Fix illegal memory access with multi_tensor_apply size above INT_MAX #1825
+1
−1
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Currently, multi_tensor_apply causes an illegal memory access due to an overflow in the `size` field of `TensorListMetadata`. This can be reproduced using the following standalone script: ```python import torch, amp_C from apex.multi_tensor_apply import multi_tensor_applier multi_tensor_adam = amp_C.multi_tensor_adam size = 2**32+1 g_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')] p_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')] m_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')] v_32 = [torch.zeros(size, dtype=torch.float32, device='cuda')] _dummy_overflow_buf = torch.zeros(1, dtype=torch.int32, device='cuda') multi_tensor_applier(multi_tensor_adam, _dummy_overflow_buf, [g_32, p_32, m_32, v_32], 0.0, 0.9, 0.95, 1e-08, 1, 1, 1, 0.1) print(g_32) ```
|
cc: @crcrpar are the following out of date: apex/csrc/multi_tensor_apply.cuh Lines 15 to 17 in b3bd26a I see the same limits in PyTorch where you already updated to use int64_t in pytorch/pytorch#101760. Otherwise, I would expect that changing to use int64_t increases the TensorListMetadata struct size and hence the kernel arg size.
(Though, it seems that CUDA 12.1 on Volta+ increased the kernel arg size limit from 4 KB to 32 KB.) |
Collaborator
Yes, but apex does not have multi-tensor-apply with a list of scalars so we might be able to dodge a tweak of depth_to_max_tensors and depth_to_max_blocks |
crcrpar
approved these changes
Aug 17, 2024
Collaborator
crcrpar
left a comment
crcrpar
left a comment
There was a problem hiding this comment.
excuse my delay, thank you
hwchen2017
pushed a commit
to deepspeedai/DeepSpeed
that referenced
this pull request
Oct 21, 2025
…7639) This change fixes an overflow issue in `TensorListMetadata` where the `sizes` array used `int` (32-bit signed integer). This caused incorrect behavior (e.g., no parameter updates) when handling tensor sizes exceeding `INT_MAX` (2^31 - 1). The change here is identical to NVIDIA/apex PR [#1825](NVIDIA/apex#1825) for `multi_tensor_apply.cuh`. For further details regarding this fix, please refer to issue [#7640](#7640). Signed-off-by: Wang Yan <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Currently, multi_tensor_apply causes an illegal memory access due to an overflow in the
sizesfield ofTensorListMetadata. This can be reproduced using the following standalone script: