[CPU] Ship triton-cpu wheel and fix several hardcoded pin_memory=True (#52092)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2026-08-13 14:40:53 +08:00
committed by GitHub
parent 89c8401c8a
commit 373592ef57
8 changed files with 50 additions and 41 deletions
+3 -1
View File
@@ -61,6 +61,7 @@ from vllm.sequence import IntermediateTensors
from vllm.tokenizers import cached_tokenizer_from_config
from vllm.transformers_utils.processor import cached_processor_from_config
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import PIN_MEMORY
from .blip2 import Blip2QFormerModel
from .interfaces import (
@@ -732,7 +733,8 @@ class GraniteSpeechForConditionalGeneration(
target_device = self.encoder.input_linear.weight.device
if target_device == input_features_mask.device:
return input_features_mask
return input_features_mask.pin_memory().to(target_device, non_blocking=True)
masked = input_features_mask.pin_memory() if PIN_MEMORY else input_features_mask
return masked.to(target_device, non_blocking=True)
def _pad_and_stack_input_features(
self,
+2 -1
View File
@@ -112,6 +112,7 @@ from vllm.tokenizers.registry import cached_tokenizer_from_config
from vllm.triton_utils import HAS_TRITON, tl, triton
from vllm.utils.collection_utils import is_list_of
from vllm.utils.math_utils import round_up
from vllm.utils.torch_utils import PIN_MEMORY
from vllm.v1.worker.encoder_cudagraph_defs import EncoderCudaGraphReplayBuffers
from ...utils.torch_utils import async_tensor_h2d
@@ -709,7 +710,7 @@ class Qwen3_VisionTransformer(nn.Module):
pinned = torch.empty(
(num_pos, pos_ids[0].shape[1]),
dtype=pos_ids[0].dtype,
pin_memory=True,
pin_memory=PIN_MEMORY,
)
pos_ids = torch.cat(pos_ids, dim=0, out=pinned).to(
self.device, non_blocking=True