Hugging Face's logo

aidendle94
/
GLM-5.2-MXFP4-Experts-GPTQ

Safetensors
vllm
glm_moe_dsa
glm
Mixture of Experts
mxfp4
fp8
gptq
mixed-precision
quantization
dgx-spark
gb10
hybrid_mxfp4_ct
Model card Files
xet
 Community
aidendle94 commited on
Commit
5285593
·
verified ·
1 Parent(s): e417136

Full standalone model: FP8 attention/shared + NVFP4 dense + GPTQ-MXFP4 experts + MTP draft + stitched index

Browse files
Files changed (18) hide show
  1. .gitattributes +1 -0
  2. LICENSE +25 -0
  3. chat_template.jinja +119 -0
  4. config.json +453 -0
  5. generation_config.json +12 -0
  6. hybrid-ct-00000.safetensors +3 -0
  7. hybrid-ct-00001.safetensors +3 -0
  8. hybrid-ct-00002.safetensors +3 -0
  9. model.safetensors.index.json +0 -0
  10. mtp-draft/config.json +425 -0
  11. mtp-draft/model-mtp-inputscales.safetensors +3 -0
  12. mtp-draft/model-mtp.safetensors +3 -0
  13. mtp-draft/model.safetensors.index.json +0 -0
  14. recipe.yaml +99 -0
  15. tokenizer.json +3 -0
  16. tokenizer_config.json +34 -0
  17. vllm_overlay/INSTALL.md +71 -0
  18. vllm_overlay/hybrid_mxfp4_ct.py +86 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ 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
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ This repository repackages, without re-quantization, tensors from:
4
+ - zai-org/GLM-5.2 (MIT) — the base model
5
+ - RedHatAI/GLM-5.2-NVFP4-FP8 (MIT) — FP8/NVFP4 compressed-tensors quantization
6
+ - festr2/GLM-5.2-BF16-AMDMXFP4experts (MIT) — MXFP4 (AMD Quark-calibrated)
7
+ routed-expert and MTP-expert tensors
8
+
9
+ Permission is hereby granted, free of charge, to any person obtaining a copy
10
+ of this software and associated documentation files (the "Software"), to deal
11
+ in the Software without restriction, including without limitation the rights
12
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
13
+ copies of the Software, and to permit persons to whom the Software is
14
+ furnished to do so, subject to the following conditions:
15
+
16
+ The above copyright notice and this permission notice shall be included in all
17
+ copies or substantial portions of the Software.
18
+
19
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
22
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
24
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
25
+ SOFTWARE.
chat_template.jinja ADDED
@@ -0,0 +1,119 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [gMASK]<sop>
2
+ {%- set effective_reasoning_effort = 'high' if reasoning_effort is defined and reasoning_effort == 'high' else 'max' -%}
3
+ {%- if (enable_thinking is not defined or enable_thinking) and effective_reasoning_effort is not none -%}<|system|>Reasoning Effort: {{ effective_reasoning_effort | capitalize }}{%- endif -%}
4
+ {%- if tools -%}
5
+ {%- macro tool_to_json(tool) -%}
6
+ {%- set ns_tool = namespace(first=true) -%}
7
+ {{ '{' -}}
8
+ {%- for k, v in tool.items() -%}
9
+ {%- if k != 'defer_loading' and k != 'strict' -%}
10
+ {%- if not ns_tool.first -%}{{- ', ' -}}{%- endif -%}
11
+ {%- set ns_tool.first = false -%}
12
+ "{{ k }}": {{ v | tojson(ensure_ascii=False) }}
13
+ {%- endif -%}
14
+ {%- endfor -%}
15
+ {{- '}' -}}
16
+ {%- endmacro -%}
17
+ <|system|>
18
+ # Tools
19
+
20
+ You may call one or more functions to assist with the user query.
21
+
22
+ You are provided with function signatures within <tools></tools> XML tags:
23
+ <tools>
24
+ {% for tool in tools %}
25
+ {%- if 'function' in tool -%}
26
+ {%- set tool = tool['function'] -%}
27
+ {%- endif -%}
28
+ {% if tool.defer_loading is not defined or not tool.defer_loading %}
29
+ {{ tool_to_json(tool) }}
30
+ {% endif %}
31
+ {% endfor %}
32
+ </tools>
33
+
34
+ For each function call, output the function name and arguments within the following XML format:
35
+ <tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>{%- endif -%}
36
+ {%- macro visible_text(content) -%}
37
+ {%- if content is string -%}
38
+ {{- content }}
39
+ {%- elif content is iterable and content is not mapping -%}
40
+ {%- for item in content -%}
41
+ {%- if item is mapping and item.type == 'text' -%}
42
+ {{- item.text }}
43
+ {%- elif item is string -%}
44
+ {{- item }}
45
+ {%- elif item is mapping and item.type in ['image', 'image_url', 'video', 'video_url', 'audio', 'audio_url', 'input_audio'] -%}
46
+ {%- set media_type = item.type | replace('_url', '') | replace('input_', '') -%}
47
+ {{- "<reminder>You are unable to process this " ~ media_type ~ " because you don't have multi-modal input ability. Try different methods.</reminder>" }}
48
+ {%- endif -%}
49
+ {%- endfor -%}
50
+ {%- else -%}
51
+ {{- content }}
52
+ {%- endif -%}
53
+ {%- endmacro -%}
54
+ {%- set ns = namespace(last_user_index=-1) -%}
55
+ {%- for m in messages %}
56
+ {%- if m.role == 'user' %}
57
+ {%- set ns.last_user_index = loop.index0 -%}
58
+ {%- endif %}
59
+ {%- endfor %}
60
+ {%- for m in messages -%}
61
+ {%- if m.role == 'user' -%}<|user|>{{ visible_text(m.content) }}
62
+ {%- elif m.role == 'assistant' -%}
63
+ <|assistant|>
64
