[Hardware][XPU] Register matmul and linear batch-invariant kernels for XPU (#49209)

Signed-off-by: tzielinski-habana <tomasz.zielinski@intel.com>
Signed-off-by: Tomasz Zielinski <85164140+tzielinski-habana@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Tomasz Zielinski
2026-09-02 13:31:38 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Claude Opus 5 Harry Mellor
parent 7894394b07
commit ffe3bb3c72
7 changed files with 272 additions and 33 deletions
+13
View File
@@ -2435,6 +2435,19 @@ class EngineArgs:
self.attention_backend
)
# Batch-invariant mode requires deterministic attention behavior.
# If no backend is explicitly requested, prefer Triton Attention.
if (
envs.VLLM_BATCH_INVARIANT
and attention_config.backend is None
and current_platform.is_xpu()
):
attention_config.backend = AttentionBackendEnum.TRITON_ATTN
logger.info(
"VLLM_BATCH_INVARIANT is enabled and no attention backend was "
"specified; defaulting to TRITON_ATTN."
)
# TurboQuant requires FlashAttention 2 — FA3 boundary layers assert
# FlashAttentionImpl which fails with TurboQuantAttentionImpl.
if resolved_cache_dtype.startswith("turboquant_") and (