apiversion: modelslim_v1
metadata:
config_id: deepseek_w4a8
score: 90
verified_model_types:
- DeepSeek-V3.2-Exp
label:
w_bit: 4
a_bit: 8
is_sparse: False
kv_cache: False
default_w4a8_dynamic: &default_w4a8_dynamic
act:
scope: "per_token"
dtype: "int8"
symmetric: True
method: "minmax"
weight:
scope: "per_channel"
dtype: "int4"
symmetric: True
method: "ssz"
default_w8a8_dynamic: &default_w8a8_dynamic
act:
scope: "per_token"
dtype: "int8"
symmetric: True
method: "minmax"
weight:
scope: "per_channel"
dtype: "int8"
symmetric: True
method: "minmax"
spec:
process:
- type: "quarot"
- type: "flex_awq_ssz"
qconfig:
act:
scope: "per_token"
dtype: "int8"
symmetric: True
method: "minmax"
weight:
scope: "per_channel"
dtype: "int4"
symmetric: True
method: "ssz"
ext:
step: 10
enable_subgraph_type:
- 'up-down'
include:
- "*"
exclude:
- "model.layers.0.*"
- "model.layers.1.*"
- "model.layers.2.*"
- "*mlp.shared_experts.*"
- type: "flex_smooth_quant"
enable_subgraph_type:
- 'norm-linear'
- 'ov'
- 'up-down'
include:
- "*self_attn*"
- "model.layers.0.mlp.*"
- "model.layers.1.mlp.*"
- "model.layers.2.mlp.*"
- "*mlp.shared_experts.*"
- "*input_layernorm*"
exclude:
- "*post_attention_layernorm*"
- type: "group"
configs:
- type: "linear_quant"
qconfig: *default_w8a8_dynamic
include:
- "*self_attn*"
exclude:
- "*q_a_proj"
- "*kv_a_proj*"
- "*kv_b_proj"
- "*wq_b"
- "*wk"
- "*weights_proj"
- type: "linear_quant"
qconfig: *default_w8a8_dynamic
include:
- "*mlp*"
exclude:
- "*gate"
- "*mlp.experts.*"
- type: "linear_quant"
qconfig: *default_w8a8_dynamic
include:
- "model.layers.61.mlp.experts.*"
- type: "linear_quant"
qconfig: *default_w4a8_dynamic
include:
- "*mlp.experts.*"
exclude:
- "model.layers.61.*"
dataset: qwen3_cot_w4a4.json
save:
- type: "ascendv1_saver"
part_file_size: 4