[Rebase] Rebase to vllm 0.26.0 (#5443)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Signed-off-by: tzhouam <tzhouam@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: tzhouam <tzhouam@gmail.com>
This commit is contained in:
authored and GitHub committed 2026-07-27 23:35:49 +08:00
1 parent fb89ab43ea
commit cbab1d4d8f
41 files changed
+334 -183

No files matched your search

@@ -71,6 +71,11 @@ class ARDiffusionRequestAdapter:
self.num_preemptions = 0
# vLLM watermark gate reads this; map the request lifecycle onto it.
self.status = RequestStatus.WAITING
# vLLM 0.26 KVCacheManager.allocate_slots reads this when computing
# remove_skipped_blocks. Always 0 here: an AR-Diffusion request only
# advances num_computed_tokens on on_chunk_committed(), so there are
# never optimistically-counted in-flight tokens to subtract.
self.num_in_flight_tokens = 0
@property
def num_computed_tokens(self) -> int: