Skip to content

Commit 4aab99c

Browse files
committed
auto detect kvcomp config file
1 parent 0c21f79 commit 4aab99c

File tree

1 file changed

+7
-3
lines changed

1 file changed

+7
-3
lines changed

ucm/sparse/kvcomp/kvcomp_hbm.py

Lines changed: 7 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -45,6 +45,10 @@ def kvcomp_config_path_for_model(vllm_config) -> str:
4545
rel = "ucm/sparse/kvcomp/configs/kvcomp_deepseek_r1_awq_config.json"
4646
elif "qwen3" in model and "32b" in model:
4747
rel = "ucm/sparse/kvcomp/configs/kvcomp_qwen3_32B_config.json"
48+
elif "qwen3" in model and "4b" in model:
49+
rel = "ucm/sparse/kvcomp/configs/kvcomp_qwen3_4B_config.json"
50+
elif "qwq" in model and "32b" in model:
51+
rel = "ucm/sparse/kvcomp/configs/kvcomp_qwq_32B_config.json"
4852
elif "deepseek" in model and "v2" in model:
4953
rel = "ucm/sparse/kvcomp/configs/kvcomp_deepseek_v2_lite_config.json"
5054
else:
@@ -115,9 +119,9 @@ def __init__(self, vllm_config: VllmConfig, role: UcmSparseRole):
115119
.get("KvCompOnDevice")
116120
)
117121

118-
kvcompOnDevice_config_path = self.kvcompOnDevice_cfg[
119-
"kvcompOnDevice_config_path"
120-
]
122+
# auto detect config file for KVCompOnDevice
123+
kvcompOnDevice_config_path = kvcomp_config_path_for_model(vllm_config)
124+
121125
self.kvcompOnDevice_config = KvCompConfig.from_json(kvcompOnDevice_config_path)
122126
logger.info(f"read kvcomp config file : {kvcompOnDevice_config_path} ")
123127
self.hash_topk_tokens = self.kvcompOnDevice_config.vllm_hash_attention_topk

0 commit comments

Comments
 (0)