[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
2025-10-05 16:48:03 +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/engine/engine.py
+++ b/fastdeploy/engine/engine.py
@@ -1099,6 +1099,7 @@ class LLMEngine:
            "enable_custom_all_reduce": self.cfg.parallel_config.enable_custom_all_reduce,
            "enable_logprob": self.cfg.enable_logprob,
            "enable_mm": self.cfg.enable_mm,
+            "lm_head_fp32": self.cfg.model_config.lm_head_fp32,
        }
        for worker_flag, value in worker_append_flag.items():
            if value: