perf: optimize GDN decode with SGLang fused recurrent kernel#727
Open
perf: optimize GDN decode with SGLang fused recurrent kernel#727
Conversation
Use SGLang's fused_sigmoid_gating_delta_rule_update for pure-decode batches, which fuses gating computation (sigmoid, softplus, dt_bias) with the delta rule state update into a single kernel call, avoiding materialization of intermediate g/beta tensors. Also fix conv_state transpose to be conditional — SGLang provides [slot, conv_dim, state_len] layout directly, while ModelRunner uses [slot, state_len, conv_dim]. Benchmark (MI308X, TP2, ISL=60381, OSL=132): Qwen3.5-27B bf16: | Version | TPOT(ms) | TTFT(ms) | Tput/GPU | |-------------------|----------|----------|----------| | before this patch | 15.0 | 12613 | 2078 | | + this patch | 14.5 | 12628 | 2079 | Qwen3.5-35B-A3B-FP8 (TP1, ISL=4096, OSL=2048, concurrency=384): | Version | TPOT(ms) | TTFT(ms) | Tput/GPU | |-------------------|----------|----------|----------| | before this patch | 133.3 | 32559 | 7722 | | + this patch | 129.4 | 32400 | 7937 | Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Contributor
There was a problem hiding this comment.
Pull request overview
This PR optimizes Gated Delta Net (GDN) attention for decode-only batches by leveraging SGLang’s fused sigmoid-gating + delta-rule update kernel to reduce intermediate tensor materialization and kernel launches, while also improving compatibility across different attention-metadata/cache layouts.
Changes:
- Adds an optional SGLang fused decode path (
fused_sigmoid_gating_delta_rule_update) and skips materializingg/betawhen that path is used. - Makes
conv_statetransposition conditional to support differing cache layouts (ModelRunner vs SGLang). - Accesses
gdn_metadataviagetattrto avoid errors when non-GDN attention metadata is used.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+25
to
+27
| sglang_fused_sigmoid_gating_delta_rule_update = importlib.import_module( | ||
| "sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent" | ||
| ).fused_sigmoid_gating_delta_rule_update |
Comment on lines
+191
to
+194
| # transpose below. SGLang already provides [slot, conv_dim, state_len], | ||
| # and the Triton kernel consumes the original conv_state strides directly. | ||
| if conv_state.size(1) != self.conv1d.weight.size(0): | ||
| # transpose for ModelRunner |
wanzhenchn
reviewed
May 9, 2026
| # and the Triton kernel consumes the original conv_state strides directly. | ||
| if conv_state.size(1) != self.conv1d.weight.size(0): | ||
| # transpose for ModelRunner | ||
| conv_state = conv_state.transpose(-1, -2) |
Contributor
There was a problem hiding this comment.
the fix about conv state layout here have been merged in PR #532, please rebase main for the changes.
wanzhenchn
reviewed
May 9, 2026
|
|
||
| try: | ||
| sglang_fused_sigmoid_gating_delta_rule_update = importlib.import_module( | ||
| "sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent" |
Contributor
There was a problem hiding this comment.
Given that this optimization is customized for the SGLang framework and for code maintainability, would it be better to implement it via inheritance in atom/plugin/sglang/attention_backend/attention_gdn.py?
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
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.
Summary
fused_sigmoid_gating_delta_rule_updatefor pure-decode batches in GDN attention, fusing gating computation (sigmoid, softplus, dt_bias) with the delta rule state update into a single kernel call[slot, conv_dim, state_len]layout directly, while ModelRunner uses[slot, state_len, conv_dim]getattrforgdn_metadataaccess to avoid AttributeError on non-GDN attention metadataBenchmark
Qwen3.5-27B bf16 (MI308X, TP2, ISL=60381, OSL=132):
Qwen3.5-35B-A3B-FP8 (MI308X, TP1, ISL=4096, OSL=2048, concurrency=384):
Files Changed
atom/model_ops/attention_gdn.py— fused decode path + conv_state fixTest plan
🤖 Generated with Claude Code