biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Refactor] Separate sequence and token pooling types (#32026)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-10 12:53:24 +08:00
committed by GitHub
parent 52d428295d
commit 583a90e005
42 changed files with 324 additions and 204 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/models/qwen2_rm.py
View File

@@ -96,7 +96,7 @@ class Qwen2RewardBaseModel(nn.Module, SupportsLoRA, SupportsPP):
return loader.load_weights(weights)
@default_pooling_type("ALL")
@default_pooling_type(tok_pooling_type="ALL")
class Qwen2ForRewardModel(Qwen2RewardBaseModel):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
vllm_config.model_config.hf_config.num_labels = 1
@@ -108,7 +108,7 @@ class Qwen2ForRewardModel(Qwen2RewardBaseModel):
self.pooler = pooler_for_token_classify(pooler_config)
@default_pooling_type("STEP")
@default_pooling_type(tok_pooling_type="STEP")
class Qwen2ForProcessRewardModel(Qwen2RewardBaseModel):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
vllm_config.model_config.hf_config.num_labels = 2