Pedro Augusto f3177cd807 sparse/config: default ATTN to flex_gemm_sparse_attn on Darwin when flex_gemm probe passes ... The flex_gemm_sparse_attn backend is now flash-attention-v2 with simdgroup_matrix_multiply_accumulate for Q@K^T and P@V plus simd-shuffle softmax row reductions, and wins 5–15× over SDPA-padded-CPU-bounce at every measured shape including max_seqlen=2048. Production decoder block on M3 Max: 5.04× wall-clock vs all-SDPA-padded baseline (3.49 ms vs 17.57 ms), entirely from this change being the default. The existing __flex_gemm_works_on_mps() probe covers the same package the attention path lives in, so the gate is identical: if conv probe passes, set both CONV='flex_gemm' and ATTN='flex_gemm_sparse_attn'. SPARSE_ATTN_BACKEND= (or ATTN_BACKEND=) env override is unchanged. Also fix benchmarks/e2e_decoder.py to inject the repo root into sys.path so it runs as `python benchmarks/e2e_decoder.py` from any cwd.		2026-07-17 07:51:05 +03:00
..
attention	sparse/attention: drop FUSED_ATTN_MAX_SEQLEN threshold to match CUDA-path precedent	2026-07-17 07:51:05 +03:00
conv	add mlx backend for apple silicon with metal gpu acceleration	2026-07-17 07:51:05 +03:00
spatial	Init Release	2025-12-16 19:00:42 +00:00
transformer	Init Release	2025-12-16 19:00:42 +00:00
__init__.py	Init Release	2025-12-16 19:00:42 +00:00
basic.py	Init Release	2025-12-16 19:00:42 +00:00
config.py	sparse/config: default ATTN to flex_gemm_sparse_attn on Darwin when flex_gemm probe passes	2026-07-17 07:51:05 +03:00
linear.py	Init Release	2025-12-16 19:00:42 +00:00
nonlinearity.py	Init Release	2025-12-16 19:00:42 +00:00
norm.py	sparse/norm: CPU-staged zeros_like workaround for MPS	2026-07-17 07:51:05 +03:00