[CPU] add CPU support for Voxtral (#53921)
Signed-off-by: Liron Kesem <52330564+LironKesem@users.noreply.github.com>
LironKesem committed
Aug 31, 2026 at 00:20 UTC
44fe2a392b71d52a8d72faf2f8278834379482c9
1 file changed
+5
-1
vllm/model_executor/models/whisper_causal.py
+5
-1
@@ -32,6 +32,7 @@ from vllm.v1.attention.backend import (
32
CommonAttentionMetadata,
33
subclass_attention_backend_with_overrides,
34
)
35
+from vllm.v1.attention.backends.cpu_attn import CPUAttentionBackend
36
from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend
37
38
try:
@@ -225,6 +226,7 @@ def create_whisper_attention_backend_with_block_pooling(
226
b
227
for b in (
228
AiterFlashAttentionBackend,
229
+ CPUAttentionBackend,
230
FlashAttentionBackend,
231
RocmAttentionBackend,
232
TritonAttentionBackend,
@@ -239,7 +241,9 @@ def create_whisper_attention_backend_with_block_pooling(
241
"appreciated."
242
)
243
242
- if not issubclass(underlying_attn_backend, FlashAttentionBackend):
244
+ if not issubclass(
245
+ underlying_attn_backend, (CPUAttentionBackend, FlashAttentionBackend)
246
+ ):
247
logger.info(
248
"Using %s for Whisper causal attention with block pooling. "
249
"This backend was recently enabled for this model. "