fix: fix deepseek accuracy when ep_size=1

Signed-off-by: zzzzwwjj <1183291235@qq.com>

Author: zzzzwwjj

vllm_ascend/ops/fused_moe.py

@@ -198,6 +198,7 @@ def fused_experts(
     num_experts = w1.shape[0]
     dtype = hidden_states.dtype
     device = hidden_states.device
+    topk_weights = topk_weights.to(dtype)
     # assert dtype in [torch.float32, torch.float16, torch.bfloat16
     #                  ], "Only float32, float16, and bfloat16 are supported"
 

vllm_ascend/quantization/w8a8_dynamic.py

@@ -342,6 +342,7 @@ def fused_experts(hidden_states: torch.Tensor,
     num_experts = w1.shape[0]
     dtype = hidden_states.dtype
     device = hidden_states.device
+    topk_weights = topk_weights.to(dtype)
 
     if expert_map is not None:
         # Generate token indices and flatten