Skip to content

[Refactor][TCPCG] Standardize diffusion eager graph wrappers (7/9) - #42291

Open
Oasis-Git wants to merge 3 commits into
sgl-project:mainfrom
Oasis-Git:refactor/tcpcg-07-diffusion-eager
Open

Oasis-Git wants to merge 3 commits into
sgl-project:mainfrom
Oasis-Git:refactor/tcpcg-07-diffusion-eager

Conversation

@Oasis-Git

@Oasis-Git Oasis-Git commented Oct 2, 2026 •

Copy link
Copy Markdown
Collaborator

This is step 7/9, split from the original TCPCG removal PR #41634.

Depends on #42286 (2/9); merge those dependencies first. This branch targets upstream main, so its diff includes unmerged prerequisites. Review this step alone using the isolated step-7 diff.

Series

  1. Relocate shared graph tensor and DSA head-gate helpers — [Refactor][TCPCG] Relocate shared graph tensor and DSA head-gate helpers (1/9) #42285.
  2. Add explicit batch support to BCG eager regions — [Refactor][TCPCG] Bind live forward batches in BCG eager calls (2/9) #42286.
  3. Retire TCPCG backend selection, configuration, documentation, and tests — [Refactor][TCPCG] Retire backend selection and configuration (3/9) #42287.
  4. Consolidate Radix attention eager regions and isolate Inkling-specific behavior — [Refactor][TCPCG] Consolidate attention eager regions and isolate Inkling policy (4/9) #42288.
  5. Migrate DSA and DeepSeek eager regions — [Refactor][TCPCG] Migrate DSA and DeepSeek graph breaks to eager methods (5/9) #42289.
  6. Migrate MoE and Mamba eager regions — [Refactor][TCPCG] Migrate MoE and Mamba graph execution to eager methods (6/9) #42290.
  7. Migrate diffusion eager wrappers — this PR. Depends only on step 2.
  8. Remove obsolete runtime contexts, registries, and split-op plumbing.
  9. Simplify quantization paths separately because of their torch.compile implications.

The disputed relocation of inactive compiler/reference files is excluded pending a separate decision.

Changes

Migrate diffusion attention and MiniMax H3 wrappers to bare @eager_on_graph and consistently named _eager_ functions. Preserve tuple boxing and the existing graph-dispatch behavior.

This is the independent diffusion slice: it depends only on the eager API introduced in step 2, not the SRT caller migrations in steps 3–6.

Validation

Pre-commit and Python syntax checks passed on the changed files. Diffusion CUDA execution and upstream CI remain to be validated; the local development machine is macOS.


CI States

Latest PR Test (Base): ❌ Run #37077968419
Latest PR Test (Extra): ❌ Run #37077968211
Latest PR Test (AMD ROCm 10): ❌ Run #37077968500

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant