Capicua25x commited on
Commit
71690bc
·
verified ·
1 Parent(s): d41b19d

Ornith-1.0-35B MXFP4 + grafted MoE MTP head (vision), for AMD RDNA4 / vLLM

Browse files
.gitattributes CHANGED
@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ assets/ornith_35b_eval.png filter=lfs diff=lfs merge=lfs -text
37
+ assets/ornith_397b_eval.png filter=lfs diff=lfs merge=lfs -text
38
+ assets/ornith_logo.png filter=lfs diff=lfs merge=lfs -text
39
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,129 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model:
4
+ - deepreinforce-ai/Ornith-1.0-35B
5
+ - Capicua25x/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill-MXFP4-Vision
6
+ base_model_relation: merge
7
+ pipeline_tag: image-text-to-text
8
+ library_name: vllm
9
+ language:
10
+ - en
11
+ tags:
12
+ - mxfp4
13
+ - compressed-tensors
14
+ - speculative-decoding
15
+ - mtp
16
+ - multi-token-prediction
17
+ - moe
18
+ - qwen3.5
19
+ - vllm
20
+ - rocm
21
+ - rdna4
22
+ - amd
23
+ - vision
24
+ ---
25
+
26
+ # Ornith-1.0-35B-A3B — MXFP4 + MTP (vision), for AMD RDNA4 / vLLM
27
+
28
+ [`deepreinforce-ai/Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B) — a
29
+ **Qwen MoE model (A3B, ~3 B active parameters per token)** — quantized to **MXFP4** with a **grafted MTP
30
+ (Multi-Token-Prediction) draft head**, packaged to run **out of the box on AMD Radeon RDNA4 (gfx1201)
31
+ under vLLM** with lossless self-speculative decoding. Vision retained.
32
+
33
+ - **Trunk** — Ornith-1.0-35B quantized to **MXFP4** (`compressed-tensors`, group size 32, symmetric).
34
+ `lm_head`, `embed_tokens`, norms, and the vision tower are kept **BF16**. The MXFP4 trunk is
35
+ byte-verified against DeepReinforce's published weights (16/16 shards SHA256-identical before quant).
36
+ - **Draft head — a cross-model graft.** The MoE MTP head (785 BF16 tensors, with per-expert weights) is
37
+ transplanted from a **sibling Qwen-MoE checkpoint**,
38
+ [`Capicua25x/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill-MXFP4-Vision`](https://huggingface.co/Capicua25x/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill-MXFP4-Vision)
39
+ (a DeepSeek-V4-Pro-Thinking distill of Qwen3.6-35B-A3B). Because the trunk and the donor share the
40
+ **same Qwen MoE architecture** — identical hidden size and expert layout — the head transfers cleanly
41
+ and drafts well on the Ornith trunk (acceptance below).
42
+ - **Speculative decoding** — vLLM native `mtp` method, `num_speculative_tokens=3`. **Lossless**: the
43
+ target model verifies every drafted token, so the output distribution is identical to the trunk with
44
+ no draft head. A mismatched head could only *slow* drafting, never change an output — which is what
45
+ makes a cross-model graft safe to ship. This one drafts at **~69% acceptance**, a clear net speed-up.
46
+
47
+ ## Measured on AMD RDNA4
48
+
49
+ 2× Radeon AI PRO R9700 (gfx1201), tensor-parallel 2, vLLM `0.19.1` (image below).
50
+
51
+ - **MTP draft acceptance (n=3): ≈69.5% average — measured on live production traffic**, not a synthetic
52
+ probe. Per-position **0.84 / 0.67 / 0.57**; mean acceptance length **~3.08**. Lossless throughout.
53
+ - **Vision**, **tool-calling** (`qwen3_xml`), and **reasoning split** (`qwen3`) all confirmed working.
54
+ - **Throughput** (TP2, 256 output tokens, per-user / aggregate tok/s):
55
+
56
+ | concurrent | short prompt | ~6k prompt |
57
+ |---:|---|---|
58
+ | 1 | 130.9 / 131 | 100.9 / 101 |
59
+ | 16 | 58.9 / 889 | 34.2 / 523 |
60
+ | 32 | 40.2 / 1219 | 20.3 / 617 |
61
+ | 64 | 29.9 / 1822 | 12.3 / 723 |
62
+ | 96 | 23.4 / 1584 | 9.2 / 717 |
63
+ | 128 | 21.4 / 1779 | 8.2 / 761 |
64
+
65
+ Usable concurrency ceiling (per-user ≥ 20 tok/s): **~128** at short context, **~32** at 6k context.
66
+ Single-stream decode is **130.9 tok/s** short / **100.9 tok/s** at 6k — the A3B sparsity (~3 B active)
67
+ makes it *faster* single-stream than a dense 9B at short context, and comparable at long context, while
68
+ carrying far more total knowledge.
69
+
70
+ ## Run it on AMD RDNA4
71
+
72
+ Uses the prebuilt RDNA4 vLLM image
73
+ [`capicua25x/vllm-rocm-rdna4`](https://hub.docker.com/r/capicua25x/vllm-rocm-rdna4) (tag `0.19.1`):
74
+
75
+ ```bash
76
+ docker run --rm --network=host \
77
+ --device=/dev/kfd --device=/dev/dri \
78
+ --group-add=video --group-add=render --ipc=host --ulimit memlock=-1 \
79
+ -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
80
+ capicua25x/vllm-rocm-rdna4:0.19.1 \
81
+ --model Capicua25x/Ornith-1.0-35B-MXFP4-Vision-MTP \
82
+ --served-model-name ornith --trust-remote-code \
83
+ --tensor-parallel-size 2 \
84
+ --gpu-memory-utilization 0.92 \
85
+ --max-model-len 262144 \
86
+ --attention-backend TRITON_ATTN \
87
+ --enable-prefix-caching --max-num-seqs 64 \
88
+ --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 \
89
+ --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
90
+ ```
91
+
92
+ ### The settings that actually matter on RDNA4
93
+ - `--attention-backend TRITON_ATTN` — required on gfx1201; the 0.19.1 default (`ROCM_ATTN`) collapses
94
+ spec-decode throughput at concurrency.
95
+ - `--speculative-config '{"method":"mtp","num_speculative_tokens":3}'` — enables the grafted MTP head.
96
+ **n=3 maximizes throughput; n=1–2 maximize per-token acceptance.** Tune per workload.
97
+ - `--tool-call-parser qwen3_xml --reasoning-parser qwen3` — Qwen-family tool-calling + reasoning split.
98
+ - `--trust-remote-code` — the Qwen MoE vision architecture.
99
+ - **Memory** — fits **2× 32 GB** at MXFP4 with 256k context (`--gpu-memory-utilization 0.92`,
100
+ `--max-num-seqs 64`). The 35 B does not fit a single 32 GB GPU at this context; use TP2.
101
+
102
+ ## How it was built (reproducible)
103
+ 1. **MXFP4 quantize** Ornith-1.0-35B with `compressed-tensors` (4-bit float, group 32, symmetric;
104
+ `lm_head` / `embed_tokens` / norms / vision tower left BF16).
105
+ 2. **Graft** the `mtp.*` MoE head (785 BF16 tensors — including the per-expert
106
+ `mtp.layers.0.mlp.experts.N.*` projections) from the donor checkpoint into a new
107
+ `model-mtp.safetensors` shard and patch `model.safetensors.index.json`.
108
+ 3. **Mark the head unquantized** — add every `mtp.*` Linear module (the `fc`, the attention projections,
109
+ and all per-expert MLP projections) to `quantization_config.ignore`, so vLLM's compressed-tensors
110
+ loader keeps the BF16 head as-is instead of expecting MXFP4 weight-scales. This is the one
111
+ mixed-precision gotcha; for a MoE head it means **hundreds** of ignore entries (derived
112
+ automatically from the head's rank-2 tensors).
113
+
114
+ Steps 2–3 are scripted in [`recipe_graft_mxfp4.py`](./recipe_graft_mxfp4.py) — it copies **every**
115
+ `mtp.*` tensor from the donor and derives the ignore list from the rank-2 (matmul) weights, so the same
116
+ script handles both a dense head and this MoE head.
117
+
118
+ ## Credits
119
+ - **DeepReinforce** — [`Ornith-1.0-35B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B), the base/trunk model (MIT).
120
+ - **nerkyor** — [`Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill`](https://huggingface.co/nerkyor/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill), which trained the MoE MTP head grafted here.
121
+ - **DeepSeek** — DeepSeek-V4-Pro-Thinking, the model that distill learned from.
122
+ - **Qwen / Alibaba** — the Qwen MoE (A3B) architecture shared by both the trunk and the grafted head.
123
+ - **vLLM** and **compressed-tensors** — serving stack and quantization format.
124
+ - **Rob Smith** (`tcclaviger`) — the [RDNA4 vLLM base image](https://hub.docker.com/r/tcclaviger/vllm-rocm-mxfp4-nvfp4) that made gfx1201 serving possible. The [`capicua25x/vllm-rocm-rdna4`](https://hub.docker.com/r/capicua25x/vllm-rocm-rdna4) image this model runs on is a forward-port of his work — without it, none of this runs.
125
+
126
+ ## License
127
+ **Apache-2.0.** This combines the Ornith-1.0-35B trunk (MIT, DeepReinforce) with a grafted MTP head
128
+ (Apache-2.0). MIT is one-way compatible with Apache-2.0, so the combined derivative is distributed under
129
+ Apache-2.0; both upstream notices are preserved above in Credits.
chat_template.jinja ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- set image_count = namespace(value=0) %}
2
+ {%- set video_count = namespace(value=0) %}
3
+ {%- macro render_content(content, do_vision_count, is_system_content=false) %}
4
+ {%- if content is string %}
5
+ {{- content }}
6
+ {%- elif content is iterable and content is not mapping %}
7
+ {%- for item in content %}
8
+ {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
9
+ {%- if is_system_content %}
10
+ {{- raise_exception('System message cannot contain images.') }}
11
+ {%- endif %}
12
+ {%- if do_vision_count %}
13
+ {%- set image_count.value = image_count.value + 1 %}
14
+ {%- endif %}
15
+ {%- if add_vision_id %}
16
+ {{- 'Picture ' ~ image_count.value ~ ': ' }}
17
+ {%- endif %}
18
+ {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
19
+ {%- elif 'video' in item or item.type == 'video' %}
20
+ {%- if is_system_content %}
21
+ {{- raise_exception('System message cannot contain videos.') }}
22
+ {%- endif %}
23
+ {%- if do_vision_count %}
24
+ {%- set video_count.value = video_count.value + 1 %}
25
+ {%- endif %}
26
+ {%- if add_vision_id %}
27
+ {{- 'Video ' ~ video_count.value ~ ': ' }}
28
+ {%- endif %}
29
+ {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
30
+ {%- elif 'text' in item %}
31
+ {{- item.text }}
32
+ {%- else %}
33
+ {{- raise_exception('Unexpected item type in content.') }}
34
+ {%- endif %}
35
+ {%- endfor %}
36
+ {%- elif content is none or content is undefined %}
37
+ {{- '' }}
38
+ {%- else %}
39
+ {{- raise_exception('Unexpected content type.') }}
40
+ {%- endif %}
41
+ {%- endmacro %}
42
+ {%- if not messages %}
43
+ {{- raise_exception('No messages provided.') }}
44
+ {%- endif %}
45
+ {%- if tools and tools is iterable and tools is not mapping %}
46
+ {{- '<|im_start|>system\n' }}
47
+ {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
48
+ {%- for tool in tools %}
49
+ {{- "\n" }}
50
+ {{- tool | tojson }}
51
+ {%- endfor %}
52
+ {{- "\n</tools>" }}
53
+ {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
54
+ {%- if messages[0].role == 'system' %}
55
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
56
+ {%- if content %}
57
+ {{- '\n\n' + content }}
58
+ {%- endif %}
59
+ {%- endif %}
60
+ {{- '<|im_end|>\n' }}
61
+ {%- else %}
62
+ {%- if messages[0].role == 'system' %}
63
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
64
+ {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
65
+ {%- endif %}
66
+ {%- endif %}
67
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
68
+ {%- for message in messages[::-1] %}
69
+ {%- set index = (messages|length - 1) - loop.index0 %}
70
+ {%- if ns.multi_step_tool and message.role == "user" %}
71
+ {%- set content = render_content(message.content, false)|trim %}
72
+ {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
73
+ {%- set ns.multi_step_tool = false %}
74
+ {%- set ns.last_query_index = index %}
75
+ {%- endif %}
76
+ {%- endif %}
77
+ {%- endfor %}
78
+ {%- if ns.multi_step_tool %}
79
+ {{- raise_exception('No user query found in messages.') }}
80
+ {%- endif %}
81
+ {%- for message in messages %}
82
+ {%- set content = render_content(message.content, true)|trim %}
83
+ {%- if message.role == "system" %}
84
+ {%- if not loop.first %}
85
+ {{- raise_exception('System message must be at the beginning.') }}
86
+ {%- endif %}
87
+ {%- elif message.role == "user" %}
88
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
89
+ {%- elif message.role == "assistant" %}
90
+ {%- set reasoning_content = '' %}
91
+ {%- if message.reasoning_content is string %}
92
+ {%- set reasoning_content = message.reasoning_content %}
93
+ {%- else %}
94
+ {%- if '</think>' in content %}
95
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
96
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
97
+ {%- endif %}
98
+ {%- endif %}
99
+ {%- set reasoning_content = reasoning_content|trim %}
100
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
101
+ {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
102
+ {%- for tool_call in message.tool_calls %}
103
