From 44fe2a392b71d52a8d72faf2f8278834379482c9 Mon Sep 17 00:00:00 2001 From: LironKesem <52330564+LironKesem@users.noreply.github.com> Date: Mon, 31 Aug 2026 00:20:37 -0400 Subject: [PATCH] [CPU] add CPU support for Voxtral (#53921) Signed-off-by: Liron Kesem <52330564+LironKesem@users.noreply.github.com> --- vllm/model_executor/models/whisper_causal.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/whisper_causal.py b/vllm/model_executor/models/whisper_causal.py index 93f2aa1741..d76e9b0d4a 100644 --- a/vllm/model_executor/models/whisper_causal.py +++ b/vllm/model_executor/models/whisper_causal.py @@ -32,6 +32,7 @@ from vllm.v1.attention.backend import ( CommonAttentionMetadata, subclass_attention_backend_with_overrides, ) +from vllm.v1.attention.backends.cpu_attn import CPUAttentionBackend from vllm.v1.attention.backends.flash_attn import FlashAttentionBackend try: @@ -225,6 +226,7 @@ def create_whisper_attention_backend_with_block_pooling( b for b in ( AiterFlashAttentionBackend, + CPUAttentionBackend, FlashAttentionBackend, RocmAttentionBackend, TritonAttentionBackend, @@ -239,7 +241,9 @@ def create_whisper_attention_backend_with_block_pooling( "appreciated." ) - if not issubclass(underlying_attn_backend, FlashAttentionBackend): + if not issubclass( + underlying_attn_backend, (CPUAttentionBackend, FlashAttentionBackend) + ): logger.info( "Using %s for Whisper causal attention with block pooling. " "This backend was recently enabled for this model. "