Hugging Face's logo

lukealonso
/
MiniMax-M2.5-NVFP4

Safetensors
minimax_m2
custom_code
8-bit precision
modelopt
Model card Files
xet
Community
7
lukealonso commited on
Commit
b110d33
·
verified ·
1 Parent(s): 078d52a

Add files using upload-large-folder tool

Browse files
Files changed (29) hide show
  1. config.json +88 -28
  2. generation_config.json +68 -7
  3. hf_quant_config.json +153 -132
  4. model-00001-of-00025.safetensors +3 -0
  5. model-00002-of-00025.safetensors +3 -0
  6. model-00003-of-00025.safetensors +3 -0
  7. model-00004-of-00025.safetensors +3 -0
  8. model-00005-of-00025.safetensors +3 -0
  9. model-00006-of-00025.safetensors +3 -0
  10. model-00007-of-00025.safetensors +3 -0
  11. model-00008-of-00025.safetensors +3 -0
  12. model-00009-of-00025.safetensors +3 -0
  13. model-00010-of-00025.safetensors +3 -0
  14. model-00011-of-00025.safetensors +3 -0
  15. model-00012-of-00025.safetensors +3 -0
  16. model-00013-of-00025.safetensors +3 -0
  17. model-00014-of-00025.safetensors +3 -0
  18. model-00015-of-00025.safetensors +3 -0
  19. model-00016-of-00025.safetensors +3 -0
  20. model-00017-of-00025.safetensors +3 -0
  21. model-00018-of-00025.safetensors +3 -0
  22. model-00019-of-00025.safetensors +3 -0
  23. model-00020-of-00025.safetensors +3 -0
  24. model-00021-of-00025.safetensors +3 -0
  25. model-00022-of-00025.safetensors +3 -0
  26. model-00023-of-00025.safetensors +3 -0
  27. model-00024-of-00025.safetensors +3 -0
  28. model-00025-of-00025.safetensors +3 -0
  29. model.safetensors.index.json +2 -2
config.json CHANGED
@@ -1,8 +1,87 @@
1
  {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  "architectures": [
3
  "MiniMaxM2ForCausalLM"
4
  ],
5
- "attention_dropout": 0.0,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6
  "attn_type_list": [
7
  1,
8
  1,
@@ -71,41 +150,17 @@
71
  "AutoConfig": "configuration_minimax_m2.MiniMaxM2Config",
72
  "AutoModelForCausalLM": "modeling_minimax_m2.MiniMaxM2ForCausalLM"
73
  },
74
- "bos_token_id": 1,
75
- "dtype": "bfloat16",
76
- "eos_token_id": 2,
77
- "head_dim": 128,
78
- "hidden_act": "silu",
79
- "hidden_size": 3072,
80
- "initializer_range": 0.02,
81
- "intermediate_size": 1536,
82
- "max_position_embeddings": 196608,
83
  "model_type": "minimax_m2",
84
  "mtp_transformer_layers": 1,
85
- "num_attention_heads": 48,
86
- "num_experts_per_tok": 8,
87
- "num_hidden_layers": 62,
88
- "num_key_value_heads": 8,
89
- "num_local_experts": 256,
90
  "num_mtp_modules": 3,
91
- "output_router_logits": false,
92
- "partial_rotary_factor": 0.5,
93
  "qk_norm_type": "per_layer",
94
- "rms_norm_eps": 1e-06,
95
- "rope_theta": 5000000,
96
- "rotary_dim": 64,
97
- "router_aux_loss_coef": 0.001,
98
- "router_jitter_noise": 0.0,
99
  "scoring_func": "sigmoid",
100
  "shared_intermediate_size": 0,
101
- "sliding_window": null,
102
- "tie_word_embeddings": false,
103
- "transformers_version": "4.57.6",
104
- "use_cache": true,
105
  "use_mtp": true,
106
- "use_qk_norm": true,
107
  "use_routing_bias": true,
108
- "vocab_size": 200064,
 
 
109
  "quantization_config": {
110
  "config_groups": {
111
  "group_0": {
@@ -254,6 +309,11 @@
254
  "model.layers.9.self_attn*"
255
  ],
256
  "quant_algo": "NVFP4",
 
 
 
 
 
257
  "producer": {
258
  "name": "modelopt",
259
  "version": "0.39.0.dev290+gf9d9a71de.d20260213"
 
1
  {
2
+ "vocab_size": 200064,
3
+ "max_position_embeddings": 196608,
4
+ "hidden_size": 3072,
5
+ "intermediate_size": 1536,
6
+ "num_hidden_layers": 62,
7
+ "num_attention_heads": 48,
8
+ "sliding_window": null,
9
+ "num_key_value_heads": 8,
10
+ "hidden_act": "silu",
11
+ "initializer_range": 0.02,
12
+ "rms_norm_eps": 1e-06,
13
+ "use_cache": true,
14
+ "rope_theta": 5000000,
15
+ "attention_dropout": 0.0,
16
+ "head_dim": 128,
17
+ "num_experts_per_tok": 8,
18
+ "num_local_experts": 256,
19
+ "output_router_logits": false,
20
+ "router_aux_loss_coef": 0.001,
21
+ "router_jitter_noise": 0.0,
22
+ "use_qk_norm": true,
23
+ "rotary_dim": 64,
24
+ "partial_rotary_factor": 0.5,
25
+ "return_dict": true,
26
+ "output_hidden_states": false,
27
+ "torchscript": false,
28
+ "dtype": "bfloat16",
29
+ "pruned_heads": {},
30
+ "tie_word_embeddings": false,
31
+ "chunk_size_feed_forward": 0,
32
+ "is_encoder_decoder": false,
33
+ "is_decoder": false,
34
+ "cross_attention_hidden_size": null,
35
+ "add_cross_attention": false,
36
+ "tie_encoder_decoder": false,
37
  "architectures": [
38
  "MiniMaxM2ForCausalLM"
39
  ],
40
+ "finetuning_task": null,
41
+ "id2label": {
42
+ "0": "LABEL_0",
43
+ "1": "LABEL_1"
44
+ },
45
+ "label2id": {
46
+ "LABEL_0": 0,
47
+ "LABEL_1": 1
48
+ },
49
+ "task_specific_params": null,
50
+ "problem_type": null,
51
+ "tokenizer_class": null,
52
+ "prefix": null,
53
+ "bos_token_id": null,
54
+ "pad_token_id": null,
55
+ "eos_token_id": null,
56
+ "sep_token_id": null,
57
+ "decoder_start_token_id": null,
58
+ "max_length": 20,
59
+ "min_length": 0,
60
+ "do_sample": false,
61
+ "early_stopping": false,
62
+ "num_beams": 1,
63
+ "temperature": 1.0,
64
+ "top_k": 50,
65
+ "top_p": 1.0,
66
+ "typical_p": 1.0,
67
+ "repetition_penalty": 1.0,
68
+ "length_penalty": 1.0,
69
+ "no_repeat_ngram_size": 0,
70
+ "encoder_no_repeat_ngram_size": 0,
71
+ "bad_words_ids": null,
72
+ "num_return_sequences": 1,
73
+ "output_scores": false,
74
+ "return_dict_in_generate": false,
75
+ "forced_bos_token_id": null,
76
+ "forced_eos_token_id": null,
77
+ "remove_invalid_values": false,
78
+ "exponential_decay_length_penalty": null,
79
+ "suppress_tokens": null,
80
+ "begin_suppress_tokens": null,
81
+ "num_beam_groups": 1,
82
+ "diversity_penalty": 0.0,
83
+ "_name_or_path": null,
84
+ "transformers_version": "4.57.6",
85
  "attn_type_list": [
86
  1,
87
  1,
 
150
  "AutoConfig": "configuration_minimax_m2.MiniMaxM2Config",
151
  "AutoModelForCausalLM": "modeling_minimax_m2.MiniMaxM2ForCausalLM"
152
  },
 
 
 
 
 
 
 
 
 
153
  "model_type": "minimax_m2",
154
  "mtp_transformer_layers": 1,
 
 
 
 
 
155
  "num_mtp_modules": 3,
 
 
156
  "qk_norm_type": "per_layer",
 
 
 
 
 
157
  "scoring_func": "sigmoid",
158
  "shared_intermediate_size": 0,
 
 
 
 
159
  "use_mtp": true,
 
160
  "use_routing_bias": true,
161
+ "tf_legacy_loss": false,
162
+ "use_bfloat16": false,
163
+ "output_attentions": false,
164
  "quantization_config": {
165
  "config_groups": {
166
  "group_0": {
 
309
  "model.layers.9.self_attn*"
310
  ],
311
  "quant_algo": "NVFP4",
312
+ "kv_cache_scheme": {
313
+ "dynamic": false,
314
+ "num_bits": 8,
315
+ "type": "float"
316
+ },
317
  "producer": {
318
  "name": "modelopt",
319
  "version": "0.39.0.dev290+gf9d9a71de.d20260213"
generation_config.json CHANGED
@@ -1,8 +1,69 @@
1
  {
2
- "bos_token_id": 200019,
3
- "do_sample": true,
4
- "eos_token_id": 200020,
5
- "top_k": 40,
6
- "top_p": 0.95,
7
- "transformers_version": "4.57.6"
8
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  {
2
+ "max_length": 20,
3
+ "max_new_tokens": null,
4
+ "min_length": 0,
5
+ "min_new_tokens": null,
6
+ "early_stopping": false,
7
+ "max_time": null,
8
+ "stop_strings": null,
9
+ "do_sample": true,
10
+ "num_beams": 1,
11
+ "use_cache": true,
12
+ "cache_implementation": null,
13
+ "cache_config": null,
14
+ "return_legacy_cache": null,
15
+ "prefill_chunk_size": null,
16
+ "temperature": 1.0,
17
+ "top_k": 40,
18
+ "top_p": 0.95,
19
+ "min_p": null,
20
+ "typical_p": 1.0,
21
+ "epsilon_cutoff": 0.0,
22
+ "eta_cutoff": 0.0,
23
+ "repetition_penalty": 1.0,
24
+ "encoder_repetition_penalty": 1.0,
25
+ "length_penalty": 1.0,
26
+ "no_repeat_ngram_size": 0,
27
+ "bad_words_ids": null,
28
+ "renormalize_logits": false,
29
+ "forced_bos_token_id": null,
30
+ "forced_eos_token_id": null,
31
+ "remove_invalid_values": false,
32
+ "exponential_decay_length_penalty": null,
33
+ "suppress_tokens": null,
34
+ "begin_suppress_tokens": null,
35
+ "sequence_bias": null,
36
+ "token_healing": false,
37
+ "guidance_scale": null,
38
+ "watermarking_config": null,
39
+ "num_return_sequences": 1,
40
+ "output_attentions": false,
41
+ "output_hidden_states": false,
42
+ "output_scores": false,
43
+ "output_logits": null,
44
+ "return_dict_in_generate": false,
45
+ "pad_token_id": null,
46
+ "bos_token_id": 200019,
47
+ "eos_token_id": 200020,
48
+ "encoder_no_repeat_ngram_size": 0,
49
+ "decoder_start_token_id": null,
50
+ "is_assistant": false,
51
+ "num_assistant_tokens": 20,
52
+ "num_assistant_tokens_schedule": "constant",
53
+ "assistant_confidence_threshold": 0.4,
54
+ "prompt_lookup_num_tokens": null,
55
+ "max_matching_ngram_size": null,
56
+ "assistant_early_exit": null,
57
+ "assistant_lookbehind": 10,
58
+ "target_lookbehind": 10,
59
+ "disable_compile": false,
60
+ "low_memory": null,
61
+ "penalty_alpha": null,
62
+ "dola_layers": null,
63
+ "diversity_penalty": 0.0,
64
+ "num_beam_groups": 1,
65
+ "constraints": null,
66
+ "force_words_ids": null,
67
+ "_from_model_config": false,
68
+ "transformers_version": "4.57.6"
69
+ }
hf_quant_config.json CHANGED
@@ -1,138 +1,159 @@
1
  {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  "producer": {
3
  "name": "modelopt",
4
  "version": "0.39.0.dev290+gf9d9a71de.d20260213"
5
  },
6
- "quantization": {
7
- "quant_algo": "NVFP4",
8
- "kv_cache_quant_algo": null,
9
- "group_size": 16,
10
- "exclude_modules": [
11
- "lm_head",
12
- "model.layers.0.block_sparse_moe.gate",
13
- "model.layers.0.self_attn*",
14
- "model.layers.1.block_sparse_moe.gate",
15
- "model.layers.1.self_attn*",
16
- "model.layers.10.block_sparse_moe.gate",
17
- "model.layers.10.self_attn*",
18
- "model.layers.11.block_sparse_moe.gate",
19
- "model.layers.11.self_attn*",
20
- "model.layers.12.block_sparse_moe.gate",
21
- "model.layers.12.self_attn*",
22
- "model.layers.13.block_sparse_moe.gate",
23
- "model.layers.13.self_attn*",
24
- "model.layers.14.block_sparse_moe.gate",
25
- "model.layers.14.self_attn*",
26
- "model.layers.15.block_sparse_moe.gate",
27
- "model.layers.15.self_attn*",
28
- "model.layers.16.block_sparse_moe.gate",
29
- "model.layers.16.self_attn*",
30
- "model.layers.17.block_sparse_moe.gate",
31
- "model.layers.17.self_attn*",
32
- "model.layers.18.block_sparse_moe.gate",
33
- "model.layers.18.self_attn*",
34
- "model.layers.19.block_sparse_moe.gate",
35
- "model.layers.19.self_attn*",
36
- "model.layers.2.block_sparse_moe.gate",
37
- "model.layers.2.self_attn*",
38
- "model.layers.20.block_sparse_moe.gate",
39
- "model.layers.20.self_attn*",
40
- "model.layers.21.block_sparse_moe.gate",
41
- "model.layers.21.self_attn*",
42
- "model.layers.22.block_sparse_moe.gate",
43
- "model.layers.22.self_attn*",
44
- "model.layers.23.block_sparse_moe.gate",
45
- "model.layers.23.self_attn*",
46
- "model.layers.24.block_sparse_moe.gate",
47
- "model.layers.24.self_attn*",
48
- "model.layers.25.block_sparse_moe.gate",
49
- "model.layers.25.self_attn*",
50
- "model.layers.26.block_sparse_moe.gate",
51
- "model.layers.26.self_attn*",
52
- "model.layers.27.block_sparse_moe.gate",
53
- "model.layers.27.self_attn*",
54
- "model.layers.28.block_sparse_moe.gate",
55
- "model.layers.28.self_attn*",
56
- "model.layers.29.block_sparse_moe.gate",
57
- "model.layers.29.self_attn*",
58
- "model.layers.3.block_sparse_moe.gate",
59
- "model.layers.3.self_attn*",
60
- "model.layers.30.block_sparse_moe.gate",
61
- "model.layers.30.self_attn*",
62
- "model.layers.31.block_sparse_moe.gate",
63
- "model.layers.31.self_attn*",
64
- "model.layers.32.block_sparse_moe.gate",
65
- "model.layers.32.self_attn*",
66
- "model.layers.33.block_sparse_moe.gate",
67
- "model.layers.33.self_attn*",
68
- "model.layers.34.block_sparse_moe.gate",
69
- "model.layers.34.self_attn*",
70
- "model.layers.35.block_sparse_moe.gate",
71
- "model.layers.35.self_attn*",
72
- "model.layers.36.block_sparse_moe.gate",
73
- "model.layers.36.self_attn*",
74
- "model.layers.37.block_sparse_moe.gate",
75
- "model.layers.37.self_attn*",
76
- "model.layers.38.block_sparse_moe.gate",
77
- "model.layers.38.self_attn*",
78
- "model.layers.39.block_sparse_moe.gate",
79
- "model.layers.39.self_attn*",
80
- "model.layers.4.block_sparse_moe.gate",
81
- "model.layers.4.self_attn*",
82
- "model.layers.40.block_sparse_moe.gate",
83
- "model.layers.40.self_attn*",
84
- "model.layers.41.block_sparse_moe.gate",
85
- "model.layers.41.self_attn*",
86
- "model.layers.42.block_sparse_moe.gate",
87
- "model.layers.42.self_attn*",
88
- "model.layers.43.block_sparse_moe.gate",
89
- "model.layers.43.self_attn*",
90
- "model.layers.44.block_sparse_moe.gate",
91
- "model.layers.44.self_attn*",
92
- "model.layers.45.block_sparse_moe.gate",
93
- "model.layers.45.self_attn*",
94
- "model.layers.46.block_sparse_moe.gate",
95
- "model.layers.46.self_attn*",
96
- "model.layers.47.block_sparse_moe.gate",
97
- "model.layers.47.self_attn*",
98
- "model.layers.48.block_sparse_moe.gate",
99
- "model.layers.48.self_attn*",
100
- "model.layers.49.block_sparse_moe.gate",
101
- "model.layers.49.self_attn*",
102
- "model.layers.5.block_sparse_moe.gate",
103
- "model.layers.5.self_attn*",
104
- "model.layers.50.block_sparse_moe.gate",
105
- "model.layers.50.self_attn*",
106
- "model.layers.51.block_sparse_moe.gate",
107
- "model.layers.51.self_attn*",
108
- "model.layers.52.block_sparse_moe.gate",
109
- "model.layers.52.self_attn*",
110
- "model.layers.53.block_sparse_moe.gate",
111
- "model.layers.53.self_attn*",
112
- "model.layers.54.block_sparse_moe.gate",
113
- "model.layers.54.self_attn*",
114
- "model.layers.55.block_sparse_moe.gate",
115
- "model.layers.55.self_attn*",
116
- "model.layers.56.block_sparse_moe.gate",
117
- "model.layers.56.self_attn*",
118
- "model.layers.57.block_sparse_moe.gate",
119
- "model.layers.57.self_attn*",
120
- "model.layers.58.block_sparse_moe.gate",
121
- "model.layers.58.self_attn*",
122
- "model.layers.59.block_sparse_moe.gate",
123
- "model.layers.59.self_attn*",
124
- "model.layers.6.block_sparse_moe.gate",
125
- "model.layers.6.self_attn*",
126
- "model.layers.60.block_sparse_moe.gate",
127
- "model.layers.60.self_attn*",
128
- "model.layers.61.block_sparse_moe.gate",
129
- "model.layers.61.self_attn*",
130
- "model.layers.7.block_sparse_moe.gate",
131
- "model.layers.7.self_attn*",
132
- "model.layers.8.block_sparse_moe.gate",
133
- "model.layers.8.self_attn*",
134
- "model.layers.9.block_sparse_moe.gate",
135
- "model.layers.9.self_attn*"
136
- ]
137
- }
138
  }
 
1
  {
2
+ "config_groups": {
3
+ "group_0": {
4
+ "input_activations": {
5
+ "dynamic": false,
6
+ "num_bits": 4,
7
+ "type": "float",
8
+ "group_size": 16
9
+ },
10
+ "weights": {
11
+ "dynamic": false,
12
+ "num_bits": 4,
13
+ "type": "float",
14
+ "group_size": 16
15
+ },
16
+ "targets": [
17
+ "Linear"
18
+ ]
19
+ }
20
+ },
21
+ "ignore": [
22
+ "lm_head",
23
+ "model.layers.0.block_sparse_moe.gate",
24
+ "model.layers.0.self_attn*",
25
+ "model.layers.1.block_sparse_moe.gate",
26
+ "model.layers.1.self_attn*",
27
+ "model.layers.10.block_sparse_moe.gate",
28
+ "model.layers.10.self_attn*",
29
+ "model.layers.11.block_sparse_moe.gate",
30
+ "model.layers.11.self_attn*",
31
+ "model.layers.12.block_sparse_moe.gate",
32
+ "model.layers.12.self_attn*",
33
+ "model.layers.13.block_sparse_moe.gate",
34
+ "model.layers.13.self_attn*",
35
+ "model.layers.14.block_sparse_moe.gate",
36
+ "model.layers.14.self_attn*",
37
+ "model.layers.15.block_sparse_moe.gate",
38
+ "model.layers.15.self_attn*",
39
+ "model.layers.16.block_sparse_moe.gate",
40
+ "model.layers.16.self_attn*",
41
+ "model.layers.17.block_sparse_moe.gate",
42
+ "model.layers.17.self_attn*",
43
+ "model.layers.18.block_sparse_moe.gate",
44
+ "model.layers.18.self_attn*",
45
+ "model.layers.19.block_sparse_moe.gate",
46
+ "model.layers.19.self_attn*",
47
+ "model.layers.2.block_sparse_moe.gate",
48
+ "model.layers.2.self_attn*",
49
+ "model.layers.20.block_sparse_moe.gate",
50
+ "model.layers.20.self_attn*",
51
+ "model.layers.21.block_sparse_moe.gate",
52
+ "model.layers.21.self_attn*",
53
+ "model.layers.22.block_sparse_moe.gate",
54
+ "model.layers.22.self_attn*",
55
+ "model.layers.23.block_sparse_moe.gate",
56
+ "model.layers.23.self_attn*",
57
+ "model.layers.24.block_sparse_moe.gate",
58
+ "model.layers.24.self_attn*",
59
+ "model.layers.25.block_sparse_moe.gate",
60
+ "model.layers.25.self_attn*",
61
+ "model.layers.26.block_sparse_moe.gate",
62
+ "model.layers.26.self_attn*",
63
+ "model.layers.27.block_sparse_moe.gate",
64
+ "model.layers.27.self_attn*",
65
+ "model.layers.28.block_sparse_moe.gate",
66
+ "model.layers.28.self_attn*",
67
+ "model.layers.29.block_sparse_moe.gate",
68
+ "model.layers.29.self_attn*",
69
+ "model.layers.3.block_sparse_moe.gate",
70
+ "model.layers.3.self_attn*",
71
+ "model.layers.30.block_sparse_moe.gate",
72
+ "model.layers.30.self_attn*",
73
+ "model.layers.31.block_sparse_moe.gate",
74
+ "model.layers.31.self_attn*",
75
+ "model.layers.32.block_sparse_moe.gate",
76
+ "model.layers.32.self_attn*",
77
+ "model.layers.33.block_sparse_moe.gate",
78
+ "model.layers.33.self_attn*",
79
+ "model.layers.34.block_sparse_moe.gate",
80
+ "model.layers.34.self_attn*",
81
+ "model.layers.35.block_sparse_moe.gate",
82
+ "model.layers.35.self_attn*",
83
+ "model.layers.36.block_sparse_moe.gate",
84
+ "model.layers.36.self_attn*",
85
+ "model.layers.37.block_sparse_moe.gate",
86
+ "model.layers.37.self_attn*",
87
+ "model.layers.38.block_sparse_moe.gate",
88
+ "model.layers.38.self_attn*",
89
+ "model.layers.39.block_sparse_moe.gate",
90
+ "model.layers.39.self_attn*",
91
+ "model.layers.4.block_sparse_moe.gate",
92
+ "model.layers.4.self_attn*",
93
+ "model.layers.40.block_sparse_moe.gate",
94
+ "model.layers.40.self_attn*",
95
+ "model.layers.41.block_sparse_moe.gate",
96
+ "model.layers.41.self_attn*",
97
+ "model.layers.42.block_sparse_moe.gate",
98
+ "model.layers.42.self_attn*",
99
+ "model.layers.43.block_sparse_moe.gate",
100
+ "model.layers.43.self_attn*",
101
+ "model.layers.44.block_sparse_moe.gate",
102
+ "model.layers.44.self_attn*",
103
+ "model.layers.45.block_sparse_moe.gate",
104
+ "model.layers.45.self_attn*",
105
+ "model.layers.46.block_sparse_moe.gate",
106
+ "model.layers.46.self_attn*",
107
+ "model.layers.47.block_sparse_moe.gate",
108
+ "model.layers.47.self_attn*",
109
+ "model.layers.48.block_sparse_moe.gate",
110
+ "model.layers.48.self_attn*",
111
+ "model.layers.49.block_sparse_moe.gate",
112
+ "model.layers.49.self_attn*",
113
+ "model.layers.5.block_sparse_moe.gate",
114
+ "model.layers.5.self_attn*",
115
+ "model.layers.50.block_sparse_moe.gate",
116
+ "model.layers.50.self_attn*",
117
+ "model.layers.51.block_sparse_moe.gate",
118
+ "model.layers.51.self_attn*",
119
+ "model.layers.52.block_sparse_moe.gate",
120
+ "model.layers.52.self_attn*",
121
+ "model.layers.53.block_sparse_moe.gate",
122
+ "model.layers.53.self_attn*",
123
+ "model.layers.54.block_sparse_moe.gate",
124
+ "model.layers.54.self_attn*",
125
+ "model.layers.55.block_sparse_moe.gate",
126
+ "model.layers.55.self_attn*",
127
+ "model.layers.56.block_sparse_moe.gate",
128
+ "model.layers.56.self_attn*",
129
+ "model.layers.57.block_sparse_moe.gate",
130
+ "model.layers.57.self_attn*",
131
+ "model.layers.58.block_sparse_moe.gate",
132
+ "model.layers.58.self_attn*",
133
+ "model.layers.59.block_sparse_moe.gate",
134
+ "model.layers.59.self_attn*",
135
+ "model.layers.6.block_sparse_moe.gate",
136
+ "model.layers.6.self_attn*",
137
+ "model.layers.60.block_sparse_moe.gate",
138
+ "model.layers.60.self_attn*",
139
+ "model.layers.61.block_sparse_moe.gate",
140
+ "model.layers.61.self_attn*",
141
+ "model.layers.7.block_sparse_moe.gate",
142
+ "model.layers.7.self_attn*",
143
+ "model.layers.8.block_sparse_moe.gate",
144
+ "model.layers.8.self_attn*",
145
+ "model.layers.9.block_sparse_moe.gate",
146
+ "model.layers.9.self_attn*"
147
+ ],
148
+ "quant_algo": "NVFP4",
149
+ "kv_cache_scheme": {
150
+ "dynamic": false,
151
+ "num_bits": 8,
152
+ "type": "float"
153
+ },
154
  "producer": {
155
  "name": "modelopt",
156
  "version": "0.39.0.dev290+gf9d9a71de.d20260213"
157
  },
158
+ "quant_method": "modelopt"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
159
  }
model-00001-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1589ba9462802a55fd5900c8c28ffba60627e681937a6001d2597672d37a976
3
+ size 5371098704
model-00002-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:451e842967ad7ec01a3814923f269afc6702f48910ece62707aeeff32e003169
3
+ size 5371409248
model-00003-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d71afa924dc4868e103ac27116b3b402c100eaee91b64da900f4012279aaf0b3
3
+ size 5371986680
model-00004-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57bed5729ddfcacb9c275915836c163aff26d661b5df478d38e4bbb29c0df9a3
3
+ size 5371409184
model-00005-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:676ad420afbe842e11f8ec0857ecd82e82a48d3d7dd0aa033f139c3e15a0c1bc
3
+ size 5371994784
model-00006-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d36c172b27440846bd5078d3841b510aff31ded8a30b2dcd2b65823d89c4e533
3
+ size 5371416560
model-00007-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5ce2f907798b8f95ce241b1922f0a1131d810a3234b29a091d0cccf9c29fead4
3
+ size 5371994936
model-00008-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c6f7c3be5516c869dca9305a4d63b5d963bbf2f5ee285b29c822f062c033d406
3
+ size 5371416408
model-00009-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee59f384bd96b9f93d9bdded3c7bc98f2679e4bfbaa302b18b7fe03e193af779
3
+ size 5371995088
model-00010-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da92ea76f8f6b785adc9f8f96fb7c932bcfe7e1259b3535c3defcbc7e5cd39e8
3
+ size 5371416256
model-00011-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5c47c49341fba980523a57f5fa5a866fb4abf2a1138897d61228db1332fbfc3
3
+ size 5371995240
model-00012-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:567f20bc35ad0132ddd0cf76769c752f97e39e9de7d2e87e62519d997d3e0c7c
3
+ size 5371416104
model-00013-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:68dbb8c429d31d1253386f4d350e5828ebd41f02a7f149e7f43148f4d3a36f0a
3
+ size 5371995392
model-00014-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:faaab72ca6ceb6b904fee186d9507dfad4cb56696a38df0f9ff92c0a5e67cfb3
3
+ size 5371415952
model-00015-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:008bbca532496ff5f745c7f1308b2a10b028116e307a0fff1818d2d6d30872e4
3
+ size 5371995544
model-00016-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b1f558dd23a8ddcecae8aa110f08263d2ae7608bf50b249f138c241b3202554
3
+ size 5371415800
model-00017-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:418c76db9d9b81941520e261304e4d84deb165dfe1f66b9ddd72f7b53f8b9b19
3
+ size 5371995696
model-00018-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a0f22842745bca21f1a002689da8cd4f5a440204ad86f51124128d29e073ab3
3
+ size 5371415680
model-00019-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eec013a8d27b62f1820ab42bf9e7f2494bca6d8a1754700e7a8652fe76db9d1b
3
+ size 5371995824
model-00020-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9875e27fed69043f393fe23810ecfeb0564c2f43cb4ae675f7598a738f704183
3
+ size 5371415680
model-00021-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:23a4498b0394924543b2a053492e1823bbf1de1a81b1698490e8cb94e2ab6504
3
+ size 5371995824
model-00022-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:285787e95748c1e4c0b0e86b11e3addb1132cd17e8e7377caa94879f79f24e53
3
+ size 5371415672
model-00023-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:12f2e8e1640da602c859dfc94c481a362f437810962e00ba0f94b7e918348f4f
3
+ size 5371416984
model-00024-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f77ce8e0cf1d0e24a6ced53dacb058ecf55093d3f2f1675d9af2b6d379c86a49
3
+ size 5371994512
model-00025-of-00025.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f1ac0d7bc40f8d174ef64a17518df4b1f1a6aa37c673d053b03c19f896fb47db
3
+ size 5505740892
model.safetensors.index.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a6a22e9afca3d4b15cf85b49617efd6be2d2242d63508cb172909fbd9bf9fb72
3
- size 19508929
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:26cb03410b71b101b6a744021d52ed07489514e935fda0d95e855cc7351d3769
3
+ size 19519036