[Enhancement] Enhance attention examples and fix bugs for several corner cases #51

Rachmanino · 2025-11-17T12:10:00Z

Keep rowmax to avoid numerical issues
Correctly fill acc_s for padded positions (rather than default value 0)
Remove legacy compile flags

gemini-code-assist · 2025-11-17T12:11:18Z

Summary of Changes

Hello @Rachmanino, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request enhances the attention examples by addressing numerical stability issues, correcting padding errors, and removing legacy compile flags. These changes improve the robustness and correctness of the attention mechanism in various corner cases.

Highlights

Numerical Stability: Ensures numerical stability by keeping rowmax to avoid potential issues in attention calculations.
Padding Correction: Corrects the filling of acc_s for padded positions, using -T.infinity(acc_s.dtype) instead of the default value of 0, which fixes bugs in corner cases.
Legacy Flag Removal: Removes legacy compile flags, streamlining the compilation process.
Code Modernization: Replaces the naming of internal functions to remove the underscore prefix, and modernizes the T.copy calls to include src=.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces several important enhancements and bug fixes related to attention mechanisms. Key improvements include better numerical stability by correctly updating scores_max across iterations, and accurate handling of padded positions in attention scores by initializing them with negative infinity. Additionally, legacy compile flags have been removed, and internal function names have been standardized by removing leading underscores for improved clarity. One minor inconsistency was noted regarding the removal of TL_ENABLE_FAST_MATH in sparse_mla.py compared to other files.

gemini-code-assist · 2025-11-17T12:14:48Z

top/kernels/flash_attn/fwd.py

+                        # We shall fill -inf for padded positions
+                        for i, j in T.Parallel(block_M, block_N):
+                            acc_s[i, j] = T.if_then_else(k * block_N + j >= seq_len, 
+                                            -T.infinity(acc_s.dtype), 0)


This change correctly initializes acc_s with -T.infinity(acc_s.dtype) for padded positions, preventing numerical issues and ensuring correct attention weight calculation. This directly addresses the PR description point "Correctly fill acc_s for padded positions".

gemini-code-assist · 2025-11-17T12:14:49Z

top/kernels/flash_decode/mha_decode.py

+                for i, j in T.Parallel(block_M, block_N):
+                    acc_s[i, j] = T.if_then_else(k * block_N + j >= seqlen_kv, 
+                                                -T.infinity(acc_s.dtype), 0)


Similar to the previous change, this correctly initializes acc_s for padded positions with negative infinity in the split MHA decode path, ensuring numerical stability.

gemini-code-assist · 2025-11-17T12:14:49Z

top/kernels/deepseek_mla/mla_decode.py

+                    for i, j in T.Parallel(block_H, block_N):
+                        acc_s[i, j] = T.if_then_else(kv_start + j >= seqlen_kv, 
+                                                -T.infinity(acc_s.dtype), 0)


Replacing T.clear(acc_s) with a conditional fill ensures that padded positions are correctly initialized with -T.infinity(acc_s.dtype). This is a critical fix for numerical correctness, as T.clear would default to zero, potentially leading to incorrect attention weights for padded elements. This directly addresses the PR description point "Correctly fill acc_s for padded positions".

gemini-code-assist · 2025-11-17T12:14:49Z

top/kernels/flash_decode/mha_decode.py

+                        for i, j in T.Parallel(block_M, block_N):
+                            acc_s[i, j] = T.if_then_else(k * block_N + j >= seqlen_kv, 
+                                            -T.infinity(acc_s.dtype), 0)


This change correctly initializes acc_s with -T.infinity(acc_s.dtype) for padded positions, preventing numerical issues and ensuring correct attention weight calculation. This directly addresses the PR description point "Correctly fill acc_s for padded positions".

gemini-code-assist · 2025-11-17T12:14:49Z

top/kernels/flash_attn/fwd.py

+                for i, j in T.Parallel(block_M, block_N):
+                    acc_s[i, j] = T.if_then_else(k * block_N + j >= seq_len, 
+                                    -T.infinity(acc_s.dtype), 0)