+ {%- if tool_call.function is defined %}
104
+ {%- set tool_call = tool_call.function %}
105
+ {%- endif %}
106
+ {%- if loop.first %}
107
+ {%- if content|trim %}
108
+ {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
109
+ {%- else %}
110
+ {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
111
+ {%- endif %}
112
+ {%- else %}
113
+ {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
114
+ {%- endif %}
115
+ {%- if tool_call.arguments is defined %}
116
+ {%- for args_name, args_value in tool_call.arguments|items %}
117
+ {{- '<parameter=' + args_name + '>\n' }}
118
+ {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
119
+ {{- args_value }}
120
+ {{- '\n</parameter>\n' }}
121
+ {%- endfor %}
122
+ {%- endif %}
123
+ {{- '</function>\n</tool_call>' }}
124
+ {%- endfor %}
125
+ {%- endif %}
126
+ {{- '<|im_end|>\n' }}
127
+ {%- elif message.role == "tool" %}
128
+ {%- if loop.previtem and loop.previtem.role != "tool" %}
129
+ {{- '<|im_start|>user' }}
130
+ {%- endif %}
131
+ {{- '\n<tool_response>\n' }}
132
+ {{- content }}
133
+ {{- '\n</tool_response>' }}
134
+ {%- if not loop.last and loop.nextitem.role != "tool" %}
135
+ {{- '<|im_end|>\n' }}
136
+ {%- elif loop.last %}
137
+ {{- '<|im_end|>\n' }}
138
+ {%- endif %}
139
+ {%- else %}
140
+ {{- raise_exception('Unexpected message role.') }}
141
+ {%- endif %}
142
+ {%- endfor %}
143
+ {%- if add_generation_prompt %}
144
+ {{- '<|im_start|>assistant\n' }}
145
+ {%- if enable_thinking is defined and enable_thinking is false %}
146
+ {{- '<think>\n\n</think>\n\n' }}
147
+ {%- else %}
148
+ {{- '<think>\n' }}
149
+ {%- endif %}
150
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,1501 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3_5MoeForConditionalGeneration"
4
+ ],
5
+ "bos_token_id": null,
6
+ "dtype": "bfloat16",
7
+ "eos_token_id": 248046,
8
+ "hidden_size": 2048,
9
+ "image_token_id": 248056,
10
+ "model_type": "qwen3_5_moe",
11
+ "pad_token_id": 248044,
12
+ "text_config": {
13
+ "attention_bias": false,
14
+ "attention_dropout": 0.0,
15
+ "attn_output_gate": true,
16
+ "bos_token_id": 248044,
17
+ "dtype": "bfloat16",
18
+ "eos_token_id": 248044,
19
+ "full_attention_interval": 4,
20
+ "head_dim": 256,
21
+ "hidden_act": "silu",
22
+ "hidden_size": 2048,
23
+ "initializer_range": 0.02,
24
+ "layer_types": [
25
+ "linear_attention",
26
+ "linear_attention",
27
+ "linear_attention",
28
+ "full_attention",
29
+ "linear_attention",
30
+ "linear_attention",
31
+ "linear_attention",
32
+ "full_attention",
33
+ "linear_attention",
34
+ "linear_attention",
35
+ "linear_attention",
36
+ "full_attention",
37
+ "linear_attention",
38
+ "linear_attention",
39
+ "linear_attention",
40
+ "full_attention",
41
+ "linear_attention",
42
+ "linear_attention",
43
+ "linear_attention",
44
+ "full_attention",
45
+ "linear_attention",
46
+ "linear_attention",
47
+ "linear_attention",
48
+ "full_attention",
49
+ "linear_attention",
50
+ "linear_attention",
51
+ "linear_attention",
52
+ "full_attention",
53
+ "linear_attention",
54
+ "linear_attention",
55
+ "linear_attention",
56
+ "full_attention",
57
+ "linear_attention",
58
+ "linear_attention",
59
+ "linear_attention",
60
+ "full_attention",
61
+ "linear_attention",
62
+ "linear_attention",
63
+ "linear_attention",
64
+ "full_attention"
65
+ ],
66
+ "linear_conv_kernel_dim": 4,
67
+ "linear_key_head_dim": 128,
68
+ "linear_num_key_heads": 16,
69
+ "linear_num_value_heads": 32,
70
+ "linear_value_head_dim": 128,
71
+ "mamba_ssm_dtype": "float32",
72
+ "max_position_embeddings": 262144,
73
+ "model_type": "qwen3_5_moe_text",
74
+ "moe_intermediate_size": 512,
75
+ "mtp_num_hidden_layers": 1,
76
+ "mtp_use_dedicated_embeddings": false,
77
+ "num_attention_heads": 16,
78
+ "num_experts": 256,
79
+ "num_experts_per_tok": 8,
80
+ "num_hidden_layers": 40,
81
+ "num_key_value_heads": 2,
82
+ "output_router_logits": false,
83
+ "pad_token_id": 248044,
84
+ "partial_rotary_factor": 0.25,
85
+ "rms_norm_eps": 1e-06,
86
+ "rope_parameters": {
87
+ "mrope_interleaved": true,
88
+ "mrope_section": [
89
+ 11,
90
+ 11,
91
+ 10
92
+ ],
93
+ "partial_rotary_factor": 0.25,
94
+ "rope_theta": 10000000,
95
+ "rope_type": "default"
96
+ },
97
+ "router_aux_loss_coef": 0.0,
98
+ "shared_expert_intermediate_size": 512,
99
+ "tie_word_embeddings": false,
100
+ "use_cache": false,
101
+ "vocab_size": 248320
102
+ },
103
+ "tie_word_embeddings": false,
104
+ "transformers_version": "5.8.1",
105
+ "use_cache": false,
106
+ "video_token_id": 248057,
107
+ "vision_config": {
108
+ "deepstack_visual_indexes": [],
109
+ "depth": 27,
110
+ "dtype": "bfloat16",
111
+ "hidden_act": "gelu_pytorch_tanh",
112
+ "hidden_size": 1152,
113
+ "in_channels": 3,
114
+ "initializer_range": 0.02,
115
+ "intermediate_size": 4304,
116
+ "model_type": "qwen3_5_moe_vision",
117
+ "num_heads": 16,
118
+ "num_position_embeddings": 2304,
119
+ "out_hidden_size": 2048,
120
+ "patch_size": 16,
121
+ "spatial_merge_size": 2,
122
+ "temporal_patch_size": 2
123
+ },
124
+ "vision_end_token_id": 248054,
125
+ "vision_start_token_id": 248053,
126
+ "quantization_config": {
127
+ "quant_method": "compressed-tensors",
128
+ "format": "mxfp4-pack-quantized",
129
+ "config_groups": {
130
+ "group_0": {
131
+ "targets": [
132
+ "Linear"
133
+ ],
134
+ "weights": {
135
+ "num_bits": 4,
136
+ "type": "float",
137
+ "strategy": "group",
138
+ "group_size": 32,
139
+ "symmetric": true
140
+ }
141
+ }
142
+ },
143
+ "ignore": [
144
+ "lm_head",
145
+ "model.language_model.embed_tokens",
146
+ "model.language_model.layers.0.input_layernorm",
147
+ "model.language_model.layers.0.linear_attn.conv1d",
148
+ "model.language_model.layers.0.linear_attn.norm",
149
+ "model.language_model.layers.0.mlp.gate",
150
+ "model.language_model.layers.0.mlp.shared_expert.down_proj",
151
+ "model.language_model.layers.0.mlp.shared_expert.gate_proj",
152
+ "model.language_model.layers.0.mlp.shared_expert.up_proj",
153
+ "model.language_model.layers.0.mlp.shared_expert_gate",
154
+ "model.language_model.layers.0.post_attention_layernorm",
155
+ "model.language_model.layers.1.input_layernorm",
156
+ "model.language_model.layers.1.linear_attn.conv1d",
157
+ "model.language_model.layers.1.linear_attn.norm",
158
+ "model.language_model.layers.1.mlp.gate",
159
+ "model.language_model.layers.1.mlp.shared_expert.down_proj",
160
+ "model.language_model.layers.1.mlp.shared_expert.gate_proj",
161
+ "model.language_model.layers.1.mlp.shared_expert.up_proj",
162
+ "model.language_model.layers.1.mlp.shared_expert_gate",
163
+ "model.language_model.layers.1.post_attention_layernorm",
164
+ "model.language_model.layers.10.input_layernorm",
165
+ "model.language_model.layers.10.linear_attn.conv1d",
166
+ "model.language_model.layers.10.linear_attn.norm",
167
+ "model.language_model.layers.10.mlp.gate",
168
+ "model.language_model.layers.10.mlp.shared_expert.down_proj",
169
+ "model.language_model.layers.10.mlp.shared_expert.gate_proj",
170
+ "model.language_model.layers.10.mlp.shared_expert.up_proj",
171
+ "model.language_model.layers.10.mlp.shared_expert_gate",
172
+ "model.language_model.layers.10.post_attention_layernorm",
173
+ "model.language_model.layers.11.input_layernorm",
174
+ "model.language_model.layers.11.mlp.gate",
175
+ "model.language_model.layers.11.mlp.shared_expert.down_proj",
176
+ "model.language_model.layers.11.mlp.shared_expert.gate_proj",
177
+ "model.language_model.layers.11.mlp.shared_expert.up_proj",
178
+ "model.language_model.layers.11.mlp.shared_expert_gate",
179
+ "model.language_model.layers.11.post_attention_layernorm",
180
+ "model.language_model.layers.11.self_attn.k_norm",
181
+ "model.language_model.layers.11.self_attn.k_proj",
182
+ "model.language_model.layers.11.self_attn.o_proj",
183
+ "model.language_model.layers.11.self_attn.q_norm",
184
+ "model.language_model.layers.11.self_attn.q_proj",
185
+ "model.language_model.layers.11.self_attn.v_proj",
186
+ "model.language_model.layers.12.input_layernorm",
187
+ "model.language_model.layers.12.linear_attn.conv1d",
188
+ "model.language_model.layers.12.linear_attn.norm",
189
+ "model.language_model.layers.12.mlp.gate",
190
+ "model.language_model.layers.12.mlp.shared_expert.down_proj",
191
+ "model.language_model.layers.12.mlp.shared_expert.gate_proj",
192
+ "model.language_model.layers.12.mlp.shared_expert.up_proj",
193
+ "model.language_model.layers.12.mlp.shared_expert_gate",
194
+ "model.language_model.layers.12.post_attention_layernorm",
195
+ "model.language_model.layers.13.input_layernorm",
196
+ "model.language_model.layers.13.linear_attn.conv1d",
197
+ "model.language_model.layers.13.linear_attn.norm",
198
+ "model.language_model.layers.13.mlp.gate",
199
+ "model.language_model.layers.13.mlp.shared_expert.down_proj",
200
+ "model.language_model.layers.13.mlp.shared_expert.gate_proj",
201
+ "model.language_model.layers.13.mlp.shared_expert.up_proj",
202
+ "model.language_model.layers.13.mlp.shared_expert_gate",
203
+ "model.language_model.layers.13.post_attention_layernorm",
204
+ "model.language_model.layers.14.input_layernorm",
205
+ "model.language_model.layers.14.linear_attn.conv1d",
206
+ "model.language_model.layers.14.linear_attn.norm",
207
+ "model.language_model.layers.14.mlp.gate",
208
+ "model.language_model.layers.14.mlp.shared_expert.down_proj",
209
+ "model.language_model.layers.14.mlp.shared_expert.gate_proj",
210
+ "model.language_model.layers.14.mlp.shared_expert.up_proj",
211
+ "model.language_model.layers.14.mlp.shared_expert_gate",
212
+ "model.language_model.layers.14.post_attention_layernorm",
213
+ "model.language_model.layers.15.input_layernorm",
214
+ "model.language_model.layers.15.mlp.gate",
215
+ "model.language_model.layers.15.mlp.shared_expert.down_proj",
216
+ "model.language_model.layers.15.mlp.shared_expert.gate_proj",
217
+ "model.language_model.layers.15.mlp.shared_expert.up_proj",
218
+ "model.language_model.layers.15.mlp.shared_expert_gate",
219
+ "model.language_model.layers.15.post_attention_layernorm",
220
+ "model.language_model.layers.15.self_attn.k_norm",
221
+ "model.language_model.layers.15.self_attn.k_proj",
222
+ "model.language_model.layers.15.self_attn.o_proj",
223
+ "model.language_model.layers.15.self_attn.q_norm",
224
+ "model.language_model.layers.15.self_attn.q_proj",
225
+ "model.language_model.layers.15.self_attn.v_proj",
226
+ "model.language_model.layers.16.input_layernorm",
227
+ "model.language_model.layers.16.linear_attn.conv1d",
228
+ "model.language_model.layers.16.linear_attn.norm",
229
+ "model.language_model.layers.16.mlp.gate",
230
+ "model.language_model.layers.16.mlp.shared_expert.down_proj",
231
+ "model.language_model.layers.16.mlp.shared_expert.gate_proj",
232
+ "model.language_model.layers.16.mlp.shared_expert.up_proj",
233
+ "model.language_model.layers.16.mlp.shared_expert_gate",
234
+ "model.language_model.layers.16.post_attention_layernorm",
235
+ "model.language_model.layers.17.input_layernorm",
236
+ "model.language_model.layers.17.linear_attn.conv1d",
237
+ "model.language_model.layers.17.linear_attn.norm",
238
+ "model.language_model.layers.17.mlp.gate",
239
+ "model.language_model.layers.17.mlp.shared_expert.down_proj",
240
+ "model.language_model.layers.17.mlp.shared_expert.gate_proj",
241
+ "model.language_model.layers.17.mlp.shared_expert.up_proj",
242
+ "model.language_model.layers.17.mlp.shared_expert_gate",
243
+ "model.language_model.layers.17.post_attention_layernorm",
244
+ "model.language_model.layers.18.input_layernorm",
245
+ "model.language_model.layers.18.linear_attn.conv1d",
246
+ "model.language_model.layers.18.linear_attn.norm",
247
+ "model.language_model.layers.18.mlp.gate",
248
+ "model.language_model.layers.18.mlp.shared_expert.down_proj",
249
+ "model.language_model.layers.18.mlp.shared_expert.gate_proj",
250
+ "model.language_model.layers.18.mlp.shared_expert.up_proj",
251
+ "model.language_model.layers.18.mlp.shared_expert_gate",
252
+ "model.language_model.layers.18.post_attention_layernorm",
253
+ "model.language_model.layers.19.input_layernorm",
254
+ "model.language_model.layers.19.mlp.gate",
255
+ "model.language_model.layers.19.mlp.shared_expert.down_proj",
256
+ "model.language_model.layers.19.mlp.shared_expert.gate_proj",
257
+ "model.language_model.layers.19.mlp.shared_expert.up_proj",
258
