docs: MULTI_GPU.md lists Pascal as unsupported for the layer split, but 2x Tesla P40 runs it (v0.1.39) #1029

Observation (v0.1.39, main = 6f32ec0, unmodified): two Tesla P40 22 GB (compute capability 6.1) ran Flash-Next IQ2_XS with the layer split, although docs/MULTI_GPU.md lists cards below compute capability 7.5 as unsupported.

How: I installed with ./setup.sh ... --cuda 12 --build --gpu 1 (CUDA 12.4 + g++-13, engine compiled for sm_61 with STRATA_EXPERIMENTAL_SM60=1), then edited the generated config by hand to "gpu": [0, 1] and "layer_split": "auto" and started serve/server.py --engine strata --config .... I did not try --gpus 0,1 through setup, so I can't say whether setup itself would refuse it.

Evidence (engine log, two-card runs):

strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)
strata generate: layer split auto: K=24 - predicted 75.8 ms per decode window; the caches hold 24576 of 24576 profiled pairs (~100.0% of the routed mass)
strata generate: layer split: CUDA1 holds its weights, session [24, 48) and the head; 18.63 GiB free

Median decode for a short prompt went from 19.6 tok/s (one P40) to 34.8 tok/s (both); prompt reading 348 / 388 tok/s (4K / 16K) on one card and 353 / 501 on both. Three runs per cell; the first run of each is much slower (warm-up). All per-run data, engine logs and the config files are in #1028 (a results-only community report).

Docs: docs/MULTI_GPU.md ("Not supported", and the example listing a GTX 1080 Ti as "not supported - older than the RTX 20 series") and docs/OLDER_GPUS.md (which allows --gpus 0,1 "with a newer card" and says a V100 can share a model with an RTX 30/40 card) read differently for two Pascal cards, which neither page mentions. If this is expected to work experimentally, a sentence in either page would help; if it only works by hand-editing the config, that may be worth saying too.

Caveats: one machine, one quantisation (IQ2_XS), 32,768 context, three runs; the two-card runs were not NUMA-pinned. Not tested: --batch with a layer split, vision, contexts above 32K, other quantisations. Happy to test a specific variant if that helps (for example --gpus 0,1 --cuda 12 through setup, or IQ3_S).

Related Pascal reports: #875, #395. The wider study (llama.cpp and Ollama on the same weights, accuracy and repeatability): https://github.com/sarge18/p40-llm-engine-bakeoff

Disclosure: prepared with an AI assistant (Claude Sonnet 5.5, medium effort) under my direction; every number above comes from the per-run data in #1028.