+ {%- set content = visible_text(m.content) %}
65
+ {%- if m.reasoning_content is string %}
66
+ {%- set reasoning_content = m.reasoning_content %}
67
+ {%- elif '</think>' in content %}
68
+ {%- set reasoning_content = content.split('</think>')[0].split('<think>')[-1] %}
69
+ {%- set content = content.split('</think>')[-1] %}
70
+ {%- endif %}
71
+ {%- if ((clear_thinking is defined and not clear_thinking) or loop.index0 > ns.last_user_index) and reasoning_content is defined -%}
72
+ {{ '<think>' + reasoning_content + '</think>'}}
73
+ {%- else -%}
74
+ {{ '<think></think>' }}
75
+ {%- endif -%}
76
+ {%- if content.strip() -%}
77
+ {{ content.strip() }}
78
+ {%- endif -%}
79
+ {% if m.tool_calls %}
80
+ {% for tc in m.tool_calls %}
81
+ {%- if tc.function %}
82
+ {%- set tc = tc.function %}
83
+ {%- endif %}
84
+ {{- '<tool_call>' + tc.name -}}
85
+ {% set _args = tc.arguments %}{% for k, v in _args.items() %}<arg_key>{{ k }}</arg_key><arg_value>{{ v | tojson(ensure_ascii=False) if v is not string else v }}</arg_value>{% endfor %}</tool_call>{% endfor %}
86
+ {% endif %}
87
+ {%- elif m.role == 'tool' -%}
88
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
89
+ {{- '<|observation|>' -}}
90
+ {%- endif %}
91
+ {%- if m.content is string -%}
92
+ {{- '<tool_response>' + m.content + '</tool_response>' -}}
93
+ {%- elif m.content is iterable and m.content is not mapping and m.content and m.content.0.type == "tool_reference" -%}
94
+ {{- '<tool_response><tools>\n' -}}
95
+ {% for tr in m.content %}
96
+ {%- for tool in tools -%}
97
+ {%- if 'function' in tool -%}
98
+ {%- set tool = tool['function'] -%}
99
+ {%- endif -%}
100
+ {%- if tool.name == tr.name -%}
101
+ {{- tool_to_json(tool) + '\n' -}}
102
+ {%- endif -%}
103
+ {%- endfor -%}
104
+ {%- endfor -%}
105
+ {{- '</tools></tool_response>' -}}
106
+ {%- elif m.content is iterable and m.content is not mapping and m.content and m.content.0 is mapping and m.content.0.output is defined -%}
107
+ {%- for tr in m.content -%}
108
+ {{- '<tool_response>' + tr.output + '</tool_response>' -}}
109
+ {%- endfor -%}
110
+ {%- else -%}
111
+ {{- '<tool_response>' + visible_text(m.content) + '</tool_response>' -}}
112
+ {% endif -%}
113
+ {%- elif m.role == 'system' -%}
114
+ <|system|>{{ visible_text(m.content) }}
115
+ {%- endif -%}
116
+ {%- endfor -%}
117
+ {%- if add_generation_prompt -%}
118
+ <|assistant|>{{- '<think></think>' if (enable_thinking is defined and not enable_thinking) else '<think>' -}}
119
+ {%- endif -%}
config.json ADDED
@@ -0,0 +1,453 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "GlmMoeDsaForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 0,
8
+ "dtype": "bfloat16",
9
+ "eos_token_id": [
10
+ 154820,
11
+ 154827,
12
+ 154829
13
+ ],
14
+ "ep_size": 1,
15
+ "first_k_dense_replace": 3,
16
+ "head_dim": 64,
17
+ "hidden_act": "silu",
18
+ "hidden_size": 6144,
19
+ "index_head_dim": 128,
20
+ "index_n_heads": 32,
21
+ "index_share_for_mtp_iteration": true,
22
+ "index_skip_topk_offset": 3,
23
+ "index_topk": 2048,
24
+ "index_topk_freq": 4,
25
+ "index_topk_pattern": null,
26
+ "indexer_rope_interleave": true,
27
+ "indexer_types": [
28
+ "full",
29
+ "full",
30
+ "full",
31
+ "shared",
32
+ "shared",
33
+ "shared",
34
+ "full",
35
+ "shared",
36
+ "shared",
37
+ "shared",
38
+ "full",
39
+ "shared",
40
+ "shared",
41
+ "shared",
42
+ "full",
43
+ "shared",
44
+ "shared",
45
+ "shared",
46
+ "full",
47
+ "shared",
48
+ "shared",
49
+ "shared",
50
+ "full",
51
+ "shared",
52
+ "shared",
53
+ "shared",
54
+ "full",
55
+ "shared",
56
+ "shared",
57
+ "shared",
58
+ "full",
59
+ "shared",
60
+ "shared",
61
+ "shared",
62
+ "full",
63
+ "shared",
64
+ "shared",
65
+ "shared",
66
+ "full",
67
+ "shared",
68
+ "shared",
69
+ "shared",
70
+ "full",
71
+ "shared",
72
+ "shared",
73
+ "shared",
74
+ "full",
75
+ "shared",
76
+ "shared",
77
+ "shared",
78
+ "full",
79
+ "shared",
80
+ "shared",
81
+ "shared",
82
+ "full",
83
+ "shared",
84
+ "shared",
85
+ "shared",
86
+ "full",
87
+ "shared",
88
+ "shared",
89
+ "shared",
90
+ "full",
91
+ "shared",
92
+ "shared",
93
+ "shared",
94
+ "full",
95
+ "shared",
96
+ "shared",
97
+ "shared",
98
+ "full",
99
+ "shared",
100
+ "shared",
101
+ "shared",
102
+ "full",
103
+ "shared",
104
+ "shared",
105
+ "shared"
106
+ ],
107
+ "initializer_range": 0.02,
108
+ "intermediate_size": 12288,
109
+ "kv_lora_rank": 512,
110
+ "max_position_embeddings": 1048576,
111
+ "mlp_bias": false,
112
+ "mlp_layer_types": [
113
+ "dense",
114
+ "dense",
115
+ "dense",
116
+ "sparse",
117
+ "sparse",
118
+ "sparse",
119
+ "sparse",
120
+ "sparse",
121
+ "sparse",
122
+ "sparse",
123
+ "sparse",
124
+ "sparse",
125
+ "sparse",
126
+ "sparse",
127
+ "sparse",
128
+ "sparse",
129
+ "sparse",
130
+ "sparse",
131
+ "sparse",
132
+ "sparse",
133
+ "sparse",
134
+ "sparse",
135
+ "sparse",
136
+ "sparse",
137
+ "sparse",
138
+ "sparse",
139
+ "sparse",
140
+ "sparse",
141
+ "sparse",
142
+ "sparse",
143
+ "sparse",
144
+ "sparse",
145
+ "sparse",
146
+ "sparse",
147
+ "sparse",
148
+ "sparse",
149
+ "sparse",
150
+ "sparse",
