The first #10485 fix (#10494) made the Blackwell physical-batch boost
per-device/context-aware, which neutralized the big compute-buffer OOM, but
the reporter's 2x16 GiB consumer Blackwell still OOM'd. Tracing the post-fix
log: the model now loads its weights, builds the main context and warms up
fine, and dies only on the *last* allocation — the MTP draft context's 800 MiB
KV cache on the tighter device.
#10411 changed only two defaults: the physical batch (now gated) and a
VRAM-scaled parallel-slot count. The KV cache is unified (n_ctx_seq == full
context proves slots share the budget, so parallel doesn't multiply KV), but
n_seq_max=4 still adds per-slot compute-graph / context-checkpoint / output
scratch. On a device packed ~99% by a 27B model spanning both cards, that
overhead is the few-hundred-MiB straw — which is why reverting #10411 (and only
#10411) restores a working load.
Gate the parallel-slot default on the same per-device headroom predicate as the
batch boost: when a large context already fills a single card
(largeContextForDevice), keep n_parallel=1. A user running one big-context model
that barely fits across two consumer GPUs is not serving four concurrent
tenants. Small contexts and large unified-memory devices (GB10) keep full
concurrency. Applied on both the single-host path and the distributed router.
Also make the auto-tuning visible and reversible (the debugging here needed
DEBUG logs and a git bisect):
- Log the effective performance-relevant runtime options at INFO once per
model load ("effective runtime tuning …": context, n_batch, n_gpu_layers,
parallel, flash_attention, f16) so an admin can see what will run and pin or
override any value in the model YAML.
- LOCALAI_DISABLE_HARDWARE_DEFAULTS=true skips the hardware auto-tuning
entirely (mirrors LOCALAI_DISABLE_GUESSING) for stock llama.cpp behavior.
Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
