[Relax][TIR] Introduce new cumsum op for gpu
#16934
Merged
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
Member
Author
tqchen
approved these changes
Apr 26, 2024
Lunderberg
requested changes
Apr 26, 2024
| size = (8, 2000) | ||
| np_data = np.random.randint(0, 10, size).astype("int32") | ||
| np_cumsum = np.cumsum(np_data, axis=-1) | ||
| for target in ["cuda", "vulkan -supports_int64=1"]: |
Contributor
There was a problem hiding this comment.
Nitpick: Use @tvm.testing.parametrize_targets("cuda", "vulkan -supports_int64=1") instead of looping over each target. This performs each test case in a separate pytest environment,
- Exercises each test in a separate pytest case. Can distinguish between failure on one specific backend as opposed to failure on every backend.
- Applies the appropriate
@tvm.testing.requires_*marks for each target. Currently, this test would fail if a developer runs it withset(USE_CUDA ON)andset(USE_VULKAN OFF).
@tvm.testing.parametrize_targets("cuda", "vulkan -supports_int64=1")
def test_dispatch_cumsum_gpu(target, dev):
...| shape = call.struct_info.shape | ||
| kwargs = {} | ||
| if ( | ||
| (axis == -1 or axis == len(shape) - 1) |
Contributor
There was a problem hiding this comment.
For tensors of unknown shape, the shape field is none. Instead of len(call.struct_info.shape), can we use call.struct_info.ndim? (Alternatively, since it looks like the implementation requires an explicit shape in order to apply a reshape, we could add shape is not None to this condition.)
Member
Author
There was a problem hiding this comment.
Thanks for the caching. Unfortunately, the original implementation does not support unknown shape. I added a check in the pass
Contributor
|
Looks like I took too long to review. I think the changes requested should probably be made in a follow-up PR. |
Hzfengsy
pushed a commit
to Hzfengsy/tvm
that referenced
this pull request
Apr 28, 2024
The current pass `ForceNarrowIndexToI32` fails to narrow dtype for let binding. This PR fixes the issue. BTW, this PR addresses the comments in apache#16934
Hzfengsy
pushed a commit
to Hzfengsy/tvm
that referenced
this pull request
Apr 29, 2024
The current pass `ForceNarrowIndexToI32` fails to narrow dtype for let binding. This PR fixes the issue. BTW, this PR addresses the comments in apache#16934
Hzfengsy
pushed a commit
to Hzfengsy/tvm
that referenced
this pull request
May 1, 2024
The current pass `ForceNarrowIndexToI32` fails to narrow dtype for let binding. This PR fixes the issue. BTW, this PR addresses the comments in apache#16934
Hzfengsy
pushed a commit
to Hzfengsy/tvm
that referenced
this pull request
May 5, 2024
The current pass `ForceNarrowIndexToI32` fails to narrow dtype for let binding. This PR fixes the issue. BTW, this PR addresses the comments in apache#16934
tqchen
pushed a commit
that referenced
this pull request
May 6, 2024
The current pass `ForceNarrowIndexToI32` fails to narrow dtype for let binding. This PR fixes the issue. BTW, this PR addresses the comments in #16934
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.
New implementation for
cumsumforaxis=-1, which is the common case in LLM sampling and MoE inference)Tested CUDA, Vulkan on RTX 3080, and Metal on Apple M1 Pro. The new implementation is faster or equal vendor-libraries on small size (
<128k), and nearly 10x faster than topi implementation in all cases