+ "model.language_model.layers.19.mlp.shared_expert_gate",
259
+ "model.language_model.layers.19.post_attention_layernorm",
260
+ "model.language_model.layers.19.self_attn.k_norm",
261
+ "model.language_model.layers.19.self_attn.k_proj",
262
+ "model.language_model.layers.19.self_attn.o_proj",
263
+ "model.language_model.layers.19.self_attn.q_norm",
264
+ "model.language_model.layers.19.self_attn.q_proj",
265
+ "model.language_model.layers.19.self_attn.v_proj",
266
+ "model.language_model.layers.2.input_layernorm",
267
+ "model.language_model.layers.2.linear_attn.conv1d",
268
+ "model.language_model.layers.2.linear_attn.norm",
269
+ "model.language_model.layers.2.mlp.gate",
270
+ "model.language_model.layers.2.mlp.shared_expert.down_proj",
271
+ "model.language_model.layers.2.mlp.shared_expert.gate_proj",
272
+ "model.language_model.layers.2.mlp.shared_expert.up_proj",
273
+ "model.language_model.layers.2.mlp.shared_expert_gate",
274
+ "model.language_model.layers.2.post_attention_layernorm",
275
+ "model.language_model.layers.20.input_layernorm",
276
+ "model.language_model.layers.20.linear_attn.conv1d",
277
+ "model.language_model.layers.20.linear_attn.norm",
278
+ "model.language_model.layers.20.mlp.gate",
279
+ "model.language_model.layers.20.mlp.shared_expert.down_proj",
280
+ "model.language_model.layers.20.mlp.shared_expert.gate_proj",
281
+ "model.language_model.layers.20.mlp.shared_expert.up_proj",
282
+ "model.language_model.layers.20.mlp.shared_expert_gate",
283
+ "model.language_model.layers.20.post_attention_layernorm",
284
+ "model.language_model.layers.21.input_layernorm",
285
+ "model.language_model.layers.21.linear_attn.conv1d",
286
+ "model.language_model.layers.21.linear_attn.norm",
287
+ "model.language_model.layers.21.mlp.gate",
288
+ "model.language_model.layers.21.mlp.shared_expert.down_proj",
289
+ "model.language_model.layers.21.mlp.shared_expert.gate_proj",
290
+ "model.language_model.layers.21.mlp.shared_expert.up_proj",
291
+ "model.language_model.layers.21.mlp.shared_expert_gate",
292
+ "model.language_model.layers.21.post_attention_layernorm",
293
+ "model.language_model.layers.22.input_layernorm",
294
+ "model.language_model.layers.22.linear_attn.conv1d",
295
+ "model.language_model.layers.22.linear_attn.norm",
296
+ "model.language_model.layers.22.mlp.gate",
297
+ "model.language_model.layers.22.mlp.shared_expert.down_proj",
298
+ "model.language_model.layers.22.mlp.shared_expert.gate_proj",
299
+ "model.language_model.layers.22.mlp.shared_expert.up_proj",
300
+ "model.language_model.layers.22.mlp.shared_expert_gate",
301
+ "model.language_model.layers.22.post_attention_layernorm",
302
+ "model.language_model.layers.23.input_layernorm",
303
+ "model.language_model.layers.23.mlp.gate",
304
+ "model.language_model.layers.23.mlp.shared_expert.down_proj",
305
+ "model.language_model.layers.23.mlp.shared_expert.gate_proj",
306
+ "model.language_model.layers.23.mlp.shared_expert.up_proj",
307
+ "model.language_model.layers.23.mlp.shared_expert_gate",
308
+ "model.language_model.layers.23.post_attention_layernorm",
309
+ "model.language_model.layers.23.self_attn.k_norm",
310
+ "model.language_model.layers.23.self_attn.k_proj",
311
+ "model.language_model.layers.23.self_attn.o_proj",
312
+ "model.language_model.layers.23.self_attn.q_norm",
313
+ "model.language_model.layers.23.self_attn.q_proj",
314
+ "model.language_model.layers.23.self_attn.v_proj",
315
+ "model.language_model.layers.24.input_layernorm",
316
+ "model.language_model.layers.24.linear_attn.conv1d",
317
+ "model.language_model.layers.24.linear_attn.norm",
318
+ "model.language_model.layers.24.mlp.gate",
319
+ "model.language_model.layers.24.mlp.shared_expert.down_proj",
320
+ "model.language_model.layers.24.mlp.shared_expert.gate_proj",
321
+ "model.language_model.layers.24.mlp.shared_expert.up_proj",
322
+ "model.language_model.layers.24.mlp.shared_expert_gate",
323
+ "model.language_model.layers.24.post_attention_layernorm",
324
+ "model.language_model.layers.25.input_layernorm",
325
+ "model.language_model.layers.25.linear_attn.conv1d",
326
+ "model.language_model.layers.25.linear_attn.norm",
327
+ "model.language_model.layers.25.mlp.gate",
328
+ "model.language_model.layers.25.mlp.shared_expert.down_proj",
329
+ "model.language_model.layers.25.mlp.shared_expert.gate_proj",
330
+ "model.language_model.layers.25.mlp.shared_expert.up_proj",
331
+ "model.language_model.layers.25.mlp.shared_expert_gate",
332
+ "model.language_model.layers.25.post_attention_layernorm",
333
+ "model.language_model.layers.26.input_layernorm",
334
+ "model.language_model.layers.26.linear_attn.conv1d",
335
+ "model.language_model.layers.26.linear_attn.norm",
336
+ "model.language_model.layers.26.mlp.gate",
337
+ "model.language_model.layers.26.mlp.shared_expert.down_proj",
338
+ "model.language_model.layers.26.mlp.shared_expert.gate_proj",
339
+ "model.language_model.layers.26.mlp.shared_expert.up_proj",
340
+ "model.language_model.layers.26.mlp.shared_expert_gate",
341
+ "model.language_model.layers.26.post_attention_layernorm",
342
+ "model.language_model.layers.27.input_layernorm",
343
+ "model.language_model.layers.27.mlp.gate",
344
+ "model.language_model.layers.27.mlp.shared_expert.down_proj",
345
+ "model.language_model.layers.27.mlp.shared_expert.gate_proj",
346
+ "model.language_model.layers.27.mlp.shared_expert.up_proj",
347
+ "model.language_model.layers.27.mlp.shared_expert_gate",
348
+ "model.language_model.layers.27.post_attention_layernorm",
349
+ "model.language_model.layers.27.self_attn.k_norm",
350
+ "model.language_model.layers.27.self_attn.k_proj",
351
+ "model.language_model.layers.27.self_attn.o_proj",
352
+ "model.language_model.layers.27.self_attn.q_norm",
353
+ "model.language_model.layers.27.self_attn.q_proj",
354
+ "model.language_model.layers.27.self_attn.v_proj",
355
+ "model.language_model.layers.28.input_layernorm",
356
+ "model.language_model.layers.28.linear_attn.conv1d",
357
+ "model.language_model.layers.28.linear_attn.norm",
358
+ "model.language_model.layers.28.mlp.gate",
359
+ "model.language_model.layers.28.mlp.shared_expert.down_proj",
360
+ "model.language_model.layers.28.mlp.shared_expert.gate_proj",
361
+ "model.language_model.layers.28.mlp.shared_expert.up_proj",
362
+ "model.language_model.layers.28.mlp.shared_expert_gate",
363
+ "model.language_model.layers.28.post_attention_layernorm",
364
+ "model.language_model.layers.29.input_layernorm",
365
+ "model.language_model.layers.29.linear_attn.conv1d",
366
+ "model.language_model.layers.29.linear_attn.norm",
367
+ "model.language_model.layers.29.mlp.gate",
368
+ "model.language_model.layers.29.mlp.shared_expert.down_proj",
369
+ "model.language_model.layers.29.mlp.shared_expert.gate_proj",
370
+ "model.language_model.layers.29.mlp.shared_expert.up_proj",
371
+ "model.language_model.layers.29.mlp.shared_expert_gate",
372
+ "model.language_model.layers.29.post_attention_layernorm",
373
+ "model.language_model.layers.3.input_layernorm",
374
+ "model.language_model.layers.3.mlp.gate",
375
+ "model.language_model.layers.3.mlp.shared_expert.down_proj",
376
+ "model.language_model.layers.3.mlp.shared_expert.gate_proj",
377
+ "model.language_model.layers.3.mlp.shared_expert.up_proj",
378
+ "model.language_model.layers.3.mlp.shared_expert_gate",
379
+ "model.language_model.layers.3.post_attention_layernorm",
380
+ "model.language_model.layers.3.self_attn.k_norm",
381
+ "model.language_model.layers.3.self_attn.k_proj",
382
+ "model.language_model.layers.3.self_attn.o_proj",
383
+ "model.language_model.layers.3.self_attn.q_norm",
384
+ "model.language_model.layers.3.self_attn.q_proj",
385
+ "model.language_model.layers.3.self_attn.v_proj",
386
+ "model.language_model.layers.30.input_layernorm",
387
+ "model.language_model.layers.30.linear_attn.conv1d",
388
+ "model.language_model.layers.30.linear_attn.norm",
389
+ "model.language_model.layers.30.mlp.gate",
390
+ "model.language_model.layers.30.mlp.shared_expert.down_proj",
391
+ "model.language_model.layers.30.mlp.shared_expert.gate_proj",
392
+ "model.language_model.layers.30.mlp.shared_expert.up_proj",
393
+ "model.language_model.layers.30.mlp.shared_expert_gate",
394
+ "model.language_model.layers.30.post_attention_layernorm",
395
+ "model.language_model.layers.31.input_layernorm",
396
+ "model.language_model.layers.31.mlp.gate",
397
+ "model.language_model.layers.31.mlp.shared_expert.down_proj",
398
+ "model.language_model.layers.31.mlp.shared_expert.gate_proj",
399
+ "model.language_model.layers.31.mlp.shared_expert.up_proj",
400
+ "model.language_model.layers.31.mlp.shared_expert_gate",
401
+ "model.language_model.layers.31.post_attention_layernorm",
402
+ "model.language_model.layers.31.self_attn.k_norm",
403
+ "model.language_model.layers.31.self_attn.k_proj",
404
+ "model.language_model.layers.31.self_attn.o_proj",
405
+ "model.language_model.layers.31.self_attn.q_norm",
406
+ "model.language_model.layers.31.self_attn.q_proj",
407
+ "model.language_model.layers.31.self_attn.v_proj",
408
+ "model.language_model.layers.32.input_layernorm",
409
+ "model.language_model.layers.32.linear_attn.conv1d",
410
+ "model.language_model.layers.32.linear_attn.norm",
411
+ "model.language_model.layers.32.mlp.gate",
412
+ "model.language_model.layers.32.mlp.shared_expert.down_proj",
413
+ "model.language_model.layers.32.mlp.shared_expert.gate_proj",
414
+ "model.language_model.layers.32.mlp.shared_expert.up_proj",
415
+ "model.language_model.layers.32.mlp.shared_expert_gate",
416
+ "model.language_model.layers.32.post_attention_layernorm",
417
+ "model.language_model.layers.33.input_layernorm",
418
+ "model.language_model.layers.33.linear_attn.conv1d",
419
+ "model.language_model.layers.33.linear_attn.norm",
420
+ "model.language_model.layers.33.mlp.gate",
421
+ "model.language_model.layers.33.mlp.shared_expert.down_proj",
422
+ "model.language_model.layers.33.mlp.shared_expert.gate_proj",
423
+ "model.language_model.layers.33.mlp.shared_expert.up_proj",
424
+ "model.language_model.layers.33.mlp.shared_expert_gate",
425
+ "model.language_model.layers.33.post_attention_layernorm",
426
+ "model.language_model.layers.34.input_layernorm",
427
+ "model.language_model.layers.34.linear_attn.conv1d",
428
+ "model.language_model.layers.34.linear_attn.norm",
429
+ "model.language_model.layers.34.mlp.gate",
430
+ "model.language_model.layers.34.mlp.shared_expert.down_proj",
431
+ "model.language_model.layers.34.mlp.shared_expert.gate_proj",
432
+ "model.language_model.layers.34.mlp.shared_expert.up_proj",
433
+ "model.language_model.layers.34.mlp.shared_expert_gate",
434
+ "model.language_model.layers.34.post_attention_layernorm",
435
+ "model.language_model.layers.35.input_layernorm",
436
+ "model.language_model.layers.35.mlp.gate",
437
+ "model.language_model.layers.35.mlp.shared_expert.down_proj",
438
+ "model.language_model.layers.35.mlp.shared_expert.gate_proj",
439
+ "model.language_model.layers.35.mlp.shared_expert.up_proj",
440
+ "model.language_model.layers.35.mlp.shared_expert_gate",
441
+ "model.language_model.layers.35.post_attention_layernorm",
442
+ "model.language_model.layers.35.self_attn.k_norm",
443
+ "model.language_model.layers.35.self_attn.k_proj",
444
+ "model.language_model.layers.35.self_attn.o_proj",
445
+ "model.language_model.layers.35.self_attn.q_norm",
446
+ "model.language_model.layers.35.self_attn.q_proj",
447
+ "model.language_model.layers.35.self_attn.v_proj",
448
+ "model.language_model.layers.36.input_layernorm",
449
+ "model.language_model.layers.36.linear_attn.conv1d",
450
+ "model.language_model.layers.36.linear_attn.norm",
451
+ "model.language_model.layers.36.mlp.gate",
452
+ "model.language_model.layers.36.mlp.shared_expert.down_proj",
453
+ "model.language_model.layers.36.mlp.shared_expert.gate_proj",
454
+ "model.language_model.layers.36.mlp.shared_expert.up_proj",
455
+ "model.language_model.layers.36.mlp.shared_expert_gate",
456
+ "model.language_model.layers.36.post_attention_layernorm",
457
+ "model.language_model.layers.37.input_layernorm",
458
+ "model.language_model.layers.37.linear_attn.conv1d",
459
+ "model.language_model.layers.37.linear_attn.norm",
460
+ "model.language_model.layers.37.mlp.gate",
461
+ "model.language_model.layers.37.mlp.shared_expert.down_proj",
462
