Hugging Face's logo

Firworks
/
Ministral-3-3B-Reasoning-2512-nvfp4

Safetensors
mistral3
8-bit precision
compressed-tensors
Model card Files
xet
Community
Firworks commited on
Commit
3ffb9ee
·
verified ·
1 Parent(s): 10bf84b

Add NVFP4 quantized checkpoint

Browse files
Files changed (7) hide show
  1. .gitattributes +1 -0
  2. README.md +23 -0
  3. config.json +282 -0
  4. generation_config.json +8 -0
  5. model.safetensors +3 -0
  6. recipe.yaml +8 -0
  7. tekken.json +3 -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
+ tekken.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ datasets:
3
+ - Rombo-Org/Optimized_Reasoning
4
+ base_model:
5
+ - mistralai/Ministral-3-3B-Reasoning-2512
6
+ ---
7
+ # Ministral-3-3B-Reasoning-2512-nvfp4
8
+
9
+ **Format:** NVFP4 — weights & activations quantized to FP4 with dual scaling.
10
+ **Base model:** `mistralai/Ministral-3-3B-Reasoning-2512`
11
+ **How it was made:** One-shot calibration with LLM Compressor (NVFP4 recipe), long-seq calibration with Rombo-Org/Optimized_Reasoning.
12
+
13
+ > Notes: Keep `lm_head` in high precision; calibrate on long, domain-relevant sequences.
14
+
15
+ Check the original model card for information about this model.
16
+
17
+ # Running the model with VLLM in Docker
18
+ ```sh
19
+ sudo docker run --runtime nvidia --gpus all -p 8000:8000 --ipc=host vllm/vllm-openai:nightly --model Firworks/Ministral-3-3B-Reasoning-2512-nvfp4 --dtype auto --max-model-len 32768
20
+ ```
21
+ This was tested on an RTX Pro 6000 Blackwell cloud instance.
22
+
23
+ If there are other models you're interested in seeing quantized to NVFP4 for use on the DGX Spark, or other modern Blackwell (or newer) cards let me know. I'm trying to make more NVFP4 models available to allow more people to try them out.
config.json ADDED
@@ -0,0 +1,282 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Mistral3ForConditionalGeneration"
4
+ ],
5
+ "dtype": "bfloat16",
6
+ "image_token_index": 10,
7
+ "model_type": "mistral3",
8
+ "multimodal_projector_bias": false,
9
+ "projector_hidden_act": "gelu",
10
+ "quantization_config": {
11
+ "config_groups": {
12
+ "group_0": {
13
+ "format": "nvfp4-pack-quantized",
14
+ "input_activations": {
15
+ "actorder": null,
16
+ "block_structure": null,
17
+ "dynamic": "local",
18
+ "group_size": 16,
19
+ "num_bits": 4,
20
+ "observer": "static_minmax",
21
+ "observer_kwargs": {},
22
+ "scale_dtype": "torch.float8_e4m3fn",
23
+ "strategy": "tensor_group",
24
+ "symmetric": true,
25
+ "type": "float",
26
+ "zp_dtype": null
27
+ },
28
+ "output_activations": null,
29
+ "targets": [
30
+ "Linear"
31
+ ],
32
+ "weights": {
33
+ "actorder": null,
34
+ "block_structure": null,
35
+ "dynamic": false,
36
+ "group_size": 16,
37
+ "num_bits": 4,
38
+ "observer": "static_minmax",
39
+ "observer_kwargs": {},
40
+ "scale_dtype": "torch.float8_e4m3fn",
41
+ "strategy": "tensor_group",
42
+ "symmetric": true,
43
+ "type": "float",
44
+ "zp_dtype": null
45
+ }
46
+ }
47
+ },
48
+ "format": "nvfp4-pack-quantized",
49
+ "global_compression_ratio": null,
50
+ "ignore": [
51
+ "model.vision_tower.transformer.layers.0.feed_forward.gate_proj",
52
+ "model.vision_tower.transformer.layers.0.feed_forward.up_proj",
53
+ "model.vision_tower.transformer.layers.0.feed_forward.down_proj",
54
+ "model.vision_tower.transformer.layers.0.attention.k_proj",
55
+ "model.vision_tower.transformer.layers.0.attention.v_proj",
56
+ "model.vision_tower.transformer.layers.0.attention.q_proj",
