diff --git "a/recipe/results_summary.json" "b/recipe/results_summary.json" new file mode 100644--- /dev/null +++ "b/recipe/results_summary.json" @@ -0,0 +1,2297 @@ +{ + "layout": "in-file MTP (one GGUF per tier; the MTP block shares the trunk's embedding and output matrices)", + "model": "Agnes-3.0-Flash-Preview", + "model_repo_std": "kingjones777/Agnes-3.0-Flash-Preview-MTP-ROCmFP4-GGUF", + "model_repo_imat": "kingjones777/Agnes-3.0-Flash-Preview-MTP-ROCmFP4-imatrix-GGUF", + "source": { + "repo": "Agnes-AI/Agnes-3.0-Flash", + "revision": "891ce4f9ffb89b22888aa7fcc2bb2f3618867684" + }, + "bf16": { + "ppl": 6.3314, + "ppl_err": 0.07632, + "ppl_paired": 6.329637, + "ppl_paired_err": 0.076239, + "afternoon_ppl": 6.3294, + "afternoon_ppl_paired": 6.32763, + "chunks": 40, + "n_ctx": 2048, + "scored_tokens": 40920, + "trunk_size_bytes": 64427096032, + "with_mtp_size_bytes": 65339462560 + }, + "binary": { + "repo": "https://github.com/charlie12345/ROCmFPX", + "commit": "d3ca53726109b46fd5cd6fa42f688c707079b44e", + "commit_in_official_main": "yes (ROCmFPX/ROCmFPX main is ahead of it, 0 behind; checked via the GitHub compare API 2026-09-16)", + "original": { + "dir": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin", + "sha256": { + "llama-server": "d077b98f3f04df4b482a5826c8922e39cda3aa8d2e2b2fbe1870f36d245cebe7", + "llama-quantize": "1559189968a30781f837e4d25ca3cdf26c8eb0351023675406535eddcbe0db13", + "llama-imatrix": "bfdce5c21a01bdc8043d69335b74a7b68b8335ae989991a1e10e21d5faeaeb8e", + "llama-perplexity": "c2175925bb0c61df89ff787845d8e03a35403f684be52719ac2987b8ca636a53" + } + }, + "patched": { + "dir": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "sha256": { + "llama-server": "14e9f1299240c891478c820b8c20811ca5721044d2e63e23799eaa9be26b30bd" + }, + "patch_file": "recipe/patches/mtp-prompt-cache-fix.patch", + "patch_sha256": "412a0173f60faa937018f0153efe9b5d9230c1606735502485e7e520fffb68dd", + "patch_source": "kingjones777/Qwen3.8-27B-ROCmFP4-STRIX-MTP-GGUF (patches/mtp-prompt-cache-fix.patch)", + "applied": "cleanly (git apply --check)", + "diffstat": [ + "common/common.cpp | 6 ++++", + "common/common.h | 5 +++", + "tools/server/server-context.cpp | 72 +++++++++++++++++++++++++++++++++--------", + "tools/server/server-task.cpp | 33 +++++++++++++++----" + ], + "built_target": "llama-server only (same CMake cache as the original build)" + } + }, + "vision_probe": { + "width": 448, + "height": 448, + "bytes": 1945, + "sha256": "53924b4715db9757d5b70614dd0e560b8f7de686ff3df9792721d157e2594f5e" + }, + "imatrix": { + "chunks": 129, + "n_ctx": 512, + "calibration": "bartowski calibration_datav3.txt", + "file": "Agnes-3.0-Flash-Preview.imatrix", + "sha256": "eb66b219d8bb057b78b933e5b78042fc47e028415d626690883c8c8f72707a09", + "size_bytes": 15937792 + }, + "tiers": { + "q106": { + "file": "Agnes-3.0-Flash-Preview-MTP-Q4_0-ROCmFP4-STRIX_LEAN.gguf", + "local_file": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "dir": "out-mtp", + "ftype": 106, + "size_bytes": 18061413280, + "graded_file": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "graded_size_bytes": 17790808032, + "quant_mib": 17214.21, + "bpw": 4.42, + "quant_seconds": 310.37778000000003, + "imatrix_entries": null, + "readback": "PASS", + "tensors": 972, + "nextn_tensors": 4, + "output_weight": "Q6_K", + "token_embd": "Q5_K", + "trunk_identity": { + "result": "IDENTICAL", + "shared": 957, + "n_type_changed": 0, + "n_bytes_changed": 0, + "only_in_old": [] + }, + "receipt_vs_afternoon": "MISMATCH", + "quality_measured": true, + "afternoon": { + "ppl": 6.470544, + "ppl_err": 0.078949, + "ppl_ratio": 1.022586, + "kld_mean": 0.040638, + "kld_err": 0.000923, + "kld_p99": 0.418962, + "kld_median": 0.015596, + "same_top_p": 91.96, + "rms_dp": 5.666 + }, + "mtp_block": { + "tensors": 15, + "bytes": 270604288, + "types": { + "attn_k.weight": "Q4_0_ROCMFP4", + "attn_output.weight": "Q4_0_ROCMFP4_FAST", + "attn_q.weight": "Q4_0_ROCMFP4_FAST", + "attn_v.weight": "Q4_0_ROCMFP4", + "ffn_down.weight": "Q4_0_ROCMFP4_FAST", + "ffn_gate.weight": "Q4_0_ROCMFP4_FAST", + "ffn_up.weight": "Q4_0_ROCMFP4_FAST", + "nextn.eh_proj.weight": "Q8_0" + } + }, + "ppl": 6.487656, + "ppl_err": 0.079193, + "ppl_ratio": 1.024965, + "kld_mean": 0.043788, + "kld_err": 0.001045, + "kld_p99": 0.426193, + "kld_median": 0.018434, + "same_top_p": 91.293, + "rms_dp": 5.96 + }, + "q102": { + "file": "Agnes-3.0-Flash-Preview-MTP-Q4_0-ROCmFP4-COHERENT.gguf", + "local_file": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_COHERENT.gguf", + "dir": "out-mtp", + "ftype": 102, + "size_bytes": 19075040160, + "graded_file": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_COHERENT.gguf", + "graded_size_bytes": 18792146912, + "quant_mib": 18180.88, + "bpw": 4.67, + "quant_seconds": 443.50016, + "imatrix_entries": null, + "readback": "PASS", + "tensors": 972, + "nextn_tensors": 4, + "output_weight": "Q6_K", + "token_embd": "Q6_K", + "trunk_identity": { + "result": "IDENTICAL", + "shared": 957, + "n_type_changed": 0, + "n_bytes_changed": 0, + "only_in_old": [] + }, + "receipt_vs_afternoon": "MISMATCH", + "quality_measured": true, + "afternoon": { + "ppl": 6.411907, + "ppl_err": 0.077742, + "ppl_ratio": 1.013319, + "kld_mean": 0.035442, + "kld_err": 0.000759, + "kld_p99": 0.367319, + "kld_median": 0.013563, + "same_top_p": 92.314, + "rms_dp": 5.297 + }, + "mtp_block": { + "tensors": 15, + "bytes": 282892288, + "types": { + "attn_k.weight": "Q4_0_ROCMFP4", + "attn_output.weight": "Q4_0_ROCMFP4", + "attn_q.weight": "Q4_0_ROCMFP4", + "attn_v.weight": "Q4_0_ROCMFP4", + "ffn_down.weight": "Q4_0_ROCMFP4", + "ffn_gate.weight": "Q4_0_ROCMFP4", + "ffn_up.weight": "Q4_0_ROCMFP4", + "nextn.eh_proj.weight": "Q8_0" + } + }, + "ppl": 6.429036, + "ppl_err": 0.077982, + "ppl_ratio": 1.015704, + "kld_mean": 0.038486, + "kld_err": 0.000887, + "kld_p99": 0.37653, + "kld_median": 0.016407, + "same_top_p": 91.386, + "rms_dp": 5.601 + }, + "q115": { + "file": "Agnes-3.0-Flash-Preview-MTP-Q8_0-ROCmFPX-AGENT.gguf", + "local_file": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX_AGENT.gguf", + "dir": "out-mtp", + "ftype": 115, + "size_bytes": 34260171680, + "graded_file": "Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX_AGENT.gguf", + "graded_size_bytes": 33777065952, + "quant_mib": 32662.55, + "bpw": 8.39, + "quant_seconds": 82.62498, + "imatrix_entries": null, + "readback": "PASS", + "tensors": 972, + "nextn_tensors": 4, + "output_weight": "Q8_0", + "token_embd": "Q8_0", + "trunk_identity": { + "result": "IDENTICAL", + "shared": 957, + "n_type_changed": 0, + "n_bytes_changed": 0, + "only_in_old": [] + }, + "receipt_vs_afternoon": "MATCH", + "quality_measured": true, + "afternoon": { + "ppl": 6.301422, + "ppl_err": 0.075776, + "ppl_ratio": 0.995858, + "kld_mean": 0.007674, + "kld_err": 0.000819, + "kld_p99": 0.055398, + "kld_median": 0.001653, + "same_top_p": 96.733, + "rms_dp": 2.412 + }, + "mtp_block": { + "tensors": 15, + "bytes": 483104768, + "types": { + "attn_k.weight": "Q8_0", + "attn_output.weight": "Q8_0", + "attn_q.weight": "Q8_0", + "attn_v.weight": "Q8_0", + "ffn_down.weight": "Q8_0", + "ffn_gate.weight": "Q8_0", + "ffn_up.weight": "Q8_0", + "nextn.eh_proj.weight": "Q8_0_ROCMFPX" + } + }, + "ppl": 6.301422, + "ppl_err": 0.075776, + "ppl_ratio": 0.995542, + "kld_mean": 0.007418, + "kld_err": 0.000652, + "kld_p99": 0.056239, + "kld_median": 0.001653, + "same_top_p": 96.735, + "rms_dp": 2.444 + }, + "q111": { + "file": "Agnes-3.0-Flash-Preview-MTP-Q8_0-ROCmFPX.gguf", + "local_file": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX.gguf", + "dir": "out-mtp", + "ftype": 111, + "size_bytes": 33781267360, + "graded_file": "Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX.gguf", + "graded_size_bytes": 33310777312, + "quant_mib": 32205.84, + "bpw": 8.27, + "quant_seconds": 97.95607000000001, + "imatrix_entries": null, + "readback": "PASS", + "tensors": 972, + "nextn_tensors": 4, + "output_weight": "Q8_0", + "token_embd": "Q8_0", + "trunk_identity": { + "result": "IDENTICAL", + "shared": 957, + "n_type_changed": 0, + "n_bytes_changed": 0, + "only_in_old": [] + }, + "receipt_vs_afternoon": "MATCH", + "quality_measured": true, + "afternoon": { + "ppl": 6.317616, + "ppl_err": 0.076081, + "ppl_ratio": 0.998417, + "kld_mean": 0.010477, + "kld_err": 0.000749, + "kld_p99": 0.085791, + "kld_median": 0.003005, + "same_top_p": 96.026, + "rms_dp": 2.987 + }, + "mtp_block": { + "tensors": 15, + "bytes": 470489088, + "types": { + "attn_k.weight": "Q8_0_ROCMFPX", + "attn_output.weight": "Q8_0_ROCMFPX", + "attn_q.weight": "Q8_0_ROCMFPX", + "attn_v.weight": "Q8_0_ROCMFPX", + "ffn_down.weight": "Q8_0_ROCMFPX", + "ffn_gate.weight": "Q8_0_ROCMFPX", + "ffn_up.weight": "Q8_0_ROCMFPX", + "nextn.eh_proj.weight": "Q8_0_ROCMFPX" + } + }, + "ppl": 6.317616, + "ppl_err": 0.076081, + "ppl_ratio": 0.998101, + "kld_mean": 0.01081, + "kld_err": 0.000803, + "kld_p99": 0.08599, + "kld_median": 0.00301, + "same_top_p": 96.043, + "rms_dp": 2.996 + }, + "q106i": { + "file": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0-ROCmFP4-STRIX_LEAN.gguf", + "local_file": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "dir": "out-mtp-imat", + "ftype": 106, + "size_bytes": 18061413504, + "graded_file": "Agnes-3.0-Flash-Preview-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "graded_size_bytes": 17790808256, + "quant_mib": 17214.21, + "bpw": 4.42, + "quant_seconds": 430.50327000000004, + "imatrix_entries": 558, + "readback": "PASS", + "tensors": 972, + "nextn_tensors": 4, + "output_weight": "Q6_K", + "token_embd": "Q5_K", + "trunk_identity": { + "result": "IDENTICAL", + "shared": 957, + "n_type_changed": 0, + "n_bytes_changed": 