apiversion: modelslim_v1
metadata:
config_id: deepseek_v4_pro_w4a8
score: 90
verified_model_types:
- DeepSeek-V4-Pro
verified_tags:
DeepSeek-V4-Pro:
- - vLLM-Ascend
- Atlas_A3_Inference
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"
block_size: 32
- 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:
- "*.ffn.shared_experts.*"
- type: "flex_smooth_quant"
enable_subgraph_type:
- 'norm-linear'
include:
- "*"
exclude:
- "*ffn_norm*"
- type: "linear_quant"
qconfig: *default_w8a8_dynamic
include:
- "*attn*"
exclude:
- "*wo_a"
- "*wo_b"
- "*compressor.wgate"
- "*compressor.wkv"
- "*indexer.weights_proj"
- "*indexer.compressor.wgate"
- "*indexer.compressor.wkv"
- type: "linear_quant"
qconfig: *default_w4a8_dynamic
include:
- "*ffn*"
exclude:
- "*gate"
- "*shared_experts*"
- type: "linear_quant"
qconfig: *default_w8a8_dynamic
include:
- "*ffn.shared_experts*"
dataset: mix_calib.jsonl
save:
- type: "ascendv1_saver"
part_file_size: 4