+ "model.language_model.layers.37.mlp.shared_expert.gate_proj",
463
+ "model.language_model.layers.37.mlp.shared_expert.up_proj",
464
+ "model.language_model.layers.37.mlp.shared_expert_gate",
465
+ "model.language_model.layers.37.post_attention_layernorm",
466
+ "model.language_model.layers.38.input_layernorm",
467
+ "model.language_model.layers.38.linear_attn.conv1d",
468
+ "model.language_model.layers.38.linear_attn.norm",
469
+ "model.language_model.layers.38.mlp.gate",
470
+ "model.language_model.layers.38.mlp.shared_expert.down_proj",
471
+ "model.language_model.layers.38.mlp.shared_expert.gate_proj",
472
+ "model.language_model.layers.38.mlp.shared_expert.up_proj",
473
+ "model.language_model.layers.38.mlp.shared_expert_gate",
474
+ "model.language_model.layers.38.post_attention_layernorm",
475
+ "model.language_model.layers.39.input_layernorm",
476
+ "model.language_model.layers.39.mlp.gate",
477
+ "model.language_model.layers.39.mlp.shared_expert.down_proj",
478
+ "model.language_model.layers.39.mlp.shared_expert.gate_proj",
479
+ "model.language_model.layers.39.mlp.shared_expert.up_proj",
480
+ "model.language_model.layers.39.mlp.shared_expert_gate",
481
+ "model.language_model.layers.39.post_attention_layernorm",
482
+ "model.language_model.layers.39.self_attn.k_norm",
483
+ "model.language_model.layers.39.self_attn.k_proj",
484
+ "model.language_model.layers.39.self_attn.o_proj",
485
+ "model.language_model.layers.39.self_attn.q_norm",
486
+ "model.language_model.layers.39.self_attn.q_proj",
487
+ "model.language_model.layers.39.self_attn.v_proj",
488
+ "model.language_model.layers.4.input_layernorm",
489
+ "model.language_model.layers.4.linear_attn.conv1d",
490
+ "model.language_model.layers.4.linear_attn.norm",
491
+ "model.language_model.layers.4.mlp.gate",
492
+ "model.language_model.layers.4.mlp.shared_expert.down_proj",
493
+ "model.language_model.layers.4.mlp.shared_expert.gate_proj",
494
+ "model.language_model.layers.4.mlp.shared_expert.up_proj",
495
+ "model.language_model.layers.4.mlp.shared_expert_gate",
496
+ "model.language_model.layers.4.post_attention_layernorm",
497
+ "model.language_model.layers.5.input_layernorm",
498
+ "model.language_model.layers.5.linear_attn.conv1d",
499
+ "model.language_model.layers.5.linear_attn.norm",
500
+ "model.language_model.layers.5.mlp.gate",
501
+ "model.language_model.layers.5.mlp.shared_expert.down_proj",
502
+ "model.language_model.layers.5.mlp.shared_expert.gate_proj",
503
+ "model.language_model.layers.5.mlp.shared_expert.up_proj",
504
+ "model.language_model.layers.5.mlp.shared_expert_gate",
505
+ "model.language_model.layers.5.post_attention_layernorm",
506
+ "model.language_model.layers.6.input_layernorm",
507
+ "model.language_model.layers.6.linear_attn.conv1d",
508
+ "model.language_model.layers.6.linear_attn.norm",
509
+ "model.language_model.layers.6.mlp.gate",
510
+ "model.language_model.layers.6.mlp.shared_expert.down_proj",
511
+ "model.language_model.layers.6.mlp.shared_expert.gate_proj",
512
+ "model.language_model.layers.6.mlp.shared_expert.up_proj",
513
+ "model.language_model.layers.6.mlp.shared_expert_gate",
514
+ "model.language_model.layers.6.post_attention_layernorm",
515
+ "model.language_model.layers.7.input_layernorm",
516
+ "model.language_model.layers.7.mlp.gate",
517
+ "model.language_model.layers.7.mlp.shared_expert.down_proj",
518
+ "model.language_model.layers.7.mlp.shared_expert.gate_proj",
519
+ "model.language_model.layers.7.mlp.shared_expert.up_proj",
520
+ "model.language_model.layers.7.mlp.shared_expert_gate",
521
+ "model.language_model.layers.7.post_attention_layernorm",
522
+ "model.language_model.layers.7.self_attn.k_norm",
523
+ "model.language_model.layers.7.self_attn.k_proj",
524
+ "model.language_model.layers.7.self_attn.o_proj",
525
+ "model.language_model.layers.7.self_attn.q_norm",
526
+ "model.language_model.layers.7.self_attn.q_proj",
527
+ "model.language_model.layers.7.self_attn.v_proj",
528
+ "model.language_model.layers.8.input_layernorm",
529
+ "model.language_model.layers.8.linear_attn.conv1d",
530
+ "model.language_model.layers.8.linear_attn.norm",
531
+ "model.language_model.layers.8.mlp.gate",
532
+ "model.language_model.layers.8.mlp.shared_expert.down_proj",
533
+ "model.language_model.layers.8.mlp.shared_expert.gate_proj",
534
+ "model.language_model.layers.8.mlp.shared_expert.up_proj",
535
+ "model.language_model.layers.8.mlp.shared_expert_gate",
536
+ "model.language_model.layers.8.post_attention_layernorm",
537
+ "model.language_model.layers.9.input_layernorm",
538
+ "model.language_model.layers.9.linear_attn.conv1d",
539
+ "model.language_model.layers.9.linear_attn.norm",
540
+ "model.language_model.layers.9.mlp.gate",
541
+ "model.language_model.layers.9.mlp.shared_expert.down_proj",
542
+ "model.language_model.layers.9.mlp.shared_expert.gate_proj",
543
+ "model.language_model.layers.9.mlp.shared_expert.up_proj",
544
+ "model.language_model.layers.9.mlp.shared_expert_gate",
545
+ "model.language_model.layers.9.post_attention_layernorm",
546
+ "model.language_model.norm",
547
+ "model.visual.blocks.0.attn.proj",
548
+ "model.visual.blocks.0.attn.qkv",
549
+ "model.visual.blocks.0.mlp.linear_fc1",
550
+ "model.visual.blocks.0.mlp.linear_fc2",
551
+ "model.visual.blocks.0.norm1",
552
+ "model.visual.blocks.0.norm2",
553
+ "model.visual.blocks.1.attn.proj",
554
+ "model.visual.blocks.1.attn.qkv",
555
+ "model.visual.blocks.1.mlp.linear_fc1",
556
+ "model.visual.blocks.1.mlp.linear_fc2",
557
+ "model.visual.blocks.1.norm1",
558
+ "model.visual.blocks.1.norm2",
559
+ "model.visual.blocks.10.attn.proj",
560
+ "model.visual.blocks.10.attn.qkv",
561
+ "model.visual.blocks.10.mlp.linear_fc1",
562
+ "model.visual.blocks.10.mlp.linear_fc2",
563
+ "model.visual.blocks.10.norm1",
564
+ "model.visual.blocks.10.norm2",
565
+ "model.visual.blocks.11.attn.proj",
566
+ "model.visual.blocks.11.attn.qkv",
567
+ "model.visual.blocks.11.mlp.linear_fc1",
568
+ "model.visual.blocks.11.mlp.linear_fc2",
569
+ "model.visual.blocks.11.norm1",
570
+ "model.visual.blocks.11.norm2",
571
+ "model.visual.blocks.12.attn.proj",
572
+ "model.visual.blocks.12.attn.qkv",
573
+ "model.visual.blocks.12.mlp.linear_fc1",
574
+ "model.visual.blocks.12.mlp.linear_fc2",
575
+ "model.visual.blocks.12.norm1",
576
+ "model.visual.blocks.12.norm2",
577
+ "model.visual.blocks.13.attn.proj",
578
+ "model.visual.blocks.13.attn.qkv",
579
+ "model.visual.blocks.13.mlp.linear_fc1",
580
+ "model.visual.blocks.13.mlp.linear_fc2",
581
+ "model.visual.blocks.13.norm1",
582
+ "model.visual.blocks.13.norm2",
583
+ "model.visual.blocks.14.attn.proj",
584
+ "model.visual.blocks.14.attn.qkv",
585
+ "model.visual.blocks.14.mlp.linear_fc1",
586
+ "model.visual.blocks.14.mlp.linear_fc2",
587
+ "model.visual.blocks.14.norm1",
588
+ "model.visual.blocks.14.norm2",
589
+ "model.visual.blocks.15.attn.proj",
590
+ "model.visual.blocks.15.attn.qkv",
591
+ "model.visual.blocks.15.mlp.linear_fc1",
592
+ "model.visual.blocks.15.mlp.linear_fc2",
593
+ "model.visual.blocks.15.norm1",
594
+ "model.visual.blocks.15.norm2",
595
+ "model.visual.blocks.16.attn.proj",
596
+ "model.visual.blocks.16.attn.qkv",
597
+ "model.visual.blocks.16.mlp.linear_fc1",
598
+ "model.visual.blocks.16.mlp.linear_fc2",
599
+ "model.visual.blocks.16.norm1",
600
+ "model.visual.blocks.16.norm2",
601
+ "model.visual.blocks.17.attn.proj",
602
+ "model.visual.blocks.17.attn.qkv",
603
+ "model.visual.blocks.17.mlp.linear_fc1",
604
+ "model.visual.blocks.17.mlp.linear_fc2",
605
+ "model.visual.blocks.17.norm1",
606
+ "model.visual.blocks.17.norm2",
607
+ "model.visual.blocks.18.attn.proj",
608
+ "model.visual.blocks.18.attn.qkv",
609
+ "model.visual.blocks.18.mlp.linear_fc1",
610
+ "model.visual.blocks.18.mlp.linear_fc2",
611
+ "model.visual.blocks.18.norm1",
612
+ "model.visual.blocks.18.norm2",
613
+ "model.visual.blocks.19.attn.proj",
614
+ "model.visual.blocks.19.attn.qkv",
615
+ "model.visual.blocks.19.mlp.linear_fc1",
616
+ "model.visual.blocks.19.mlp.linear_fc2",
617
+ "model.visual.blocks.19.norm1",
618
+ "model.visual.blocks.19.norm2",
619
+ "model.visual.blocks.2.attn.proj",
620
+ "model.visual.blocks.2.attn.qkv",
621
+ "model.visual.blocks.2.mlp.linear_fc1",
622
+ "model.visual.blocks.2.mlp.linear_fc2",
623
+ "model.visual.blocks.2.norm1",
624
+ "model.visual.blocks.2.norm2",
625
+ "model.visual.blocks.20.attn.proj",
626
+ "model.visual.blocks.20.attn.qkv",
627
+ "model.visual.blocks.20.mlp.linear_fc1",
628
+ "model.visual.blocks.20.mlp.linear_fc2",
629
+ "model.visual.blocks.20.norm1",
630
+ "model.visual.blocks.20.norm2",
631
+ "model.visual.blocks.21.attn.proj",
632
+ "model.visual.blocks.21.attn.qkv",
633
+ "model.visual.blocks.21.mlp.linear_fc1",
634
+ "model.visual.blocks.21.mlp.linear_fc2",
635
+ "model.visual.blocks.21.norm1",
636
+ "model.visual.blocks.21.norm2",
637
+ "model.visual.blocks.22.attn.proj",
638
+ "model.visual.blocks.22.attn.qkv",
639
+ "model.visual.blocks.22.mlp.linear_fc1",
640
+ "model.visual.blocks.22.mlp.linear_fc2",
641
+ "model.visual.blocks.22.norm1",
642
+ "model.visual.blocks.22.norm2",
643
+ "model.visual.blocks.23.attn.proj",
644
+ "model.visual.blocks.23.attn.qkv",
645
+ "model.visual.blocks.23.mlp.linear_fc1",
646
+ "model.visual.blocks.23.mlp.linear_fc2",
647
+ "model.visual.blocks.23.norm1",
648
+ "model.visual.blocks.23.norm2",
649
+ "model.visual.blocks.24.attn.proj",
650
+ "model.visual.blocks.24.attn.qkv",
651
+ "model.visual.blocks.24.mlp.linear_fc1",
652
+ "model.visual.blocks.24.mlp.linear_fc2",
653
+ "model.visual.blocks.24.norm1",
654
+ "model.visual.blocks.24.norm2",
655
+ "model.visual.blocks.25.attn.proj",
656
+ "model.visual.blocks.25.attn.qkv",
657
+ "model.visual.blocks.25.mlp.linear_fc1",
658
+ "model.visual.blocks.25.mlp.linear_fc2",
659
+ "model.visual.blocks.25.norm1",
660
+ "model.visual.blocks.25.norm2",
661
+ "model.visual.blocks.26.attn.proj",
662
+ "model.visual.blocks.26.attn.qkv",
663
+ "model.visual.blocks.26.mlp.linear_fc1",
664
+ "model.visual.blocks.26.mlp.linear_fc2",
665
+ "model.visual.blocks.26.norm1",
666
+ "model.visual.blocks.26.norm2",
667
+ "model.visual.blocks.3.attn.proj",
668
+ "model.visual.blocks.3.attn.qkv",
669
+ "model.visual.blocks.3.mlp.linear_fc1",
670
+ "model.visual.blocks.3.mlp.linear_fc2",
671
+ "model.visual.blocks.3.norm1",
672
+ "model.visual.blocks.3.norm2",
673
+ "model.visual.blocks.4.attn.proj",
674
+ "model.visual.blocks.4.attn.qkv",
675
+ "model.visual.blocks.4.mlp.linear_fc1",
676
+ "model.visual.blocks.4.mlp.linear_fc2",
677
+ "model.visual.blocks.4.norm1",
678
+ "model.visual.blocks.4.norm2",
679
+ "model.visual.blocks.5.attn.proj",
680
+ "model.visual.blocks.5.attn.qkv",
681
+ "model.visual.blocks.5.mlp.linear_fc1",
682
+ "model.visual.blocks.5.mlp.linear_fc2",
683
+ "model.visual.blocks.5.norm1",
684
+ "model.visual.blocks.5.norm2",
685
+ "model.visual.blocks.6.attn.proj",
686
+ "model.visual.blocks.6.attn.qkv",
687
+ "model.visual.blocks.6.mlp.linear_fc1",
688
+ "model.visual.blocks.6.mlp.linear_fc2",
689
+ "model.visual.blocks.6.norm1",
690
+ "model.visual.blocks.6.norm2",
691
+ "model.visual.blocks.7.attn.proj",
692
+ "model.visual.blocks.7.attn.qkv",
693
+ "model.visual.blocks.7.mlp.linear_fc1",
694
+ "model.visual.blocks.7.mlp.linear_fc2",
695
+ "model.visual.blocks.7.norm1",
696
+ "model.visual.blocks.7.norm2",
697
+ "model.visual.blocks.8.attn.proj",
698
+ "model.visual.blocks.8.attn.qkv",
699
+ "model.visual.blocks.8.mlp.linear_fc1",
700
+ "model.visual.blocks.8.mlp.linear_fc2",
701
+ "model.visual.blocks.8.norm1",
702
+ "model.visual.blocks.8.norm2",
703
+ "model.visual.blocks.9.attn.proj",
704
+ "model.visual.blocks.9.attn.qkv",
705
+ "model.visual.blocks.9.mlp.linear_fc1",
706
+ "model.visual.blocks.9.mlp.linear_fc2",
707
+ "model.visual.blocks.9.norm1",
708
+ "model.visual.blocks.9.norm2",
709
+ "model.visual.merger.linear_fc1",
710
+ "model.visual.merger.linear_fc2",
711
+ "model.visual.merger.norm",
712
+ "model.visual.patch_embed.proj",
713
+ "model.visual.pos_embed",
714
+ "mtp.fc",
715
+ "mtp.layers.0.input_layernorm",
716
+ "mtp.layers.0.mlp.experts.0.down_proj",
717
+ "mtp.layers.0.mlp.experts.0.gate_proj",
718
+ "mtp.layers.0.mlp.experts.0.up_proj",
719
+ "mtp.layers.0.mlp.experts.1.down_proj",
720
+ "mtp.layers.0.mlp.experts.1.gate_proj",
721