0, + "only_in_old": [] + }, + "receipt_vs_afternoon": "MISMATCH", + "quality_measured": true, + "afternoon": { + "ppl": 6.413088, + "ppl_err": 0.077536, + "ppl_ratio": 1.013506, + "kld_mean": 0.035791, + "kld_err": 0.001001, + "kld_p99": 0.381905, + "kld_median": 0.012275, + "same_top_p": 92.309, + "rms_dp": 5.328 + }, + "mtp_block": { + "tensors": 15, + "bytes": 270604288, + "types": { + "attn_k.weight": "Q4_0_ROCMFP4", + "attn_output.weight": "Q4_0_ROCMFP4_FAST", + "attn_q.weight": "Q4_0_ROCMFP4_FAST", + "attn_v.weight": "Q4_0_ROCMFP4", + "ffn_down.weight": "Q4_0_ROCMFP4_FAST", + "ffn_gate.weight": "Q4_0_ROCMFP4_FAST", + "ffn_up.weight": "Q4_0_ROCMFP4_FAST", + "nextn.eh_proj.weight": "Q8_0" + } + }, + "ppl": 6.426891, + "ppl_err": 0.07772, + "ppl_ratio": 1.015365, + "kld_mean": 0.03872, + "kld_err": 0.000987, + "kld_p99": 0.386349, + "kld_median": 0.015312, + "same_top_p": 91.522, + "rms_dp": 5.622, + "same_tensor_types_as_standard": true, + "file_size_delta_bytes": 224, + "differs_from_standard": true + }, + "q102i": { + "file": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0-ROCmFP4-COHERENT.gguf", + "local_file": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "dir": "out-mtp-imat", + "ftype": 102, + "size_bytes": 19075040384, + "graded_file": "Agnes-3.0-Flash-Preview-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "graded_size_bytes": 18792147136, + "quant_mib": 18180.88, + "bpw": 4.67, + "quant_seconds": 443.67743, + "imatrix_entries": 558, + "readback": "PASS", + "tensors": 972, + "nextn_tensors": 4, + "output_weight": "Q6_K", + "token_embd": "Q6_K", + "trunk_identity": { + "result": "IDENTICAL", + "shared": 957, + "n_type_changed": 0, + "n_bytes_changed": 0, + "only_in_old": [] + }, + "receipt_vs_afternoon": "MISMATCH", + "quality_measured": true, + "afternoon": { + "ppl": 6.388763, + "ppl_err": 0.077065, + "ppl_ratio": 1.009661, + "kld_mean": 0.029416, + "kld_err": 0.0008, + "kld_p99": 0.30392, + "kld_median": 0.010466, + "same_top_p": 92.972, + "rms_dp": 4.919 + }, + "mtp_block": { + "tensors": 15, + "bytes": 282892288, + "types": { + "attn_k.weight": "Q4_0_ROCMFP4", + "attn_output.weight": "Q4_0_ROCMFP4", + "attn_q.weight": "Q4_0_ROCMFP4", + "attn_v.weight": "Q4_0_ROCMFP4", + "ffn_down.weight": "Q4_0_ROCMFP4", + "ffn_gate.weight": "Q4_0_ROCMFP4", + "ffn_up.weight": "Q4_0_ROCMFP4", + "nextn.eh_proj.weight": "Q8_0" + } + }, + "ppl": 6.400447, + "ppl_err": 0.077237, + "ppl_ratio": 1.011187, + "kld_mean": 0.032188, + "kld_err": 0.000766, + "kld_p99": 0.313102, + "kld_median": 0.013397, + "same_top_p": 92.107, + "rms_dp": 5.252, + "same_tensor_types_as_standard": true, + "file_size_delta_bytes": 224, + "differs_from_standard": true + } + }, + "separate_head_replaced": { + "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf": { + "size_bytes": 2051963392, + "tensors": 18, + "tensor_bytes": 2041018368, + "vocab_matrices_bytes": 1758105600, + "vocab_matrix_types": { + "output.weight": "Q6_K", + "token_embd.weight": "Q4_0" + } + }, + "mtp-Agnes-3.0-Flash-Preview-Q8_0.gguf": { + "size_bytes": 3197430272, + "tensors": 18, + "tensor_bytes": 3186485248, + "vocab_matrices_bytes": 2701721600, + "vocab_matrix_types": { + "output.weight": "Q8_0", + "token_embd.weight": "Q8_0" + } + } + }, + "aux": { + "mmproj-Agnes-3.0-Flash-Preview-BF16.gguf": 931146048 + }, + "bench": [ + { + "label": "p-nospec-rocm", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": null, + "nmax": null, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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": 11.04, + "tg_min": 11.04, + "tg_max": 11.05, + "pp_median": 278.9, + "prompt_n": 7100, + "accept": null + }, + { + "label": "p-infile-rocm-n3", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 3, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.1, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 26.32, + "tg_min": 26.31, + "tg_max": 26.63, + "pp_median": 261.8, + "prompt_n": 7101, + "accept": 0.797 + }, + { + "label": "p-infile-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.1, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 26.64, + "tg_min": 26.63, + "tg_max": 26.99, + "pp_median": 261.3, + "prompt_n": 7102, + "accept": 0.724 + }, + { + "label": "p-infile-rocm-n5", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 5, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.1, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 5 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 26.11, + "tg_min": 25.62, + "tg_max": 26.11, + "pp_median": 260.4, + "prompt_n": 7097, + "accept": 0.638 + }, + { + "label": "p-infile-strict-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": true, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.1, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0 --spec-mtp-strict-qwen", + "tg_median": 27.02, + "tg_min": 26.66, + "tg_max": 27.03, + "pp_median": 261.2, + "prompt_n": 7097, + "accept": 0.731 + }, + { + "label": "p-nospec-vk", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": null, + "nmax": null, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 20.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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": 11.2, + "tg_min": 11.15, + "tg_max": 11.21, + "pp_median": 200.0, + "prompt_n": 7098, + "accept": null + }, + { + "label": "p-infile-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 4.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 27.31, + "tg_min": 27.19, + "tg_max": 28.18, + "pp_median": 201.2, + "prompt_n": 7096, + "accept": 0.766 + }, + { + "label": "p-infile-strict-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": true, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 4.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0 --spec-mtp-strict-qwen", + "tg_median": 27.86, + "tg_min": 27.44, + "tg_max": 28.21, + "pp_median": 197.8, + "prompt_n": 7100, + "accept": 0.78 + }, + { + "label": "t-q106i-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp-imat/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 26.31, + "tg_min": 25.56, + "tg_max": 26.42, + "pp_median": 260.8, + "prompt_n": 7096, + "accept": 0.697 + }, + { + "label": "t-q106i-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 4.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp-imat/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 25.37, + "tg_min": 25.27, + "tg_max": 25.69, + "pp_median": 197.2, + "prompt_n": 7099, + "accept": 0.684 + }, + { + "label": "t-q102-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_COHERENT.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 24.76, + "tg_min": 23.03, + "tg_max": 26.12, + "pp_median": 231.8, + "prompt_n": 7097, + "accept": 0.66 + }, + { + "label": "t-q102-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_COHERENT.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 23.51, + "tg_min": 23.5, + "tg_max": 24.41, + "pp_median": 197.7, + "prompt_n": 7097, + "accept": 0.653 + }, + { + "label": "t-q102i-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp-imat/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 23.3, + "tg_min": 23.29, + "tg_max": 23.66, + "pp_median": 231.9, + "prompt_n": 7099, + "accept": 0.618 + }, + { + "label": "t-q102i-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 4.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp-imat/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 25.46, + "tg_min": 25.28, + "tg_max": 25.85, + "pp_median": 197.7, + "prompt_n": 7096, + "accept": 0.715 + }, + { + "label": "t-q115-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX_AGENT.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 38.1, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX_AGENT.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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 16.97, + "tg_min": 16.05, + "tg_max": 17.0, + "pp_median": 239.6, + "prompt_n": 7095, + "accept": 0.673 + }, + { + "label": "t-q115-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX_AGENT.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 38.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX_AGENT.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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 14.97, + "tg_min": 14.96, + "tg_max": 15.16, + "pp_median": 211.7, + "prompt_n": 7098, + "accept": 0.622 + }, + { + "label": "t-q111-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 38.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX.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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 16.42, + "tg_min": 16.42, + "tg_max": 16.42, + "pp_median": 229.0, + "prompt_n": 7097, + "accept": 0.642 + }, + { + "label": "t-q111-vk-n4", + "model": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 36.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp/Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX.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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 16.17, + "tg_min": 16.16, + "tg_max": 16.17, + "pp_median": 211.0, + "prompt_n": 7096, + "accept": 0.618 + }, + { + "label": "p-infile-rocm-n4-prose", + "model": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "in-file", + "nmax": 4, + "strict": false, + "bin": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin", + "dev": "ROCm0", + "ctx": 65536, + "workload": "prose", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-mtp-imat/Agnes-3.0-Flash-Preview-MTP-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 --spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 16.57, + "tg_min": 15.54, + "tg_max": 17.54, + "pp_median": 260.0, + "prompt_n": 7450, + "accept": 0.352 + } + ], + "gates": [ + { + "label": "id-nospec-orig", + "reps": 3, + "sha256": [ + "1b6a728e0487efea", + "1b6a728e0487efea", + "1b6a728e0487efea" + ], + "self_consistent": true + }, + { + "label": "id-nospec-patched", + "reps": 3, + "sha256": [ + "1b6a728e0487efea", + "1b6a728e0487efea", + "1b6a728e0487efea" + ], + "self_consistent": true + }, + { + "label": "id-mtp-default", + "reps": 3, + "sha256": [ + "1b6a728e0487efea", + "1b6a728e0487efea", + "1b6a728e0487efea" + ], + "self_consistent": true + }, + { + "label": "id-mtp-strict", + "reps": 3, + "sha256": [ + "1b6a728e0487efea", + "1b6a728e0487efea", + "1b6a728e0487efea" + ], + "self_consistent": true + }, + { + "label": "cmp-id-nospec-patched", + "identical_to_plain": true, + "first_diff_char": null + }, + { + "label": "cmp-id-mtp-default", + "identical_to_plain": true, + "first_diff_char": null + }, + { + "label": "cmp-id-mtp-strict", + "identical_to_plain": true, + "first_diff_char": null + }, + { + "label": "c-orig-infile", + "turn1_prompt_n": 7062, + "turn2_prompt_n": 7067, + "turn2_cache_n": 0, + "reuse_fraction": 0.0, + "result": "FAIL" + }, + { + "label": "c2-patched-infile", + "rows": [ + { + "variant": 0, + "warm_prompt_n": 1028, + "warm_cache_n": 6007, + "cold_prompt_n": 7035, + "cold_cache_n": 0, + "identical": true, + "warm_sha": "99c26cb1ecd9", + "cold_sha": "99c26cb1ecd9", + "warm_prompt_ms": 6756.08, + "cold_prompt_ms": 41181.18 + }, + { + "variant": 1, + "warm_prompt_n": 1029, + "warm_cache_n": 6007, + "cold_prompt_n": 7036, + "cold_cache_n": 0, + "identical": false, + "warm_sha": "ae4446830948", + "cold_sha": "a76e8e8ec803", + "warm_prompt_ms": 6792.08, + "cold_prompt_ms": 38897.38 + }, + { + "variant": 2, + "warm_prompt_n": 1027, + "warm_cache_n": 6008, + "cold_prompt_n": 7035, + "cold_cache_n": 0, + "identical": true, + "warm_sha": "18e167a90d9c", + "cold_sha": "18e167a90d9c", + "warm_prompt_ms": 6209.49, + "cold_prompt_ms": 38506.94 + } + ], + "all_reused": true, + "all_identical": false, + "result": "FAIL" + }, + { + "label": "c2-patched-infile-strict", + "rows": [ + { + "variant": 0, + "warm_prompt_n": 1028, + "warm_cache_n": 6007, + "cold_prompt_n": 7035, + "cold_cache_n": 0, + "identical": true, + "warm_sha": "99c26cb1ecd9", + "cold_sha": "99c26cb1ecd9", + "warm_prompt_ms": 6866.85, + "cold_prompt_ms": 41081.95 + }, + { + "variant": 1, + "warm_prompt_n": 1029, + "warm_cache_n": 6007, + "cold_prompt_n": 7036, + "cold_cache_n": 0, + "identical": false, + "warm_sha": "ae4446830948", + "cold_sha": "a76e8e8ec803", + "warm_prompt_ms": 6883.41, + "cold_prompt_ms": 41200.94 + }, + { + "variant": 2, + "warm_prompt_n": 1027, + "warm_cache_n": 6008, + "cold_prompt_n": 7035, + "cold_cache_n": 0, + "identical": true, + "warm_sha": "18e167a90d9c", + "cold_sha": "18e167a90d9c", + "warm_prompt_ms": 6637.3, + "cold_prompt_ms": 41090.32 + } + ], + "all_reused": true, + "all_identical": false, + "result": "FAIL" + }, + { + "label": "c2-patched-nospec", + "rows": [ + { + "variant": 0, + "warm_prompt_n": 1028, + "warm_cache_n": 6007, + "cold_prompt_n": 7035, + "cold_cache_n": 0, + "identical": true, + "warm_sha": "99c26cb1ecd9", + "cold_sha": "99c26cb1ecd9", + "warm_prompt_ms": 5817.37, + "cold_prompt_ms": 25211.32 + }, + { + "variant": 1, + "warm_prompt_n": 1029, + "warm_cache_n": 6007, + "cold_prompt_n": 7036, + "cold_cache_n": 0, + "identical": false, + "warm_sha": "ae4446830948", + "cold_sha": "a76e8e8ec803", + "warm_prompt_ms": 4166.25, + "cold_prompt_ms": 25266.35 + }, + { + "variant": 2, + "warm_prompt_n": 1027, + "warm_cache_n": 6008, + "cold_prompt_n": 7035, + "cold_cache_n": 0, + "identical": false, + "warm_sha": "edc7c6599ed6", + "cold_sha": "bafffafea107", + "warm_prompt_ms": 3984.65, + "cold_prompt_ms": 25249.11 + } + ], + "all_reused": true, + "all_identical": false, + "result": "FAIL" + }, + { + "label": "c3-patched-infile", + "gate": "cachegate3-aligned", + "n": 10, + "L": 7039, + "expect_cache_n": 6011, + "all_reused": true, + "all_aligned": true, + "identical": 10, + "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": [ + 119, + 286 + ], + "cold_draft": [ + 119, + 286 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "f8e04e49df3c", + "cold_sha": "f8e04e49df3c", + "warm_text": "Based on the code provided, here is a Python function that parses the file and extracts all tensor names.\n\nSince the file contains logic for mapping Hugging Face tensor names to GGUF tensor names (via `tensor_map` and `filter_tensors`), and also defines specific tensor handling in `dequant_model`, the most robust way to \"list every tensor name\" is to extract the **GGUF canonical tensor names** defined in the `gguf` library (which this script imports) and the **source tensor patterns** found in the logic.\n\nHowever, since `gguf.MODEL_TENSORS` is a dictionary defined in the imported `gguf` package (not in this file), we cannot list them directly from this file's text without importing the package.\n\n**But**, we can list the **tensor name patterns and suffixes** explicitly referenced in this file's logic.\n\nHere is a function that scans the file content for tensor", + "cold_text": "Based on the code provided, here is a Python function that parses the file and extracts all tensor names.\n\nSince the file contains logic for mapping Hugging Face tensor names to GGUF tensor names (via `tensor_map` and `filter_tensors`), and also defines specific tensor handling in `dequant_model`, the most robust way to \"list every tensor name\" is to extract the **GGUF canonical tensor names** defined in the `gguf` library (which this script imports) and the **source tensor patterns** found in the logic.\n\nHowever, since `gguf.MODEL_TENSORS` is a dictionary defined in the imported `gguf` package (not in this file), we cannot list them directly from this file's text without importing the package.\n\n**But**, we can list the **tensor name patterns and suffixes** explicitly referenced in this file's logic.\n\nHere is a function that scans the file content for tensor", + "warm_prompt_ms": 4203.46, + "cold_prompt_ms": 26517.26 + }, + { + "variant": 1, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 141, + 197 + ], + "cold_draft": [ + 141, + 197 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "1dc870095e26", + "cold_sha": "1dc870095e26", + "warm_text": "Based on the code provided, here is a function that counts the number of model classes registered in the `ModelBase._model_classes` dictionary.\n\n```python\ndef count_registered_model_classes() -> int:\n \"\"\"\n Counts the total number of model classes registered in ModelBase._model_classes.\n \n Returns:\n int: The total count of registered model classes across all ModelTypes.\n \"\"\"\n total_count = 0\n for model_type, classes_dict in ModelBase._model_classes.items():\n total_count += len(classes_dict)\n return total_count\n```\n\n### Explanation:\n1. **`ModelBase._model_classes`** is a dictionary where the keys are `ModelType` enums (`TEXT`, `MMPROJ`) and the values are dictionaries mapping class names to their corresponding `type[ModelBase]` subclasses.\n2. The function iterates over each `", + "cold_text": "Based on the code provided, here is a function that counts the number of model classes registered in the `ModelBase._model_classes` dictionary.\n\n```python\ndef count_registered_model_classes() -> int:\n \"\"\"\n Counts the total number of model classes registered in ModelBase._model_classes.\n \n Returns:\n int: The total count of registered model classes across all ModelTypes.\n \"\"\"\n total_count = 0\n for model_type, classes_dict in ModelBase._model_classes.items():\n total_count += len(classes_dict)\n return total_count\n```\n\n### Explanation:\n1. **`ModelBase._model_classes`** is a dictionary where the keys are `ModelType` enums (`TEXT`, `MMPROJ`) and the values are dictionaries mapping class names to their corresponding `type[ModelBase]` subclasses.\n2. The function iterates over each `", + "warm_prompt_ms": 4231.79, + "cold_prompt_ms": 26741.07 + }, + { + "variant": 2, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 136, + 214 + ], + "cold_draft": [ + 136, + 214 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "790e4205bfa3", + "cold_sha": "790e4205bfa3", + "warm_text": "To find the longest method in the provided Python file, we need to parse the code, identify all method definitions (functions defined inside classes), and calculate their length based on the number of lines or characters they span.\n\nHere is a robust Python script that performs this task using the `ast` (Abstract Syntax Tree) module. It will:\n1. Parse the source code.\n2. Identify all `FunctionDef` and `AsyncFunctionDef` nodes that are methods (i.e., defined inside a `ClassDef`).\n3. Calculate the length of each method (number of lines in the source code).\n4. Return the name of the longest method.\n\n### Python Script to Find the Longest Method\n\n```python\nimport ast\nimport sys\nfrom pathlib import Path\n\ndef find_longest_method(file_path: str) -> str:\n \"\"\"\n Parses a Python file and finds the longest method", + "cold_text": "To find the longest method in the provided Python file, we need to parse the code, identify all method definitions (functions defined inside classes), and calculate their length based on the number of lines or characters they span.\n\nHere is a robust Python script that performs this task using the `ast` (Abstract Syntax Tree) module. It will:\n1. Parse the source code.\n2. Identify all `FunctionDef` and `AsyncFunctionDef` nodes that are methods (i.e., defined inside a `ClassDef`).\n3. Calculate the length of each method (number of lines in the source code).\n4. Return the name of the longest method.