diff options
Diffstat (limited to 'AGENTS.md')
| -rw-r--r-- | AGENTS.md | 32 |
1 files changed, 32 insertions, 0 deletions
@@ -31,9 +31,41 @@ reimplement anything upstream's scripts already do. 12 GB card, so close other programs before starting it. - Gyro on Arc was tested upstream only on an A750 (`i915`). The B580 is untested natively; one WSL2 B580 report saw device loss. +- Measured here (v0.2.2, B580, 31 GiB RAM): on `xe` the engine runs with + `--stream-experts` and `STRATA_VERIFY_NO_HOST`, so every expert must sit in + the VRAM cache or a pinned RAM mirror (capped at MemAvailable minus 4 GiB). + Gyro-M does not fit: 4.6 GiB VRAM cache + 21.6 GiB mirror left 5,309 + experts out and the engine REFUSED. setup's "35 GB of VRAM + RAM: fits" + check is wrong for this path. Gyro-S fits: 3,062 experts in VRAM, 21,514 + mirrored (21 GiB), about 6 GiB of RAM left, decode 22-29 tok/s. +- Gyro-S is not stable here: 2 of about 6 requests hit + `UR_RESULT_ERROR_DEVICE_LOST` in the verify (speculative decoding) step, + the engine segfaulted (139) and the chat showed "verify error". The server + restarts the engine and the card recovered each time. +- Likely cause, not verified: device spins on host flags are capped at + 20,000 reads on B-series cards (`sycl/include/strata/sycl_doorbell.hpp`), + tuned on a B70 that holds every expert in VRAM. Here 21 GiB of experts + come from the RAM mirror over PCIe, so a wait can outrun the cap. Upstream + documents the same 139 crash pattern for a cold first request + (`docs/INTEL.md`, `STRATA_SYCL_SPIN_MAX`). +- Untried next steps: (1) `"env": {"STRATA_SPIN_MAX": "2000000"}` in the + config, read at engine start. Risk: on `xe` an overlong spin can wedge the + GPU until reboot, and the B580 drives the desktop. (2) Drop `--spec`, + `--spec-min-p` and `--mtp` from the config to avoid the draft verify, at + lower decode speed. +- Status 2026-10-22: testing stopped. Gyro-M deleted. Gyro-S (55 GiB) and + the shared draft layer remain in `Strata-data/`. - On `xe`, asking for more VRAM than the card has does not fail: buffers spill into RAM and the host can livelock. Never raise `--context` or lower `--vram-reserve-mib` past what setup's fit rule picked. +- Upstream bug (v0.2.2): `setup_intel.py` tells `setup.py` the host has + 1024 GB of RAM (`S.ram_gb = lambda: fake_ram`), so on the `xe` path + `setup.py` adds `--ple-io ram` to the Gyro config. That loads the 54 GB + n-gram table into RAM and would OOM this host. The Intel code itself says + Gyro's table must stay on the SSD (it only guards the `i915` branch). After + every `setup` run, check `Strata/strata-gyro-gyro-m.json` and drop + `--ple-io ram` from `args`. Start the model with + `Strata/run-gyro-gyro-m.sh`, which does not rewrite the config. - Pictures are off by default on Battlemage (they reset a B70). Do not pass `--vision yes`. - `docker build --pull` matters: a stale base can carry an ocloc that crashes |