+ "mtp.layers.0.mlp.experts.1.up_proj",
722
+ "mtp.layers.0.mlp.experts.10.down_proj",
723
+ "mtp.layers.0.mlp.experts.10.gate_proj",
724
+ "mtp.layers.0.mlp.experts.10.up_proj",
725
+ "mtp.layers.0.mlp.experts.100.down_proj",
726
+ "mtp.layers.0.mlp.experts.100.gate_proj",
727
+ "mtp.layers.0.mlp.experts.100.up_proj",
728
+ "mtp.layers.0.mlp.experts.101.down_proj",
729
+ "mtp.layers.0.mlp.experts.101.gate_proj",
730
+ "mtp.layers.0.mlp.experts.101.up_proj",
731
+ "mtp.layers.0.mlp.experts.102.down_proj",
732
+ "mtp.layers.0.mlp.experts.102.gate_proj",
733
+ "mtp.layers.0.mlp.experts.102.up_proj",
734
+ "mtp.layers.0.mlp.experts.103.down_proj",
735
+ "mtp.layers.0.mlp.experts.103.gate_proj",
736
+ "mtp.layers.0.mlp.experts.103.up_proj",
737
+ "mtp.layers.0.mlp.experts.104.down_proj",
738
+ "mtp.layers.0.mlp.experts.104.gate_proj",
739
+ "mtp.layers.0.mlp.experts.104.up_proj",
740
+ "mtp.layers.0.mlp.experts.105.down_proj",
741
+ "mtp.layers.0.mlp.experts.105.gate_proj",
742
+ "mtp.layers.0.mlp.experts.105.up_proj",
743
+ "mtp.layers.0.mlp.experts.106.down_proj",
744
+ "mtp.layers.0.mlp.experts.106.gate_proj",
745
+ "mtp.layers.0.mlp.experts.106.up_proj",
746
+ "mtp.layers.0.mlp.experts.107.down_proj",
747
+ "mtp.layers.0.mlp.experts.107.gate_proj",
748
+ "mtp.layers.0.mlp.experts.107.up_proj",
749
+ "mtp.layers.0.mlp.experts.108.down_proj",
750
+ "mtp.layers.0.mlp.experts.108.gate_proj",
751
+ "mtp.layers.0.mlp.experts.108.up_proj",
752
+ "mtp.layers.0.mlp.experts.109.down_proj",
753
+ "mtp.layers.0.mlp.experts.109.gate_proj",
754
+ "mtp.layers.0.mlp.experts.109.up_proj",
755
+ "mtp.layers.0.mlp.experts.11.down_proj",
756
+ "mtp.layers.0.mlp.experts.11.gate_proj",
757
+ "mtp.layers.0.mlp.experts.11.up_proj",
758
+ "mtp.layers.0.mlp.experts.110.down_proj",
759
+ "mtp.layers.0.mlp.experts.110.gate_proj",
760
+ "mtp.layers.0.mlp.experts.110.up_proj",
761
+ "mtp.layers.0.mlp.experts.111.down_proj",
762
+ "mtp.layers.0.mlp.experts.111.gate_proj",
763
+ "mtp.layers.0.mlp.experts.111.up_proj",
764
+ "mtp.layers.0.mlp.experts.112.down_proj",
765
+ "mtp.layers.0.mlp.experts.112.gate_proj",
766
+ "mtp.layers.0.mlp.experts.112.up_proj",
767
+ "mtp.layers.0.mlp.experts.113.down_proj",
768
+ "mtp.layers.0.mlp.experts.113.gate_proj",
769
+ "mtp.layers.0.mlp.experts.113.up_proj",
770
+ "mtp.layers.0.mlp.experts.114.down_proj",
771
+ "mtp.layers.0.mlp.experts.114.gate_proj",
772
+ "mtp.layers.0.mlp.experts.114.up_proj",
773
+ "mtp.layers.0.mlp.experts.115.down_proj",
774
+ "mtp.layers.0.mlp.experts.115.gate_proj",
775
+ "mtp.layers.0.mlp.experts.115.up_proj",
776
+ "mtp.layers.0.mlp.experts.116.down_proj",
777
+ "mtp.layers.0.mlp.experts.116.gate_proj",
778
+ "mtp.layers.0.mlp.experts.116.up_proj",
779
+ "mtp.layers.0.mlp.experts.117.down_proj",
780
+ "mtp.layers.0.mlp.experts.117.gate_proj",
781
+ "mtp.layers.0.mlp.experts.117.up_proj",
782
+ "mtp.layers.0.mlp.experts.118.down_proj",
783
+ "mtp.layers.0.mlp.experts.118.gate_proj",
784
+ "mtp.layers.0.mlp.experts.118.up_proj",
785
+ "mtp.layers.0.mlp.experts.119.down_proj",
786
+ "mtp.layers.0.mlp.experts.119.gate_proj",
787
+ "mtp.layers.0.mlp.experts.119.up_proj",
788
+ "mtp.layers.0.mlp.experts.12.down_proj",
789
+ "mtp.layers.0.mlp.experts.12.gate_proj",
790
+ "mtp.layers.0.mlp.experts.12.up_proj",
791
+ "mtp.layers.0.mlp.experts.120.down_proj",
792
+ "mtp.layers.0.mlp.experts.120.gate_proj",
793
+ "mtp.layers.0.mlp.experts.120.up_proj",
794
+ "mtp.layers.0.mlp.experts.121.down_proj",
795
+ "mtp.layers.0.mlp.experts.121.gate_proj",
796
+ "mtp.layers.0.mlp.experts.121.up_proj",
797
+ "mtp.layers.0.mlp.experts.122.down_proj",
798
+ "mtp.layers.0.mlp.experts.122.gate_proj",
799
+ "mtp.layers.0.mlp.experts.122.up_proj",
800
+ "mtp.layers.0.mlp.experts.123.down_proj",
801
+ "mtp.layers.0.mlp.experts.123.gate_proj",
802
+ "mtp.layers.0.mlp.experts.123.up_proj",
803
+ "mtp.layers.0.mlp.experts.124.down_proj",
804
+ "mtp.layers.0.mlp.experts.124.gate_proj",
805
+ "mtp.layers.0.mlp.experts.124.up_proj",
806
+ "mtp.layers.0.mlp.experts.125.down_proj",
807
+ "mtp.layers.0.mlp.experts.125.gate_proj",
808
+ "mtp.layers.0.mlp.experts.125.up_proj",
809
+ "mtp.layers.0.mlp.experts.126.down_proj",
810
+ "mtp.layers.0.mlp.experts.126.gate_proj",
811
+ "mtp.layers.0.mlp.experts.126.up_proj",
812
+ "mtp.layers.0.mlp.experts.127.down_proj",
813
+ "mtp.layers.0.mlp.experts.127.gate_proj",
814
+ "mtp.layers.0.mlp.experts.127.up_proj",
815
+ "mtp.layers.0.mlp.experts.128.down_proj",
816
+ "mtp.layers.0.mlp.experts.128.gate_proj",
817
+ "mtp.layers.0.mlp.experts.128.up_proj",
818
+ "mtp.layers.0.mlp.experts.129.down_proj",
819
+ "mtp.layers.0.mlp.experts.129.gate_proj",
820
+ "mtp.layers.0.mlp.experts.129.up_proj",
821
+ "mtp.layers.0.mlp.experts.13.down_proj",
822
+ "mtp.layers.0.mlp.experts.13.gate_proj",
823
+ "mtp.layers.0.mlp.experts.13.up_proj",
824
+ "mtp.layers.0.mlp.experts.130.down_proj",
825
+ "mtp.layers.0.mlp.experts.130.gate_proj",
826
+ "mtp.layers.0.mlp.experts.130.up_proj",
827
+ "mtp.layers.0.mlp.experts.131.down_proj",
828
+ "mtp.layers.0.mlp.experts.131.gate_proj",
829
+ "mtp.layers.0.mlp.experts.131.up_proj",
830
+ "mtp.layers.0.mlp.experts.132.down_proj",
831
+ "mtp.layers.0.mlp.experts.132.gate_proj",
832
+ "mtp.layers.0.mlp.experts.132.up_proj",
833
+ "mtp.layers.0.mlp.experts.133.down_proj",
834
+ "mtp.layers.0.mlp.experts.133.gate_proj",
835
+ "mtp.layers.0.mlp.experts.133.up_proj",
836
+ "mtp.layers.0.mlp.experts.134.down_proj",
837
+ "mtp.layers.0.mlp.experts.134.gate_proj",
838
+ "mtp.layers.0.mlp.experts.134.up_proj",
839
+ "mtp.layers.0.mlp.experts.135.down_proj",
840
+ "mtp.layers.0.mlp.experts.135.gate_proj",
841
+ "mtp.layers.0.mlp.experts.135.up_proj",
842
+ "mtp.layers.0.mlp.experts.136.down_proj",
843
+ "mtp.layers.0.mlp.experts.136.gate_proj",
844
+ "mtp.layers.0.mlp.experts.136.up_proj",
845
+ "mtp.layers.0.mlp.experts.137.down_proj",
846
+ "mtp.layers.0.mlp.experts.137.gate_proj",
847
+ "mtp.layers.0.mlp.experts.137.up_proj",
848
+ "mtp.layers.0.mlp.experts.138.down_proj",
849
+ "mtp.layers.0.mlp.experts.138.gate_proj",
850
+ "mtp.layers.0.mlp.experts.138.up_proj",
851
+ "mtp.layers.0.mlp.experts.139.down_proj",
852
+ "mtp.layers.0.mlp.experts.139.gate_proj",
853
+ "mtp.layers.0.mlp.experts.139.up_proj",
854
+ "mtp.layers.0.mlp.experts.14.down_proj",
855
+ "mtp.layers.0.mlp.experts.14.gate_proj",
856
+ "mtp.layers.0.mlp.experts.14.up_proj",
857
+ "mtp.layers.0.mlp.experts.140.down_proj",
858
+ "mtp.layers.0.mlp.experts.140.gate_proj",
859
+ "mtp.layers.0.mlp.experts.140.up_proj",
860
+ "mtp.layers.0.mlp.experts.141.down_proj",
861
+ "mtp.layers.0.mlp.experts.141.gate_proj",
862
+ "mtp.layers.0.mlp.experts.141.up_proj",
863
+ "mtp.layers.0.mlp.experts.142.down_proj",
864
+ "mtp.layers.0.mlp.experts.142.gate_proj",
865
+ "mtp.layers.0.mlp.experts.142.up_proj",
866
+ "mtp.layers.0.mlp.experts.143.down_proj",
867
+ "mtp.layers.0.mlp.experts.143.gate_proj",
868
+ "mtp.layers.0.mlp.experts.143.up_proj",
869
+ "mtp.layers.0.mlp.experts.144.down_proj",
870
+ "mtp.layers.0.mlp.experts.144.gate_proj",
871
+ "mtp.layers.0.mlp.experts.144.up_proj",
872
+ "mtp.layers.0.mlp.experts.145.down_proj",
873
+ "mtp.layers.0.mlp.experts.145.gate_proj",
874
+ "mtp.layers.0.mlp.experts.145.up_proj",
875
+ "mtp.layers.0.mlp.experts.146.down_proj",
876
+ "mtp.layers.0.mlp.experts.146.gate_proj",
877
+ "mtp.layers.0.mlp.experts.146.up_proj",
878
+ "mtp.layers.0.mlp.experts.147.down_proj",
879
+ "mtp.layers.0.mlp.experts.147.gate_proj",
880
+ "mtp.layers.0.mlp.experts.147.up_proj",
881
+ "mtp.layers.0.mlp.experts.148.down_proj",
882
+ "mtp.layers.0.mlp.experts.148.gate_proj",
883
+ "mtp.layers.0.mlp.experts.148.up_proj",
884
+ "mtp.layers.0.mlp.experts.149.down_proj",
885
+ "mtp.layers.0.mlp.experts.149.gate_proj",
886
+ "mtp.layers.0.mlp.experts.149.up_proj",
887
+ "mtp.layers.0.mlp.experts.15.down_proj",
888
+ "mtp.layers.0.mlp.experts.15.gate_proj",
889
+ "mtp.layers.0.mlp.experts.15.up_proj",
890
+ "mtp.layers.0.mlp.experts.150.down_proj",
891
+ "mtp.layers.0.mlp.experts.150.gate_proj",
892
+ "mtp.layers.0.mlp.experts.150.up_proj",
893
+ "mtp.layers.0.mlp.experts.151.down_proj",
894
+ "mtp.layers.0.mlp.experts.151.gate_proj",
895
+ "mtp.layers.0.mlp.experts.151.up_proj",
896
+ "mtp.layers.0.mlp.experts.152.down_proj",
897
+ "mtp.layers.0.mlp.experts.152.gate_proj",
898
+ "mtp.layers.0.mlp.experts.152.up_proj",
899
+ "mtp.layers.0.mlp.experts.153.down_proj",
900
+ "mtp.layers.0.mlp.experts.153.gate_proj",
901
+ "mtp.layers.0.mlp.experts.153.up_proj",
902
+ "mtp.layers.0.mlp.experts.154.down_proj",
903
+ "mtp.layers.0.mlp.experts.154.gate_proj",
904
+ "mtp.layers.0.mlp.experts.154.up_proj",
905
+ "mtp.layers.0.mlp.experts.155.down_proj",
906
+ "mtp.layers.0.mlp.experts.155.gate_proj",
907
+ "mtp.layers.0.mlp.experts.155.up_proj",
908
+ "mtp.layers.0.mlp.experts.156.down_proj",
909
+ "mtp.layers.0.mlp.experts.156.gate_proj",
910
+ "mtp.layers.0.mlp.experts.156.up_proj",
911
+ "mtp.layers.0.mlp.experts.157.down_proj",
912
+ "mtp.layers.0.mlp.experts.157.gate_proj",
913
+ "mtp.layers.0.mlp.experts.157.up_proj",
914
+ "mtp.layers.0.mlp.experts.158.down_proj",
915
+ "mtp.layers.0.mlp.experts.158.gate_proj",
916
+ "mtp.layers.0.mlp.experts.158.up_proj",
917
+ "mtp.layers.0.mlp.experts.159.down_proj",
918
+ "mtp.layers.0.mlp.experts.159.gate_proj",
919
+ "mtp.layers.0.mlp.experts.159.up_proj",
920
+ "mtp.layers.0.mlp.experts.16.down_proj",
921
+ "mtp.layers.0.mlp.experts.16.gate_proj",
922
+ "mtp.layers.0.mlp.experts.16.up_proj",
923
+ "mtp.layers.0.mlp.experts.160.down_proj",
924
+ "mtp.layers.0.mlp.experts.160.gate_proj",
925
+ "mtp.layers.0.mlp.experts.160.up_proj",
926
+ "mtp.layers.0.mlp.experts.161.down_proj",
927
+ "mtp.layers.0.mlp.experts.161.gate_proj",
928
+ "mtp.layers.0.mlp.experts.161.up_proj",
929
+ "mtp.layers.0.mlp.experts.162.down_proj",
930
+ "mtp.layers.0.mlp.experts.162.gate_proj",
931
+ "mtp.layers.0.mlp.experts.162.up_proj",
932
+ "mtp.layers.0.mlp.experts.163.down_proj",
933
+ "mtp.layers.0.mlp.experts.163.gate_proj",
934
+ "mtp.layers.0.mlp.experts.163.up_proj",
935
+ "mtp.layers.0.mlp.experts.164.down_proj",
936
+ "mtp.layers.0.mlp.experts.164.gate_proj",
937
+ "mtp.layers.0.mlp.experts.164.up_proj",
938
+ "mtp.layers.0.mlp.experts.165.down_proj",
939
+ "mtp.layers.0.mlp.experts.165.gate_proj",
940
+ "mtp.layers.0.mlp.experts.165.up_proj",
941
+ "mtp.layers.0.mlp.experts.166.down_proj",
942
+ "mtp.layers.0.mlp.experts.166.gate_proj",
943
+ "mtp.layers.0.mlp.experts.166.up_proj",
944
+ "mtp.layers.0.mlp.experts.167.down_proj",
945
+ "mtp.layers.0.mlp.experts.167.gate_proj",
946
+ "mtp.layers.0.mlp.experts.167.up_proj",
947
+ "mtp.layers.0.mlp.experts.168.down_proj",
948
+ "mtp.layers.0.mlp.experts.168.gate_proj",
949
+ "mtp.layers.0.mlp.experts.168.up_proj",
950
+ "mtp.layers.0.mlp.experts.169.down_proj",
951
+ "mtp.layers.0.mlp.experts.169.gate_proj",
952
+ "mtp.layers.0.mlp.experts.169.up_proj",
953
+ "mtp.layers.0.mlp.experts.17.down_proj",
954
+ "mtp.layers.0.mlp.experts.17.gate_proj",
955
+ "mtp.layers.0.mlp.experts.17.up_proj",
956
+ "mtp.layers.0.mlp.experts.170.down_proj",
957
+ "mtp.layers.0.mlp.experts.170.gate_proj",
958
+ "mtp.layers.0.mlp.experts.170.up_proj",
959
+ "mtp.layers.0.mlp.experts.171.down_proj",
960
+ "mtp.layers.0.mlp.experts.171.gate_proj",
961
+ "mtp.layers.0.mlp.experts.171.up_proj",
962
+ "mtp.layers.0.mlp.experts.172.down_proj",
963
+ "mtp.layers.0.mlp.experts.172.gate_proj",
964
+ "mtp.layers.0.mlp.experts.172.up_proj",
965
+ "mtp.layers.0.mlp.experts.173.down_proj",
966
+ "mtp.layers.0.mlp.experts.173.gate_proj",
967
+ "mtp.layers.0.mlp.experts.173.up_proj",
968
+ "mtp.layers.0.mlp.experts.174.down_proj",
969
+ "mtp.layers.0.mlp.experts.174.gate_proj",
970
+ "mtp.layers.0.mlp.experts.174.up_proj",
971
+ "mtp.layers.0.mlp.experts.175.down_proj",
972
+ "mtp.layers.0.mlp.experts.175.gate_proj",
973
+ "mtp.layers.0.mlp.experts.175.up_proj",
974
+ "mtp.layers.0.mlp.experts.176.down_proj",
975
+ "mtp.layers.0.mlp.experts.176.gate_proj",
976
+ "mtp.layers.0.mlp.experts.176.up_proj",
977
+ "mtp.layers.0.mlp.experts.177.down_proj",
978
+ "mtp.layers.0.mlp.experts.177.gate_proj",
979
+ "mtp.layers.0.mlp.experts.177.up_proj",
980
+ "mtp.layers.0.mlp.experts.178.down_proj",
981
+ "mtp.layers.0.mlp.experts.178.gate_proj",
982
