[jhen0409]

Overview

Related to my finding in #30218, test-backend-ops -o FLASH_ATTN_EXT was 2541/2604 (63 failed on sinks=1) on single NPU, 2549/2604 on row split.

In this PR:

• QK output scales: from 16-bit splat / bias = scale to 32-bit splat / bias = 0
• sinks: l += 2^(sink * log2e - m), move to fa_build_d_diag_inv_l

Additional information

test-backend-ops -o FLASH_ATTN_EXT (IQ-9075)

config	master	patch
single NPU (HTP0:0), passed	2541/2604	2604/2604
single NPU, failures with sinks=1	63	0
row split (HTP0[0-1]), passed	2549/2604	2604/2604
row split, failures with sinks=1	55	0

No performance regression found. Also tested on gpt-oss 20b mxfp4 (IQ-9075 with NPU, 8-session layer split), prefill KLD vs CPU 0.0410 -> 0.0293.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, Claude Code. The agent also wrote a temp code to verify the NMSE result. I've reviewed and simplified the code.

[max-krasnyansky]

Nice. I've been meaning to revisit those accuracy mismatches with sinks. The HVX FA is passing all tests cleanly (you can force it with GGML_HEXAGON_FA_SELECT=1) but the HMX always had a slight mismatch ever since I added the sinks (specifically for better GPT-OSS-20B support). Will check it out on my setups asap.

[max-krasnyansky]

Vectorized sinks logic and removed register spills.
Looks good to me. And we're passing all FA tests now.
@lhez for another ack