[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/jamba.py
+++ b/vllm/model_executor/models/jamba.py
@@ -613,7 +613,7 @@ class JambaForSequenceClassification(JambaForCausalLM):
            config.hidden_size,
            num_labels,
            bias=score_bias,
-            dtype=torch.float32,
+            dtype=vllm_config.model_config.head_dtype,
        )

        pooler_config = vllm_config.model_config.pooler_config