vLLM engine core stability and performance, specifically under WSL2 with large chunked-prefill requests using KVarN and DFlash2 speculative decoding.

## Platform

**WSL2 (Ubuntu, kernel 6.18.33.2-microsoft-standard-WSL2, NVIDIA driver 616.56,
single RTX 3090 24GB).** Everything in this report was observed on WSL2, and our
leading suspicion involves the WSL2 GPU stack, so the platform is a first-class
part of the bug report, not a footnote. If this cannot be reproduced on bare
metal, WSL2 itself is a primary suspect.

## Summary

With this repo's `SPEC=dflash2 CTX=huge PREFIX_CACHE=1` single-user config
(KVarN 4/2-bit KV, DFlash2 7-token speculative decoding; vLLM 0.27.1 at commit
2ae239f), the vLLM engine core can **wedge on the first large chunked-prefill
request after a fresh boot**:

- the unconditional 10-second `Engine 000: … Running:` stats lines go silent;
- the API process stays up — the port answers, requests are accepted, but none
is ever scheduled;
- in one production incident (2026-08-28) this persisted **~4.5h** with the GPU
at 100% (~24.1/24 GB resident) producing nothing, until the core spontaneously
unwedged; the operator then stopped the engine.
- over the following ~36h the same failure recurred **5 times**, each minutes
after a fresh boot, always during a large-prompt (10k+ token) request.

A one-request reproduction harness, an A/B across three configurations, and a
py-spy capture of the wedged core are below. We did not root-cause it to a
single line, so this reports the phenomenon + evidence rather than a
speculative fix, and ends with two questions for the maintainer.

## Environment

- WSL2 ...