\n\n### Python Script to Find the Longest Method\n\n```python\nimport ast\nimport sys\nfrom pathlib import Path\n\ndef find_longest_method(file_path: str) -> str:\n \"\"\"\n Parses a Python file and finds the longest method", + "warm_prompt_ms": 4242.21, + "cold_prompt_ms": 26771.6 + }, + { + "variant": 3, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 133, + 230 + ], + "cold_draft": [ + 133, + 230 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "5699f0214b3a", + "cold_sha": "5699f0214b3a", + "warm_text": "Based on the code provided, there is only **one** regular expression explicitly defined and used in the file. It is located in the `dequant_packed` function within the `ModelBase` class.\n\nHere is the function that extracts every regular expression pattern found in the file:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef extract_regex_patterns(file_path: str) -> list[str]:\n \"\"\"\n Scans a Python file and returns a list of all regular expression patterns\n found within `re.compile()` or `re.search()`/`re.match()` calls.\n \n Args:\n file_path: Path to the Python file to scan.\n \n Returns:\n A list of strings, each representing a regex pattern found in the file.\n \"\"\"\n path = Path(file_path)\n if not path.is_file():\n raise FileNotFoundError(f\"File not found: {", + "cold_text": "Based on the code provided, there is only **one** regular expression explicitly defined and used in the file. It is located in the `dequant_packed` function within the `ModelBase` class.\n\nHere is the function that extracts every regular expression pattern found in the file:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef extract_regex_patterns(file_path: str) -> list[str]:\n \"\"\"\n Scans a Python file and returns a list of all regular expression patterns\n found within `re.compile()` or `re.search()`/`re.match()` calls.\n \n Args:\n file_path: Path to the Python file to scan.\n \n Returns:\n A list of strings, each representing a regex pattern found in the file.\n \"\"\"\n path = Path(file_path)\n if not path.is_file():\n raise FileNotFoundError(f\"File not found: {", + "warm_prompt_ms": 4248.32, + "cold_prompt_ms": 26769.67 + }, + { + "variant": 4, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 133, + 230 + ], + "cold_draft": [ + 133, + 230 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "625a5d797269", + "cold_sha": "625a5d797269", + "warm_text": "Based on the code provided, here is the mapping of each defined class to its base classes.\n\nNote: The file defines several classes. Some inherit from standard library types (`IntEnum`, `TypeVar`), while others inherit from custom base classes defined in the same file (`ModelBase`).\n\n### Class Hierarchy Mapping\n\n| Class Name | Base Classes | Notes |\n| :--- | :--- | :--- |\n| `SentencePieceTokenTypes` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `ModelType` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `AnyModel` | `TypeVar` | A `TypeVar` bound to `type[ModelBase]`. |\n| `ModelBase` | *(None / `object`)* | The root abstract base class. It does not explicitly inherit from any other class in this file. |\n|", + "cold_text": "Based on the code provided, here is the mapping of each defined class to its base classes.\n\nNote: The file defines several classes. Some inherit from standard library types (`IntEnum`, `TypeVar`), while others inherit from custom base classes defined in the same file (`ModelBase`).\n\n### Class Hierarchy Mapping\n\n| Class Name | Base Classes | Notes |\n| :--- | :--- | :--- |\n| `SentencePieceTokenTypes` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `ModelType` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `AnyModel` | `TypeVar` | A `TypeVar` bound to `type[ModelBase]`. |\n| `ModelBase` | *(None / `object`)* | The root abstract base class. It does not explicitly inherit from any other class in this file. |\n|", + "warm_prompt_ms": 4245.74, + "cold_prompt_ms": 26731.0 + }, + { + "variant": 5, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 142, + 195 + ], + "cold_draft": [ + 142, + 195 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "3a8b9a882949", + "cold_sha": "3a8b9a882949", + "warm_text": "Here is a Python function that uses the `ast` module to parse the provided code and extract all string constants.\n\n```python\nimport ast\nfrom typing import List, Dict, Any\n\ndef extract_string_constants(code: str) -> List[str]:\n \"\"\"\n Extracts all string constants from the given Python source code.\n \n Args:\n code: The source code string to analyze.\n \n Returns:\n A list of all string literals found in the code.\n \"\"\"\n try:\n tree = ast.parse(code)\n except SyntaxError as e:\n raise ValueError(f\"Failed to parse code: {e}\")\n \n strings = []\n \n class StringVisitor(ast.NodeVisitor):\n def visit_Constant(self, node: ast.Constant) -> None:\n # Check if the constant is a string\n if isinstance(node.value, str):\n ", + "cold_text": "Here is a Python function that uses the `ast` module to parse the provided code and extract all string constants.\n\n```python\nimport ast\nfrom typing import List, Dict, Any\n\ndef extract_string_constants(code: str) -> List[str]:\n \"\"\"\n Extracts all string constants from the given Python source code.\n \n Args:\n code: The source code string to analyze.\n \n Returns:\n A list of all string literals found in the code.\n \"\"\"\n try:\n tree = ast.parse(code)\n except SyntaxError as e:\n raise ValueError(f\"Failed to parse code: {e}\")\n \n strings = []\n \n class StringVisitor(ast.NodeVisitor):\n def visit_Constant(self, node: ast.Constant) -> None:\n # Check if the constant is a string\n if isinstance(node.value, str):\n ", + "warm_prompt_ms": 4253.03, + "cold_prompt_ms": 26778.71 + }, + { + "variant": 6, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 130, + 244 + ], + "cold_draft": [ + 130, + 244 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "b92757254095", + "cold_sha": "b92757254095", + "warm_text": "To determine which imports are unused, we need to analyze the provided code snippet and check if each imported module or symbol is actually referenced in the code.\n\nHere is the analysis of the imports in the file:\n\n### Imports Analyzed\n\n1. **`from __future__ import annotations`**\n * **Usage:** Used implicitly to enable PEP 563 (deferred evaluation of annotations). It is standard practice and affects how type hints are processed.\n * **Status:** **Used** (implicitly, as it changes the behavior of all type annotations in the file).\n\n2. **`import ast`**\n * **Usage:** Searched the code for `ast.`. No usage found.\n * **Status:** **Unused**\n\n3. **`import logging`**\n * **Usage:** `logger = logging.getLogger(\"hf-to-gguf", + "cold_text": "To determine which imports are unused, we need to analyze the provided code snippet and check if each imported module or symbol is actually referenced in the code.\n\nHere is the analysis of the imports in the file:\n\n### Imports Analyzed\n\n1. **`from __future__ import annotations`**\n * **Usage:** Used implicitly to enable PEP 563 (deferred evaluation of annotations). It is standard practice and affects how type hints are processed.\n * **Status:** **Used** (implicitly, as it changes the behavior of all type annotations in the file).\n\n2. **`import ast`**\n * **Usage:** Searched the code for `ast.`. No usage found.\n * **Status:** **Unused**\n\n3. **`import logging`**\n * **Usage:** `logger = logging.getLogger(\"hf-to-gguf", + "warm_prompt_ms": 4246.47, + "cold_prompt_ms": 26790.78 + }, + { + "variant": 7, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 145, + 181 + ], + "cold_draft": [ + 145, + 181 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "eb0ac6ecb887", + "cold_sha": "eb0ac6ecb887", + "warm_text": "Based on the code provided, here is the list of methods that explicitly raise an exception.\n\n### Methods Raising Exceptions\n\n1. **`ModelBase.__init__`**\n * **Exception:** `TypeError`\n * **Condition:** Raised if the class is instantiated directly as `ModelBase`, `TextModel`, or `MmprojModel` (abstract base classes).\n * **Code:**\n ```python\n if type(self) is ModelBase or \\\n type(self) is TextModel or \\\n type(self) is MmprojModel:\n raise TypeError(f\"{type(self).__name__!r} should not be directly instantiated\")\n ```\n * **Exception:** `ImportError`\n * **Condition:** Raised if `self.is_mistral_format` is `True` but `mistral_common` is not installed.\n ", + "cold_text": "Based on the code provided, here is the list of methods that explicitly raise an exception.\n\n### Methods Raising Exceptions\n\n1. **`ModelBase.__init__`**\n * **Exception:** `TypeError`\n * **Condition:** Raised if the class is instantiated directly as `ModelBase`, `TextModel`, or `MmprojModel` (abstract base classes).\n * **Code:**\n ```python\n if type(self) is ModelBase or \\\n type(self) is TextModel or \\\n type(self) is MmprojModel:\n raise TypeError(f\"{type(self).__name__!r} should not be directly instantiated\")\n ```\n * **Exception:** `ImportError`\n * **Condition:** Raised if `self.is_mistral_format` is `True` but `mistral_common` is not installed.\n ", + "warm_prompt_ms": 4250.5, + "cold_prompt_ms": 26778.97 + }, + { + "variant": 8, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 136, + 217 + ], + "cold_draft": [ + 136, + 217 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "f818f816424b", + "cold_sha": "f818f816424b", + "warm_text": "Here is a Python script that parses the provided code, identifies all class definitions, and counts the lines of code associated with each class.\n\n```python\nimport ast\nimport sys\nfrom collections import defaultdict\n\ndef count_lines_per_class(code: str) -> dict[str, int]:\n \"\"\"\n Parses the given Python code string and counts the number of lines \n belonging to each class definition.