A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[XPU] Whisper model support on XPU Platform (#25123)
Signed-off-by: chzhang <chaojun.zhang@intel.com>
Chaojun Zhang committed
Sep 18, 2025 at 12:30 UTC
3bc18127ff1c644257abcf84a1a56fab8c0d3f0c
2 files changed
+3
-3
vllm/attention/layer.py
+2
-2
@@ -391,8 +391,8 @@ class MultiHeadAttention(nn.Module):
391
backend = _Backend.FLASH_ATTN
392
use_upstream_fa = True
393
394
- if current_platform.is_rocm():
395
- # currently, only torch_sdpa is supported on rocm
394
+ if current_platform.is_rocm() or current_platform.is_xpu():
395
+ # currently, only torch_sdpa is supported on rocm/xpu
396
self.attn_backend = _Backend.TORCH_SDPA
397
else:
398
vllm/v1/worker/utils.py
+1
-1
@@ -282,7 +282,7 @@ def bind_kv_cache(
282
# TODO - analyze where runner_kv_caches is used and the right
283
# way to ensure it properly reflects multiple attention layers
284
# in the same decoder block.
285
- if current_platform.is_cuda():
285
+ if current_platform.is_cuda() or current_platform.is_xpu():
286
# We know that the GPU runner is not impacted by this
287
# case. Some test code depends on runner_kv_caches, but
288
# not in a way that's impacted by ignoring this.