#5636 [Engine] sm_70/75: verify-window capture dies at `copy_i32_from_mapped launch: invalid argument` on first request

open · @archadrian · 4 Kommentare · Auf GitHub

Server & APINVIDIA / CUDAModels & quants

Beschreibung

**Build:** main @ 36ced1ac, CUDA 12.8, `-DCMAKE_CUDA_ARCHITECTURES="70;75"` with `-DSTRATA_EXPERIMENTAL_SM60=ON`, `-DSTRATA_NATIVE_EXPERTS=ON -DSTRATA_Q6K_EXPERTS=ON -DSTRATA_MMQ_KQUANTS=ON`.
**Cards:** Tesla V100S-32 (sm_70) and RTX 2080 Ti (sm_75), each alone, inside a libvirt guest (VFIO).
**Model:** Qwen3-Next-80B-A3B GyroM native pack (`--pack ... --native ... --no-ple --spec 4 --kv k8v4 --max-context 8192`).

Everything loads cleanly (weights, cache, buffers; `free with everything loaded` printed), then the first request dies at:

```
strata verify: capturing the 12-token window (2366 MiB of VRAM free)
copy_i32_from_mapped launch: invalid argument
```

The engine exits; serve returns 503. Repro is deterministic on the 2080 Ti (sm_75): every attempt fails — `--spec 4` and `--spec 2`, expert-cache 10160 (auto) and 8000, `STRATA_VERIFY_DEBUG=1` on and off, 644 MiB and 2366 MiB free at capture (so not headroom). On the V100 (sm_70) the first launch crashed the same way; a relaunch with `STRATA_VERIFY_DEBUG=1` captured fine and then ran correctly (75-77 t/s decode, 100% cache hits) — so the failure may be a race in the capture preamble that the debug prints happen to dodge, and sm_75 loses the race every time.

`copy_i32_from_mapped_kernel` itself is a plain grid<=64 x 128 copy (elementwise.cu:393), so the invalid argument looks like it comes from an earlier op invalidating the capture rather than the copy launch itself.

**Expected:** the window graph captures on sm_70/75 (the same tree ran generation fine on sm_86/sm_89+cards). **Actual:** capture dies at the first windowed launch.

Happy to test patches or a debug build on both cards.

Mehr auf der Site

Links zu Install, Modellen, Releases.