Conversation
added 3 commits
July 26, 2026 17:51
Signed-off-by: xiaogang.zhou <xiaogang.zhou@bytedance.com>
Signed-off-by: xiaogang.zhou <xiaogang.zhou@bytedance.com>
Signed-off-by: xiaogang.zhou <xiaogang.zhou@bytedance.com>
zhougit86
requested review from
aleozlx,
bkryu,
cyx-6,
dhiraj113,
jiahanc,
jimmyzho,
kahyunnam,
nv-yunzheq,
qiching,
qsang-nv,
saltyminty,
samuellees,
sricketts,
yongwww,
yyihuang and
yzh119
as code owners
July 26, 2026 10:07
Contributor
|
Caution The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased. |
Contributor
📝 WalkthroughWalkthroughSM120 NVFP4 attention now supports grouped-query attention by tracking query and key/value head counts separately through preprocessing, validation, quantization, CUDA parameter packing, KV tile loading, and accuracy tests. ChangesGrouped-query attention support
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant QKV as QKV tensors
participant Python as nvfp4_attention_sm120_quantize_qkv
participant Binding as nvfp4_attention_sm120_fwd
participant Mainloop as CollectiveMainloopFwd
participant KVLoaders as KLoader and VLoader
QKV->>Python: q, k, v with separate head counts
Python->>Binding: quantized tensors, scales, qk_correction
Binding->>Mainloop: h, h_k, h_h_k_ratio
Mainloop->>KVLoaders: load KV tiles using bidkv
KVLoaders-->>Binding: attention output and lse
Possibly related PRs
Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
tests/attention/test_nvfp4_attention_sm120.py (1)
194-202: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick winGQA test config doesn't exercise multi-KV-head remapping.
With
num_kv_heads=1,h_h_k_ratio = num_q_heads // num_kv_heads = 4, sobidkv = bidh // h_h_k_ratiois0for every validbidh(0..3). This effectively tests MQA (single shared KV head), not general GQA — the newbidkvindexing logic inmainloop.cuh/load_k.cuh/load_v.cuhforbidkv > 0(multiple distinct KV head groups) is left unverified by this test.✅ Suggested config change to exercise multiple KV-head groups
- batch, num_q_heads, num_kv_heads, seq_len, head_dim = 1, 4, 1, 256, 128 + batch, num_q_heads, num_kv_heads, seq_len, head_dim = 1, 6, 2, 256, 128🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@tests/attention/test_nvfp4_attention_sm120.py` around lines 194 - 202, Update the test configuration around the manual seed and tensor setup so num_kv_heads is greater than 1 and still divides num_q_heads, such as selecting two KV heads for four query heads. Preserve the existing tensor shapes and dtypes while ensuring valid query heads map to both bidkv 0 and bidkv 1, exercising multi-KV-head GQA remapping.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@flashinfer/nvfp4_attention_sm120.py`:
- Around line 377-379: Clarify the docstring describing q_fp4 and k_fp4 layouts
so q_fp4 uses num_q_heads while k_fp4 uses num_kv_heads, without grouping them
under a shared num_q_heads shape. Preserve the existing v_fp4_t and v_scale_t
transposed layout description.
---
Nitpick comments:
In `@tests/attention/test_nvfp4_attention_sm120.py`:
- Around line 194-202: Update the test configuration around the manual seed and
tensor setup so num_kv_heads is greater than 1 and still divides num_q_heads,
such as selecting two KV heads for four query heads. Preserve the existing
tensor shapes and dtypes while ensuring valid query heads map to both bidkv 0
and bidkv 1, exercising multi-KV-head GQA remapping.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 496e80b1-db38-4a5f-aa77-27edc0699601
📒 Files selected for processing (7)
csrc/nvfp4_attention_sm120/nvfp4_attention_sm120_binding.cuflashinfer/nvfp4_attention_sm120.pyinclude/flashinfer/attention/sm120/nvfp4_attention_sm120/api/launcher.hinclude/flashinfer/attention/sm120/nvfp4_attention_sm120/compute/mainloop.cuhinclude/flashinfer/attention/sm120/nvfp4_attention_sm120/compute/producer/load_k.cuhinclude/flashinfer/attention/sm120/nvfp4_attention_sm120/compute/producer/load_v.cuhtests/attention/test_nvfp4_attention_sm120.py
Comment on lines
+377
to
+379
| ``[batch, num_q_heads, seq_len, head_dim / 2]``; ``k_fp4`` uses | ||
| ``num_kv_heads``. ``v_fp4_t`` and ``v_scale_t`` are stored transposed as | ||
| ``[batch, num_kv_heads, head_dim, packed_seq_len]``. |
Contributor
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win
Confusing docstring wording for q_fp4/k_fp4 layout.
The sentence says "q_fp4 and k_fp4 use layout [batch, num_q_heads, ...]; k_fp4 uses num_kv_heads" — it groups k_fp4 with the num_q_heads layout and then immediately contradicts that in the next clause.
✏️ Suggested docstring fix
The packed tensors should be produced by
- :func:`nvfp4_attention_sm120_quantize_qkv`. ``q_fp4`` and ``k_fp4`` use layout
- ``[batch, num_q_heads, seq_len, head_dim / 2]``; ``k_fp4`` uses
- ``num_kv_heads``. ``v_fp4_t`` and ``v_scale_t`` are stored transposed as
+ :func:`nvfp4_attention_sm120_quantize_qkv`. ``q_fp4`` uses layout
+ ``[batch, num_q_heads, seq_len, head_dim / 2]``; ``k_fp4`` uses the same
+ layout with ``num_kv_heads`` in place of ``num_q_heads``. ``v_fp4_t`` and
+ ``v_scale_t`` are stored transposed as
``[batch, num_kv_heads, head_dim, packed_seq_len]``.📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| ``[batch, num_q_heads, seq_len, head_dim / 2]``; ``k_fp4`` uses | |
| ``num_kv_heads``. ``v_fp4_t`` and ``v_scale_t`` are stored transposed as | |
| ``[batch, num_kv_heads, head_dim, packed_seq_len]``. | |
| The packed tensors should be produced by | |
| :func:`nvfp4_attention_sm120_quantize_qkv`. ``q_fp4`` uses layout | |
| ``[batch, num_q_heads, seq_len, head_dim / 2]``; ``k_fp4`` uses the same | |
| layout with ``num_kv_heads`` in place of ``num_q_heads``. ``v_fp4_t`` and | |
| ``v_scale_t`` are stored transposed as | |
| ``[batch, num_kv_heads, head_dim, packed_seq_len]``. |
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@flashinfer/nvfp4_attention_sm120.py` around lines 377 - 379, Clarify the
docstring describing q_fp4 and k_fp4 layouts so q_fp4 uses num_q_heads while
k_fp4 uses num_kv_heads, without grouping them under a shared num_q_heads shape.
Preserve the existing v_fp4_t and v_scale_t transposed layout description.
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
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.
📌 Description
Add GQA support to the SM120/SM121 NVFP4 attention path.
This PR allows
qto usenum_q_headswhilek/vusenum_kv_heads, withnum_q_heads % num_kv_heads == 0. The kernel now maps each Q head to its shared KV head usingq_head / h_h_k_ratio, while keeping outputs, LSE, Q scale, andqk_correctionaligned with Q heads.🔍 Related Issues
Related to #3809
🚀 Pull Request Checklist
Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete.
✅ Pre-commit Checks
pre-commitby runningpip install pre-commit(or used your preferred method).pre-commit install.pre-commit run --all-filesand fixed any reported issues.🧪 Tests
unittest, etc.).Added a GQA coverage case for
nvfp4_attention_sm120.Reviewer Notes
The main logic to review is the Q-head to KV-head mapping:
q,q_scale,qk_correction,out, andlsestay aligned with Q heads;k,v,k_scale, andv_scaleare aligned with KV heads.Summary by CodeRabbit
New Features
Bug Fixes
Tests