+ "mtp.layers.0.mlp.experts.178.up_proj",
983
+ "mtp.layers.0.mlp.experts.179.down_proj",
984
+ "mtp.layers.0.mlp.experts.179.gate_proj",
985
+ "mtp.layers.0.mlp.experts.179.up_proj",
986
+ "mtp.layers.0.mlp.experts.18.down_proj",
987
+ "mtp.layers.0.mlp.experts.18.gate_proj",
988
+ "mtp.layers.0.mlp.experts.18.up_proj",
989
+ "mtp.layers.0.mlp.experts.180.down_proj",
990
+ "mtp.layers.0.mlp.experts.180.gate_proj",
991
+ "mtp.layers.0.mlp.experts.180.up_proj",
992
+ "mtp.layers.0.mlp.experts.181.down_proj",
993
+ "mtp.layers.0.mlp.experts.181.gate_proj",
994
+ "mtp.layers.0.mlp.experts.181.up_proj",
995
+ "mtp.layers.0.mlp.experts.182.down_proj",
996
+ "mtp.layers.0.mlp.experts.182.gate_proj",
997
+ "mtp.layers.0.mlp.experts.182.up_proj",
998
+ "mtp.layers.0.mlp.experts.183.down_proj",
999
+ "mtp.layers.0.mlp.experts.183.gate_proj",
1000
+ "mtp.layers.0.mlp.experts.183.up_proj",
1001
+ "mtp.layers.0.mlp.experts.184.down_proj",
1002
+ "mtp.layers.0.mlp.experts.184.gate_proj",
1003
+ "mtp.layers.0.mlp.experts.184.up_proj",
1004
+ "mtp.layers.0.mlp.experts.185.down_proj",
1005
+ "mtp.layers.0.mlp.experts.185.gate_proj",
1006
+ "mtp.layers.0.mlp.experts.185.up_proj",
1007
+ "mtp.layers.0.mlp.experts.186.down_proj",
1008
+ "mtp.layers.0.mlp.experts.186.gate_proj",
1009
+ "mtp.layers.0.mlp.experts.186.up_proj",
1010
+ "mtp.layers.0.mlp.experts.187.down_proj",
1011
+ "mtp.layers.0.mlp.experts.187.gate_proj",
1012
+ "mtp.layers.0.mlp.experts.187.up_proj",
1013
+ "mtp.layers.0.mlp.experts.188.down_proj",
1014
+ "mtp.layers.0.mlp.experts.188.gate_proj",
1015
+ "mtp.layers.0.mlp.experts.188.up_proj",
1016
+ "mtp.layers.0.mlp.experts.189.down_proj",
1017
+ "mtp.layers.0.mlp.experts.189.gate_proj",
1018
+ "mtp.layers.0.mlp.experts.189.up_proj",
1019
+ "mtp.layers.0.mlp.experts.19.down_proj",
1020
+ "mtp.layers.0.mlp.experts.19.gate_proj",
1021
+ "mtp.layers.0.mlp.experts.19.up_proj",
1022
+ "mtp.layers.0.mlp.experts.190.down_proj",
1023
+ "mtp.layers.0.mlp.experts.190.gate_proj",
1024
+ "mtp.layers.0.mlp.experts.190.up_proj",
1025
+ "mtp.layers.0.mlp.experts.191.down_proj",
1026
+ "mtp.layers.0.mlp.experts.191.gate_proj",
1027
+ "mtp.layers.0.mlp.experts.191.up_proj",
1028
+ "mtp.layers.0.mlp.experts.192.down_proj",
1029
+ "mtp.layers.0.mlp.experts.192.gate_proj",
1030
+ "mtp.layers.0.mlp.experts.192.up_proj",
1031
+ "mtp.layers.0.mlp.experts.193.down_proj",
1032
+ "mtp.layers.0.mlp.experts.193.gate_proj",
1033
+ "mtp.layers.0.mlp.experts.193.up_proj",
1034
+ "mtp.layers.0.mlp.experts.194.down_proj",
1035
+ "mtp.layers.0.mlp.experts.194.gate_proj",
1036
+ "mtp.layers.0.mlp.experts.194.up_proj",
1037
+ "mtp.layers.0.mlp.experts.195.down_proj",
1038
+ "mtp.layers.0.mlp.experts.195.gate_proj",
1039
+ "mtp.layers.0.mlp.experts.195.up_proj",
1040
+ "mtp.layers.0.mlp.experts.196.down_proj",
1041
+ "mtp.layers.0.mlp.experts.196.gate_proj",
1042
+ "mtp.layers.0.mlp.experts.196.up_proj",
1043
+ "mtp.layers.0.mlp.experts.197.down_proj",
1044
+ "mtp.layers.0.mlp.experts.197.gate_proj",
1045
+ "mtp.layers.0.mlp.experts.197.up_proj",
1046
+ "mtp.layers.0.mlp.experts.198.down_proj",
1047
+ "mtp.layers.0.mlp.experts.198.gate_proj",
1048
+ "mtp.layers.0.mlp.experts.198.up_proj",
1049
+ "mtp.layers.0.mlp.experts.199.down_proj",
1050
+ "mtp.layers.0.mlp.experts.199.gate_proj",
1051
+ "mtp.layers.0.mlp.experts.199.up_proj",
1052
+ "mtp.layers.0.mlp.experts.2.down_proj",
1053
+ "mtp.layers.0.mlp.experts.2.gate_proj",
1054
+ "mtp.layers.0.mlp.experts.2.up_proj",
1055
+ "mtp.layers.0.mlp.experts.20.down_proj",
1056
+ "mtp.layers.0.mlp.experts.20.gate_proj",
1057
+ "mtp.layers.0.mlp.experts.20.up_proj",
1058
+ "mtp.layers.0.mlp.experts.200.down_proj",
1059
+ "mtp.layers.0.mlp.experts.200.gate_proj",
1060
+ "mtp.layers.0.mlp.experts.200.up_proj",
1061
+ "mtp.layers.0.mlp.experts.201.down_proj",
1062
+ "mtp.layers.0.mlp.experts.201.gate_proj",
1063
+ "mtp.layers.0.mlp.experts.201.up_proj",
1064
+ "mtp.layers.0.mlp.experts.202.down_proj",
1065
+ "mtp.layers.0.mlp.experts.202.gate_proj",
1066
+ "mtp.layers.0.mlp.experts.202.up_proj",
1067
+ "mtp.layers.0.mlp.experts.203.down_proj",
1068
+ "mtp.layers.0.mlp.experts.203.gate_proj",
1069
+ "mtp.layers.0.mlp.experts.203.up_proj",
1070
+ "mtp.layers.0.mlp.experts.204.down_proj",
1071
+ "mtp.layers.0.mlp.experts.204.gate_proj",
1072
+ "mtp.layers.0.mlp.experts.204.up_proj",
1073
+ "mtp.layers.0.mlp.experts.205.down_proj",
1074
+ "mtp.layers.0.mlp.experts.205.gate_proj",
1075
+ "mtp.layers.0.mlp.experts.205.up_proj",
1076
+ "mtp.layers.0.mlp.experts.206.down_proj",
1077
+ "mtp.layers.0.mlp.experts.206.gate_proj",
1078
+ "mtp.layers.0.mlp.experts.206.up_proj",
1079
+ "mtp.layers.0.mlp.experts.207.down_proj",
1080
+ "mtp.layers.0.mlp.experts.207.gate_proj",
1081
+ "mtp.layers.0.mlp.experts.207.up_proj",
1082
+ "mtp.layers.0.mlp.experts.208.down_proj",
1083
+ "mtp.layers.0.mlp.experts.208.gate_proj",
1084
+ "mtp.layers.0.mlp.experts.208.up_proj",
1085
+ "mtp.layers.0.mlp.experts.209.down_proj",
1086
+ "mtp.layers.0.mlp.experts.209.gate_proj",
1087
+ "mtp.layers.0.mlp.experts.209.up_proj",
1088
+ "mtp.layers.0.mlp.experts.21.down_proj",
1089
+ "mtp.layers.0.mlp.experts.21.gate_proj",
1090
+ "mtp.layers.0.mlp.experts.21.up_proj",
1091
+ "mtp.layers.0.mlp.experts.210.down_proj",
1092
+ "mtp.layers.0.mlp.experts.210.gate_proj",
1093
+ "mtp.layers.0.mlp.experts.210.up_proj",
1094
+ "mtp.layers.0.mlp.experts.211.down_proj",
1095
+ "mtp.layers.0.mlp.experts.211.gate_proj",
1096
+ "mtp.layers.0.mlp.experts.211.up_proj",
1097
+ "mtp.layers.0.mlp.experts.212.down_proj",
1098
+ "mtp.layers.0.mlp.experts.212.gate_proj",
1099
+ "mtp.layers.0.mlp.experts.212.up_proj",
1100
+ "mtp.layers.0.mlp.experts.213.down_proj",
1101
+ "mtp.layers.0.mlp.experts.213.gate_proj",
1102
+ "mtp.layers.0.mlp.experts.213.up_proj",
1103
+ "mtp.layers.0.mlp.experts.214.down_proj",
1104
+ "mtp.layers.0.mlp.experts.214.gate_proj",
1105
+ "mtp.layers.0.mlp.experts.214.up_proj",
1106
+ "mtp.layers.0.mlp.experts.215.down_proj",
1107
+ "mtp.layers.0.mlp.experts.215.gate_proj",
1108
+ "mtp.layers.0.mlp.experts.215.up_proj",
1109
+ "mtp.layers.0.mlp.experts.216.down_proj",
1110
+ "mtp.layers.0.mlp.experts.216.gate_proj",
1111
+ "mtp.layers.0.mlp.experts.216.up_proj",
1112
+ "mtp.layers.0.mlp.experts.217.down_proj",
1113
+ "mtp.layers.0.mlp.experts.217.gate_proj",
1114
+ "mtp.layers.0.mlp.experts.217.up_proj",
1115
+ "mtp.layers.0.mlp.experts.218.down_proj",
1116
+ "mtp.layers.0.mlp.experts.218.gate_proj",
1117
+ "mtp.layers.0.mlp.experts.218.up_proj",
1118
+ "mtp.layers.0.mlp.experts.219.down_proj",
1119
+ "mtp.layers.0.mlp.experts.219.gate_proj",
1120
+ "mtp.layers.0.mlp.experts.219.up_proj",
1121
+ "mtp.layers.0.mlp.experts.22.down_proj",
1122
+ "mtp.layers.0.mlp.experts.22.gate_proj",
1123
+ "mtp.layers.0.mlp.experts.22.up_proj",
1124
+ "mtp.layers.0.mlp.experts.220.down_proj",
1125
+ "mtp.layers.0.mlp.experts.220.gate_proj",
1126
+ "mtp.layers.0.mlp.experts.220.up_proj",
1127
+ "mtp.layers.0.mlp.experts.221.down_proj",
1128
+ "mtp.layers.0.mlp.experts.221.gate_proj",
1129
+ "mtp.layers.0.mlp.experts.221.up_proj",
1130
+ "mtp.layers.0.mlp.experts.222.down_proj",
1131
+ "mtp.layers.0.mlp.experts.222.gate_proj",
1132
+ "mtp.layers.0.mlp.experts.222.up_proj",
1133
+ "mtp.layers.0.mlp.experts.223.down_proj",
1134
+ "mtp.layers.0.mlp.experts.223.gate_proj",
1135
+ "mtp.layers.0.mlp.experts.223.up_proj",
1136
+ "mtp.layers.0.mlp.experts.224.down_proj",
1137
+ "mtp.layers.0.mlp.experts.224.gate_proj",
1138
+ "mtp.layers.0.mlp.experts.224.up_proj",
1139
+ "mtp.layers.0.mlp.experts.225.down_proj",
1140
+ "mtp.layers.0.mlp.experts.225.gate_proj",
1141
+ "mtp.layers.0.mlp.experts.225.up_proj",
1142
+ "mtp.layers.0.mlp.experts.226.down_proj",
1143
+ "mtp.layers.0.mlp.experts.226.gate_proj",
1144
+ "mtp.layers.0.mlp.experts.226.up_proj",
1145
+ "mtp.layers.0.mlp.experts.227.down_proj",
1146
+ "mtp.layers.0.mlp.experts.227.gate_proj",
1147
+ "mtp.layers.0.mlp.experts.227.up_proj",
1148
+ "mtp.layers.0.mlp.experts.228.down_proj",
1149
+ "mtp.layers.0.mlp.experts.228.gate_proj",
1150
+ "mtp.layers.0.mlp.experts.228.up_proj",
1151
+ "mtp.layers.0.mlp.experts.229.down_proj",
1152
+ "mtp.layers.0.mlp.experts.229.gate_proj",
1153
+ "mtp.layers.0.mlp.experts.229.up_proj",
1154
+ "mtp.layers.0.mlp.experts.23.down_proj",
1155
+ "mtp.layers.0.mlp.experts.23.gate_proj",
1156
+ "mtp.layers.0.mlp.experts.23.up_proj",
1157
+ "mtp.layers.0.mlp.experts.230.down_proj",
1158
+ "mtp.layers.0.mlp.experts.230.gate_proj",
1159
+ "mtp.layers.0.mlp.experts.230.up_proj",
1160
+ "mtp.layers.0.mlp.experts.231.down_proj",
1161
+ "mtp.layers.0.mlp.experts.231.gate_proj",
1162
+ "mtp.layers.0.mlp.experts.231.up_proj",
1163
+ "mtp.layers.0.mlp.experts.232.down_proj",
1164
+ "mtp.layers.0.mlp.experts.232.gate_proj",
1165
+ "mtp.layers.0.mlp.experts.232.up_proj",
1166
+ "mtp.layers.0.mlp.experts.233.down_proj",
1167
+ "mtp.layers.0.mlp.experts.233.gate_proj",
1168
+ "mtp.layers.0.mlp.experts.233.up_proj",
1169
+ "mtp.layers.0.mlp.experts.234.down_proj",
1170
+ "mtp.layers.0.mlp.experts.234.gate_proj",
1171
+ "mtp.layers.0.mlp.experts.234.up_proj",
1172
+ "mtp.layers.0.mlp.experts.235.down_proj",
1173
+ "mtp.layers.0.mlp.experts.235.gate_proj",
1174
+ "mtp.layers.0.mlp.experts.235.up_proj",
1175
+ "mtp.layers.0.mlp.experts.236.down_proj",
1176
+ "mtp.layers.0.mlp.experts.236.gate_proj",
1177
+ "mtp.layers.0.mlp.experts.236.up_proj",
1178
+ "mtp.layers.0.mlp.experts.237.down_proj",
1179
+ "mtp.layers.0.mlp.experts.237.gate_proj",
1180
+ "mtp.layers.0.mlp.experts.237.up_proj",
1181
+ "mtp.layers.0.mlp.experts.238.down_proj",
1182
+ "mtp.layers.0.mlp.experts.238.gate_proj",
1183
+ "mtp.layers.0.mlp.experts.238.up_proj",
1184
+ "mtp.layers.0.mlp.experts.239.down_proj",
1185
+ "mtp.layers.0.mlp.experts.239.gate_proj",
1186
+ "mtp.layers.0.mlp.experts.239.up_proj",
1187
+ "mtp.layers.0.mlp.experts.24.down_proj",
1188
+ "mtp.layers.0.mlp.experts.24.gate_proj",
1189
+ "mtp.layers.0.mlp.experts.24.up_proj",
1190
+ "mtp.layers.0.mlp.experts.240.down_proj",
1191
+ "mtp.layers.0.mlp.experts.240.gate_proj",
1192
+ "mtp.layers.0.mlp.experts.240.up_proj",
1193
+ "mtp.layers.0.mlp.experts.241.down_proj",
1194
+ "mtp.layers.0.mlp.experts.241.gate_proj",
1195
+ "mtp.layers.0.mlp.experts.241.up_proj",
1196
+ "mtp.layers.0.mlp.experts.242.down_proj",
1197
+ "mtp.layers.0.mlp.experts.242.gate_proj",
1198
+ "mtp.layers.0.mlp.experts.242.up_proj",
1199
+ "mtp.layers.0.mlp.experts.243.down_proj",
1200
+ "mtp.layers.0.mlp.experts.243.gate_proj",
1201
+ "mtp.layers.0.mlp.experts.243.up_proj",
1202
+ "mtp.layers.0.mlp.experts.244.down_proj",
1203
+ "mtp.layers.0.mlp.experts.244.gate_proj",
1204
+ "mtp.layers.0.mlp.experts.244.up_proj",
1205
+ "mtp.layers.0.mlp.experts.245.down_proj",
1206
+ "mtp.layers.0.mlp.experts.245.gate_proj",
1207
+ "mtp.layers.0.mlp.experts.245.up_proj",
1208
+ "mtp.layers.0.mlp.experts.246.down_proj",
1209
+ "mtp.layers.0.mlp.experts.246.gate_proj",
1210
+ "mtp.layers.0.mlp.experts.246.up_proj",
1211
+ "mtp.layers.0.mlp.experts.247.down_proj",
1212
+ "mtp.layers.0.mlp.experts.247.gate_proj",
1213
+ "mtp.layers.0.mlp.experts.247.up_proj",
1214
+ "mtp.layers.0.mlp.experts.248.down_proj",
1215
+ "mtp.layers.0.mlp.experts.248.gate_proj",
1216
+ "mtp.layers.0.mlp.experts.248.up_proj",
1217
+ "mtp.layers.0.mlp.experts.249.down_proj",
1218
+ "mtp.layers.0.mlp.experts.249.gate_proj",
1219
+ "mtp.layers.0.mlp.experts.249.up_proj",
1220
+ "mtp.layers.0.mlp.experts.25.down_proj",
1221
+ "mtp.layers.0.mlp.experts.25.gate_proj",
1222
+ "mtp.layers.0.mlp.experts.25.up_proj",
1223
+ "mtp.layers.0.mlp.experts.250.down_proj",
1224
+ "mtp.layers.0.mlp.experts.250.gate_proj",
1225
+ "mtp.layers.0.mlp.experts.250.up_proj",
1226
+ "mtp.layers.0.mlp.experts.251.down_proj",
1227
+ "mtp.layers.0.mlp.experts.251.gate_proj",
1228
+ "mtp.layers.0.mlp.experts.251.up_proj",
1229
+ "mtp.layers.0.mlp.experts.252.down_proj",
1230
+ "mtp.layers.0.mlp.experts.252.gate_proj",
1231
+ "mtp.layers.0.mlp.experts.252.up_proj",
1232
+ "mtp.layers.0.mlp.experts.253.down_proj",
1233
+ "mtp.layers.0.mlp.experts.253.gate_proj",
1234
+ "mtp.layers.0.mlp.experts.253.up_proj",
1235
+ "mtp.layers.0.mlp.experts.254.down_proj",
1236
+ "mtp.layers.0.mlp.experts.254.gate_proj",
1237
+ "mtp.layers.0.mlp.experts.254.up_proj",
1238
+ "mtp.layers.0.mlp.experts.255.down_proj",
