[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
+29 -25
View File
@@ -115,6 +115,7 @@ 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.gpu_sync_debug import gpu_sync_allowed
from ...utils.torch_utils import async_tensor_h2d
from .interfaces import (
MultiModalEmbeddings,
@@ -2368,19 +2369,17 @@ class Qwen3VLForConditionalGeneration(
spatial_merge_size=self.visual.spatial_merge_size,
q=self.video_pruning_rate,
)
# Apply retention mask.
emb = emb[retention_mask]
# Calculate the actual number of retained tokens per frame.
num_frames, rows, cols = (
t,
h // merge_size,
w // merge_size,
)
retention_mask_thw = retention_mask.reshape(num_frames, rows, cols)
num_tokens_per_frame = (
retention_mask_thw.sum(dim=(1, 2)).long().tolist()
)
with gpu_sync_allowed():
# Apply retention mask.
emb = emb[retention_mask]
# Calculate the actual number of retained tokens per frame.
num_frames, rows, cols = t, h // merge_size, w // merge_size
retention_mask_thw = retention_mask.reshape(num_frames, rows, cols)
num_tokens_per_frame = (
retention_mask_thw.sum(dim=(1, 2)).long().tolist()
)
else:
feature_size = emb.shape[0] // num_frames
num_tokens_per_frame = [feature_size] * num_frames
@@ -2548,10 +2547,14 @@ class Qwen3VLForConditionalGeneration(
.permute(1, 0)
)
full_is_video_embed = unpruned_token_ids_tensor == embed_token_id
expanded_positions[is_video_embed, :3] = original_mrope[full_is_video_embed][
retention_mask
]
expanded_positions[~is_video_embed, :3] = original_mrope[~full_is_video_embed]
with gpu_sync_allowed():
expanded_positions[is_video_embed, :3] = original_mrope[
full_is_video_embed
][retention_mask]
expanded_positions[~is_video_embed, :3] = original_mrope[
~full_is_video_embed
]
expanded_positions[..., 3] = is_vision_start
expanded_positions[..., 4] = is_video_embed
@@ -2814,15 +2817,16 @@ class Qwen3VLForConditionalGeneration(
torch.empty(5, 0, device=device, dtype=torch.long)
)
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