test op-cumsum

Author: puppy

vllm_ascend/sample/sampler.py

@@ -51,7 +51,10 @@ def _apply_top_k_top_p(
             logits.masked_fill_(elements_to_discard, -float("inf"))
 
         if p is not None:
+            old_dtype = probs_sort.dtype
+            probs_sort = probs_sort.to(torch.float32)
             cumprob = torch.cumsum(probs_sort, dim=-1)
+            cumprob = cumprob.to(old_dtype)
             top_p_mask = cumprob <= 1 - p.unsqueeze(dim=1)
             top_p_mask[:, -1] = False  # at least one