xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-08 14:28:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix incorrect original shape in hashing (#23672)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: Lukas Geiger <lukas.geiger94@gmail.com>
This commit is contained in:
Cyrus Leung 2025-08-27 03:01:25 +08:00 committed by GitHub
parent 98aa16ff41
commit 9715f7bb0f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 12 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
tests/multimodal/test_hasher.py
View File

@ -45,10 +45,11 @@ def test_hash_collision_image_transpose():
assert hasher.hash_kwargs(image=image1) != hasher.hash_kwargs(image=image2) assert hasher.hash_kwargs(image=image1) != hasher.hash_kwargs(image=image2)
def test_hash_collision_tensor_shape(): @pytest.mark.parametrize("dtype", [torch.float32, torch.bfloat16])
def test_hash_collision_tensor_shape(dtype):
# The hash should be different though the data is the same when flattened # The hash should be different though the data is the same when flattened
arr1 = torch.zeros((5, 10, 20, 3)) arr1 = torch.zeros((5, 10, 20, 3), dtype=dtype)
arr2 = torch.zeros((10, 20, 5, 3)) arr2 = torch.zeros((10, 20, 5, 3), dtype=dtype)
hasher = MultiModalHasher hasher = MultiModalHasher
assert hasher.hash_kwargs(data=arr1) != hasher.hash_kwargs(data=arr2) assert hasher.hash_kwargs(data=arr1) != hasher.hash_kwargs(data=arr2)

10
vllm/multimodal/hasher.py
View File

@ -45,16 +45,22 @@ class MultiModalHasher:
if isinstance(obj, torch.Tensor): if isinstance(obj, torch.Tensor):
tensor_obj: torch.Tensor = obj.cpu() tensor_obj: torch.Tensor = obj.cpu()
tensor_dtype = tensor_obj.dtype tensor_dtype = tensor_obj.dtype
tensor_shape = tensor_obj.shape
# NumPy does not support bfloat16.
# Workaround: View the tensor as a contiguous 1D array of bytes
if tensor_dtype == torch.bfloat16: if tensor_dtype == torch.bfloat16:
tensor_obj = tensor_obj.contiguous() tensor_obj = tensor_obj.contiguous()
tensor_obj = tensor_obj.view( tensor_obj = tensor_obj.view(
(tensor_obj.numel(), )).view(torch.uint8) (tensor_obj.numel(), )).view(torch.uint8)
return cls.item_to_bytes( return cls.item_to_bytes(
"tensor", { "tensor", {
"original_dtype": str(tensor_dtype), "original_dtype": str(tensor_dtype),
"original_shape": tuple(tensor_obj.shape), "original_shape": tuple(tensor_shape),
"data": tensor_obj.numpy() "data": tensor_obj.numpy(),
}) })
return cls.item_to_bytes("tensor", tensor_obj.numpy()) return cls.item_to_bytes("tensor", tensor_obj.numpy())
if isinstance(obj, np.ndarray): if isinstance(obj, np.ndarray):
# If the array is non-contiguous, we need to copy it first # If the array is non-contiguous, we need to copy it first