[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:
co-authored by
Claude Opus 4.8
parent
7b544ecb52
commit
615d4cfade
+5
-6
@@ -9,6 +9,7 @@
|
||||
# ruff: noqa: E501
|
||||
import torch
|
||||
|
||||
from vllm.utils.gpu_sync_debug import gpu_sync_allowed
|
||||
from vllm.triton_utils import triton
|
||||
|
||||
from .utils import tensor_cache
|
||||
@@ -21,12 +22,10 @@ def prepare_lens(cu_seqlens: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
@tensor_cache
|
||||
def prepare_chunk_indices(cu_seqlens: torch.Tensor, chunk_size: int) -> torch.Tensor:
|
||||
indices = torch.cat(
|
||||
[
|
||||
torch.arange(n)
|
||||
for n in triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()
|
||||
]
|
||||
)
|
||||
# This will be fixed by https://github.com/vllm-project/vllm/pull/51540.
|
||||
with gpu_sync_allowed():
|
||||
chunk_counts = triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()
|
||||
indices = torch.cat([torch.arange(n) for n in chunk_counts])
|
||||
chunk_indices = torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1)
|
||||
return chunk_indices.to(
|
||||
device=cu_seqlens.device, dtype=cu_seqlens.dtype, non_blocking=True
|
||||
|
||||
Reference in New Issue
Block a user