Fix noaux_tc cuda Error 700 in CUDAGraph (#4174)

2025-10-07 01:22:59 +08:00 · 2025-09-23 18:41:33 +08:00
parent c96a535a5d
commit 1a6283424e
12 changed files with 334 additions and 148 deletions
--- a/fastdeploy/model_executor/models/deepseek_v3.py
+++ b/fastdeploy/model_executor/models/deepseek_v3.py
@@ -121,6 +121,7 @@ class DeepSeekV3MoE(nn.Layer):
        super().__init__()

        self.tp_size = fd_config.parallel_config.tensor_parallel_size
+        self.norm_topk_prob = fd_config.model_config.norm_topk_prob

        weight_key_map = {
            "gate_correction_bias_key": f"{prefix}.gate.e_score_correction_bias",
@@ -150,6 +151,7 @@ class DeepSeekV3MoE(nn.Layer):
        self.experts = FusedMoE(
            fd_config=fd_config,
            reduce_results=False,
+            renormalize=self.norm_topk_prob,
            moe_intermediate_size=fd_config.model_config.moe_intermediate_size,
            num_experts=fd_config.model_config.n_routed_experts,
            top_k=fd_config.model_config.num_experts_per_tok,