false|single-word": {
+        "content": "ready",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "enable_thinking=false|multi-arg": {
+        "content": "",
+        "reasoning_len": 0,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": []
+      },
+      "reasoning_effort=high|correct-decline": {
+        "content": "We need answer directly. 391.\n</think>\n\n391",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": [
+          "</think>"
+        ]
+      },
+      "reasoning_effort=high|single-word": {
+        "content": "We need need output exactly ready.\n</think>\n\nready",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": [
+          "</think>"
+        ]
+      },
+      "reasoning_effort=high|multi-arg": {
+        "content": "We need need tool. Current weather Paris celsius.\n</think>\n\n",
+        "reasoning_len": 0,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": [
+          "</think>"
+        ]
+      },
+      "reasoning_effort=medium|correct-decline": {
+        "content": "\n\n</think>\n\n391",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": [
+          "</think>"
+        ]
+      },
+      "reasoning_effort=medium|single-word": {
+        "content": "\n\n</think>\n\nready",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": [
+          "</think>"
+        ]
+      },
+      "reasoning_effort=medium|multi-arg": {
+        "content": "\n\n</think>\n\n",
+        "reasoning_len": 0,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": [
+          "</think>"
+        ]
+      },
+      "reasoning_effort=none|correct-decline": {
+        "content": "391",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "reasoning_effort=none|single-word": {
+        "content": "ready",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "reasoning_effort=none|multi-arg": {
+        "content": "",
+        "reasoning_len": 0,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": []
+      }
+    },
+    "probes_high_default_on": {
+      "no-kwargs|correct-decline": {
+        "content": "391",
+        "reasoning_len": 30,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "no-kwargs|multi-arg": {
+        "content": "",
+        "reasoning_len": 50,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": []
+      },
+      "reasoning_effort=medium|correct-decline": {
+        "content": "391",
+        "reasoning_len": 0,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "reasoning_effort=medium|multi-arg": {
+        "content": "",
+        "reasoning_len": 0,
+        "tool_calls": [
+          "get_weather"
+        ],
+        "leaks": []
+      },
+      "reasoning_effort=none|correct-decline": {
+        "content": "",
+        "reasoning_len": 3,
+        "tool_calls": [],
+        "leaks": []
+      },
+      "reasoning_effort=none|multi-arg": {
+        "content": "",
+        "reasoning_len": 133,
+        "tool_calls": [],
+        "leaks": []
+      }
+    },
+    "medium_mapping_label": "n-tools-q106-tpl-medium"
+  },
+  "tools_diag": {
+    "stock_on_replies": 7,
+    "stock_on_leaks": 7,
+    "stock_on_reasoning_extracted": 0,
+    "nested_off_attempts": 4,
+    "nested_off_http500": 2,
+    "gate_http500_logged": true,
+    "flag_probes": {
+      "default": {
+        "leaks": 3,
+        "n": 3
+      },
+      "fmt-deepseek": {
+        "leaks": 3,
+        "n": 3
+      },
+      "srv-kwargs-high": {
+        "leaks": 3,
+        "n": 3
+      },
+      "reasoning-on": {
+        "leaks": 3,
+        "n": 3
+      },
+      "tpl-enable-thinking": {
+        "leaks": 0,
+        "n": 3
+      }
+    }
+  },
+  "seats": {
+    "max1-nex-fast": {
+      "unit": "max1-nex-fast",
+      "port": 8097,
+      "load_s": 25,
+      "time": "2026-09-17T01:21:10Z",
+      "direct_reply": "ready",
+      "direct_tg": 41.41386950489719,
+      "default_reply": "ready",
+      "default_reasoning_len": 0,
+      "default_leak": false,
+      "thinking_reply": "",
+      "thinking_reasoning_len": 5,
+      "thinking_leak": false,
+      "gateway_model": "nex-n2.5-mini-fast@max1",
+      "gateway_reply": "ready",
+      "result": "PASS"
+    },
+    "max1-nex-fast-imat": {
+      "unit": "max1-nex-fast-imat",
+      "port": 8098,
+      "load_s": 25,
+      "time": "2026-09-17T01:21:43Z",
+      "direct_reply": "ready",
+      "direct_tg": 41.54290343352097,
+      "default_reply": "ready",
+      "default_reasoning_len": 0,
+      "default_leak": false,
+      "thinking_reply": "",
+      "thinking_reasoning_len": 5,
+      "thinking_leak": false,
+      "gateway_model": "nex-n2.5-mini-fast-imatrix@max1",
+      "gateway_reply": "ready",
+      "result": "PASS"
+    }
+  },
+  "measured": "2026-09-16",
+  "measured_range": [
+    "2026-09-16",
+    "2026-09-17"
+  ]
+}

recipe/templates/chat_template_enable_thinking_medium.jinja

@@ -0,0 +1,155 @@
+{%- if reasoning_effort is not defined and enable_thinking is defined %}{%- set reasoning_effort = 'medium' if enable_thinking else 'none' %}{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if reasoning_effort is not defined or reasoning_effort is none %}
+        {{- '<think>' }}
+    {%- elif reasoning_effort == 'none' %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- elif reasoning_effort == 'high' %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>' }}
+    {%- endif %}
+{%- endif %}

recipe/templates/chat_template_enable_thinking_v2.jinja

@@ -0,0 +1,155 @@
+{%- if enable_thinking is defined and not enable_thinking %}{%- set reasoning_effort = 'none' %}{%- elif reasoning_effort is not defined and enable_thinking is defined %}{%- set reasoning_effort = 'high' %}{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if reasoning_effort is not defined or reasoning_effort is none %}
+        {{- '<think>' }}
+    {%- elif reasoning_effort == 'none' %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- elif reasoning_effort == 'high' %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>' }}
+    {%- endif %}
+{%- endif %}