[CANN] fix flash-attn mask: use BOOL attenMask instead of F16 -Inf pseShift

[July-h5kf3]

Overview

FusedInferAttentionScoreV2 was receiving llama.cpp's additive F16 attention mask (-Inf for masked positions) via pseShiftOptional. Ascend documents pseShift as position encoding (F16/BFLOAT16) and attenMask as BOOL/INT8/UINT8.On long KV lengths (e.g. multi-image MiniCPM-o, Sk≈768), putting -Inf into pseShift produced NaNs in the fused op and garbled generation (? floods).

This change:

• converts the hard-mask part of src3 (-Inf) to a contiguous BOOL attenMask (True = do not attend) and passes it via attenMaskOptional
• keeps finite additive biases in pseShift (clamped so PSE never contains -Inf)
• when max_bias != 0 (ALiBi), multiplies the finite PSE by the ALiBi slopes

Additional information

• Ascend aclnnFusedInferAttentionScoreV2:
  • attenMaskOptional: BOOL / INT8 / UINT8
  • pseShiftOptional: F16 / BF16 position encoding / additive bias
• Related prior workaround: CANN AUTO flash-attn force-off in llama-context.cpp (can revisit in a follow-up after this lands)
• Same pse-as-mask pattern also exists in upstream ggml-org/llama.cpp

Test plan

• test-backend-ops -b CANN0 -o FLASH_ATTN_EXT -p "mask=1,sinks=0,max_bias=0" → 320/320 OK
• MiniCPM-o, multi-image, --flash-attn on: no NaN / no ? flood
• text-only + --flash-attn on (llama-cli --single-turn): OK
• Rebuild ggml-cann with this change

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - used AI to help isolate the CANN FA NaN (mask vs pseShift) and draft the attenMask conversion; I reviewed the final patch line-by-line and can explain / maintain it