pppeng 9a0b786f2b [bugfix][0.18.0] Fix race in non-blocking num_accepted_tokens (#8764) ...

### What this PR does / why we need it?
The same fix from https://github.com/vllm-project/vllm/pull/36013.
In _update_states_after_model_execute, num_accepted_tokens is copied
from GPU to pinned CPU memory with non_blocking=True. The CPU-side numpy
view is later read in _build_attention_metadata during the next
execute_model call. With async scheduling, _bookkeeping_sync
deliberately avoids any CUDA synchronization (the whole point of async
scheduling), so there is no guarantee the DMA has landed before the CPU
read.

Signed-off-by: ppppeng <zepengliu912@qq.com>

2026-04-27 23:28:52 +08:00

..

v2

[v0.18.0][CI] Fix releases/v0.18.0 ci test only support vllm v0.18.0 (#7686)

2026-03-26 18:36:04 +08:00

__init__.py

[Misc][V0 Deprecation] Remove Cache Engine Used for V0 Worker (#1878)

2025-07-19 09:42:32 +08:00

block_table.py

[Hybrid] support prefix cache for Qwen3.5/Next with --mamba-cache-mode align (#7103)

2026-03-15 09:44:09 +08:00

model_runner_v1.py

[bugfix][0.18.0] Fix race in non-blocking num_accepted_tokens (#8764)

2026-04-27 23:28:52 +08:00

npu_input_batch.py

[Hybrid] support prefix cache for Qwen3.5/Next with --mamba-cache-mode align (#7103)

2026-03-15 09:44:09 +08:00

pcp_utils.py

feat(attention_cp): support chunked prefill for Qwen3Next with PCP&DCP (#6900)

2026-03-09 17:55:09 +08:00

worker.py

[0.18.0][profiler] profile AICore and MTE time with torch profiler (#7730)

2026-03-27 16:37:54 +08:00