[Gluon] Implement attention kernels for d64 and d128 #7009
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
784d43d to
2096971
Compare
b8cde4d to
9fed043
Compare
97ed0fd to
984460f
Compare
40eea56 to
e98c5ff
Compare
peterbell10
reviewed
Jun 13, 2025
| self.ready_bars = ready_bars | ||
| self.empty_bars = empty_bars | ||
| self.num_buffers = gl.constexpr(num_buffers) | ||
| self.num_consumers = gl.constexpr(num_consumers) |
Contributor
There was a problem hiding this comment.
I wonder, would an annotated assignment work here?
Collaborator
Author
There was a problem hiding this comment.
No because this is being executed as Python code :/
Collaborator
Author
There was a problem hiding this comment.
Would love ideas on how to improve the syntax here. There are a couple of other places where explicit wrapping of values with gl.constexpr is needed, even inside Gluon code
| @gluon.jit | ||
| def release(self): | ||
| if isinstance(self.mem, gl.shared_memory_descriptor): | ||
| self.mem._keep_alive() |
Contributor
There was a problem hiding this comment.
Do we need to add _keep_alive for tensor memory descriptors as well?
Collaborator
Author
There was a problem hiding this comment.
There is no corresponding operation for it, but thinking about it, we might need one
```
fused-attention-batch4-head32-d64-fwd-causal=True-config=Config(warp_specialize=True, use_gluon=False):
N_CTX Triton [FP16] Triton [FP8]
0 1024.0 206.903414 226.941886
1 2048.0 397.637820 397.894749
2 4096.0 538.967275 537.617627
3 8192.0 633.806800 630.562116
4 16384.0 690.005849 685.848451
fused-attention-batch4-head32-d64-fwd-causal=True-config=Config(warp_specialize=False, use_gluon=True):
N_CTX Triton [FP16] Triton [FP8]
0 1024.0 158.854438 167.127676
1 2048.0 430.994165 445.445875
2 4096.0 574.152492 593.125123
3 8192.0 669.977292 692.156473
4 16384.0 725.290898 749.413157
```
```
fused-attention-batch4-head32-d128-fwd-causal=True-config=Config(warp_specialize=True, use_gluon=False):
N_CTX Triton [FP16] Triton [FP8]
0 1024.0 168.101257 180.147833
1 2048.0 232.412164 251.379866
2 4096.0 282.099851 306.706495
3 8192.0 313.280068 342.515038
4 16384.0 331.734557 362.578482
fused-attention-batch4-head32-d128-fwd-causal=True-config=Config(warp_specialize=False, use_gluon=True):
N_CTX Triton [FP16] Triton [FP8]
0 1024.0 153.825955 303.351930
1 2048.0 504.991716 532.393691
2 4096.0 723.003324 750.103002
3 8192.0 888.283770 924.001097
4 16384.0 979.523873 1045.632286
```
Collaborator
Author
Final perf numbers for this. The gap for D64 is quite big now since cublas appears to have improved perf a lot from 720->920. For D128, most of the gap esp for small ctx is making the kernel persistent. |
tie-pilot-qxw
pushed a commit
to tie-pilot-qxw/triton
that referenced
this pull request
Aug 30, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
No description provided.