#5747 [Engine] SYCL layer split: serve plans DFlash2 per the CUDA/HIP rule, the fork refuses (generate.cpp:4425) but exits 139 SIGSEGV - fallback-none crash-loops the service
closed · @Fett2 · 2 comentários · No GitHub
Setup & installServer & APIMulti-GPUAMD / HIPNVIDIA / CUDAModels & quantsWindows
Descrição
### Area
Engine
### Desktop app, Terminal or source?
Source (native `serve.server` under systemd; SYCL engine built from tag via the `strata-sycl-dev` container)
### What happened
On a 2x Arc layer-split install, once serve's drafting decision actually *can* see the engine (PR #5745: on stock v0.2.12+ the `strata-sycl.sh` route can never plan DFlash2 at all, because `engine_form()` probes the launcher script), serve plans DFlash2 per its CUDA/HIP rule - and the SYCL fork then refuses the start but **exits 139 (SIGSEGV)**, where its code says `return 2`. With `--draft-fallback none` that refusal is fatal: the unit's restart policy looped ~6 failed engine boots (~25 s each) before we edited the config by hand. With `Auto`, serve's safety net recovered the service - at the cost of one failed engine boot per restart (`Drafting: MTP (DFlash2 didn't start): starting again with the MTP`). Measured on our box 7 Nov, v0.2.12, engine 0.1.58:
- the engine's refusal line: `strata generate: --draft dflash2 is for one card and one request at a time (not --batch, --pipeline-windows or a layer split) in this build` - the guard is in the SYCL fork, `sycl/src/program/generate.cpp:4425`: `if (multi_gpu || o.batch > 0 || o.pipeline_windows > 0) { fprintf(...); return 2; }`
- what serve recorded for the same process: exit **139**. The conversion point (the launcher wrapper, or a crash after the printed refusal) is **unconfirmed** - the exit code is measured, the mechanism is not;
- serve's plan-time rule (Auto is "DFlash2 where the engine has it", named df2 allowed on a layer split - `src/program/generate.cpp` on CUDA/HIP even says `--draft dflash2 runs on a layer split`, RESULTS-DF2MGPU.md) is true for CUDA/HIP and structurally false for the SYCL fork;
- Arc B-series defaults to `pipeline_windows=2`, so the fork's guard also blocks the single-card case unless `--pipeline-windows 0` is added: "the binary contains the flags" is not "the drafter can run in this configuration".
Two asks, both small:
1. gate drafting **planning** on the sycl backend by the fork's guard (one card, `batch == 0`, `pipeline_windows == 0`), or keep Auto at the MTP for SYCL entirely - otherwise every layer-split Intel install that has the df2 pack on disk plans a drafter its engine refuses, as soon as #5745 lands;
2. make the refusal reach serve as exit code 2 rather than SIGSEGV/139. `engine_crashed()` treats 139 as "nothing a second try fixes" (read of current `main`), so what we measured at service level is a restart loop and a dead unit until a human edits the config.
### Smallest steps or command that reproduce it
1. A 2-card SYCL layer-split install: `gpu [1,0]`, `layer_split 14` pinned, Qwen3.8-Flash-Next IQ3_S at `--context 262144 --stream-experts`, stock DFlash2 pack on disk (`strata drafting fetch`; we verified it byte-identical to our own repack of the public HF source, SHA-256 `cda40d62...`).
2. With PR #5745's one-line fix applied to serve (or equivalently an `exe` that is the engine binary itself - **unconfirmed** whether stock native-route installs reach this without the patch, we got there through the fixed decision layer).
3. Set the config `"drafting": "dflash2"` and restart the service: the engine prints the refusal above and exits 139; systemd loops ~6 attempts. Set `"drafting": "auto"` instead: Auto picks DFlash2, the same death happens once, and serve retries with the MTP and recovers.
### Reproduction status
Reproduced repeatedly on one machine (7 Nov: one named-df2 start that restart-looped ~6 times, then an `Auto` instance after the config edit - both exit 139 with the same refusal line). Not re-run on engine 0.1.59/0.1.59.1; not tried on a single-card SYCL install (the `pipeline_windows > 0` clause says one would be needed there too - reading of the fork guard, unconfirmed).
### Strata version
serve/Terminal v0.2.12 + the one-line form-probe fix now in PR #5745; engine 0.1.58 built from tag via `sycl/tools/build.sh` (oneAPI 2026.1, AOT bmg-g31).
### Graphics card, memory and system
EPYC 7443 (8-ch DDR4-3200, 123 GiB RAM), Arc Pro B70 32 GB (Gen4 x16) + Arc Pro B60 (Gen4 x8), `xe` driver, layer split k14 with expert caches 7,168 / 12,110 slots, native serve under systemd (no Docker), Ubuntu-family host.
Refs PR #5745 (fixes the blind probe that hid this path; it also makes it reachable) and #5138, where the blind probe was first reported. Related but distinct: #5159 is a SYCL segfault *during* one-shot generation, not this start-time refusal - likely different root causes, named only so the exit-139 ask (2) can be seen in context.
No site
Links install, modelos, releases.