[Precision] Change lm_head layer running in float32 (#3596)

* support lm_head fp32 bf16 fp16 * delete print * code check * check * check * code check * check * check
2025-10-07 09:31:35 +08:00 · 2025-08-26 20:20:06 +08:00
parent 2136990144
commit d233e3c97c
14 changed files with 85 additions and 49 deletions
--- a/fastdeploy/model_executor/models/qwen2.py
+++ b/fastdeploy/model_executor/models/qwen2.py
@@ -326,7 +326,7 @@ class Qwen2ForCausalLM(ModelForCasualLM):
    def compute_logits(self, hidden_states: paddle.Tensor):
        """ """
        logits = self.lm_head(hidden_states)
-        logits = paddle.cast(logits, paddle.float32)
+        logits = logits.astype(paddle.float32)
        logits[:, self.ori_vocab_size :] = -float("inf")

        return logits