[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
+10 -4
View File
@@ -98,6 +98,7 @@ from vllm.transformers_utils.processor import (
)
from vllm.transformers_utils.utils import convert_model_repo_to_path
from vllm.utils.collection_utils import flatten_2d_lists
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from vllm.utils.torch_utils import set_default_torch_dtype
@@ -1548,7 +1549,8 @@ class MiniCPMV2_5(MiniCPMVBaseModel, SupportsLoRA):
all_pixel_values[i, ..., :L_item] = pixel_values_item
num_patches = tgt_sizes.prod(-1)
max_patches = num_patches.max().item()
with gpu_sync_allowed():
max_patches = num_patches.max().item()
assert isinstance(max_patches, int)
patch_attn_mask = torch.zeros((B, max_patches), dtype=torch.bool, device=device)
@@ -1640,7 +1642,9 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
all_pixel_values[i, ..., :L_item] = pixel_values_item
num_patches = tgt_sizes.prod(-1)
max_patches = num_patches.max().item()
# Needed as a Python int to size the mask below.
with gpu_sync_allowed():
max_patches = num_patches.max().item()
assert isinstance(max_patches, int)
patch_attn_mask = torch.zeros((B, max_patches), dtype=torch.bool, device=device)
@@ -1737,7 +1741,8 @@ class MiniCPMV4_0(MiniCPMVBaseModel, SupportsLoRA):
all_pixel_values[i, ..., :L_item] = pixel_values_item
num_patches = tgt_sizes.prod(-1)
max_patches = num_patches.max().item()
with gpu_sync_allowed():
max_patches = num_patches.max().item()
assert isinstance(max_patches, int)
patch_attn_mask = torch.zeros((B, max_patches), dtype=torch.bool, device=device)
@@ -1839,7 +1844,8 @@ class MiniCPMV4_5(MiniCPMVBaseModel, SupportsLoRA):
all_pixel_values[i, ..., :L_item] = pixel_values_item
num_patches = tgt_sizes.prod(-1)
max_patches = num_patches.max().item()
with gpu_sync_allowed():
max_patches = num_patches.max().item()
assert isinstance(max_patches, int)
patch_attn_mask = torch.zeros((B, max_patches), dtype=torch.bool, device=device)