Community mixed-3090 report: CUDA graph headroom matters more than filling VRAM
An October 29 QEngine report identifies a CUDA graph allocation failure. Increasing the expert-cache margin from 256 to 1024MB allowed a full run, at a speed cost. This is one community report, not an AraoAI reproduction.
At 18:37 UTC on October 29, a new comment on club-3090 issue #1846 supplied a fatal log and a same-day rerun. A mixed RTX 3090 and RTX 3090 Ti setup serving Qwen3.8-Flash-Next through QEngine exhausted GPU memory while instantiating CUDA graphs. The issue body contains the previous day's measurements; this article uses the new comment within the current window.
The author kept the image digest and FLASH_COMM=ce FLASH_MAX_SEQS=4 FLASH_KV_ARCHIVE_GB=0 fixed. The engine first logged graph instantiate out of memory, other graphs dropped, then reported an allocation failure at flash_seq.cu:298 and aborted. This is distinct from the non-finite-value failure under FLASH_COMM=auto. Changing communication mode does not guarantee enough memory for later graph allocations.
The controlled setting was FLASH_EC_MARGIN_MB, the per-card margin beside the expert cache. After increasing it from the default 256 to 1024MB and recreating the container, the author reports a complete matrix, both thinking settings and a soak run with zero restarts. Narrative decode changed from 141.8 to 127.8 tok/s; 10K prefill changed from 1159 to 1096 tok/s. These are different test phases, not a single paired benchmark record. They demonstrate a speed cost for headroom, not a free optimization.
The result does not establish 1024MB as a universal setting. It is one reporter on one mixed-card rig. A maintainer saw the same warning without a crash, so the warning alone is not proof that failure is inevitable. Static free VRAM also reflects boot-time KV-pool and expert-cache allocations. PR #1856 was unmerged when checked and changes documentation only, not engine defaults.
Keep the exact image digest, environment, fatal logs and restart counter. Automatic restart can conceal a failed evaluation: a responsive endpoint afterward does not prove that the full run completed. QEngine currently uses PolyForm Noncommercial 1.0.0, with noncommercial restrictions that must be checked before deployment. AraoAI has not reproduced this report and does not present it as purchasing advice.
Sources: same-day rerun; documentation PR; engine license.