已合并
【msmodelslim】【feature】Qwen3-Coder-480B-A35B w4a8 量化 #58
【msmodelslim】【feature】Qwen3-Coder-480B-A35B w4a8 量化 #58
已合并
ylzzz创建于 1月14日
3 个文件变更+103-1
@@ -11,7 +11,7 @@ qwen2 = Qwen2-7B, Qwen2-72B
11qwen2_5 = Qwen2.5-7B-Instruct, Qwen2.5-32B-Instruct, Qwen2.5-72B-Instruct, Qwen2.5-Coder-7B-Instruct,11qwen2_5 = Qwen2.5-7B-Instruct, Qwen2.5-32B-Instruct, Qwen2.5-72B-Instruct, Qwen2.5-Coder-7B-Instruct,
12 DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B12 DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B
13qwen3 = Qwen3-8B, Qwen3-14B, Qwen3-32B13qwen3 = Qwen3-8B, Qwen3-14B, Qwen3-32B
14-qwen3_moe = Qwen3-30B, Qwen3-235B14+qwen3_moe = Qwen3-30B, Qwen3-235B, Qwen3-Coder-480B-A35B
15qwq = Qwen-QwQ-32B, QwQ-32B15qwq = Qwen-QwQ-32B, QwQ-32B
16wan2_1 = Wan2_1, Wan2.116wan2_1 = Wan2_1, Wan2.1
17qwen3_next = Qwen3-Next-80B-A3B-Instruct17qwen3_next = Qwen3-Next-80B-A3B-Instruct
@@ -0,0 +1,59 @@
1+apiversion: modelslim_v1
2+metadata:
3+ config_id: qwen3-coder-480b-w4a8-v1
4+ score: 90
5+ verified_model_types:
6+ - Qwen3-Coder-480B-A35B
7+ label:
8+ w_bit: 4
9+ a_bit: 8
10+ is_sparse: False
11+ kv_cache: False
12+ 
13+default_w8a8_static: &default_w8a8_static
14+ act:
15+ scope: "per_tensor"
16+ dtype: "int8"
17+ symmetric: False
18+ method: "minmax"
19+ weight:
20+ scope: "per_channel"
21+ dtype: "int8"
22+ symmetric: True
23+ method: "minmax"
24+ 
25+default_w4a8_dynamic: &default_w4a8_dynamic
26+ act:
27+ scope: "per_token"
28+ dtype: "int8"
29+ symmetric: True
30+ method: "minmax"
31+ weight:
32+ scope: "per_channel"
33+ dtype: "int4"
34+ symmetric: True
35+ method: "ssz"
36+ 
37+spec:
38+ process:
39+ - type : "quarot"
40+ - type: "flex_smooth_quant"
41+ enable_subgraph_type:
42+ - 'norm-linear'
43+ include:
44+ - "*"
45+ - type: "group"
46+ configs:
47+ - type: "linear_quant"
48+ qconfig: *default_w8a8_static
49+ include: ["*self_attn*"]
50+ - type: "linear_quant"
51+ qconfig: *default_w4a8_dynamic
52+ include: ["*mlp.experts*"]
53+ exclude: ["model.layers.0.*", "model.layers.1.*", "model.layers.2.*", "model.layers.3.*", "model.layers.4.*", "model.layers.57.*", "model.layers.58.*", "model.layers.59.*", "model.layers.60.*", "model.layers.61.*"]
54+ 
55+ save:
56+ - type: "ascendv1_saver"
57+ part_file_size: 4
58+ dataset: autocodebench.jsonl
59+