Issues / #776
#776 --batch on a 100%-resident layer split: engine dies at "captured the batch window over slots 0,1" (zero-doorbell path?)
closed · @Seen-Tomorrow · 7 commentaires · Sur GitHub
BenchmarksMulti-GPUNVIDIA / CUDAModels & quantsWindows
Description
Engine: 0.1.39, Windows, strata-windows-x64.zip Rig: 2× RTX PRO 6000 Blackwell (96 GB), PCIe 5.0 x8/x8 under one switch (PXB), 256 GB RAM Config: --layer-split 24 --expert-cache auto --kv int8 --kv-resident 32768 --max-context 256000 --remote-expert-opt --trim-stage-weights --batch 4 — learned profile, 100% of experts resident on both cards (zero-doorbell graph active on every solo capture) Repro: engine serves solo requests perfectly (222–280 tok/s, ~7000tok/s prefill). Send 4 concurrent greedy requests (~80-token prompts, 200 tokens out). All four admit and copy into slots (~50 ms each); engine exits 1, last log line: strata verify: captured the batch window over slots 0,1 Same with and without --batch-groups 2. --no-capture and --no-token-graph both refuse batch at init, so there is no non-graphed shape to A/B against. Theory: the zero-doorbell verify graph (#646, "runs only when every expert of a layer is in VRAM… on big cards was not measured here") combined with batch-window capture over a layer split. Possibly the first rig where both are live. Happy to test builds — full logs attached. [issue-batch-zerodoorbell-excerpt.log](https://github.com/user-attachments/files/33026669/issue-batch-zerodoorbell-excerpt.log) [issue-batch-zerodoorbell.md](https://github.com/user-attachments/files/33026668/issue-batch-zerodoorbell.md)
Sur le site
Liens install, modèles, releases.