\n \n Returns a dictionary mapping class names to their line counts.\n \"\"\"\n try:\n tree = ast.parse(code)\n except SyntaxError as e:\n print(f\"Error parsing code: {e}\", file=sys.stderr)\n return {}\n\n # We need to map AST nodes back to their line numbers in the source code.\n # The AST nodes have 'lineno' and 'end_lineno' attributes.\n \n class_line_counts = defaultdict(int)\n \n # Walk the AST", + "cold_text": "Here is a Python script that parses the provided code, identifies all class definitions, and counts the lines of code associated with each class.\n\n```python\nimport ast\nimport sys\nfrom collections import defaultdict\n\ndef count_lines_per_class(code: str) -> dict[str, int]:\n \"\"\"\n Parses the given Python code string and counts the number of lines \n belonging to each class definition.\n \n Returns a dictionary mapping class names to their line counts.\n \"\"\"\n try:\n tree = ast.parse(code)\n except SyntaxError as e:\n print(f\"Error parsing code: {e}\", file=sys.stderr)\n return {}\n\n # We need to map AST nodes back to their line numbers in the source code.\n # The AST nodes have 'lineno' and 'end_lineno' attributes.\n \n class_line_counts = defaultdict(int)\n \n # Walk the AST", + "warm_prompt_ms": 4247.28, + "cold_prompt_ms": 26797.84 + }, + { + "variant": 9, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 140, + 202 + ], + "cold_draft": [ + 140, + 202 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "a3a5d95ad2de", + "cold_sha": "a3a5d95ad2de", + "warm_text": "Here is a Python script that parses the provided code using the `ast` module to identify duplicate method names across different classes.\n\n```python\nimport ast\nimport sys\nfrom collections import defaultdict\n\ndef find_duplicate_methods(code: str) -> dict[str, list[str]]:\n \"\"\"\n Parses the given Python code string and finds method names that are \n defined in multiple different classes.\n \n Returns:\n A dictionary where keys are method names and values are lists of \n class names where that method is defined.\n \"\"\"\n tree = ast.parse(code)\n \n # Map: method_name -> list of class names\n method_to_classes = defaultdict(list)\n \n for node in ast.walk(tree):\n if isinstance(node, ast.ClassDef):\n class_name = node.name\n for item in node.body:\n if isinstance(item, ast.FunctionDef):\n method_name", + "cold_text": "Here is a Python script that parses the provided code using the `ast` module to identify duplicate method names across different classes.\n\n```python\nimport ast\nimport sys\nfrom collections import defaultdict\n\ndef find_duplicate_methods(code: str) -> dict[str, list[str]]:\n \"\"\"\n Parses the given Python code string and finds method names that are \n defined in multiple different classes.\n \n Returns:\n A dictionary where keys are method names and values are lists of \n class names where that method is defined.\n \"\"\"\n tree = ast.parse(code)\n \n # Map: method_name -> list of class names\n method_to_classes = defaultdict(list)\n \n for node in ast.walk(tree):\n if isinstance(node, ast.ClassDef):\n class_name = node.name\n for item in node.body:\n if isinstance(item, ast.FunctionDef):\n method_name", + "warm_prompt_ms": 4252.48, + "cold_prompt_ms": 26825.89 + } + ], + "result": "PASS" + }, + { + "label": "c3-patched-infile-strict", + "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": [ + 119, + 284 + ], + "cold_draft": [ + 119, + 284 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "f8e04e49df3c", + "cold_sha": "f8e04e49df3c", + "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": "Based on the code provided, here is a Python function that parses the file and extracts all tensor names.\n\nSince the file contains logic for mapping Hugging Face tensor names to GGUF tensor names (via `tensor_map` and `filter_tensors`), and also defines specific tensor handling in `dequant_model`, the most robust way to \"list every tensor name\" is to extract the **GGUF canonical tensor names** defined in the `gguf` library (which this script imports) and the **source tensor patterns** found in the logic.\n\nHowever, since `gguf.MODEL_TENSORS` is a dictionary defined in the imported `gguf` package (not in this file), we cannot list them directly from this file's text without importing the package.\n\n**But**, we can list the **tensor name patterns and suffixes** explicitly referenced in this file's logic.\n\nHere is a function that scans the file content for tensor", + "cold_text": "Based on the code provided, here is a Python function that parses the file and extracts all tensor names.\n\nSince the file contains logic for mapping Hugging Face tensor names to GGUF tensor names (via `tensor_map` and `filter_tensors`), and also defines specific tensor handling in `dequant_model`, the most robust way to \"list every tensor name\" is to extract the **GGUF canonical tensor names** defined in the `gguf` library (which this script imports) and the **source tensor patterns** found in the logic.\n\nHowever, since `gguf.MODEL_TENSORS` is a dictionary defined in the imported `gguf` package (not in this file), we cannot list them directly from this file's text without importing the package.\n\n**But**, we can list the **tensor name patterns and suffixes** explicitly referenced in this file's logic.\n\nHere is a function that scans the file content for tensor", + "warm_prompt_ms": 4249.71, + "cold_prompt_ms": 26798.61 + }, + { + "variant": 1, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 139, + 200 + ], + "cold_draft": [ + 139, + 200 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "7bf19aba991c", + "cold_sha": "7bf19aba991c", + "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": "Based on the code provided, here is a function that counts the number of model classes registered in the `ModelBase._model_classes` dictionary.\n\n```python\ndef count_registered_model_classes() -> int:\n \"\"\"\n Counts the total number of model classes registered in ModelBase._model_classes.\n \n Returns:\n int: The total count of registered model classes across all ModelTypes.\n \"\"\"\n total_count = 0\n for model_type, classes_dict in ModelBase._model_classes.items():\n total_count += len(classes_dict)\n return total_count\n```\n\n### Explanation:\n1. **`ModelBase._model_classes`** is a dictionary where:\n - Keys are `ModelType` enum values (`TEXT`, `MMPROJ`).\n - Values are dictionaries mapping model name strings to their corresponding `type[ModelBase]` classes.\n2. The", + "cold_text": "Based on the code provided, here is a function that counts the number of model classes registered in the `ModelBase._model_classes` dictionary.\n\n```python\ndef count_registered_model_classes() -> int:\n \"\"\"\n Counts the total number of model classes registered in ModelBase._model_classes.\n \n Returns:\n int: The total count of registered model classes across all ModelTypes.\n \"\"\"\n total_count = 0\n for model_type, classes_dict in ModelBase._model_classes.items():\n total_count += len(classes_dict)\n return total_count\n```\n\n### Explanation:\n1. **`ModelBase._model_classes`** is a dictionary where:\n - Keys are `ModelType` enum values (`TEXT`, `MMPROJ`).\n - Values are dictionaries mapping model name strings to their corresponding `type[ModelBase]` classes.\n2. The", + "warm_prompt_ms": 4259.8, + "cold_prompt_ms": 26867.64 + }, + { + "variant": 2, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 135, + 214 + ], + "cold_draft": [ + 135, + 214 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "790e4205bfa3", + "cold_sha": "790e4205bfa3", + "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": "To find the longest method in the provided Python file, we need to parse the code, identify all method definitions (functions defined inside classes), and calculate their length based on the number of lines or characters they span.\n\nHere is a robust Python script that performs this task using the `ast` (Abstract Syntax Tree) module. It will:\n1. Parse the source code.\n2. Identify all `FunctionDef` and `AsyncFunctionDef` nodes that are methods (i.e., defined inside a `ClassDef`).\n3. Calculate the length of each method (number of lines in the source code).\n4. Return the name of the longest method.\n\n### Python Script to Find the Longest Method\n\n```python\nimport ast\nimport sys\nfrom pathlib import Path\n\ndef find_longest_method(file_path: str) -> str:\n \"\"\"\n Parses a Python file and finds the longest method", + "cold_text": "To find the longest method in the provided Python file, we need to parse the code, identify all method definitions (functions defined inside classes), and calculate their length based on the number of lines or characters they span.\n\nHere is a robust Python script that performs this task using the `ast` (Abstract Syntax Tree) module. It will:\n1. Parse the source code.\n2. Identify all `FunctionDef` and `AsyncFunctionDef` nodes that are methods (i.e., defined inside a `ClassDef`).\n3. Calculate the length of each method (number of lines in the source code).\n4. Return the name of the longest method.\n\n### Python Script to Find the Longest Method\n\n```python\nimport ast\nimport sys\nfrom pathlib import Path\n\ndef find_longest_method(file_path: str) -> str:\n \"\"\"\n Parses a Python file and finds the longest method", + "warm_prompt_ms": 4264.45, + "cold_prompt_ms": 26898.73 + }, + { + "variant": 3, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 133, + 229 + ], + "cold_draft": [ + 133, + 229 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "5699f0214b3a", + "cold_sha": "5699f0214b3a", + "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": "Based on the code provided, there is only **one** regular expression explicitly defined and used in the file. It is located in the `dequant_packed` function within the `ModelBase` class.\n\nHere is the function that extracts every regular expression pattern found in the file:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef extract_regex_patterns(file_path: str) -> list[str]:\n \"\"\"\n Scans a Python file and returns a list of all regular expression patterns\n found within `re.compile()` or `re.search()`/`re.match()` calls.\n \n Args:\n file_path: Path to the Python file to scan.\n \n Returns:\n A list of strings, each representing a regex pattern found in the file.\n \"\"\"\n path = Path(file_path)\n if not path.is_file():\n raise FileNotFoundError(f\"File not found: {", + "cold_text": "Based on the code provided, there is only **one** regular expression explicitly defined and used in the file. It is located in the `dequant_packed` function within the `ModelBase` class.\n\nHere is the function that extracts every regular expression pattern found in the file:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef extract_regex_patterns(file_path: str) -> list[str]:\n \"\"\"\n Scans a Python file and returns a list of all regular expression patterns\n found within `re.compile()` or `re.search()`/`re.match()` calls.\n \n Args:\n file_path: Path to the Python file to scan.\n \n Returns:\n A list of strings, each representing a regex pattern found in the file.\n \"\"\"\n path = Path(file_path)\n if not path.is_file():\n raise FileNotFoundError(f\"File not found: {", + "warm_prompt_ms": 4273.77, + "cold_prompt_ms": 26832.74 + }, + { + "variant": 4, + "L": 7039, + "warm_prompt_n": 1028, + "warm_cache_n": 6011, + "cold_prompt_n": 7039, + "cold_cache_n": 0, + "aligned": true, + "warm_draft": [ + 133, + 230 + ], + "cold_draft": [ + 133, + 230 + ], + "identical": true, + "first_diff_char": null, + "warm_sha": "625a5d797269", + "cold_sha": "625a5d797269", + "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": "Based on the code provided, here is the mapping of each defined class to its base classes.\n\nNote: The file defines several classes. Some inherit from standard library types (`IntEnum`, `TypeVar`), while others inherit from custom base classes defined in the same file (`ModelBase`).\n\n### Class Hierarchy Mapping\n\n| Class Name | Base Classes | Notes |\n| :--- | :--- | :--- |\n| `SentencePieceTokenTypes` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `ModelType` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `AnyModel` | `TypeVar` | A `TypeVar` bound to `type[ModelBase]`. |\n| `ModelBase` | *(None / `object`)* | The root abstract base class. It does not explicitly inherit from any other class in this file. |\n|", + "cold_text": "Based on the code provided, here is the mapping of each defined class to its base classes.\n\nNote: The file defines several classes. Some inherit from standard library types (`IntEnum`, `TypeVar`), while others inherit from custom base classes defined in the same file (`ModelBase`).\n\n### Class Hierarchy Mapping\n\n| Class Name | Base Classes | Notes |\n| :--- | :--- | :--- |\n| `SentencePieceTokenTypes` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `ModelType` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `AnyModel` | `TypeVar` | A `TypeVar` bound to `type[ModelBase]`. |\n| `ModelBase` | *(None / `object`)* | The root abstract base class. It does not explicitly inherit from any other class in this file. |\n|", + "warm_prompt_ms": 4260.6, + "cold_prompt_ms": 26846.8 + } + ], + "result": "PASS" + }, + { + "label": "c3-patched-nospec", + "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": "b0714d79c46a", + "cold_sha": "b0714d79c46a", + "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": "Based on the code provided, here is a Python function that parses the file content to extract all tensor names.\n\nSince the file contains logic for mapping Hugging Face tensor names to GGUF tensor names (via `tensor_map` and `filter_tensors`), and also defines specific quantization dequantization logic that creates new tensor names, a robust approach is to scan the source code for string literals that look like tensor names or suffixes.\n\nHere is the function:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef list_tensor_names(file_path: str) -> list[str]:\n \"\"\"\n Lists every tensor name found in the provided Python file.\n \n This function scans the source code for:\n 1. String literals that match common tensor naming patterns (e.g., '.weight', '.bias').\n 2. References to `gguf.MODEL_TENSOR` enum values which define the canonical names.", + "cold_text": "Based on the code provided, here is a Python function that parses the file content to extract all tensor names.\n\nSince the file contains logic for mapping Hugging Face tensor names to GGUF tensor names (via `tensor_map` and `filter_tensors`), and also defines specific quantization dequantization logic that creates new tensor names, a robust approach is to scan the source code for string literals that look like tensor names or suffixes.\n\nHere is the function:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef list_tensor_names(file_path: str) -> list[str]:\n \"\"\"\n Lists every tensor name found in the provided Python file.\n \n This function scans the source code for:\n 1. String literals that match common tensor naming patterns (e.g., '.weight', '.bias').\n 2. References to `gguf.MODEL_TENSOR` enum values which define the canonical names.", + "warm_prompt_ms": 3970.36, + "cold_prompt_ms": 25076.06 + }, + { + "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": "1dc870095e26", + "cold_sha": "1dc870095e26", + "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": "Based on the code provided, here is a function that counts the number of model classes registered in the `ModelBase._model_classes` dictionary.\n\n```python\ndef count_registered_model_classes() -> int:\n \"\"\"\n Counts the total number of model classes registered in ModelBase._model_classes.\n \n Returns:\n int: The total count of registered model classes across all ModelTypes.\n \"\"\"\n total_count = 0\n for model_type, classes_dict in ModelBase._model_classes.items():\n total_count += len(classes_dict)\n return total_count\n```\n\n### Explanation:\n1. **`ModelBase._model_classes`** is a dictionary where the keys are `ModelType` enums (`TEXT`, `MMPROJ`) and the values are dictionaries mapping class names to their corresponding `type[ModelBase]` subclasses.\n2. The function iterates over each `", + "cold_text": "Based on the code provided, here is a function that counts the number of model classes registered in the `ModelBase._model_classes` dictionary.\n\n```python\ndef count_registered_model_classes() -> int:\n \"\"\"\n Counts the total number of model classes registered in ModelBase._model_classes.\n \n Returns:\n int: The total count of registered model classes across all ModelTypes.\n \"\"\"\n total_count = 0\n for model_type, classes_dict in ModelBase._model_classes.items():\n total_count += len(classes_dict)\n return total_count\n```\n\n### Explanation:\n1. **`ModelBase._model_classes`** is a dictionary where the keys are `ModelType` enums (`TEXT`, `MMPROJ`) and the values are dictionaries mapping class names to their corresponding `type[ModelBase]` subclasses.\n2. The function iterates over each `", + "warm_prompt_ms": 3981.56, + "cold_prompt_ms": 25120.34 + }, + { + "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": "34d7bda444b5", + "cold_sha": "34d7bda444b5", + "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": "To find the longest method in the provided Python file, we need to parse the code, identify all method definitions (functions defined inside classes), and calculate their length based on the number of lines or characters they span.\n\nHere is a robust Python script that performs this task using the `ast` (Abstract Syntax Tree) module. It will:\n1. Parse the source code.\n2. Identify all `FunctionDef` and `AsyncFunctionDef` nodes that are methods (i.e., defined within a `ClassDef`).\n3. Calculate the length of each method (number of lines in the source code).\n4. Return the name of the longest method.\n\n### Python Script to Find the Longest Method\n\n```python\nimport ast\nimport sys\nfrom pathlib import Path\n\ndef find_longest_method(file_path: str) -> str:\n \"\"\"\n Parses a Python file and finds the longest method", + "cold_text": "To find the longest method in the provided Python file, we need to parse the code, identify all method definitions (functions defined inside classes), and calculate their length based on the number of lines or characters they span.\n\nHere is a robust Python script that performs this task using the `ast` (Abstract Syntax Tree) module. It will:\n1. Parse the source code.\n2. Identify all `FunctionDef` and `AsyncFunctionDef` nodes that are methods (i.e., defined within a `ClassDef`).\n3. Calculate the length of each method (number of lines in the source code).\n4. Return the name of the longest method.\n\n### Python Script to Find the Longest Method\n\n```python\nimport ast\nimport sys\nfrom pathlib import Path\n\ndef find_longest_method(file_path: str) -> str:\n \"\"\"\n Parses a Python file and finds the longest method", + "warm_prompt_ms": 3981.28, + "cold_prompt_ms": 25149.93 + }, + { + "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": "5699f0214b3a", + "cold_sha": "5699f0214b3a", + "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": "Based on the code provided, there is only **one** regular expression explicitly defined and used in the file. It is located in the `dequant_packed` function within the `ModelBase` class.