[Core] Check for GPU<->CPU syncs during CI (#43107)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nick Hill
2026-08-14 17:42:05 -07:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 7b544ecb52
commit 615d4cfade
44 changed files with 600 additions and 357 deletions
+14 -11
View File
@@ -87,6 +87,7 @@ from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.worker.encoder_cudagraph_defs import EncoderCudaGraphReplayBuffers
from ...utils.gpu_sync_debug import gpu_sync_allowed
from .interfaces import (
MultiModalEmbeddings,
SupportsEagle,
@@ -1569,8 +1570,9 @@ class Qwen2_5_VLForConditionalGeneration(
video_second_per_grid=video_second_per_grid_t.item(),
).to(emb.device, non_blocking=True)
emb = emb[retention_mask]
positions = positions[retention_mask]
with gpu_sync_allowed():
emb = emb[retention_mask]
positions = positions[retention_mask]
emb = torch.cat([emb, positions], dim=1)
video_embeds_out.append(emb)
return tuple(video_embeds_out)
@@ -1624,15 +1626,16 @@ class Qwen2_5_VLForConditionalGeneration(
mm[:, -4:].permute(1, 0).long() for mm in multimodal_embeddings
]
positions, mrope_positions_delta = recompute_mrope_positions(
input_ids_t,
mm_embeddings_pos,
mrope_positions,
num_computed_tokens,
vision_start_token_id,
image_token_id,
video_token_id,
)
with gpu_sync_allowed():
positions, mrope_positions_delta = recompute_mrope_positions(
input_ids_t,
mm_embeddings_pos,
mrope_positions,
num_computed_tokens,
vision_start_token_id,
image_token_id,
video_token_id,
)
return mm_embeddings_out, positions, mrope_positions_delta