57
+ "model.vision_tower.transformer.layers.0.attention.o_proj",
58
+ "model.vision_tower.transformer.layers.1.feed_forward.gate_proj",
59
+ "model.vision_tower.transformer.layers.1.feed_forward.up_proj",
60
+ "model.vision_tower.transformer.layers.1.feed_forward.down_proj",
61
+ "model.vision_tower.transformer.layers.1.attention.k_proj",
62
+ "model.vision_tower.transformer.layers.1.attention.v_proj",
63
+ "model.vision_tower.transformer.layers.1.attention.q_proj",
64
+ "model.vision_tower.transformer.layers.1.attention.o_proj",
65
+ "model.vision_tower.transformer.layers.2.feed_forward.gate_proj",
66
+ "model.vision_tower.transformer.layers.2.feed_forward.up_proj",
67
+ "model.vision_tower.transformer.layers.2.feed_forward.down_proj",
68
+ "model.vision_tower.transformer.layers.2.attention.k_proj",
69
+ "model.vision_tower.transformer.layers.2.attention.v_proj",
70
+ "model.vision_tower.transformer.layers.2.attention.q_proj",
71
+ "model.vision_tower.transformer.layers.2.attention.o_proj",
72
+ "model.vision_tower.transformer.layers.3.feed_forward.gate_proj",
73
+ "model.vision_tower.transformer.layers.3.feed_forward.up_proj",
74
+ "model.vision_tower.transformer.layers.3.feed_forward.down_proj",
75
+ "model.vision_tower.transformer.layers.3.attention.k_proj",
76
+ "model.vision_tower.transformer.layers.3.attention.v_proj",
77
+ "model.vision_tower.transformer.layers.3.attention.q_proj",
78
+ "model.vision_tower.transformer.layers.3.attention.o_proj",
79
+ "model.vision_tower.transformer.layers.4.feed_forward.gate_proj",
80
+ "model.vision_tower.transformer.layers.4.feed_forward.up_proj",
81
+ "model.vision_tower.transformer.layers.4.feed_forward.down_proj",
82
+ "model.vision_tower.transformer.layers.4.attention.k_proj",
83
+ "model.vision_tower.transformer.layers.4.attention.v_proj",
84
+ "model.vision_tower.transformer.layers.4.attention.q_proj",
85
+ "model.vision_tower.transformer.layers.4.attention.o_proj",
86
+ "model.vision_tower.transformer.layers.5.feed_forward.gate_proj",
87
+ "model.vision_tower.transformer.layers.5.feed_forward.up_proj",
88
+ "model.vision_tower.transformer.layers.5.feed_forward.down_proj",
89
+ "model.vision_tower.transformer.layers.5.attention.k_proj",
90
+ "model.vision_tower.transformer.layers.5.attention.v_proj",
91
+ "model.vision_tower.transformer.layers.5.attention.q_proj",
92
+ "model.vision_tower.transformer.layers.5.attention.o_proj",
93
+ "model.vision_tower.transformer.layers.6.feed_forward.gate_proj",
94
+ "model.vision_tower.transformer.layers.6.feed_forward.up_proj",
95
+ "model.vision_tower.transformer.layers.6.feed_forward.down_proj",
96
+ "model.vision_tower.transformer.layers.6.attention.k_proj",
97
+ "model.vision_tower.transformer.layers.6.attention.v_proj",
98
+ "model.vision_tower.transformer.layers.6.attention.q_proj",
99
+ "model.vision_tower.transformer.layers.6.attention.o_proj",
100
+ "model.vision_tower.transformer.layers.7.feed_forward.gate_proj",
101
+ "model.vision_tower.transformer.layers.7.feed_forward.up_proj",
102
+ "model.vision_tower.transformer.layers.7.feed_forward.down_proj",
103
+ "model.vision_tower.transformer.layers.7.attention.k_proj",
104
+ "model.vision_tower.transformer.layers.7.attention.v_proj",
105
+ "model.vision_tower.transformer.layers.7.attention.q_proj",
106
+ "model.vision_tower.transformer.layers.7.attention.o_proj",
107
+ "model.vision_tower.transformer.layers.8.feed_forward.gate_proj",
108
+ "model.vision_tower.transformer.layers.8.feed_forward.up_proj",
109
+ "model.vision_tower.transformer.layers.8.feed_forward.down_proj",
110
+ "model.vision_tower.transformer.layers.8.attention.k_proj",
111
+ "model.vision_tower.transformer.layers.8.attention.v_proj",
112
+ "model.vision_tower.transformer.layers.8.attention.q_proj",
113
+ "model.vision_tower.transformer.layers.8.attention.o_proj",
114
+ "model.vision_tower.transformer.layers.9.feed_forward.gate_proj",
115
+ "model.vision_tower.transformer.layers.9.feed_forward.up_proj",
116
+ "model.vision_tower.transformer.layers.9.feed_forward.down_proj",
117
+ "model.vision_tower.transformer.layers.9.attention.k_proj",
118
+ "model.vision_tower.transformer.layers.9.attention.v_proj",
119
+ "model.vision_tower.transformer.layers.9.attention.q_proj",
120
+ "model.vision_tower.transformer.layers.9.attention.o_proj",
121
+ "model.vision_tower.transformer.layers.10.feed_forward.gate_proj",
122
+ "model.vision_tower.transformer.layers.10.feed_forward.up_proj",
123
+ "model.vision_tower.transformer.layers.10.feed_forward.down_proj",
124
+ "model.vision_tower.transformer.layers.10.attention.k_proj",
125
+ "model.vision_tower.transformer.layers.10.attention.v_proj",
126
+ "model.vision_tower.transformer.layers.10.attention.q_proj",
127
+ "model.vision_tower.transformer.layers.10.attention.o_proj",
128
+ "model.vision_tower.transformer.layers.11.feed_forward.gate_proj",
129
+ "model.vision_tower.transformer.layers.11.feed_forward.up_proj",
130
+ "model.vision_tower.transformer.layers.11.feed_forward.down_proj",
131
+ "model.vision_tower.transformer.layers.11.attention.k_proj",
132
+ "model.vision_tower.transformer.layers.11.attention.v_proj",
133
+ "model.vision_tower.transformer.layers.11.attention.q_proj",
134
+ "model.vision_tower.transformer.layers.11.attention.o_proj",
135
+ "model.vision_tower.transformer.layers.12.feed_forward.gate_proj",
136
+ "model.vision_tower.transformer.layers.12.feed_forward.up_proj",
137
+ "model.vision_tower.transformer.layers.12.feed_forward.down_proj",
138
+ "model.vision_tower.transformer.layers.12.attention.k_proj",
139
+ "model.vision_tower.transformer.layers.12.attention.v_proj",
140
+ "model.vision_tower.transformer.layers.12.attention.q_proj",
141
+ "model.vision_tower.transformer.layers.12.attention.o_proj",
142
+ "model.vision_tower.transformer.layers.13.feed_forward.gate_proj",
143
+ "model.vision_tower.transformer.layers.13.feed_forward.up_proj",
144
+ "model.vision_tower.transformer.layers.13.feed_forward.down_proj",
145
+ "model.vision_tower.transformer.layers.13.attention.k_proj",
146
+ "model.vision_tower.transformer.layers.13.attention.v_proj",
147
+ "model.vision_tower.transformer.layers.13.attention.q_proj",
148
+ "model.vision_tower.transformer.layers.13.attention.o_proj",
149
+ "model.vision_tower.transformer.layers.14.feed_forward.gate_proj",
150
+ "model.vision_tower.transformer.layers.14.feed_forward.up_proj",
151
+ "model.vision_tower.transformer.layers.14.feed_forward.down_proj",
152
+ "model.vision_tower.transformer.layers.14.attention.k_proj",
153
+ "model.vision_tower.transformer.layers.14.attention.v_proj",
154
+ "model.vision_tower.transformer.layers.14.attention.q_proj",
155
+ "model.vision_tower.transformer.layers.14.attention.o_proj",
156
+ "model.vision_tower.transformer.layers.15.feed_forward.gate_proj",
157
+ "model.vision_tower.transformer.layers.15.feed_forward.up_proj",
158
+ "model.vision_tower.transformer.layers.15.feed_forward.down_proj",
159
+ "model.vision_tower.transformer.layers.15.attention.k_proj",
160
+ "model.vision_tower.transformer.layers.15.attention.v_proj",
161
+ "model.vision_tower.transformer.layers.15.attention.q_proj",
162
+ "model.vision_tower.transformer.layers.15.attention.o_proj",
163
+ "model.vision_tower.transformer.layers.16.feed_forward.gate_proj",
164
+ "model.vision_tower.transformer.layers.16.feed_forward.up_proj",
165
+ "model.vision_tower.transformer.layers.16.feed_forward.down_proj",
166
+ "model.vision_tower.transformer.layers.16.attention.k_proj",
167
+ "model.vision_tower.transformer.layers.16.attention.v_proj",
168
+ "model.vision_tower.transformer.layers.16.attention.q_proj",
169
+ "model.vision_tower.transformer.layers.16.attention.o_proj",
170
+ "model.vision_tower.transformer.layers.17.feed_forward.gate_proj",
171
+ "model.vision_tower.transformer.layers.17.feed_forward.up_proj",
172
+ "model.vision_tower.transformer.layers.17.feed_forward.down_proj",
173
+ "model.vision_tower.transformer.layers.17.attention.k_proj",
174
+ "model.vision_tower.transformer.layers.17.attention.v_proj",
175
+ "model.vision_tower.transformer.layers.17.attention.q_proj",
176
+ "model.vision_tower.transformer.layers.17.attention.o_proj",
177
+ "model.vision_tower.transformer.layers.18.feed_forward.gate_proj",
178
+ "model.vision_tower.transformer.layers.18.feed_forward.up_proj",
179
+ "model.vision_tower.transformer.layers.18.feed_forward.down_proj",
180
+ "model.vision_tower.transformer.layers.18.attention.k_proj",
181
+ "model.vision_tower.transformer.layers.18.attention.v_proj",
182
+ "model.vision_tower.transformer.layers.18.attention.q_proj",
183
+ "model.vision_tower.transformer.layers.18.attention.o_proj",
184
+ "model.vision_tower.transformer.layers.19.feed_forward.gate_proj",
185
+ "model.vision_tower.transformer.layers.19.feed_forward.up_proj",
186
+ "model.vision_tower.transformer.layers.19.feed_forward.down_proj",
187
+ "model.vision_tower.transformer.layers.19.attention.k_proj",
188
+ "model.vision_tower.transformer.layers.19.attention.v_proj",
189
+ "model.vision_tower.transformer.layers.19.attention.q_proj",
190
+ "model.vision_tower.transformer.layers.19.attention.o_proj",
191
+ "model.vision_tower.transformer.layers.20.feed_forward.gate_proj",
192
+ "model.vision_tower.transformer.layers.20.feed_forward.up_proj",
193
+ "model.vision_tower.transformer.layers.20.feed_forward.down_proj",
194
+ "model.vision_tower.transformer.layers.20.attention.k_proj",
195
+ "model.vision_tower.transformer.layers.20.attention.v_proj",
196
+ "model.vision_tower.transformer.layers.20.attention.q_proj",
197
+ "model.vision_tower.transformer.layers.20.attention.o_proj",
198
+ "model.vision_tower.transformer.layers.21.feed_forward.gate_proj",
199
+ "model.vision_tower.transformer.layers.21.feed_forward.up_proj",
200
+ "model.vision_tower.transformer.layers.21.feed_forward.down_proj",
201
+ "model.vision_tower.transformer.layers.21.attention.k_proj",
202
+ "model.vision_tower.transformer.layers.21.attention.v_proj",
203
+ "model.vision_tower.transformer.layers.21.attention.q_proj",
204
+ "model.vision_tower.transformer.layers.21.attention.o_proj",
205
+ "model.vision_tower.transformer.layers.22.feed_forward.gate_proj",
206
+ "model.vision_tower.transformer.layers.22.feed_forward.up_proj",
207
+ "model.vision_tower.transformer.layers.22.feed_forward.down_proj",
208
+ "model.vision_tower.transformer.layers.22.attention.k_proj",
209
+ "model.vision_tower.transformer.layers.22.attention.v_proj",
210
+ "model.vision_tower.transformer.layers.22.attention.q_proj",
211
+ "model.vision_tower.transformer.layers.22.attention.o_proj",
212
+ "model.vision_tower.transformer.layers.23.feed_forward.gate_proj",
213
+ "model.vision_tower.transformer.layers.23.feed_forward.up_proj",
214
+ "model.vision_tower.transformer.layers.23.feed_forward.down_proj",
215
+ "model.vision_tower.transformer.layers.23.attention.k_proj",
216
+ "model.vision_tower.transformer.layers.23.attention.v_proj",
217
+ "model.vision_tower.transformer.layers.23.attention.q_proj",
218
+ "model.vision_tower.transformer.layers.23.attention.o_proj",
219
+ "model.multi_modal_projector.patch_merger.merging_layer",
220
+ "model.multi_modal_projector.linear_1",
221
+ "model.multi_modal_projector.linear_2",
222
+ "lm_head"
223
+ ],
224
+ "kv_cache_scheme": null,
225
+ "quant_method": "compressed-tensors",
226
+ "quantization_status": "compressed",
227
+ "sparsity_config": {},
228
+ "transform_config": {},
229
+ "version": "0.12.3.a20251114"
230
+ },
231
+ "spatial_merge_size": 2,
232
+ "text_config": {
233
+ "attention_dropout": 0.0,
234
+ "head_dim": 128,
235
+ "hidden_act": "silu",
236
+ "hidden_size": 3072,
237
+ "initializer_range": 0.02,
238
+ "intermediate_size": 9216,
239
+ "max_position_embeddings": 262144,
240
+ "model_type": "ministral3",
241
+ "num_attention_heads": 32,
242
+ "num_hidden_layers": 26,
243
+ "num_key_value_heads": 8,
244
+ "rms_norm_eps": 1e-05,
245
+ "rope_parameters": {
246
+ "beta_fast": 32.0,
247
+ "beta_slow": 1.0,
248
+ "factor": 16.0,
249
+ "llama_4_scaling_beta": 0.1,
250
+ "mscale": 1.0,
251
+ "mscale_all_dim": 1.0,
252
+ "original_max_position_embeddings": 16384,
253
+ "rope_theta": 1000000.0,
254
+ "rope_type": "yarn",
255
+ "type": "yarn"
256
+ },
257
+ "sliding_window": null,
258
+ "tie_word_embeddings": true,
259
+ "use_cache": true,
260
+ "vocab_size": 131072
261
+ },
262
+ "transformers_version": "5.0.0.dev0",
263
+ "vision_config": {
264
+ "attention_dropout": 0.0,
265
+ "head_dim": 64,
266
+ "hidden_act": "silu",
267
+ "hidden_size": 1024,
268
+ "image_size": 1540,
269
+ "initializer_range": 0.02,
270
+ "intermediate_size": 4096,
271
+ "model_type": "pixtral",
272
+ "num_attention_heads": 16,
273
+ "num_channels": 3,
274
+ "num_hidden_layers": 24,
275
+ "patch_size": 14,
276
+ "rope_parameters": {
277
+ "rope_theta": 10000.0,
278
+ "rope_type": "default"
279
+ }
280
+ },
281
+ "vision_feature_layer": -1
282
+ }
generation_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 1,
3
+ "do_sample": true,
4
+ "eos_token_id": 2,
5
+ "max_length": 262144,
6
+ "pad_token_id": 11,
7
+ "transformers_version": "5.0.0.dev0"
8
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8942b9951c342c34038b1e55ac91525d2116164ba534d7468c3b8b1799cc589d
3
+ size 4153470712
recipe.yaml ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ default_stage:
2
+ default_modifiers:
3
+ QuantizationModifier:
4
+ targets: [Linear]
5
+ ignore: [lm_head, 're:.*vision_tower.*', 're:.*vision_model.*', 're:.*mm_projector.*',
6
+ 're:.*multi_modal_projector.*', 're:.*multimodal_projector.*', 're:.*video_tower.*',
7
+ 're:.*audio_tower.*']
8
+ scheme: NVFP4
tekken.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e29d19ea32eb7e26e6c0572d57cb7f9eca0f4420e0e0fe6ae1cf3be94da1c0d6
3
+ size 16753777