Daniel Han 82e14e7ec8 flex: auto-detect FA4 prefill on Hopper / Blackwell ... --fa4_prefill now accepts three states: True (force on, warn + fall back on sub-Hopper), False (force off), None / default (auto-enable where supported). Argparse switches to BooleanOptionalAction so both --fa4_prefill and --no-fa4_prefill work, with the default being auto-detect from torch.cuda.get_device_capability. Adds a cu13 / cu12 install section and a per-GPU support matrix to scripts/benchmarks/README.md. Adds tests/test_fa4_capability_guard.py covering the nine combinations of (explicit-on / auto / explicit-off) x (sm_80 / sm_90 / sm_100 / sm_120). Monkey-patches get_device_capability and stubs PageTable / patch_qwen3_model so it runs without CUDA.		2026-04-21 05:40:09 +00:00
..
benchmarks	flex: auto-detect FA4 prefill on Hopper / Blackwell	2026-04-21 05:40:09 +00:00
enforce_kwargs_spacing.py	Formatting & bug fixes (#3563)	2025-11-07 06:00:22 -08:00
install_gemma4_mlx.sh	Move gemma4 script (#4994)	2026-04-12 23:41:15 -07:00
install_qwen3_6_mlx.sh	Add qwen3.6 script (#5084)	2026-04-17 01:21:30 -07:00
run_ruff_format.py	Formatting & bug fixes (#3563)	2025-11-07 06:00:22 -08:00