Gahow Wang
b6591950bc
patches/0001-fix-kv-transfer-abort-race.patch: Fix scheduler assert crash when KV transfer callback arrives after request abort in PD-disaggregated serving. patches/README.md: How to apply patches to source tree or installed package. Per-patch description with problem/fix/impact. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
31 lines
1.0 KiB
Markdown
31 lines
1.0 KiB
Markdown
# vLLM Patches
|
|
|
|
Patches against vLLM v0.18.1. Apply to either the source tree (`third_party/vllm/`) or the installed package.
|
|
|
|
## Applying
|
|
|
|
```bash
|
|
# To source tree (for rebuilding)
|
|
cd third_party/vllm && git apply ../../patches/*.patch
|
|
|
|
# To installed package (quick, no rebuild)
|
|
SITE=$(python -c "import vllm; print(vllm.__path__[0])")
|
|
for p in patches/*.patch; do
|
|
patch -p1 -d "$(dirname $SITE)" < "$p"
|
|
done
|
|
```
|
|
|
|
## Patches
|
|
|
|
### 0001-fix-kv-transfer-abort-race.patch
|
|
|
|
**File**: `vllm/v1/core/sched/scheduler.py`
|
|
|
|
**Problem**: When a client disconnects (timeout/abort) during PD-disaggregated serving, the Mooncake KV transfer callback arrives after the request has been removed from the scheduler. The `assert req_id in self.requests` kills the engine process.
|
|
|
|
**Fix**: Replace fatal assert with graceful skip + warning log.
|
|
|
|
**Impact**: Without this patch, decode instances crash after ~200 requests under sustained load with concurrent KV transfers.
|
|
|
|
**Upstream**: Not yet submitted. Could be upstreamed to vllm-project/vllm.
|