151
+ "sparse",
152
+ "sparse",
153
+ "sparse",
154
+ "sparse",
155
+ "sparse",
156
+ "sparse",
157
+ "sparse",
158
+ "sparse",
159
+ "sparse",
160
+ "sparse",
161
+ "sparse",
162
+ "sparse",
163
+ "sparse",
164
+ "sparse",
165
+ "sparse",
166
+ "sparse",
167
+ "sparse",
168
+ "sparse",
169
+ "sparse",
170
+ "sparse",
171
+ "sparse",
172
+ "sparse",
173
+ "sparse",
174
+ "sparse",
175
+ "sparse",
176
+ "sparse",
177
+ "sparse",
178
+ "sparse",
179
+ "sparse",
180
+ "sparse",
181
+ "sparse",
182
+ "sparse",
183
+ "sparse",
184
+ "sparse",
185
+ "sparse",
186
+ "sparse",
187
+ "sparse",
188
+ "sparse",
189
+ "sparse",
190
+ "sparse"
191
+ ],
192
+ "model_type": "glm_moe_dsa",
193
+ "moe_intermediate_size": 2048,
194
+ "moe_layer_freq": 1,
195
+ "n_group": 1,
196
+ "n_routed_experts": 256,
197
+ "n_shared_experts": 1,
198
+ "norm_topk_prob": true,
199
+ "num_attention_heads": 64,
200
+ "num_experts_per_tok": 8,
201
+ "num_hidden_layers": 78,
202
+ "num_key_value_heads": 64,
203
+ "num_nextn_predict_layers": 1,
204
+ "pad_token_id": 154820,
205
+ "pretraining_tp": 1,
206
+ "q_lora_rank": 2048,
207
+ "qk_head_dim": 256,
208
+ "qk_nope_head_dim": 192,
209
+ "qk_rope_head_dim": 64,
210
+ "quantization_config": {
211
+ "quant_method": "hybrid_mxfp4_ct",
212
+ "moe": {
213
+ "quant_method": "mxfp4"
214
+ },
215
+ "linear": {
216
+ "config_groups": {
217
+ "group_0": {
218
+ "format": "float-quantized",
219
+ "input_activations": {
220
+ "actorder": null,
221
+ "block_structure": null,
222
+ "dynamic": true,
223
+ "group_size": 128,
224
+ "num_bits": 8,
225
+ "observer": null,
226
+ "observer_kwargs": {},
227
+ "scale_dtype": null,
228
+ "strategy": "group",
229
+ "symmetric": true,
230
+ "type": "float",
231
+ "zp_dtype": null
232
+ },
233
+ "output_activations": null,
234
+ "targets": [
235
+ "re:.*self_attn\\..*"
236
+ ],
237
+ "weights": {
238
+ "actorder": null,
239
+ "block_structure": [
240
+ 128,
241
+ 128
242
+ ],
243
+ "dynamic": false,
244
+ "group_size": null,
245
+ "num_bits": 8,
246
+ "observer": "memoryless_minmax",
247
+ "observer_kwargs": {},
248
+ "scale_dtype": null,
249
+ "strategy": "block",
250
+ "symmetric": true,
251
+ "type": "float",
252
+ "zp_dtype": null
253
+ }
254
+ },
255
+ "group_1": {
256
+ "format": "nvfp4-pack-quantized",
257
+ "input_activations": {
258
+ "actorder": null,
259
+ "block_structure": null,
260
+ "dynamic": "local",
261
+ "group_size": 16,
262
+ "num_bits": 4,
263
+ "observer": "static_minmax",
264
+ "observer_kwargs": {},
265
+ "scale_dtype": "torch.float8_e4m3fn",
266
+ "strategy": "tensor_group",
267
+ "symmetric": true,
268
+ "type": "float",
269
+ "zp_dtype": null
270
+ },
271
+ "output_activations": null,
272
+ "targets": [
273
+ "re:.*\\.mlp\\.(gate_proj|up_proj|down_proj)$"
274
+ ],
275
+ "weights": {
276
+ "actorder": null,
277
+ "block_structure": null,
278
+ "dynamic": false,
279
+ "group_size": 16,
280
+ "num_bits": 4,
281
+ "observer": "memoryless_minmax",
282
+ "observer_kwargs": {},
283
+ "scale_dtype": "torch.float8_e4m3fn",
284
+ "strategy": "tensor_group",
285
+ "symmetric": true,
286
+ "type": "float",
287
+ "zp_dtype": null
288
+ }
289
+ },
290
+ "group_2": {
291
+ "format": "float-quantized",
292
+ "input_activations": {
293
+ "actorder": null,
294
+ "block_structure": null,
295
+ "dynamic": true,
296
+ "group_size": 128,
297
+ "num_bits": 8,
298
+ "observer": null,
299
+ "observer_kwargs": {},
300
+ "scale_dtype": null,
301
+ "strategy": "group",
302
+ "symmetric": true,
303
+ "type": "float",
304
+ "zp_dtype": null
305
+ },
306
+ "output_activations": null,
307
+ "targets": [
308
+ "re:.*\\.mlp\\.shared_experts\\.(gate_proj|up_proj|down_proj)$"
309
+ ],
310
+ "weights": {
311
+ "actorder": null,
312
+ "block_structure": [
313
+ 128,
314
+ 128
315
+ ],
316
+ "dynamic": false,
317
+ "group_size": null,
318
+ "num_bits": 8,
319
+ "observer": "memoryless_minmax",
320
+ "observer_kwargs": {},
321
+ "scale_dtype": null,
322
+ "strategy": "block",
323
+ "symmetric": true,
324
+ "type": "float",
325
+ "zp_dtype": null
326
+ }
327
+ }
328
+ },
329
+ "format": "mixed-precision",
330
+ "global_compression_ratio": null,
331
+ "ignore": [
332
+ "model.layers.0.self_attn.indexer.weights_proj",
333
+ "model.layers.1.self_attn.indexer.weights_proj",
334
+ "model.layers.2.self_attn.indexer.weights_proj",
335
+ "model.layers.3.mlp.gate",
336
+ "model.layers.4.mlp.gate",
337
+ "model.layers.5.mlp.gate",
338
+ "model.layers.6.self_attn.indexer.weights_proj",
339
+ "model.layers.6.mlp.gate",
340
+ "model.layers.7.mlp.gate",
341
+ "model.layers.8.mlp.gate",
342
+ "model.layers.9.mlp.gate",
343
+ "model.layers.10.self_attn.indexer.weights_proj",
344
+ "model.layers.10.mlp.gate",
345
+ "model.layers.11.mlp.gate",
346
+ "model.layers.12.mlp.gate",
347
+ "model.layers.13.mlp.gate",
348
+ "model.layers.14.self_attn.indexer.weights_proj",
349
+ "model.layers.14.mlp.gate",
350
+ "model.layers.15.mlp.gate",
351
+ "model.layers.16.mlp.gate",
352
+ "model.layers.17.mlp.gate",
353
+ "model.layers.18.self_attn.indexer.weights_proj",
354
+ "model.layers.18.mlp.gate",
355
+ "model.layers.19.mlp.gate",
356
+ "model.layers.20.mlp.gate",
357
+ "model.layers.21.mlp.gate",
358
+ "model.layers.22.self_attn.indexer.weights_proj",
359
+ "model.layers.22.mlp.gate",
360
+ "model.layers.23.mlp.gate",
361
+ "model.layers.24.mlp.gate",
362
+ "model.layers.25.mlp.gate",
363
+ "model.layers.26.self_attn.indexer.weights_proj",
364
+ "model.layers.26.mlp.gate",
365
+ "model.layers.27.mlp.gate",
366
+ "model.layers.28.mlp.gate",
367
+ "model.layers.29.mlp.gate",
368
+ "model.layers.30.self_attn.indexer.weights_proj",
369
+ "model.layers.30.mlp.gate",
370
+ "model.layers.31.mlp.gate",
371
+ "model.layers.32.mlp.gate",
372
+ "model.layers.33.mlp.gate",
373
+ "model.layers.34.self_attn.indexer.weights_proj",
374
+ "model.layers.34.mlp.gate",
375
+ "model.layers.35.mlp.gate",
376
+ "model.layers.36.mlp.gate",
377
+ "model.layers.37.mlp.gate",
378
+ "model.layers.38.self_attn.indexer.weights_proj",
379
+ "model.layers.38.mlp.gate",
380
+ "model.layers.39.mlp.gate",
381
+ "model.layers.40.mlp.gate",
382
+ "model.layers.41.mlp.gate",
383
+ "model.layers.42.self_attn.indexer.weights_proj",
384
+ "model.layers.42.mlp.gate",
385
+ "model.layers.43.mlp.gate",
386
+ "model.layers.44.mlp.gate",
387
+ "model.layers.45.mlp.gate",
388
+ "model.layers.46.self_attn.indexer.weights_proj",
389
+ "model.layers.46.mlp.gate",
390
+ "model.layers.47.mlp.gate",
391
+ "model.layers.48.mlp.gate",
392
+ "model.layers.49.mlp.gate",
393
+ "model.layers.50.self_attn.indexer.weights_proj",
394
+ "model.layers.50.mlp.gate",
395
+ "model.layers.51.mlp.gate",
396
+ "model.layers.52.mlp.gate",
397
+ "model.layers.53.mlp.gate",
398
+ "model.layers.54.self_attn.indexer.weights_proj",
399
+ "model.layers.54.mlp.gate",
400
+ "model.layers.55.mlp.gate",
401
+ "model.layers.56.mlp.gate",
402
+ "model.layers.57.mlp.gate",
403
+ "model.layers.58.self_attn.indexer.weights_proj",
404
+ "model.layers.58.mlp.gate",
405
+ "model.layers.59.mlp.gate",
406
+ "model.layers.60.mlp.gate",
407
+ "model.layers.61.mlp.gate",
408
+ "model.layers.62.self_attn.indexer.weights_proj",
409
+ "model.layers.62.mlp.gate",
410
+ "model.layers.63.mlp.gate",
411
+ "model.layers.64.mlp.gate",
412
+ "model.layers.65.mlp.gate",
413
+ "model.layers.66.self_attn.indexer.weights_proj",
414
+ "model.layers.66.mlp.gate",
415
+ "model.layers.67.mlp.gate",
416
+ "model.layers.68.mlp.gate",
417
+ "model.layers.69.mlp.gate",
418
+ "model.layers.70.self_attn.indexer.weights_proj",
419
+ "model.layers.70.mlp.gate",
420
+ "model.layers.71.mlp.gate",
421
+ "model.layers.72.mlp.gate",
422
+ "model.layers.73.mlp.gate",
423
+ "model.layers.74.self_attn.indexer.weights_proj",
424
+ "model.layers.74.mlp.gate",
425
+ "model.layers.75.mlp.gate",
426
+ "model.layers.76.mlp.gate",
427
+ "model.layers.77.mlp.gate",
428
+ "lm_head"
429
+ ],
430
+ "kv_cache_scheme": null,
431
+ "quant_method": "compressed-tensors",
432
+ "quantization_status": "compressed",
433
+ "sparsity_config": {},
434
+ "transform_config": {},
435
+ "version": "0.15.1.dev22+g792acec"
436
+ }
437
+ },
438
+ "rms_norm_eps": 1e-05,
439
+ "rope_interleave": true,
440
+ "rope_parameters": {
441
+ "rope_theta": 8000000,
442
+ "rope_type": "default"
443
+ },
444
+ "routed_scaling_factor": 2.5,
445
+ "scoring_func": "sigmoid",
446
+ "tie_word_embeddings": false,
447
+ "topk_group": 1,
448
+ "topk_method": "noaux_tc",
449
+ "transformers_version": "5.13.0.dev0",
450
+ "use_cache": true,
451
+ "v_head_dim": 256,
452
+ "vocab_size": 154880
453
+ }
generation_config.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "eos_token_id": [
4
+ 154820,
5
+ 154827,
6
+ 154829
7
+ ],
8
+ "pad_token_id": 154820,
9
+ "temperature": 1.0,
10
+ "transformers_version": "5.13.0.dev0",
11
+ "top_p": 0.95
12
+ }
hybrid-ct-00000.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec5d602153f78fc2d725a1c05a1284bdfb91e91733569eb27bf46e66191fee9c
3
+ size 8939483840
hybrid-ct-00001.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ae035e0d7df1c54fbdb94bdc739b04dbdc45089ee5eead3bd4021c6a66f9e7a1
3
+ size 9319166416
hybrid-ct-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1dffb512bbff15d9b15e8d81d493d6f052d24abf1153b9ccaccb3b273832059d
3
+ size 2654458848
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
mtp-draft/config.json ADDED
@@ -0,0 +1,425 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "GlmMoeDsaForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "dtype": "bfloat16",
8
+ "eos_token_id": [
9
+ 154820,
10
+ 154827,
11
+ 154829
12
+ ],
13
+ "ep_size": 1,
14
+ "first_k_dense_replace": 3,
15
+ "head_dim": 192,
16
+ "hidden_act": "silu",
17
+ "hidden_size": 6144,
18
+ "index_head_dim": 128,
19
+ "index_n_heads": 32,
20
+ "index_share_for_mtp_iteration": true,
21
+ "index_skip_topk_offset": 3,
22
+ "index_topk": 2048,
23
+ "index_topk_freq": 4,
24
+ "index_topk_pattern": null,
25
+ "indexer_rope_interleave": true,
26
+ "indexer_types": [
27
+ "full",
28
+ "full",
29
+ "full",
30
+ "shared",
31
+ "shared",
32
+ "shared",
33
+ "full",
34
+ "shared",
35
+ "shared",
36
+ "shared",
37
+ "full",
38
+ "shared",
39
+ "shared",
40
+ "shared",
41
+ "full",
42
+ "shared",
43
+ "shared",
44
+ "shared",
45
+ "full",
46
+ "shared",
47
+ "shared",
48
+ "shared",
49
+ "full",
50
+ "shared",
51
+ "shared",
52
+ "shared",
53
+ "full",
54
+ "shared",
55
+ "shared",
56
+ "shared",
57
+ "full",
58
+ "shared",
59
+ "shared",
60
+ "shared",
61
+ "full",
62
+ "shared",
63
+ "shared",
64
+ "shared",
65
+ "full",
66
+ "shared",
67
+ "shared",
68
+ "shared",
69
+ "full",
70
+ "shared",
71
+ "shared",
72
+ "shared",
73
+ "full",
74
+ "shared",
75
+ "shared",
76
+ "shared",
77
+ "full",
78
+ "shared",
79
+ "shared",
80
+ "shared",
81
+ "full",
82
+ "shared",
83
+ "shared",
84
+ "shared",
85
+ "full",
86
+ "shared",
87
+ "shared",
88
+ "shared",
89
+ "full",
90
+ "shared",
91
+ "shared",
92
+ "shared",
93
+ "full",
94
+ "shared",
95
+ "shared",
96
+ "shared",
97
+ "full",
98
+ "shared",
99
+ "shared",
100
+ "shared",
101
+ "full",
102
+ "shared",
103
+ "shared",
104
+ "shared"
105
+ ],
106
+ "initializer_range": 0.02,
107
+ "intermediate_size": 12288,
108
+ "kv_lora_rank": 512,
109
+ "max_position_embeddings": 1048576,
110
+ "mlp_layer_types": [
111
+ "dense",
112
+ "dense",
113
+ "dense",
114
+ "sparse",
115
+ "sparse",
116
+ "sparse",
117
+ "sparse",
118
+ "sparse",
119
+ "sparse",
120
+ "sparse",
121
+ "sparse",
122
+ "sparse",
123
+ "sparse",
124
+ "sparse",
125
+ "sparse",
126
+ "sparse",
127
+ "sparse",
128
+ "sparse",
129
+ "sparse",
130
+ "sparse",
131
+ "sparse",
132
+ "sparse",
133
+ "sparse",
134
+ "sparse",
135
+ "sparse",
136
+ "sparse",
137
+ "sparse",
138
+ "sparse",
139
+ "sparse",
140
+ "sparse",
141
+ "sparse",
142
+ "sparse",
143
+ "sparse",
144
+ "sparse",
145
+ "sparse",
146
+ "sparse",
147
+ "sparse",
148
+ "sparse",
149
+ "sparse",
150
+ "sparse",
151
+ "sparse",
152
+ "sparse",
153
+ "sparse",
154
+ "sparse",
155
+ "sparse",
156
+ "sparse",
157
+ "sparse",
158
+ "sparse",
159
+ "sparse",
160
+ "sparse",
161
+ "sparse",
162
+ "sparse",
163
+ "sparse",
164
+ "sparse",
165
+ "sparse",
166
+ "sparse",
167
+ "sparse",
168
+ "sparse",
169
+ "sparse",
170
+ "sparse",
171
+ "sparse",
172
+ "sparse",
173
+ "sparse",
174
+ "sparse",
175
+ "sparse",
176
+ "sparse",
177
+ "sparse",
178
+ "sparse",
179
+ "sparse",
180
+ "sparse",
181
+ "sparse",
182
+ "sparse",
183
+ "sparse",
184
+ "sparse",
185
+ "sparse",
186
+ "sparse",
187
+ "sparse",
188
+ "sparse"
189
+ ],
190
+ "model_type": "glm_moe_dsa",
191
+ "moe_intermediate_size": 2048,
192
+ "moe_layer_freq": 1,
193
+ "n_group": 1,
194
+ "n_routed_experts": 256,
195
+ "n_shared_experts": 1,
196
+ "norm_topk_prob": true,
197
+ "num_attention_heads": 64,
198
+ "num_experts_per_tok": 8,
199
+ "num_hidden_layers": 78,
200
+ "num_key_value_heads": 64,
201
+ "num_nextn_predict_layers": 1,
202
+ "pad_token_id": 154820,
203
+ "pretraining_tp": 1,
204
+ "q_lora_rank": 2048,
205
+ "qk_head_dim": 256,
206
+ "qk_nope_head_dim": 192,
207
+ "qk_rope_head_dim": 64,
208
+ "rms_norm_eps": 1e-05,
209
+ "rope_interleave": true,
210
+ "rope_parameters": {
211
+ "rope_theta": 8000000,
212
+ "rope_type": "default"
213
+ },
214
+ "routed_scaling_factor": 2.5,
215
+ "scoring_func": "sigmoid",
216
+ "tie_word_embeddings": false,
217
+ "topk_group": 1,
218
+ "topk_method": "noaux_tc",
219
+ "transformers_version": "5.12.0",
220
+ "use_cache": true,
221
+ "v_head_dim": 256,
222
+ "vocab_size": 154880,
223
+ "quantization_config": {
224
+ "config_groups": {
225
+ "group_0": {
226
+ "input_activations": {
227
+ "dynamic": false,
228
+ "num_bits": 4,
229
+ "type": "float",
230
+ "group_size": 16
231
+ },
232
+ "weights": {
233
+ "dynamic": false,
234
+ "num_bits": 4,
235
+ "type": "float",
236
+ "group_size": 16
237
+ },
238
+ "targets": [
239
+ "Linear"
240
+ ]
241
+ }
242
+ },
243
+ "ignore": [
244
+ "lm_head",
245
+ "model.layers.0.mlp*",
246
+ "model.layers.0.self_attn*",
247
+ "model.layers.0.self_attn.indexer*",
248
+ "model.layers.1.mlp*",
249
+ "model.layers.1.self_attn*",
250
+ "model.layers.1.self_attn.indexer*",
251
+ "*shared_experts*",
252
+ "model.layers.10.self_attn*",
253
+ "model.layers.10.self_attn.indexer*",
254
+ "model.layers.11.self_attn*",
255
+ "model.layers.11.self_attn.indexer*",
256
+ "model.layers.12.self_attn*",
257
+ "model.layers.12.self_attn.indexer*",
258
+ "model.layers.13.self_attn*",
259
+ "model.layers.13.self_attn.indexer*",
260
+ "model.layers.14.self_attn*",
261
+ "model.layers.14.self_attn.indexer*",
262
+ "model.layers.15.self_attn*",
263
+ "model.layers.15.self_attn.indexer*",
264
+ "model.layers.16.self_attn*",
265
+ "model.layers.16.self_attn.indexer*",
266
+ "model.layers.17.self_attn*",
267
+ "model.layers.17.self_attn.indexer*",
268
+ "model.layers.18.self_attn*",
269
+ "model.layers.18.self_attn.indexer*",
270
+ "model.layers.19.self_attn*",
271
+ "model.layers.19.self_attn.indexer*",
272
+ "model.layers.2.mlp*",
273
+ "model.layers.2.self_attn*",
274
+ "model.layers.2.self_attn.indexer*",
275
+ "model.layers.20.self_attn*",
276
+ "model.layers.20.self_attn.indexer*",
277
+ "model.layers.21.self_attn*",
278
+ "model.layers.21.self_attn.indexer*",
279
+ "model.layers.22.self_attn*",
280
+ "model.layers.22.self_attn.indexer*",
281
+ "model.layers.23.self_attn*",
282
+ "model.layers.23.self_attn.indexer*",
283
+ "model.layers.24.self_attn*",
284
+ "model.layers.24.self_attn.indexer*",
285
+ "model.layers.25.self_attn*",
286
+ "model.layers.25.self_attn.indexer*",
287
+ "model.layers.26.self_attn*",
288
+ "model.layers.26.self_attn.indexer*",
289
+ "model.layers.27.self_attn*",
290
+ "model.layers.27.self_attn.indexer*",
291
+ "model.layers.28.self_attn*",
292
+ "model.layers.28.self_attn.indexer*",
293
+ "model.layers.29.self_attn*",
294
+ "model.layers.29.self_attn.indexer*",
295
+ "model.layers.3.self_attn*",
296
+ "model.layers.3.self_attn.indexer*",
297
+ "model.layers.30.self_attn*",
298
+ "model.layers.30.self_attn.indexer*",
299
+ "model.layers.31.self_attn*",
300
+ "model.layers.31.self_attn.indexer*",
301
+ "model.layers.32.self_attn*",
302
+ "model.layers.32.self_attn.indexer*",
303
+ "model.layers.33.self_attn*",
304
+ "model.layers.33.self_attn.indexer*",
305
+ "model.layers.34.self_attn*",
306
+ "model.layers.34.self_attn.indexer*",
307
+ "model.layers.35.self_attn*",
308
+ "model.layers.35.self_attn.indexer*",
309
+ "model.layers.36.self_attn*",
310
+ "model.layers.36.self_attn.indexer*",
311
+ "model.layers.37.self_attn*",
312
+ "model.layers.37.self_attn.indexer*",
313
+ "model.layers.38.self_attn*",
314
+ "model.layers.38.self_attn.indexer*",
315
+ "model.layers.39.self_attn*",
316
+ "model.layers.39.self_attn.indexer*",
317
+ "model.layers.4.self_attn*",
318
+ "model.layers.4.self_attn.indexer*",
319
+ "model.layers.40.self_attn*",
320
+ "model.layers.40.self_attn.indexer*",
321
+ "model.layers.41.self_attn*",
322
+ "model.layers.41.self_attn.indexer*",
323
+ "model.layers.42.self_attn*",
324
+ "model.layers.42.self_attn.indexer*",
325
+ "model.layers.43.self_attn*",
326
+ "model.layers.43.self_attn.indexer*",
327
+ "model.layers.44.self_attn*",
328
+ "model.layers.44.self_attn.indexer*",
329
+ "model.layers.45.self_attn*",
330
+ "model.layers.45.self_attn.indexer*",
331
+ "model.layers.46.self_attn*",
332
+ "model.layers.46.self_attn.indexer*",
333
+ "model.layers.47.self_attn*",
334
+ "model.layers.47.self_attn.indexer*",
335
+ "model.layers.48.self_attn*",
336
+ "model.layers.48.self_attn.indexer*",
337
+ "model.layers.49.self_attn*",
338
+ "model.layers.49.self_attn.indexer*",
339
+ "model.layers.5.self_attn*",
340
+ "model.layers.5.self_attn.indexer*",
341
+ "model.layers.50.self_attn*",
342
+ "model.layers.50.self_attn.indexer*",
343
+ "model.layers.51.self_attn*",
344
+ "model.layers.51.self_attn.indexer*",
345
+ "model.layers.52.self_attn*",
346
+ "model.layers.52.self_attn.indexer*",
347
+ "model.layers.53.self_attn*",
348
+ "model.layers.53.self_attn.indexer*",
349
+ "model.layers.54.self_attn*",
350
+ "model.layers.54.self_attn.indexer*",
351
+ "model.layers.55.self_attn*",
352
+ "model.layers.55.self_attn.indexer*",
353
+ "model.layers.56.self_attn*",
354
+ "model.layers.56.self_attn.indexer*",
355
+ "model.layers.57.self_attn*",
356
+ "model.layers.57.self_attn.indexer*",
357
+ "model.layers.58.self_attn*",
358
+ "model.layers.58.self_attn.indexer*",
359
+ "model.layers.59.self_attn*",
360
+ "model.layers.59.self_attn.indexer*",
361
+ "model.layers.6.self_attn*",
362
+ "model.layers.6.self_attn.indexer*",
363
+ "model.layers.60.self_attn*",
364
+ "model.layers.60.self_attn.indexer*",
365
+ "model.layers.61.self_attn*",
366
+ "model.layers.61.self_attn.indexer*",
367
+ "model.layers.62.self_attn*",
368
+ "model.layers.62.self_attn.indexer*",
369
+ "model.layers.63.self_attn*",
370
+ "model.layers.63.self_attn.indexer*",
371
+ "model.layers.64.self_attn*",
372
+ "model.layers.64.self_attn.indexer*",
373
+ "model.layers.65.self_attn*",
374
+ "model.layers.65.self_attn.indexer*",
375
+ "model.layers.66.self_attn*",
376
+ "model.layers.66.self_attn.indexer*",
377
+ "model.layers.67.self_attn*",
378
+ "model.layers.67.self_attn.indexer*",
379
+ "model.layers.68.self_attn*",
380
+ "model.layers.68.self_attn.indexer*",
381
+ "model.layers.69.self_attn*",
382
+ "model.layers.69.self_attn.indexer*",
383
+ "model.layers.7.self_attn*",
384
+ "model.layers.7.self_attn.indexer*",
385
+ "model.layers.70.self_attn*",
386
+ "model.layers.70.self_attn.indexer*",
387
+ "model.layers.71.self_attn*",
388
+ "model.layers.71.self_attn.indexer*",
389
+ "model.layers.72.self_attn*",
390
+ "model.layers.72.self_attn.indexer*",
391
+ "model.layers.73.self_attn*",
392
+ "model.layers.73.self_attn.indexer*",
393
+ "model.layers.74.self_attn*",
394
+ "model.layers.74.self_attn.indexer*",
395
+ "model.layers.75.self_attn*",
396
+ "model.layers.75.self_attn.indexer*",
397
+ "model.layers.76.self_attn*",
398
+ "model.layers.76.self_attn.indexer*",
399
+ "model.layers.77.self_attn*",
400
+ "model.layers.77.self_attn.indexer*",
401
+ "model.layers.8.self_attn*",
402
+ "model.layers.8.self_attn.indexer*",
403
+ "model.layers.9.self_attn*",
404
+ "model.layers.9.self_attn.indexer*",
405
+ "model.layers.78.eh_proj*",
406
+ "model.layers.78.self_attn*",
407
+ "model.layers.78.self_attn.indexer*",
408
+ "model.layers.78.mlp.gate*",
409
+ "model.layers.78.mlp.shared_experts*",
410
+ "*embed_tokens*",
411
+ "*shared_head*",
412
+ "*mtp_block.self_attn*",
413
+ "*mtp_block.mlp.gate",
414
+ "*eh_proj*",
415
+ "*enorm*",
416
+ "*hnorm*"
417
+ ],
418
+ "quant_algo": "NVFP4",
419
+ "producer": {
420
+ "name": "modelopt",
421
+ "version": "0.39.0.dev290+gf9d9a71de.d20260214"
422
+ },
423
+ "quant_method": "modelopt"
424
+ }
425
+ }
mtp-draft/model-mtp-inputscales.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b324090fb2ae84803015c454e6161b7da802b1fb6a16b89e8fa79f3f9767762f
3
+ size 86168
mtp-draft/model-mtp.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ade0e3da08e7e6c7b1f20e4c4e8d5d3b26b81103cea22f2ead9909c7d3d0732
3
+ size 6014594896
mtp-draft/model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
recipe.yaml ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ default_stage:
2
+ default_modifiers:
3
+ QuantizationModifier:
4
+ config_groups:
5
+ attention_shared_experts:
6
+ targets: ['re:.*self_attn\..*']
7
+ weights:
8
+ num_bits: 8
9
+ type: float
10
+ symmetric: true
11
+ group_size: null
12
+ strategy: block
13
+ block_structure: [128, 128]
14
+ dynamic: false
15
+ actorder: null
16
+ scale_dtype: null
17
+ zp_dtype: null
18
+ observer: memoryless_minmax
19
+ observer_kwargs: {}
20
+ input_activations:
21
+ num_bits: 8
22
+ type: float
23
+ symmetric: true
24
+ group_size: 128
25
+ strategy: group
26
+ block_structure: null
27
+ dynamic: true
28
+ actorder: null
29
+ scale_dtype: null
30
+ zp_dtype: null
31
+ observer: null
32
+ observer_kwargs: {}
33
+ output_activations: null
34
+ format: null
35
+ dense_mlp:
36
+ targets: ['re:.*\.mlp\.(gate_proj|up_proj|down_proj)$']
37
+ weights:
38
+ num_bits: 4
39
+ type: float
40
+ symmetric: true
41
+ group_size: 16
42
+ strategy: tensor_group
43
+ block_structure: null
44
+ dynamic: false
45
+ actorder: null
46
+ scale_dtype: torch.float8_e4m3fn
47
+ zp_dtype: null
48
+ observer: memoryless_minmax
49
+ observer_kwargs: {}
50
+ input_activations:
51
+ num_bits: 4
52
+ type: float
53
+ symmetric: true
54
+ group_size: 16
55
+ strategy: tensor_group
56
+ block_structure: null
57
+ dynamic: local
58
+ actorder: null
59
+ scale_dtype: torch.float8_e4m3fn
60
+ zp_dtype: null
61
+ observer: static_minmax
62
+ observer_kwargs: {}
63
+ output_activations: null
64
+ format: null
65
+ # v2: shared experts re-quantized to block-FP8 (from BF16), was NVFP4 in v1
66
+ shared_experts_fp8:
67
+ targets: ['re:.*\.mlp\.shared_experts\.(gate_proj|up_proj|down_proj)$']
68
+ weights:
69
+ num_bits: 8
70
+ type: float
71
+ symmetric: true
72
+ group_size: null
73
+ strategy: block
74
+ block_structure: [128, 128]
75
+ dynamic: false
76
+ actorder: null
77
+ scale_dtype: null
78
+ zp_dtype: null
79
+ observer: memoryless_minmax
80
+ observer_kwargs: {}
81
+ input_activations:
82
+ num_bits: 8
83
+ type: float
84
+ symmetric: true
85
+ group_size: 128
86
+ strategy: group
87
+ block_structure: null
88
+ dynamic: true
89
+ actorder: null
90
+ scale_dtype: null
91
+ zp_dtype: null
92
+ observer: null
93
+ observer_kwargs: {}
94
+ output_activations: null
95
+ format: null
96
+ targets: [Linear]
97
+ ignore: ['re:^model\.layers\.[0-2]\..*re:.*mlp\.gate.*', 're:.*indexer\.weights_proj$',
98
+ lm_head]
99
+ bypass_divisibility_checks: false
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d
3
+ size 20217442
tokenizer_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "clean_up_tokenization_spaces": false,
4
+ "do_lower_case": false,
5
+ "eos_token": "<|endoftext|>",
6
+ "extra_special_tokens": [
7
+ "<|endoftext|>",
8
+ "[MASK]",
9
+ "[gMASK]",
10
+ "[sMASK]",
11
+ "<sop>",
12
+ "<eop>",
13
+ "<|system|>",
14
+ "<|user|>",
15
+ "<|assistant|>",
16
+ "<|observation|>",
17
+ "<|begin_of_image|>",
18
+ "<|end_of_image|>",
19
+ "<|begin_of_video|>",
20
+ "<|end_of_video|>",
21
+ "<|begin_of_audio|>",
22
+ "<|end_of_audio|>",
23
+ "<|begin_of_transcription|>",
24
+ "<|end_of_transcription|>"
25
+ ],
26
+ "is_local": false,
27
+ "local_files_only": false,
28
+ "model_max_length": 1048576,
29
+ "model_specific_special_tokens": {},
30
+ "pad_token": "<|endoftext|>",
31
+ "padding_side": "left",
32
+ "remove_space": false,
33
+ "tokenizer_class": "TokenizersBackend"
34
+ }
vllm_overlay/INSTALL.md ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Loading this checkpoint in vLLM
2
+
3
+ Two small extensions are required (tested on vLLM ~0.11.x lineages that have
4
+ `Mxfp4Config` with a MoE method and `CompressedTensorsConfig`).
5
+
6
+ ## 1. Register the hybrid quant config
7
+
8
+ Copy `hybrid_mxfp4_ct.py` into
9
+ `vllm/model_executor/layers/quantization/hybrid_mxfp4_ct.py`
10
+ and add one import at the END of that package's `__init__.py`:
11
+
12
+ ```python
13
+ from . import hybrid_mxfp4_ct # noqa: F401
14
+ ```
15
+
16
+ (Or import it from any plugin/startup hook that runs before engine init —
17
+ `@register_quantization_config` does the rest.)
18
+
19
+ ## 2. Index-authoritative tensor filter (required)
20
+
21
+ This checkpoint's MXFP4 expert shard files also contain BF16 dense tensors that
22
+ `model.safetensors.index.json` deliberately maps to the `hybrid-ct-*.safetensors`
23
+ files instead. vLLM's safetensors iterator yields every tensor in every file it
24
+ opens, which would KeyError on the duplicates — it must skip tensors the index
25
+ maps elsewhere. In `vllm/model_executor/model_loader/weight_utils.py`, inside
26
+ `safetensors_weights_iterator(...)` after the files list is built:
27
+
28
+ ```python
29
+ # Build {file -> allowed tensor names} from the index (no-op for normal
30
+ # checkpoints, where every in-file tensor maps to its own file).
31
+ _allowed_by_file = None
32
+ if sorted_files:
33
+ _idx = os.path.join(os.path.dirname(sorted_files[0]),
34
+ "model.safetensors.index.json")
35
+ if os.path.isfile(_idx):
36
+ try:
37
+ with open(_idx) as fh:
38
+ _wm = json.load(fh)["weight_map"]
39
+ _allowed_by_file = {}
40
+ for _n, _f in _wm.items():
41
+ _allowed_by_file.setdefault(_f, set()).add(_n)
42
+ except Exception:
43
+ _allowed_by_file = None
44
+
45
+ def _index_excludes(st_file, name):
46
+ if _allowed_by_file is None:
47
+ return False
48
+ return name not in _allowed_by_file.get(os.path.basename(st_file), ())
49
+ ```
50
+
51
+ then, in the per-tensor loop(s), first thing:
52
+
53
+ ```python
54
+ if _index_excludes(st_file, name):
55
+ continue
56
+ ```
57
+
58
+ ## 3. Serve
59
+
60
+ ```bash
61
+ vllm serve /path/to/this/repo --trust-remote-code --tensor-parallel-size 4 \
62
+ --kv-cache-dtype fp8
63
+ ```
64
+
65
+ Notes:
66
+ - The MoE path needs an MXFP4-capable MoE backend. Verified end-to-end with the
67
+ B12X backend on GB10/SM121 (`VLLM_USE_B12X_MOE=1`); other backends untested.
68
+ - Speculative decoding: the checkpoint carries its native layer-78 MTP block
69
+ (MXFP4 draft experts). Enable with your usual `--speculative-config`.
70
+ - Expert `weight_scale` tensors are raw **uint8** e8m0 bytes — if your loader
71
+ casts scales by dtype, ensure it treats them as uint8 (a float cast corrupts).
vllm_overlay/hybrid_mxfp4_ct.py ADDED
@@ -0,0 +1,86 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: Apache-2.0
2
+ """Hybrid quantization config: festr2 MXFP4 routed experts + RedHat
3
+ compressed-tensors (block-FP8 self_attn / NVFP4 mlp+shared) everything-else,
4
+ in one GLM-5.2 checkpoint.
5
+
6
+ Selected when config.json declares quant_method: "hybrid_mxfp4_ct".
7
+ Mounted at vllm/model_executor/layers/quantization/hybrid_mxfp4_ct.py, imported
8
+ via a 2-line append to that package's __init__ overlay.
9
+
10
+ Design (verified against production-3.75 sources):
11
+ - RoutedExperts -> Mxfp4Config.get_quant_method -> Mxfp4MoEMethod (b12x
12
+ fp4_e8m0_k32 kernel via VLLM_USE_B12X_MOE, backend auto).
13
+ - Everything else (LinearBase incl. MLA projections, Attention KV method,
14
+ dense/shared mlp) -> CompressedTensorsConfig, fed RedHat's verbatim
15
+ quantization_config as the "linear" sub-dict.
16
+ - Branch order matters: intercept RoutedExperts BEFORE delegating, else the CT
17
+ half would claim the MoE and build an NVFP4 MoE method.
18
+ """
19
+
20
+ import torch
21
+
22
+ from vllm.model_executor.layers.quantization import register_quantization_config
23
+ from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
24
+
25
+
26
+ def _routed_experts_cls():
27
+ # Lazy import (avoid import cycles at package-init time).
28
+ from vllm.model_executor.layers.fused_moe import RoutedExperts
29
+ return RoutedExperts
30
+
31
+
32
+ @register_quantization_config("hybrid_mxfp4_ct")
33
+ class HybridMxfp4CtConfig(QuantizationConfig):
34
+ """Compose Mxfp4 (routed experts) with compressed-tensors (the rest)."""
35
+
36
+ def __init__(self, moe, linear):
37
+ super().__init__()
38
+ self.moe = moe
39
+ self.linear = linear
40
+
41
+ @classmethod
42
+ def get_name(cls) -> str:
43
+ return "hybrid_mxfp4_ct"
44
+
45
+ @classmethod
46
+ def get_min_capability(cls) -> int:
47
+ return 80 # mxfp4 floor; CT is lower
48
+
49
+ def get_supported_act_dtypes(self):
50
+ return [torch.bfloat16]
51
+
52
+ @classmethod
53
+ def get_config_filenames(cls):
54
+ return []
55
+
56
+ def is_mxfp4_quant(self, prefix, layer):
57
+ # hidden-size rounding helpers treat MoE as mxfp4
58
+ return isinstance(layer, _routed_experts_cls())
59
+
60
+ @classmethod
61
+ def from_config(cls, config):
62
+ from vllm.model_executor.layers.quantization.mxfp4 import Mxfp4Config
63
+ from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
64
+ CompressedTensorsConfig,
65
+ )
66
+
67
+ moe = Mxfp4Config.from_config(config.get("moe", {}))
68
+ lin_cfg = dict(config["linear"])
69
+ # CT head-piping is skipped when top-level quant_method != CT
70
+ # (weight_utils.py:259-274); forward head counts defensively.
71
+ for k in ("total_num_heads", "total_num_kv_heads"):
72
+ if k in config:
73
+ lin_cfg.setdefault(k, config[k])
74
+ return cls(moe, CompressedTensorsConfig.from_config(lin_cfg))
75
+
76
+ def get_quant_method(self, layer, prefix):
77
+ # Propagate the model-supplied fused-module map to both halves.
78
+ self.moe.packed_modules_mapping = self.packed_modules_mapping
79
+ self.linear.packed_modules_mapping = self.packed_modules_mapping
80
+ if isinstance(layer, _routed_experts_cls()):
81
+ return self.moe.get_quant_method(layer, prefix)
82
+ return self.linear.get_quant_method(layer, prefix)
83
+
84
+ def apply_vllm_mapper(self, hf_to_vllm_mapper):
85
+ # Keep CT target/ignore remapping intact.
86
+ self.linear.apply_vllm_mapper(hf_to_vllm_mapper)