[BugFix] Remove incorrect assert in split_decodes_and_prefills (vllm-project#36553)

WoosukKwon · web-flow · commit 006aea17d7de · 2026-03-09T20:02:02.000-07:00
Signed-off-by: Woosuk Kwon &lt;woosuk@inferact.ai&gt;
diff --git a/vllm/v1/attention/backends/utils.py b/vllm/v1/attention/backends/utils.py
@@ -528,7 +528,6 @@ def split_decodes_and_prefills(
         # requests may have a query length of 0 but since they are padding its fine
         # to treat them as decodes (ensures num_decodes matches the captured size)
         if torch.all((query_lens == query_lens[0]) | (query_lens == 0)):
-            assert num_reqs * query_lens[0] == num_tokens, "tokens not padded correctly"
             return num_reqs, 0, num_tokens, 0  # all decodes
         is_prefill = query_lens != query_lens[0]
     else:
