[Model] Systematic support for fp32 head, pooling models part (#23810)

Signed-off-by: wang.yuqi <noooop@126.com>
2025-09-09 22:29:50 +08:00
parent a55cf41a09
commit 19332c0479
14 changed files with 166 additions and 61 deletions
--- a/vllm/model_executor/models/bert.py
+++ b/vllm/model_executor/models/bert.py
@@ -562,7 +562,9 @@ class BertForSequenceClassification(nn.Module, SupportsCrossEncoding,
        self.bert = BertPoolingModel(vllm_config=vllm_config,
                                     prefix=maybe_prefix(prefix, "bert"),
                                     embedding_class=BertEmbedding)
-        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+        self.classifier = nn.Linear(config.hidden_size,
+                                    config.num_labels,
+                                    dtype=vllm_config.model_config.head_dtype)

        pooler_config = vllm_config.model_config.pooler_config
        assert pooler_config is not None