[BUGFIX]解决推测解码内核类型不匹配

vllm/v1/spec_decode/utils.py

@@ -79,7 +79,8 @@ def eagle_prepare_next_token_padded_kernel(
 
     if is_discarded:
         backup_token = tl.load(backup_next_token_ids_ptr + req_idx)
-        valid_count = tl.full((), 0, dtype=tl.uint32)
+        # valid_count = tl.full((), 0, dtype=tl.uint32)
+        valid_count = tl.cast(0, tl.uint32)
         tl.store(next_token_ids_ptr + req_idx, backup_token)
         tl.store(valid_sampled_tokens_count_ptr + req_idx, valid_count)
     else:
@@ -92,7 +93,8 @@ def eagle_prepare_next_token_padded_kernel(
 
         # Rejected tokens are -1, valid tokens are in [0, vocab_size)
         is_valid_mask = (token_ids != -1) & (token_ids < vocab_size) & token_mask
-        valid_count = tl.sum(is_valid_mask)
+        # valid_count = tl.sum(is_valid_mask)
+        valid_count = tl.cast(tl.sum(is_valid_mask), tl.uint32)
 
         if valid_count > 0:
             # Guaranteed to be well-defined since