\n\nHere is the function that extracts every regular expression pattern found in the file:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef extract_regex_patterns(file_path: str) -> list[str]:\n \"\"\"\n Scans a Python file and returns a list of all regular expression patterns\n found within `re.compile()` or `re.search()`/`re.match()` calls.\n \n Args:\n file_path: Path to the Python file to scan.\n \n Returns:\n A list of strings, each representing a regex pattern found in the file.\n \"\"\"\n path = Path(file_path)\n if not path.is_file():\n raise FileNotFoundError(f\"File not found: {", + "cold_text": "Based on the code provided, there is only **one** regular expression explicitly defined and used in the file. It is located in the `dequant_packed` function within the `ModelBase` class.\n\nHere is the function that extracts every regular expression pattern found in the file:\n\n```python\nimport re\nfrom pathlib import Path\n\ndef extract_regex_patterns(file_path: str) -> list[str]:\n \"\"\"\n Scans a Python file and returns a list of all regular expression patterns\n found within `re.compile()` or `re.search()`/`re.match()` calls.\n \n Args:\n file_path: Path to the Python file to scan.\n \n Returns:\n A list of strings, each representing a regex pattern found in the file.\n \"\"\"\n path = Path(file_path)\n if not path.is_file():\n raise FileNotFoundError(f\"File not found: {", + "warm_prompt_ms": 3984.16, + "cold_prompt_ms": 25146.31 + }, + { + "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": "625a5d797269", + "cold_sha": "625a5d797269", + "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": "Based on the code provided, here is the mapping of each defined class to its base classes.\n\nNote: The file defines several classes. Some inherit from standard library types (`IntEnum`, `TypeVar`), while others inherit from custom base classes defined in the same file (`ModelBase`).\n\n### Class Hierarchy Mapping\n\n| Class Name | Base Classes | Notes |\n| :--- | :--- | :--- |\n| `SentencePieceTokenTypes` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `ModelType` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `AnyModel` | `TypeVar` | A `TypeVar` bound to `type[ModelBase]`. |\n| `ModelBase` | *(None / `object`)* | The root abstract base class. It does not explicitly inherit from any other class in this file. |\n|", + "cold_text": "Based on the code provided, here is the mapping of each defined class to its base classes.\n\nNote: The file defines several classes. Some inherit from standard library types (`IntEnum`, `TypeVar`), while others inherit from custom base classes defined in the same file (`ModelBase`).\n\n### Class Hierarchy Mapping\n\n| Class Name | Base Classes | Notes |\n| :--- | :--- | :--- |\n| `SentencePieceTokenTypes` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `ModelType` | `IntEnum` | Inherits from `enum.IntEnum`. |\n| `AnyModel` | `TypeVar` | A `TypeVar` bound to `type[ModelBase]`. |\n| `ModelBase` | *(None / `object`)* | The root abstract base class. It does not explicitly inherit from any other class in this file. |\n|", + "warm_prompt_ms": 3984.8, + "cold_prompt_ms": 25151.36 + } + ], + "result": "PASS" + }, + { + "label": "tools-infile-patched", + "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": "vision-q106-infile", + "fa": "on", + "mtp": true, + "expected": "red,blue,circle,square", + "answer": "", + "hits": [], + "error": "RemoteDisconnected: Remote end closed connection without response", + "server_died": true, + "server_log_errors": [ + "0.35.751.905 E process: missing MTP boundary for seq_id=0 pos=17 (current=3/1 previous=2/1)", + "0.35.751.911 E srv update_slots: failed to process speculative batch", + "0.35.907.029 E process: missing MTP boundary for seq_id=0 pos=34 (current=3/1 previous=2/1)", + "0.35.907.035 E srv update_slots: failed to process speculative batch", + "#2 0x000074c11333d86b in ggml_abort () from /opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/libggml-base.so.0" + ], + "result": "FAIL" + }, + { + "label": "vision-q106-infile-faoff", + "fa": "off", + "mtp": true, + "expected": "red,blue,circle,square", + "answer": "", + "hits": [], + "error": "RemoteDisconnected: Remote end closed connection without response", + "server_died": true, + "server_log_errors": [ + "0.25.875.875 E process: missing MTP boundary for seq_id=0 pos=17 (current=3/1 previous=2/1)", + "0.25.875.882 E srv update_slots: failed to process speculative batch", + "0.26.042.119 E process: missing MTP boundary for seq_id=0 pos=34 (current=3/1 previous=2/1)", + "0.26.042.125 E srv update_slots: failed to process speculative batch", + "#2 0x000075867313d86b in ggml_abort () from /opt/llama-rocm/rocmfpx-724-mtpcache/build-hipvk/bin/libggml-base.so.0" + ], + "result": "FAIL" + }, + { + "label": "vision-q106-nomtp", + "fa": "on", + "mtp": false, + "expected": "red,blue,circle,square", + "answer": "The image displays two distinct geometric shapes arranged side by side on a plain white background. On the left is a solid red circle, and on the right is a solid blue square. Both shapes are filled with their respective colors \u2014 vibrant red for the circle and deep blue for the square \u2014 and are clea", + "hits": [ + "red", + "blue", + "circle", + "square" + ], + "error": null, + "server_died": false, + "server_log_errors": [], + "result": "PASS" + }, + { + "label": "vision-q106-nomtp-faoff", + "fa": "off", + "mtp": false, + "expected": "red,blue,circle,square", + "answer": "The image displays two distinct geometric shapes arranged side by side on a plain white background. On the left is a solid red circle, and on the right is a solid blue square. Both shapes are filled with their respective colors and have no outlines or additional details. The red circle is positioned", + "hits": [ + "red", + "blue", + "circle", + "square" + ], + "error": null, + "server_died": false, + "server_log_errors": [], + "result": "PASS" + }, + { + "label": "vision-q106-infile-unpatched", + "fa": "on", + "mtp": true, + "expected": "red,blue,circle,square", + "answer": "", + "hits": [], + "error": "RemoteDisconnected: Remote end closed connection without response", + "server_died": true, + "server_log_errors": [ + "0.23.506.262 E process: missing MTP boundary for seq_id=0 pos=17 (current=3/1 previous=2/1)", + "0.23.506.270 E srv update_slots: failed to process speculative batch", + "0.23.659.691 E process: missing MTP boundary for seq_id=0 pos=34 (current=3/1 previous=2/1)", + "0.23.659.697 E srv update_slots: failed to process speculative batch", + "#2 0x0000773a9c93d86b in ggml_abort () from /opt/llama-rocm/rocmfpx-724/build-hipvk/bin/libggml-base.so.0" + ], + "result": "FAIL" + } + ], + "knee": 4, + "tier_bench_nmax": 4, + "history_separate_head": { + "note": "Measured 2026-09-16 with the separate 18-tensor head file (not shipped any more); same trunk bytes.", + "bench": [ + { + "label": "sweep-rocm-off", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": null, + "nmax": null, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 22.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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": 11.06, + "tg_min": 11.06, + "tg_max": 11.06, + "pp_median": 280.3, + "prompt_n": 7097, + "accept": null + }, + { + "label": "sweep-rocm-n1", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 1, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 1 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 17.9, + "tg_min": 17.9, + "tg_max": 17.9, + "pp_median": 263.9, + "prompt_n": 7101, + "accept": 0.875 + }, + { + "label": "sweep-rocm-n2", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 2, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 2 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 22.84, + "tg_min": 22.78, + "tg_max": 22.85, + "pp_median": 263.4, + "prompt_n": 7100, + "accept": 0.832 + }, + { + "label": "sweep-rocm-n3", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 3, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 3 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 25.85, + "tg_min": 25.32, + "tg_max": 25.96, + "pp_median": 262.7, + "prompt_n": 7094, + "accept": 0.768 + }, + { + "label": "sweep-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 26.25, + "tg_min": 25.76, + "tg_max": 26.57, + "pp_median": 262.2, + "prompt_n": 7099, + "accept": 0.704 + }, + { + "label": "sweep-rocm-n5", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 5, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 5 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 25.28, + "tg_min": 25.25, + "tg_max": 26.03, + "pp_median": 261.4, + "prompt_n": 7099, + "accept": 0.624 + }, + { + "label": "sweep-rocm-n6", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 6, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 6 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 24.59, + "tg_min": 23.81, + "tg_max": 24.59, + "pp_median": 259.5, + "prompt_n": 7099, + "accept": 0.552 + }, + { + "label": "sweep-vk-off", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": null, + "nmax": null, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 20.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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": 11.21, + "tg_min": 11.21, + "tg_max": 11.22, + "pp_median": 203.4, + "prompt_n": 7101, + "accept": null + }, + { + "label": "sweep-vk-n3", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 3, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 6.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 3 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 26.06, + "tg_min": 25.73, + "tg_max": 26.37, + "pp_median": 200.5, + "prompt_n": 7099, + "accept": 0.794 + }, + { + "label": "sweep-vk-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 6.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 28.25, + "tg_min": 27.8, + "tg_max": 28.25, + "pp_median": 198.1, + "prompt_n": 7100, + "accept": 0.785 + }, + { + "label": "sweep-vk-n5", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 5, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 4.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 5 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 27.09, + "tg_min": 27.08, + "tg_max": 28.02, + "pp_median": 198.7, + "prompt_n": 7097, + "accept": 0.664 + }, + { + "label": "head-q8-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q8_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q8_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 25.15, + "tg_min": 25.13, + "tg_max": 25.15, + "pp_median": 262.1, + "prompt_n": 7100, + "accept": 0.719 + }, + { + "label": "tier-q102-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_COHERENT.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 24.09, + "tg_min": 23.35, + "tg_max": 24.1, + "pp_median": 233.4, + "prompt_n": 7098, + "accept": 0.634 + }, + { + "label": "tier-q102-vk-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_COHERENT.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 6.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 23.73, + "tg_min": 23.49, + "tg_max": 23.82, + "pp_median": 199.5, + "prompt_n": 7100, + "accept": 0.647 + }, + { + "label": "tier-q115-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX_AGENT.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 40.1, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX_AGENT.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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 17.48, + "tg_min": 17.46, + "tg_max": 17.66, + "pp_median": 240.8, + "prompt_n": 7099, + "accept": 0.681 + }, + { + "label": "tier-q115-vk-n4", + "model": "Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX_AGENT.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 38.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX_AGENT.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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 15.81, + "tg_min": 15.41, + "tg_max": 15.82, + "pp_median": 212.0, + "prompt_n": 7094, + "accept": 0.622 + }, + { + "label": "tier-q111-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 40.1, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX.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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 16.66, + "tg_min": 16.66, + "tg_max": 17.36, + "pp_median": 230.2, + "prompt_n": 7098, + "accept": 0.63 + }, + { + "label": "tier-q111-vk-n4", + "model": "Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "code", + "load_s": 38.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-Q8_0_ROCMFPX.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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 16.88, + "tg_min": 16.66, + "tg_max": 16.89, + "pp_median": 211.5, + "prompt_n": 7097, + "accept": 0.614 + }, + { + "label": "tier-q106i-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "code", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out-imat/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 25.52, + "tg_min": 24.54, + "tg_max": 26.26, + "pp_median": 262.2, + "prompt_n": 7099, + "accept": 0.675 + }, + { + "label": "prose-q106-rocm-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "ROCm0", + "ctx": 65536, + "workload": "prose", + "load_s": 24.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device ROCm0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 17.04, + "tg_min": 16.41, + "tg_max": 17.69, + "pp_median": 261.7, + "prompt_n": 7456, + "accept": 0.367 + }, + { + "label": "prose-q106-vk-n4", + "model": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "draft": "mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf", + "nmax": 4, + "dev": "Vulkan0", + "ctx": 65536, + "workload": "prose", + "load_s": 8.0, + "cmd": "/opt/llama-rocm/rocmfpx-724/build-hipvk/bin/llama-server -m /mnt/models/agnes-3.0-flash/out/Agnes-3.0-Flash-Preview-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 --spec-type draft-mtp --model-draft /mnt/models/agnes-3.0-flash/out/mtp-Agnes-3.0-Flash-Preview-Q4_0.gguf --spec-draft-ngl 99 --spec-draft-device Vulkan0 --spec-draft-n-max 4 --spec-draft-n-min 0 --spec-draft-p-min 0.0", + "tg_median": 17.48, + "tg_min": 17.19, + "tg_max": 17.63, + "pp_median": 199.3, + "prompt_n": 7454, + "accept": 0.39 + } + ], + "gates": [ + { + "label": "cache-mtp", + "turn1_prompt_n": 7062, + "turn2_prompt_n": 7067, + "turn2_cache_n": 0, + "result": "FAIL" + }, + { + "label": "cache-nodraft", + "turn1_prompt_n": 7062, + "turn2_prompt_n": 1033, + "turn2_cache_n": 6034, + "result": "FAIL" + }, + { + "label": "tools-q106-mtp", + "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": "vision-q106", + "answer": "The image displays two distinct geometric shapes arranged side by side on a plain white background. On the left is a solid red circle, and on the right is a solid blue square. Both shapes are filled with their respective colors \u2014 vibrant red for the circle and deep blue for the square \u2014 and are clea", + "expected": "red,blue,circle,square", + "hits": [ + "red", + "blue", + "circle", + "square" + ], + "result": "PASS" + } + ], + "knee": 4, + "seat_sizing": [ + { + "ctx": 65536, + "model": "Agnes-3.0-Flash-Preview-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "avail_before": 122.11, + "avail_loaded": 93.2, + "avail_after_8k": 86.84, + "footprint_loaded_gib": 28.91, + "footprint_after_8k_gib": 35.27 + }, + { + "ctx": 131072, + "model": "Agnes-3.0-Flash-Preview-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "avail_before": 122.32, + "avail_loaded": 88.39, + "avail_after_8k": 82.14, + "footprint_loaded_gib": 33.93, + "footprint_after_8k_gib": 40.18 + }, + { + "ctx": 262144, + "model": "Agnes-3.0-Flash-Preview-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "avail_before": 122.27, + "avail_loaded": 77.3, + "avail_after_8k": 70.84, + "footprint_loaded_gib": 44.97, + "footprint_after_8k_gib": 51.43 + } + ] + }, + "seat_sizing2": [ + { + "label": "separate-head", + "ctx": 65536, + "avail_before": 122.21, + "footprint_loaded_gib": 27.32, + "footprint_after_8k_gib": 33.57, + "concurrent_download": "inactive" + }, + { + "label": "in-file", + "ctx": 65536, + "avail_before": 122.29, + "footprint_loaded_gib": 25.47, + "footprint_after_8k_gib": 31.92, + "concurrent_download": "inactive" + }, + { + "label": "in-file", + "ctx": 131072, + "avail_before": 122.32, + "footprint_loaded_gib": 28.8, + "footprint_after_8k_gib": 35.08, + "concurrent_download": "inactive" + }, + { + "label": "in-file", + "ctx": 262144, + "avail_before": 122.33, + "footprint_loaded_gib": 36.23, + "footprint_after_8k_gib": 42.12, + "concurrent_download": "inactive" + }, + { + "label": "in-file-lean", + "ctx": 65536, + "avail_before": 122.32, + "footprint_loaded_gib": 24.67, + "footprint_after_8k_gib": 31.05, + "concurrent_download": "inactive" + } + ], + "trunk_receipt": [ + { + "tag": "q106", + "file": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "rows_compared": 4, + "result": "MISMATCH" + }, + { + "tag": "q102", + "file": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_COHERENT.gguf", + "rows_compared": 4, + "result": "MISMATCH" + }, + { + "tag": "q115", + "file": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX_AGENT.gguf", + "rows_compared": 4, + "result": "MATCH" + }, + { + "tag": "q111", + "file": "Agnes-3.0-Flash-Preview-MTP-Q8_0_ROCMFPX.gguf", + "rows_compared": 4, + "result": "MATCH" + }, + { + "tag": "q106i", + "file": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "rows_compared": 4, + "result": "MISMATCH" + }, + { + "tag": "q102i", + "file": "Agnes-3.0-Flash-Preview-MTP-imatrix-Q4_0_ROCMFP4_COHERENT.gguf", + "rows_compared": 4, + "result": "MISMATCH" + } + ], + "regrade": { + "R1_cpu_burn": { + "step": "R1", + "burn_seen": 0, + "result": "SAME_AS_TODAY" + }, + "R2_base_logits": { + "step": "R2", + "base_logits_vs_afternoon": "DIFFERENT_BASE_LOGITS" + }, + "R4_repeat": { + "step": "R4", + "rows": 40, + "result": "MATCH" + }, + "kld_control": [ + { + "label": "c1-old-q106", + "file": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "reference": "B5_kld_q106.log", + "result": "MISMATCH" + }, + { + "label": "c1-old-q102", + "file": "Agnes-3.0-Flash-Preview-Q4_0_ROCMFP4_COHERENT.gguf", + "reference": "B5_kld_q102.log", + "result": "MISMATCH" + }, + { + "label": "c1-new-q106-again", + "file": "Agnes-3.0-Flash-Preview-MTP-Q4_0_ROCMFP4_STRIX_LEAN.gguf", + "reference": "R_kld4_q106.log", + "result": "MATCH" + } + ], + "R1_pass_seconds": { + "cpu_burn": 9.49, + "quiet_today": 7.53, + "afternoon": 9.66 + }, + "R1_burn_started": true, + "note": "R1: 4-bit rows under a concurrent 16-thread CPU burn vs today's quiet run and the afternoon run. The burn_seen field in regrade.jsonl is a grep bug (searched 'quantize', the log says 'quantizing'); R1_burn_started and R1_pass_seconds are the evidence that the load was real." + }, + "fold": { + "source_repo": "Agnes-AI/Agnes-3.0-Flash", + "source_dir": "hf", + "main_ffn": 17408, + "parallel_ffn": 2048, + "folded_ffn": 19456, + "stats": { + "fold": 216, + "pad": 3, + "pass": 1086 + }, + "tensors_out": 1305, + "bytes_out": 66243917920, + "fold_script_sha256": "b1e6dbcf16bbac48fbc95b763efeb85efb21ce1d1fdfa825a6ae01187f044486" + }, + "fold_verify": { + "result": "PASS", + "g3": [ + [ + "layer 0", + "8.535e-10" + ], + [ + "layer 36", + "5.425e-08" + ], + [ + "layer 71", + "2.481e-08" + ], + [ + "mtp", + "3.030e-08" + ] + ] + } +} \ No newline at end of file