[ROCm] Add Torch SDPA and xFormers optimization for FireRedASR - #102

Open

sammysun0711 wants to merge 13 commits into

FireRedTeam:mainfrom

sammysun0711:asr_opt_rocm_upstream

Commits on Oct 29, 2025

Add FireRedASR optimization on ROCm
sammysun0711
committed
974386f
Minor update
sammysun0711
committed
c367305

Commits on Nov 3, 2025

Add attention mask support in batch size > 1 condition for Torch SDPA and xFormers attention
sammysun0711
committed
d6de6ba
Refactor xFormers bias_attn handeling with BlockDiagonalMask for batch size > 1 case
sammysun0711
committed
e210318
Fix dtype
sammysun0711
committed
4dbda0b

Commits on Nov 4, 2025

Fix attn_bias creation in cross attention mask for audio input with variable length
sammysun0711
committed
aa3f757

Commits on Nov 6, 2025

Refactor xFormers backend
sammysun0711
committed
7efe26f
Update benchmark script to load audio from directory
sammysun0711
committed
2d6af13

Commits on Nov 7, 2025

Refactor benchmark run benchmark with audio directory with variable length data
sammysun0711
committed
b9bf68e

Commits on Nov 12, 2025

Algin beam search parameters, set model default precision as FP16
sammysun0711
committed
e91314d
Bugfix for xFormers attention backend
sammysun0711
committed
f63eeb4

Commits on Nov 17, 2025