Similar to other forward passes, this correctly initializes acc_s for padded positions with negative infinity in pipelined GQA, ensuring numerical stability.

gemini-code-assist · 2025-11-17T12:14:50Z

top/kernels/flash_attn/fwd.py

                        order=[-1, 0, 3, 1, -1, 2],
                        stage=[-1, 0, 0, 1, -1, 1],
-                        group=[[0], [1, 2], [3, 4, 5, 6, 7, 8, 9, 10], [11], [12], [13]]):
+                        group=[[0], [1, 2], [3, 4, 5, 6, 7, 8, 9, 10, 11], [12], [13], [14]]):


The group list has been updated, likely reflecting changes in pipelining stages or memory access patterns due to other optimizations. This is a functional change that should be verified for correctness and performance impact.

gemini-code-assist · 2025-11-17T12:14:51Z

top/kernels/deepseek_mla/mla_decode.py

+    @tilelang.jit(out_idx=[6], pass_configs={tilelang.PassConfigKey.TL_ENABLE_FAST_MATH: True})
+    def mla_decode_func(block_H, block_N, num_split, num_stages, threads=128):

        VALID_BLOCK_H = min(block_H, kv_group_num)

        @T.macro


The removal of compile_flags is a good cleanup, aligning with the PR description to remove legacy flags. This simplifies the JIT decorator configuration.

gemini-code-assist · 2025-11-17T12:14:51Z

top/kernels/flash_attn/fwd.py

+    @tilelang.jit(out_idx=[3, 4], pass_configs={tilelang.PassConfigKey.TL_ENABLE_FAST_MATH: True})
+    def mha_fwd_func(block_M, block_N, num_stages, threads):

        @T.prim_func
-        def _mha_fwd_main(
+        def mha_fwd_main(
                Q: T.Tensor(shape, dtype),  # type: ignore


Removal of compile_flags and renaming of _mha_fwd_func to mha_fwd_func and _mha_fwd_main to mha_fwd_main are consistent with the PR's objectives.

gemini-code-assist · 2025-11-17T12:14:51Z

top/kernels/flash_attn/bwd.py

+    @tilelang.jit(out_idx=[7, 8], pass_configs={tilelang.PassConfigKey.TL_ENABLE_FAST_MATH: True})
+    def mha_bwd_func(block_M, block_N, num_stages, threads):

        @T.prim_func
-        def _mha_bwd_main(
+        def mha_bwd_main(
                Q: T.Tensor(shape, dtype),  # type: ignore
                K: T.Tensor(shape, dtype),  # type: ignore


Removal of compile_flags and renaming of _mha_bwd_func to mha_bwd_func and _mha_bwd_main to mha_bwd_main are consistent with the PR's objectives.

gemini-code-assist · 2025-11-17T12:14:51Z

top/kernels/gemm/gemm.py

+    @tilelang.jit(out_idx=[-1])
+    def gemm_func(block_M, block_N, block_K, threads, num_stages, enable_rasteration):


Removal of compile_flags and renaming of _gemm_func to gemm_func and _gemm_main to gemm_main are consistent with the PR's objectives.

[Enhancement] Enhance attention examples

cbd2cd0

Rachmanino marked this pull request as draft November 17, 2025 12:10

gemini-code-assist bot reviewed Nov 17, 2025

View reviewed changes

Rachmanino marked this pull request as ready for review November 18, 2025 02:29

		@tilelang.jit(out_idx=[-1])
		def gemm_func(block_M, block_N, block_K, threads, num_stages, enable_rasteration):

[Enhancement] Enhance attention examples and fix bugs for several corner cases #51

Are you sure you want to change the base?

[Enhancement] Enhance attention examples and fix bugs for several corner cases #51

Uh oh!

Conversation

Rachmanino commented Nov 17, 2025

Uh oh!

gemini-code-assist bot commented Nov 17, 2025

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Nov 17, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants