Skip to content

Rocm changes to tests/e2e/megatron/test_glm5_744b_a40b_4layer.py#1172

Open
sreerohi wants to merge 1 commit into
radixark:mainfrom
sreerohi:rocm/glm5_744b_a440_4layer_changes
Open

Rocm changes to tests/e2e/megatron/test_glm5_744b_a40b_4layer.py#1172
sreerohi wants to merge 1 commit into
radixark:mainfrom
sreerohi:rocm/glm5_744b_a440_4layer_changes

Conversation

@sreerohi
Copy link
Copy Markdown

@sreerohi sreerohi commented May 21, 2026

  • Drop DeepEP; use alltoall dispatcher
  • Swap flashmla_sparse NSA backend for tilelang
  • Set sglang-page-size to 1 (required by tilelang NSA)

Depends on #1122 and #1123

Relates to #1105

@sreerohi sreerohi changed the title Rocm changes to tests/e2e/megatron/test_quick_ start_glm4_9B.py Rocm changes to tests/e2e/megatron/test_quick_start_glm4_9B.py May 21, 2026
@sreerohi sreerohi changed the title Rocm changes to tests/e2e/megatron/test_quick_start_glm4_9B.py Rocm changes to tests/e2e/megatron/test_glm5_744b_a40b_4layer.py May 21, 2026
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

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

Code Review

This pull request introduces ROCm support for the GLM-5 4-layer model end-to-end test. It implements environment detection using torch.version.hip to conditionally configure execution parameters, including switching the NSA backend to tilelang, reducing the SGLang page size, and utilizing the alltoall token dispatcher for MoE when running on ROCm. I have no feedback to provide.

…ark#1122 , radixark#1123):

- Drop DeepEP; use alltoall dispatcher
- Swap flashmla_sparse NSA backend for tilelang
- Set sglang-page-size to 1 (required by tilelang NSA)
@sreerohi sreerohi force-pushed the rocm/glm5_744b_a440_4layer_changes branch from 589c5f0 to 5540524 Compare May 21, 2026 20:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant