Like 0
Fork 0

Last active 1786537955

en2zmax's Avatar en2zmax revised this gist 1786537955. Go to revision

1 file changed, 151 insertions

config.json(file created)

@@ -0,0 +1,151 @@
1 + {
2 + "architectures": ["Qwen3_5MoeForConditionalGeneration"],
3 + "image_token_id": 248056,
4 + "model_type": "qwen3_5_moe",
5 + "text_config": {
6 + "attention_bias": false,
7 + "attention_dropout": 0.0,
8 + "attn_output_gate": true,
9 + "bos_token_id": 248044,
10 + "dtype": "bfloat16",
11 + "eos_token_id": 248044,
12 + "full_attention_interval": 4,
13 + "head_dim": 256,
14 + "hidden_act": "silu",
15 + "hidden_size": 2048,
16 + "initializer_range": 0.02,
17 + "layer_types": [
18 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
19 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
20 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
21 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
22 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
23 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
24 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
25 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
26 + "linear_attention", "linear_attention", "linear_attention", "full_attention",
27 + "linear_attention", "linear_attention", "linear_attention", "full_attention"
28 + ],
29 + "linear_conv_kernel_dim": 4,
30 + "linear_key_head_dim": 128,
31 + "linear_num_key_heads": 16,
32 + "linear_num_value_heads": 32,
33 + "linear_value_head_dim": 128,
34 + "mamba_ssm_dtype": "float32",
35 + "max_position_embeddings": 262144,
36 + "model_type": "qwen3_5_moe_text",
37 + "moe_intermediate_size": 512,
38 + "mtp_num_hidden_layers": 1,
39 + "mtp_use_dedicated_embeddings": false,
40 + "num_attention_heads": 16,
41 + "num_experts": 256,
42 + "num_experts_per_tok": 8,
43 + "num_hidden_layers": 40,
44 + "num_key_value_heads": 2,
45 + "output_router_logits": false,
46 + "pad_token_id": null,
47 + "partial_rotary_factor": 0.25,
48 + "rms_norm_eps": 1e-06,
49 + "rope_parameters": {
50 + "mrope_interleaved": true,
51 + "mrope_section": [11, 11, 10],
52 + "partial_rotary_factor": 0.25,
53 + "rope_theta": 10000000,
54 + "rope_type": "default"
55 + },
56 + "router_aux_loss_coef": 0.001,
57 + "shared_expert_intermediate_size": 512,
58 + "tie_word_embeddings": false,
59 + "use_cache": true,
60 + "vocab_size": 248320
61 + },
62 + "tie_word_embeddings": false,
63 + "transformers_version": "4.57.1",
64 + "video_token_id": 248057,
65 + "vision_config": {
66 + "deepstack_visual_indexes": [],
67 + "depth": 27,
68 + "hidden_act": "gelu_pytorch_tanh",
69 + "hidden_size": 1152,
70 + "in_channels": 3,
71 + "initializer_range": 0.02,
72 + "intermediate_size": 4304,
73 + "model_type": "qwen3_5_moe",
74 + "num_heads": 16,
75 + "num_position_embeddings": 2304,
76 + "out_hidden_size": 2048,
77 + "patch_size": 16,
78 + "spatial_merge_size": 2,
79 + "temporal_patch_size": 2
80 + },
81 + "vision_end_token_id": 248054,
82 + "vision_start_token_id": 248053,
83 + "quantization_config": {
84 + "activation_scheme": "dynamic",
85 + "fmt": "e4m3",
86 + "quant_method": "fp8",
87 + "modules_to_not_convert": [
88 + "model.visual.blocks.0.attn.proj", "model.visual.blocks.0.attn.qkv",
89 + "model.visual.blocks.0.mlp.linear_fc1", "model.visual.blocks.0.mlp.linear_fc2",
90 + "visual.blocks.0.attn.proj", "visual.blocks.0.attn.qkv_proj",
91 + "visual.blocks.0.mlp.linear_fc1", "visual.blocks.0.mlp.linear_fc2",
92 + "model.visual.blocks.1.attn.proj", "model.visual.blocks.1.attn.qkv",
93 + "model.visual.blocks.1.mlp.linear_fc1", "model.visual.blocks.1.mlp.linear_fc2",
94 + "visual.blocks.1.attn.proj", "visual.blocks.1.attn.qkv_proj",
95 + "visual.blocks.1.mlp.linear_fc1", "visual.blocks.1.mlp.linear_fc2",
96 + ... (блоки 2..26 — по тому же шаблону: model.visual.blocks.N.attn.proj/qkv, .mlp.linear_fc1/fc2, visual.blocks.N.attn.proj/qkv_proj, .mlp.linear_fc1/fc2),
97 + "model.visual.deepstack_merger_list.0.linear_fc1",
98 + "model.visual.deepstack_merger_list.0.linear_fc2",
99 + "model.visual.deepstack_merger_list.0.norm",
100 + "visual.deepstack_merger_list.0.linear_fc1",
101 + "visual.deepstack_merger_list.0.linear_fc2",
102 + "visual.deepstack_merger_list.0.norm",
103 + "model.visual.deepstack_merger_list.1.linear_fc1",
104 + "model.visual.deepstack_merger_list.1.linear_fc2",
105 + "model.visual.deepstack_merger_list.1.norm",
106 + "visual.deepstack_merger_list.1.linear_fc1",
107 + "visual.deepstack_merger_list.1.linear_fc2",
108 + "visual.deepstack_merger_list.1.norm",
109 + "model.visual.deepstack_merger_list.2.linear_fc1",
110 + "model.visual.deepstack_merger_list.2.linear_fc2",
111 + "model.visual.deepstack_merger_list.2.norm",
112 + "visual.deepstack_merger_list.2.linear_fc1",
113 + "visual.deepstack_merger_list.2.linear_fc2",
114 + "visual.deepstack_merger_list.2.norm",
115 + "model.visual.merger.linear_fc1", "model.visual.merger.linear_fc2", "model.visual.merger.norm",
116 + "model.visual.patch_embed.proj", "model.visual.pos_embed",
117 + "visual.merger.linear_fc1", "visual.merger.linear_fc2", "visual.merger.norm",
118 + "visual.patch_embed.proj", "visual.pos_embed",
119 + "visual", "model.visual", "lm_head", "model.embed_tokens",
120 + "model.language_model.layers.0.input_layernorm",
121 + "model.language_model.layers.0.mlp.shared_expert_gate",
122 + "model.language_model.layers.0.post_attention_layernorm",
123 + "model.language_model.layers.0.mlp.gate",
124 + "model.language_model.layers.0.linear_attn.A_log",
125 + "model.language_model.layers.0.linear_attn.conv1d",
126 + "model.language_model.layers.0.linear_attn.dt_bias",
127 + "model.language_model.layers.0.linear_attn.in_proj_ba",
128 + "model.language_model.layers.0.linear_attn.in_proj_b",
129 + "model.language_model.layers.0.linear_attn.in_proj_a",
130 + "model.language_model.layers.0.linear_attn.norm",
131 + "model.language_model.layers.1.input_layernorm",
132 + "model.language_model.layers.1.mlp.shared_expert_gate",
133 + "model.language_model.layers.1.post_attention_layernorm",
134 + "model.language_model.layers.1.mlp.gate",
135 + "model.language_model.layers.1.linear_attn.A_log",
136 + "model.language_model.layers.1.linear_attn.conv1d",
137 + "model.language_model.layers.1.linear_attn.dt_bias",
138 + "model.language_model.layers.1.linear_attn.in_proj_ba",
139 + "model.language_model.layers.1.linear_attn.in_proj_b",
140 + "model.language_model.layers.1.linear_attn.in_proj_a",
141 + "model.language_model.layers.1.linear_attn.norm",
142 + ... (слои 2..39 — по тому же шаблону; на слоях-«full_attention» добавляются "self_attn.k_norm" и "self_attn.q_norm"),
143 + "model.language_model.layers.39.self_attn.q_norm",
144 + "mtp.layers.0.input_layernorm", "mtp.layers.0.mlp.gate",
145 + "mtp.layers.0.mlp.shared_expert_gate", "mtp.layers.0.post_attention_layernorm",
146 + "mtp.layers.0.self_attn.k_norm", "mtp.layers.0.self_attn.q_norm",
147 + "mtp.fc", "mtp.norm", "mtp.pre_fc_norm_embedding", "mtp.pre_fc_norm_hidden"
148 + ],
149 + "weight_block_size": [128, 128]
150 + }
151 + }
Newer Older

Powered by Opengist ⋅ Load: 25ms⋅

English
⋅ Gitea