Skip to content

update gated delta rule state layout#4636

Open
grimoire wants to merge 2 commits into
InternLM:mainfrom
grimoire:change-gdr-state-layout
Open

update gated delta rule state layout#4636
grimoire wants to merge 2 commits into
InternLM:mainfrom
grimoire:change-gdr-state-layout

Conversation

@grimoire
Copy link
Copy Markdown
Collaborator

Transposed state is more friendly than default layout.

Copilot AI review requested due to automatic review settings May 28, 2026 10:13
Copy link
Copy Markdown
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR adds support for a transposed recurrent-state memory layout in the CUDA fused gated-delta-rule kernel and updates Qwen3.5/Qwen3-Next integration to use that layout (intended to be more hardware-friendly than the previous default layout).

Changes:

  • Add transpose_state_layout support to the TileLang fused recurrent gated-delta-rule kernel, including shape validation and a specialized direct load/store path for [V, K] state layout.
  • Update Qwen3.5/Qwen3-Next recurrent state shapes and the GatedDelta call sites to use the transposed layout.
  • Extend kernel tests to cover transposed state, transposed state + state_indices, and transposed circular-buffer behavior; adjust one circular-buffer tolerance.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file
File Description
tests/pytorch/kernel/test_gated_delta_rule.py Adds coverage for transposed-state layout (including state_indices and circular buffer) and adjusts tolerance.
lmdeploy/pytorch/nn/gated_delta.py Always enables transposed-state layout when invoking gated-delta kernels.
lmdeploy/pytorch/kernels/cuda/gated_delta_rule.py Implements transposed-state layout path and input shape validation in the fused recurrent kernel.
lmdeploy/pytorch/configurations/qwen3_next.py Swaps recurrent state shape to [V, K] for Qwen3-Next.
lmdeploy/pytorch/configurations/qwen3_5.py Swaps recurrent state shape to [V, K] for Qwen3.5 (including spec-decoding shape).
lmdeploy/pytorch/backends/gated_delta_rule.py Extends backend API to accept transpose_state_layout (and cache_seqlens for fused recurrent).
lmdeploy/pytorch/backends/cuda/gated_delta_rule.py Threads transpose_state_layout through CUDA backend calls.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment on lines 188 to 200
core_attn_out, last_state = self.chunk_func(
q,
k,
v,
g=g,
beta=beta,
scale=scale,
initial_state=init_state,
output_final_state=output_final_state,
use_qk_l2norm_in_kernel=False,
cu_seqlens=cu_seqlens,
transpose_state_layout=transpose_state_layout,
)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants