hexagon: hmx flash attention #22347
max-krasnyansky
merged
41 commits into
ggml-org:master
from
aizip:feat/hmx-fa
chraac
commented
on 2026-04-30
hmx: extract shared interleave headers and unify matmul batched
80987e6d
hmx: add HMX-accelerated flash attention for prefill
3d4f4951
hmx: replace asm wrappers with Q6_ intrinsics in hmx-utils.h
3057aa44
hmx: drop the duplicate interleave_fp16_weight_chunk_to_tiles
4b2e21a1
hmx: apply upstream optimization to hmx-flash-attn-ops.c
05f3918f
hmx: unify interleave helper
3766b5c7
hmx: multi-thread Q load / O store and enable prefill FA dispatch
288a12c8
hmx: relax matmul pipeline gate to cover k > n shapes (e.g. FFN_down)
fc60e7c0
hmx: optimize FA softmax mask phase (no-ALiBi fast path + GQA dedup)
2715ece0
hmx: Add an asm memory clobber at the phase boundary to prevent reord…
437b2a80
[experimental]: fp16 softmax (EXP2_HF) to accelerate fa
1d681b02
hmx flash-attn: refine cost model coefficients based on profiling data
d9b851e6
hmx flash-attn: replace asm clobber with targeted volatile reads on v…
92f85724
hmx flash-attn: fix prefill correctness (dst indexing, softmax reduce…
4df22427
hmx flash-attn: fix p_tiles dual-tile OOB race; enable MT + pipeline
89d1ab8a
hmx flash-attn: preserve additive mask bias in no-ALiBi fast path
32f2b601
hmx: fix softcap+EXP2_HF interaction, tighten matmul pipeline gate, a…
b4afd05b
[Help Wanted]: refactor D matrix computation into separate function f…
4b072adb
format code
2540d8cc
hexagon: looks like -O3 is causing issues with the large code base, s…
f82d4732
hexagon: use hex_ prefix for swap_ptr
e3a36b3d
hexagon: move vtcm_seq_alloc into vtcm-utils.h
7ab07be9
hmx-utils: add hmx_prefix for layout converters
1289238b
hmx-mm: move main hmx_mm functions to the end, remove unused fwd decl…
2e480560
hmx-mm: remove unused qweight_fetch_task_state_t and minor alignment …
156e55c2
hmx-fa: minor alignment fixes
be1a24df
hmx-fa: move hmx_flash_atten into hmx-ops.h
5a10173a
hmx-fa: remove redundant workpool pointer in the hmx_fa_ctx, plus min…
9940257c
hmx-fa: minor alignment and simplifications
f0af71e1
hexagon: move FA_EXP_F16 option to hostside CMake file
a7f6c72f
hmx-fa: use hvx_vec_splat_f16 instead of fp16_to_bits
431a2859
hmx-fa: add hvx_splat_u16/u8 and use that in the fa instead custom hv…
0513d862
hmx-fa: some more alignment updates in the core fa function
ec1bd5c7
hmx-fa: keep slopes in vtcm in fp16
1e10d609
hexagon: consistent noinline usage (after static)
a1f48785
hex-hmx: consistent use FARF_HIGH to enable debug output
65ed5003
hmx-utils: no need for always_inline attr
024a4328
hex-hmx: consistent noinline usage (static noinline ...)
a9967848
hex-hmx: simplify init_col_scales
59dcf9e1
hexagon: fix editorconfig errors
93f9cb6d
hmx-mm: minor alignment fixes
7f16fece
max-krasnyansky
approved these changes
on 2026-05-02
lhez
approved these changes
on 2026-05-02
Reviewers
Assignees
No one assigned
Labels
testing
ggml
Hexagon
Milestone
No milestone
Login to write a write a comment.
Login via GitHub