1239
+ "mtp.layers.0.mlp.experts.255.gate_proj",
1240
+ "mtp.layers.0.mlp.experts.255.up_proj",
1241
+ "mtp.layers.0.mlp.experts.26.down_proj",
1242
+ "mtp.layers.0.mlp.experts.26.gate_proj",
1243
+ "mtp.layers.0.mlp.experts.26.up_proj",
1244
+ "mtp.layers.0.mlp.experts.27.down_proj",
1245
+ "mtp.layers.0.mlp.experts.27.gate_proj",
1246
+ "mtp.layers.0.mlp.experts.27.up_proj",
1247
+ "mtp.layers.0.mlp.experts.28.down_proj",
1248
+ "mtp.layers.0.mlp.experts.28.gate_proj",
1249
+ "mtp.layers.0.mlp.experts.28.up_proj",
1250
+ "mtp.layers.0.mlp.experts.29.down_proj",
1251
+ "mtp.layers.0.mlp.experts.29.gate_proj",
1252
+ "mtp.layers.0.mlp.experts.29.up_proj",
1253
+ "mtp.layers.0.mlp.experts.3.down_proj",
1254
+ "mtp.layers.0.mlp.experts.3.gate_proj",
1255
+ "mtp.layers.0.mlp.experts.3.up_proj",
1256
+ "mtp.layers.0.mlp.experts.30.down_proj",
1257
+ "mtp.layers.0.mlp.experts.30.gate_proj",
1258
+ "mtp.layers.0.mlp.experts.30.up_proj",
1259
+ "mtp.layers.0.mlp.experts.31.down_proj",
1260
+ "mtp.layers.0.mlp.experts.31.gate_proj",
1261
+ "mtp.layers.0.mlp.experts.31.up_proj",
1262
+ "mtp.layers.0.mlp.experts.32.down_proj",
1263
+ "mtp.layers.0.mlp.experts.32.gate_proj",
1264
+ "mtp.layers.0.mlp.experts.32.up_proj",
1265
+ "mtp.layers.0.mlp.experts.33.down_proj",
1266
+ "mtp.layers.0.mlp.experts.33.gate_proj",
1267
+ "mtp.layers.0.mlp.experts.33.up_proj",
1268
+ "mtp.layers.0.mlp.experts.34.down_proj",
1269
+ "mtp.layers.0.mlp.experts.34.gate_proj",
1270
+ "mtp.layers.0.mlp.experts.34.up_proj",
1271
+ "mtp.layers.0.mlp.experts.35.down_proj",
1272
+ "mtp.layers.0.mlp.experts.35.gate_proj",
1273
+ "mtp.layers.0.mlp.experts.35.up_proj",
1274
+ "mtp.layers.0.mlp.experts.36.down_proj",
1275
+ "mtp.layers.0.mlp.experts.36.gate_proj",
1276
+ "mtp.layers.0.mlp.experts.36.up_proj",
1277
+ "mtp.layers.0.mlp.experts.37.down_proj",
1278
+ "mtp.layers.0.mlp.experts.37.gate_proj",
1279
+ "mtp.layers.0.mlp.experts.37.up_proj",
1280
+ "mtp.layers.0.mlp.experts.38.down_proj",
1281
+ "mtp.layers.0.mlp.experts.38.gate_proj",
1282
+ "mtp.layers.0.mlp.experts.38.up_proj",
1283
+ "mtp.layers.0.mlp.experts.39.down_proj",
1284
+ "mtp.layers.0.mlp.experts.39.gate_proj",
1285
+ "mtp.layers.0.mlp.experts.39.up_proj",
1286
+ "mtp.layers.0.mlp.experts.4.down_proj",
1287
+ "mtp.layers.0.mlp.experts.4.gate_proj",
1288
+ "mtp.layers.0.mlp.experts.4.up_proj",
1289
+ "mtp.layers.0.mlp.experts.40.down_proj",
1290
+ "mtp.layers.0.mlp.experts.40.gate_proj",
1291
+ "mtp.layers.0.mlp.experts.40.up_proj",
1292
+ "mtp.layers.0.mlp.experts.41.down_proj",
1293
+ "mtp.layers.0.mlp.experts.41.gate_proj",
1294
+ "mtp.layers.0.mlp.experts.41.up_proj",
1295
+ "mtp.layers.0.mlp.experts.42.down_proj",
1296
+ "mtp.layers.0.mlp.experts.42.gate_proj",
1297
+ "mtp.layers.0.mlp.experts.42.up_proj",
1298
+ "mtp.layers.0.mlp.experts.43.down_proj",
1299
+ "mtp.layers.0.mlp.experts.43.gate_proj",
1300
+ "mtp.layers.0.mlp.experts.43.up_proj",
1301
+ "mtp.layers.0.mlp.experts.44.down_proj",
1302
+ "mtp.layers.0.mlp.experts.44.gate_proj",
1303
+ "mtp.layers.0.mlp.experts.44.up_proj",
1304
+ "mtp.layers.0.mlp.experts.45.down_proj",
1305
+ "mtp.layers.0.mlp.experts.45.gate_proj",
1306
+ "mtp.layers.0.mlp.experts.45.up_proj",
1307
+ "mtp.layers.0.mlp.experts.46.down_proj",
1308
+ "mtp.layers.0.mlp.experts.46.gate_proj",
1309
+ "mtp.layers.0.mlp.experts.46.up_proj",
1310
+ "mtp.layers.0.mlp.experts.47.down_proj",
1311
+ "mtp.layers.0.mlp.experts.47.gate_proj",
1312
+ "mtp.layers.0.mlp.experts.47.up_proj",
1313
+ "mtp.layers.0.mlp.experts.48.down_proj",
1314
+ "mtp.layers.0.mlp.experts.48.gate_proj",
1315
+ "mtp.layers.0.mlp.experts.48.up_proj",
1316
+ "mtp.layers.0.mlp.experts.49.down_proj",
1317
+ "mtp.layers.0.mlp.experts.49.gate_proj",
1318
+ "mtp.layers.0.mlp.experts.49.up_proj",
1319
+ "mtp.layers.0.mlp.experts.5.down_proj",
1320
+ "mtp.layers.0.mlp.experts.5.gate_proj",
1321
+ "mtp.layers.0.mlp.experts.5.up_proj",
1322
+ "mtp.layers.0.mlp.experts.50.down_proj",
1323
+ "mtp.layers.0.mlp.experts.50.gate_proj",
1324
+ "mtp.layers.0.mlp.experts.50.up_proj",
1325
+ "mtp.layers.0.mlp.experts.51.down_proj",
1326
+ "mtp.layers.0.mlp.experts.51.gate_proj",
1327
+ "mtp.layers.0.mlp.experts.51.up_proj",
1328
+ "mtp.layers.0.mlp.experts.52.down_proj",
1329
+ "mtp.layers.0.mlp.experts.52.gate_proj",
1330
+ "mtp.layers.0.mlp.experts.52.up_proj",
1331
+ "mtp.layers.0.mlp.experts.53.down_proj",
1332
+ "mtp.layers.0.mlp.experts.53.gate_proj",
1333
+ "mtp.layers.0.mlp.experts.53.up_proj",
1334
+ "mtp.layers.0.mlp.experts.54.down_proj",
1335
+ "mtp.layers.0.mlp.experts.54.gate_proj",
1336
+ "mtp.layers.0.mlp.experts.54.up_proj",
1337
+ "mtp.layers.0.mlp.experts.55.down_proj",
1338
+ "mtp.layers.0.mlp.experts.55.gate_proj",
1339
+ "mtp.layers.0.mlp.experts.55.up_proj",
1340
+ "mtp.layers.0.mlp.experts.56.down_proj",
1341
+ "mtp.layers.0.mlp.experts.56.gate_proj",
1342
+ "mtp.layers.0.mlp.experts.56.up_proj",
1343
+ "mtp.layers.0.mlp.experts.57.down_proj",
1344
+ "mtp.layers.0.mlp.experts.57.gate_proj",
1345
+ "mtp.layers.0.mlp.experts.57.up_proj",
1346
+ "mtp.layers.0.mlp.experts.58.down_proj",
1347
+ "mtp.layers.0.mlp.experts.58.gate_proj",
1348
+ "mtp.layers.0.mlp.experts.58.up_proj",
1349
+ "mtp.layers.0.mlp.experts.59.down_proj",
1350
+ "mtp.layers.0.mlp.experts.59.gate_proj",
1351
+ "mtp.layers.0.mlp.experts.59.up_proj",
1352
+ "mtp.layers.0.mlp.experts.6.down_proj",
1353
+ "mtp.layers.0.mlp.experts.6.gate_proj",
1354
+ "mtp.layers.0.mlp.experts.6.up_proj",
1355
+ "mtp.layers.0.mlp.experts.60.down_proj",
1356
+ "mtp.layers.0.mlp.experts.60.gate_proj",
1357
+ "mtp.layers.0.mlp.experts.60.up_proj",
1358
+ "mtp.layers.0.mlp.experts.61.down_proj",
1359
+ "mtp.layers.0.mlp.experts.61.gate_proj",
1360
+ "mtp.layers.0.mlp.experts.61.up_proj",
1361
+ "mtp.layers.0.mlp.experts.62.down_proj",
1362
+ "mtp.layers.0.mlp.experts.62.gate_proj",
1363
+ "mtp.layers.0.mlp.experts.62.up_proj",
1364
+ "mtp.layers.0.mlp.experts.63.down_proj",
1365
+ "mtp.layers.0.mlp.experts.63.gate_proj",
1366
+ "mtp.layers.0.mlp.experts.63.up_proj",
1367
+ "mtp.layers.0.mlp.experts.64.down_proj",
1368
+ "mtp.layers.0.mlp.experts.64.gate_proj",
1369
+ "mtp.layers.0.mlp.experts.64.up_proj",
1370
+ "mtp.layers.0.mlp.experts.65.down_proj",
1371
+ "mtp.layers.0.mlp.experts.65.gate_proj",
1372
+ "mtp.layers.0.mlp.experts.65.up_proj",
1373
+ "mtp.layers.0.mlp.experts.66.down_proj",
1374
+ "mtp.layers.0.mlp.experts.66.gate_proj",
1375
+ "mtp.layers.0.mlp.experts.66.up_proj",
1376
+ "mtp.layers.0.mlp.experts.67.down_proj",
1377
+ "mtp.layers.0.mlp.experts.67.gate_proj",
1378
+ "mtp.layers.0.mlp.experts.67.up_proj",
1379
+ "mtp.layers.0.mlp.experts.68.down_proj",
1380
+ "mtp.layers.0.mlp.experts.68.gate_proj",
1381
+ "mtp.layers.0.mlp.experts.68.up_proj",
1382
+ "mtp.layers.0.mlp.experts.69.down_proj",
1383
+ "mtp.layers.0.mlp.experts.69.gate_proj",
1384
+ "mtp.layers.0.mlp.experts.69.up_proj",
1385
+ "mtp.layers.0.mlp.experts.7.down_proj",
1386
+ "mtp.layers.0.mlp.experts.7.gate_proj",
1387
+ "mtp.layers.0.mlp.experts.7.up_proj",
1388
+ "mtp.layers.0.mlp.experts.70.down_proj",
1389
+ "mtp.layers.0.mlp.experts.70.gate_proj",
1390
+ "mtp.layers.0.mlp.experts.70.up_proj",
1391
+ "mtp.layers.0.mlp.experts.71.down_proj",
1392
+ "mtp.layers.0.mlp.experts.71.gate_proj",
1393
+ "mtp.layers.0.mlp.experts.71.up_proj",
1394
+ "mtp.layers.0.mlp.experts.72.down_proj",
1395
+ "mtp.layers.0.mlp.experts.72.gate_proj",
1396
+ "mtp.layers.0.mlp.experts.72.up_proj",
1397
+ "mtp.layers.0.mlp.experts.73.down_proj",
1398
+ "mtp.layers.0.mlp.experts.73.gate_proj",
1399
+ "mtp.layers.0.mlp.experts.73.up_proj",
1400
+ "mtp.layers.0.mlp.experts.74.down_proj",
1401
+ "mtp.layers.0.mlp.experts.74.gate_proj",
1402
+ "mtp.layers.0.mlp.experts.74.up_proj",
1403
+ "mtp.layers.0.mlp.experts.75.down_proj",
1404
+ "mtp.layers.0.mlp.experts.75.gate_proj",
1405
+ "mtp.layers.0.mlp.experts.75.up_proj",
1406
+ "mtp.layers.0.mlp.experts.76.down_proj",
1407
+ "mtp.layers.0.mlp.experts.76.gate_proj",
1408
+ "mtp.layers.0.mlp.experts.76.up_proj",
1409
+ "mtp.layers.0.mlp.experts.77.down_proj",
1410
+ "mtp.layers.0.mlp.experts.77.gate_proj",
1411
+ "mtp.layers.0.mlp.experts.77.up_proj",
1412
+ "mtp.layers.0.mlp.experts.78.down_proj",
1413
+ "mtp.layers.0.mlp.experts.78.gate_proj",
1414
+ "mtp.layers.0.mlp.experts.78.up_proj",
1415
+ "mtp.layers.0.mlp.experts.79.down_proj",
1416
+ "mtp.layers.0.mlp.experts.79.gate_proj",
1417
+ "mtp.layers.0.mlp.experts.79.up_proj",
1418
+ "mtp.layers.0.mlp.experts.8.down_proj",
1419
+ "mtp.layers.0.mlp.experts.8.gate_proj",
1420
+ "mtp.layers.0.mlp.experts.8.up_proj",
1421
+ "mtp.layers.0.mlp.experts.80.down_proj",
1422
+ "mtp.layers.0.mlp.experts.80.gate_proj",
1423
+ "mtp.layers.0.mlp.experts.80.up_proj",
1424
+ "mtp.layers.0.mlp.experts.81.down_proj",
1425
+ "mtp.layers.0.mlp.experts.81.gate_proj",
1426
+ "mtp.layers.0.mlp.experts.81.up_proj",
1427
+ "mtp.layers.0.mlp.experts.82.down_proj",
1428
+ "mtp.layers.0.mlp.experts.82.gate_proj",
1429
+ "mtp.layers.0.mlp.experts.82.up_proj",
1430
+ "mtp.layers.0.mlp.experts.83.down_proj",
1431
+ "mtp.layers.0.mlp.experts.83.gate_proj",
1432
+ "mtp.layers.0.mlp.experts.83.up_proj",
1433
+ "mtp.layers.0.mlp.experts.84.down_proj",
1434
+ "mtp.layers.0.mlp.experts.84.gate_proj",
1435
+ "mtp.layers.0.mlp.experts.84.up_proj",
1436
+ "mtp.layers.0.mlp.experts.85.down_proj",
1437
+ "mtp.layers.0.mlp.experts.85.gate_proj",
1438
+ "mtp.layers.0.mlp.experts.85.up_proj",
1439
+ "mtp.layers.0.mlp.experts.86.down_proj",
1440
+ "mtp.layers.0.mlp.experts.86.gate_proj",
1441
+ "mtp.layers.0.mlp.experts.86.up_proj",
1442
+ "mtp.layers.0.mlp.experts.87.down_proj",
1443
+ "mtp.layers.0.mlp.experts.87.gate_proj",
1444
+ "mtp.layers.0.mlp.experts.87.up_proj",
1445
+ "mtp.layers.0.mlp.experts.88.down_proj",
1446
+ "mtp.layers.0.mlp.experts.88.gate_proj",
1447
+ "mtp.layers.0.mlp.experts.88.up_proj",
1448
+ "mtp.layers.0.mlp.experts.89.down_proj",
1449
+ "mtp.layers.0.mlp.experts.89.gate_proj",
1450
+ "mtp.layers.0.mlp.experts.89.up_proj",
1451
+ "mtp.layers.0.mlp.experts.9.down_proj",
1452
+ "mtp.layers.0.mlp.experts.9.gate_proj",
1453
+ "mtp.layers.0.mlp.experts.9.up_proj",
1454
+ "mtp.layers.0.mlp.experts.90.down_proj",
1455
+ "mtp.layers.0.mlp.experts.90.gate_proj",
1456
+ "mtp.layers.0.mlp.experts.90.up_proj",
1457
+ "mtp.layers.0.mlp.experts.91.down_proj",
1458
+ "mtp.layers.0.mlp.experts.91.gate_proj",
1459
+ "mtp.layers.0.mlp.experts.91.up_proj",
1460
+ "mtp.layers.0.mlp.experts.92.down_proj",
1461
+ "mtp.layers.0.mlp.experts.92.gate_proj",
1462
+ "mtp.layers.0.mlp.experts.92.up_proj",
1463
+ "mtp.layers.0.mlp.experts.93.down_proj",
1464
+ "mtp.layers.0.mlp.experts.93.gate_proj",
1465
+ "mtp.layers.0.mlp.experts.93.up_proj",
1466
+ "mtp.layers.0.mlp.experts.94.down_proj",
1467
+ "mtp.layers.0.mlp.experts.94.gate_proj",
1468
+ "mtp.layers.0.mlp.experts.94.up_proj",
1469
+ "mtp.layers.0.mlp.experts.95.down_proj",
1470
+ "mtp.layers.0.mlp.experts.95.gate_proj",
1471
+ "mtp.layers.0.mlp.experts.95.up_proj",
1472
+ "mtp.layers.0.mlp.experts.96.down_proj",
1473
+ "mtp.layers.0.mlp.experts.96.gate_proj",
1474
+ "mtp.layers.0.mlp.experts.96.up_proj",
1475
+ "mtp.layers.0.mlp.experts.97.down_proj",
1476
+ "mtp.layers.0.mlp.experts.97.gate_proj",
1477
+ "mtp.layers.0.mlp.experts.97.up_proj",
1478
+ "mtp.layers.0.mlp.experts.98.down_proj",
1479
+ "mtp.layers.0.mlp.experts.98.gate_proj",
1480
+ "mtp.layers.0.mlp.experts.98.up_proj",
1481
+ "mtp.layers.0.mlp.experts.99.down_proj",
1482
+ "mtp.layers.0.mlp.experts.99.gate_proj",
1483
+ "mtp.layers.0.mlp.experts.99.up_proj",
1484
+ "mtp.layers.0.mlp.gate",
1485
+ "mtp.layers.0.mlp.shared_expert.down_proj",
1486
+ "mtp.layers.0.mlp.shared_expert.gate_proj",
1487
+ "mtp.layers.0.mlp.shared_expert.up_proj",
1488
+ "mtp.layers.0.mlp.shared_expert_gate",
1489
+ "mtp.layers.0.post_attention_layernorm",
1490
+ "mtp.layers.0.self_attn.k_norm",
1491
+ "mtp.layers.0.self_attn.k_proj",
1492
+ "mtp.layers.0.self_attn.o_proj",
1493
+ "mtp.layers.0.self_attn.q_norm",
1494
+ "mtp.layers.0.self_attn.q_proj",
1495
+ "mtp.layers.0.self_attn.v_proj",
1496
+ "mtp.norm",
1497
+ "mtp.pre_fc_norm_embedding",
1498
+ "mtp.pre_fc_norm_hidden"
1499
+ ]
1500
+ }
1501
+ }
generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 248044,
3
+ "do_sample": true,
4
+ "eos_token_id": [
5
+ 248046,
6
+ 248044
7
+ ],
8
+ "pad_token_id": 248044,
9
+ "temperature": 1.0,
10
+ "top_k": 20,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.8.1"
13
+ }
model-00001-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2180e67afbb6cd625092df0d5f3c96f4e455d3bf4dd8476ded85f96e41c1db2f
3
+ size 2648239472
model-00002-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e894e4bd2400014c0b99888d82d77c1d6bffca6082304145a02afcb95f34b08
3
+ size 1253941328
model-00003-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:59390bb912e8c45b988b444bbc1bb773d0f318360893fe2c839d3e07ce0e2820
3
+ size 1342092768
model-00004-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7d5d92ab94b0f9a943855004ba5315f795b60d743c8a330460850fc9faf8fb16
3
+ size 1103909792
model-00005-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ca5f6124133cc97d4b93d948c701d1f55d34fb85f0a4fd4c6b0c45d26977ace3
3
+ size 1253944640
model-00006-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6704385cd9f263b22d3bbbd2d551f8b30b825c894a2a870446df6771b3e2e481
3
+ size 1342097424
model-00007-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2c6ccc4625f2a0fec61f7b9b5f3d2b896cb4def15081a811e37c56e406cbf2d5
3
+ size 1103913424
model-00008-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70aa70bbf7687448fe83c3b1c6fefd8c82970dd132ffc48608946d3ba0009dcb
3
+ size 1253945480
model-00009-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e009b234686b26b47e7d555b196b4d584aa1db69076a41bfc786e2a4aa0d359
3
+ size 1342097424
model-00010-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3b71ac62a30e7c137d776c57e9af1b5423983975c8a1631b5a8d0287cc2871a2
3
+ size 1103913424
model-00011-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1b0f182ef7d1c90f3f709821d0c3cb9728ce666745521b2aa631bdc1bc361352
3
+ size 1253945480
model-00012-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d286eaf8497173d0a7bbee6feb5ec71ef546e4543b30727fe4bb6779d50c3a25
3
+ size 1342097424
model-00013-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cad3fc59628e81019bc07719b638a9f2ab7cc0eff0be4dd2ee198127f6bc61cc
3
+ size 1103913424
model-00014-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c3b6f2b4b0c3372facdf26318ee1609a117e36ddc649a3d84ca2c14010d828d
3
+ size 1253945480
model-00015-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c1409db098265dfb02ba1b6fee96d723060ce72a3d78cbd44b8976669cea947b
3
+ size 1342097424
model-00016-of-00016.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4095466c30abcd185f5a1255cae1a42a07bab9f1c08796d1c701f9fe255167c8
3
+ size 1383099712
model-mtp.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6c8f36002c1db3906a771f613a7295486402f5c39e782df1607f01bce5f2e3f
3
+ size 1689376064
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
preprocessor_config.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": {
3
+ "longest_edge": 16777216,
4
+ "shortest_edge": 65536
5
+ },
6
+ "patch_size": 16,
7
+ "temporal_patch_size": 2,
8
+ "merge_size": 2,
9
+ "image_mean": [
10
+ 0.5,
11
+ 0.5,
12
+ 0.5
13
+ ],
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "processor_class": "Qwen3VLProcessor",
20
+ "image_processor_type": "Qwen2VLImageProcessorFast"
21
+ }
processor_config.json ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "image_processor": {
3
+ "do_convert_rgb": true,
4
+ "do_normalize": true,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Qwen2VLImageProcessor",
13
+ "image_std": [
14
+ 0.5,
15
+ 0.5,
16
+ 0.5
17
+ ],
18
+ "merge_size": 2,
19
+ "patch_size": 16,
20
+ "resample": 3,
21
+ "rescale_factor": 0.00392156862745098,
22
+ "size": {
23
+ "longest_edge": 16777216,
24
+ "shortest_edge": 65536
25
+ },
26
+ "temporal_patch_size": 2
27
+ },
28
+ "processor_class": "Qwen3VLProcessor",
29
+ "video_processor": {
30
+ "do_convert_rgb": true,
31
+ "do_normalize": true,
32
+ "do_rescale": true,
33
+ "do_resize": true,
34
+ "do_sample_frames": true,
35
+ "fps": 2,
36
+ "image_mean": [
37
+ 0.5,
38
+ 0.5,
39
+ 0.5
40
+ ],
41
+ "image_std": [
42
+ 0.5,
43
+ 0.5,
44
+ 0.5
45
+ ],
46
+ "max_frames": 768,
47
+ "merge_size": 2,
48
+ "min_frames": 4,
49
+ "patch_size": 16,
50
+ "resample": 3,
51
+ "rescale_factor": 0.00392156862745098,
52
+ "return_metadata": false,
53
+ "size": {
54
+ "longest_edge": 25165824,
55
+ "shortest_edge": 4096
56
+ },
57
+ "temporal_patch_size": 2,
58
+ "video_processor_type": "Qwen3VLVideoProcessor"
59
+ }
60
+ }
recipe_graft_mxfp4.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ """Graft a Qwen MoE MTP head into an MXFP4 (compressed-tensors) trunk.
3
+
4
+ Produces this repo: an MXFP4-quantized Ornith-1.0-35B with an MTP draft head transplanted in, so
5
+ vLLM does lossless self-speculative decoding (`--speculative-config
6
+ '{"method":"mtp","num_speculative_tokens":3}'`).
7
+
8
+ Generalized over head shape — it copies EVERY ``mtp.*`` tensor from the donor and derives the
9
+ unquantized-ignore list from the rank-2 (matmul) tensors, so the same script handles a small dense
10
+ head (a handful of Linears) and a MoE head (the `fc`, attention projections, AND hundreds of
11
+ per-expert MLP projections) with no per-model tensor list.
12
+
13
+ It does two things:
14
+ 1. Copies all ``mtp.*`` head tensors (BF16) from a donor checkpoint into a new
15
+ ``model-mtp.safetensors`` shard and patches ``model.safetensors.index.json``. Big trunk shards
16
+ are hard-linked (no multi-GB copy); small files are copied.
17
+ 2. Adds each ``mtp.*`` Linear module to ``quantization_config.ignore`` so the compressed-tensors
18
+ loader keeps the BF16 head as-is instead of expecting MXFP4 weight-scales. A tensor is treated
19
+ as a Linear (and ignored) iff it is rank-2 and ends in ``.weight``; rank-1 norms are left out.
20
+ This is the one mixed-precision gotcha — without it the model won't load.
21
+
22
+ Inputs:
23
+ --donor a checkpoint that SHIPS the mtp.* head (here: the Qwen MoE thinking distill, which keeps
24
+ its trained MoE MTP head in BF16).
25
+ --target an MXFP4 compressed-tensors trunk of Ornith-1.0-35B (lm_head/embed_tokens/vision BF16).
26
+ --out output checkpoint dir.
27
+
28
+ Credits: trunk — DeepReinforce (Ornith-1.0-35B, MIT); MTP head — nerkyor's
29
+ Qwen3.6-35B-A3B-DSV4Pro-Thinking distill (Apache-2.0), a DeepSeek-V4-Pro-Thinking distill of
30
+ Qwen3.6-35B-A3B; architecture — Qwen MoE. Serving — vLLM + compressed-tensors. RDNA4 base image —
31
+ Rob Smith / tcclaviger. License of the combined work: Apache-2.0.
32
+
33
+ Usage:
34
+ python recipe_graft_mxfp4_35b.py \
35
+ --donor Capicua25x/Qwen3.6-35B-A3B-DSV4Pro-Thinking-Distill-MXFP4-Vision \
36
+ --target ./Ornith-1.0-35B-MXFP4 \
37
+ --out ./Ornith-1.0-35B-MXFP4-Vision-MTP
38
+ """
39
+ import argparse, glob, json, os, shutil
40
+ from safetensors import safe_open
41
+ from safetensors.torch import save_file
42
+
43
+ MTP_SHARD = "model-mtp.safetensors"
44
+
45
+
46
+ def resolve(path_or_repo: str) -> str:
47
+ """Local dir with weights -> itself; else download the HF snapshot."""
48
+ if os.path.isdir(path_or_repo) and glob.glob(os.path.join(path_or_repo, "*.safetensors")):
49
+ return path_or_repo
50
+ from huggingface_hub import snapshot_download
51
+ return snapshot_download(path_or_repo)
52
+
53
+
54
+ def main() -> int:
55
+ ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
56
+ ap.add_argument("--donor", required=True, help="checkpoint that ships the mtp.* head")
57
+ ap.add_argument("--target", required=True, help="MXFP4 compressed-tensors trunk to graft into")
58
+ ap.add_argument("--out", required=True, help="output checkpoint dir")
59
+ ap.add_argument("--force", action="store_true", help="overwrite --out if it exists")
60
+ args = ap.parse_args()
61
+
62
+ donor, target = resolve(args.donor), resolve(args.target)
63
+
64
+ # 1. pull EVERY mtp.* tensor from the donor; record the Linear (rank-2) modules to ignore
65
+ head, ignore = {}, []
66
+ for shard in glob.glob(os.path.join(donor, "*.safetensors")):
67
+ with safe_open(shard, framework="pt") as f:
68
+ for k in f.keys():
69
+ if not k.startswith("mtp."):
70
+ continue
71
+ t = f.get_tensor(k).contiguous()
72
+ head[k] = t
73
+ if k.endswith(".weight") and t.dim() == 2: # a Linear weight -> mark unquantized
74
+ ignore.append(k[: -len(".weight")])
75
+ assert head, "donor ships no mtp.* tensors"
76
+ ignore = sorted(set(ignore))
77
+ print(f"head: {len(head)} tensors, {len(ignore)} Linear modules to ignore "
78
+ f"(all BF16: {all(str(t.dtype) == 'torch.bfloat16' for t in head.values())})")
79
+
80
+ # 2. validate the target can host the head
81
+ idx = json.load(open(os.path.join(target, "model.safetensors.index.json")))
82
+ wm = idx["weight_map"]
83
+ assert "lm_head.weight" in wm, "target missing lm_head.weight"
84
+ assert any(k.endswith("embed_tokens.weight") for k in wm), "target missing embed_tokens"
85
+ assert not any(k.startswith("mtp.") for k in wm), "target already has an mtp head"
86
+
87
+ # 3. build output: hard-link big shards, copy small files
88
+ if os.path.exists(args.out):
89
+ assert args.force, f"{args.out} exists (use --force)"
90
+ shutil.rmtree(args.out)
91
+ os.makedirs(args.out)
92
+ for fn in os.listdir(target):
93
+ src = os.path.realpath(os.path.join(target, fn))
94
+ if not os.path.isfile(src):
95
+ continue
96
+ dst = os.path.join(args.out, fn)
97
+ if fn.endswith(".safetensors"):
98
+ try:
99
+ os.link(src, dst)
100
+ except OSError:
101
+ shutil.copy2(src, dst)
102
+ else:
103
+ shutil.copy2(src, dst)
104
+
105
+ # 4. write the mtp shard + patch the index
106
+ save_file(head, os.path.join(args.out, MTP_SHARD), metadata={"format": "pt"})
107
+ added = 0
108
+ for k, t in head.items():
109
+ wm[k] = MTP_SHARD
110
+ added += t.numel() * t.element_size()
111
+ idx.setdefault("metadata", {})
112
+ if "total_size" in idx["metadata"]:
113
+ idx["metadata"]["total_size"] += added
114
+ json.dump(idx, open(os.path.join(args.out, "model.safetensors.index.json"), "w"), indent=2)
115
+
116
+ # 5. mark the head unquantized in the compressed-tensors config
117
+ cfg = json.load(open(os.path.join(args.out, "config.json")))
118
+ ig = cfg["quantization_config"]["ignore"]
119
+ for m in ignore:
120
+ if m not in ig:
121
+ ig.append(m)
122
+ json.dump(cfg, open(os.path.join(args.out, "config.json"), "w"), indent=2)
123
+
124
+ print(f"OK -> {args.out} (+{MTP_SHARD} {added/1e6:.1f} MB, +{len(ignore)} ignore entries)")
125
+ print("serve: vllm serve <out> --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":3}'")
126
+ return 0
127
+
128
+
129
+ if __name__ == "__main__":
130
+ raise SystemExit(main())
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
3
+ size 19989325
tokenizer_config.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "audio_bos_token": "<|audio_start|>",
4
+ "audio_eos_token": "<|audio_end|>",
5
+ "audio_token": "<|audio_pad|>",
6
+ "backend": "tokenizers",
7
+ "bos_token": null,
8
+ "clean_up_tokenization_spaces": false,
9
+ "eos_token": "<|im_end|>",
10
+ "errors": "replace",
11
+ "image_token": "<|image_pad|>",
12
+ "is_local": true,
13
+ "local_files_only": false,
14
+ "model_max_length": 262144,
15
+ "model_specific_special_tokens": {
16
+ "audio_bos_token": "<|audio_start|>",
17
+ "audio_eos_token": "<|audio_end|>",
18
+ "audio_token": "<|audio_pad|>",
19
+ "image_token": "<|image_pad|>",
20
+ "video_token": "<|video_pad|>",
21
+ "vision_bos_token": "<|vision_start|>",
22
+ "vision_eos_token": "<|vision_end|>"
23
+ },
24
+ "pad_token": "<|endoftext|>",
25
+ "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
26
+ "processor_class": "Qwen3VLProcessor",
27
+ "split_special_tokens": false,
28
+ "tokenizer_class": "Qwen2Tokenizer",
29
+ "unk_token": null,
30
+ "video_token": "<|video_pad|>",
31
+ "vision_bos_token": "<|vision_start|>",
32
+ "vision_eos_token": "<|vision_end|>"
33
+ }
video_preprocessor_config.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "size": {
3
+ "longest_edge": 25165824,
4
+ "shortest_edge": 4096
5
+ },
6
+ "patch_size": 16,
7
+ "temporal_patch_size": 2,
8
+ "merge_size": 2,
9
+ "image_mean": [
10
+ 0.5,
11
+ 0.5,
12
+ 0.5
13
+ ],
14
+ "image_std": [
15
+ 0.5,
16
+ 0.5,
17
+ 0.5
18
+ ],
19
+ "processor_class": "Qwen3VLProcessor",
20
+ "video_processor_type": "Qwen3VLVideoProcessor"
21
+ }
vocab.json ADDED
The diff for this file is too large to render. See raw diff