Skip to content

[ROCm] Enable shfl xor sync style reduce on amd - #3701

Closed
yiakwy-xpu-ml-framework-team wants to merge 11 commits into
sgl-project:mainfrom
yiakwy-xpu-ml-framework-team:enable_shfl_xor_sync_style_reduce_on_amd
Closed

yiakwy-xpu-ml-framework-team wants to merge 11 commits into
sgl-project:mainfrom
yiakwy-xpu-ml-framework-team:enable_shfl_xor_sync_style_reduce_on_amd

Conversation

@yiakwy-xpu-ml-framework-team

@yiakwy-xpu-ml-framework-team yiakwy-xpu-ml-framework-team commented Feb 19, 2025 •

Copy link
Copy Markdown
Contributor

Motivation

This is follow up of PR#3664

Modifications

  • enable shlf_xor_sync
  • enable flashinfer vec_t (should be removed once flashinfer-rocm [POC] fully finished)

Checklist

thanhhao98 pushed a commit to thanhhao98/sglang that referenced this pull request Sep 7, 2026
init_fi_a2a_workspace() rejected every device whose
is_mnnvl_fabric_supported() is False, i.e. every Blackwell box without an
IMEX/NVL72 fabric (x86 B200/B300), even though FlashInfer can serve that case:
since 0.6.16 (flashinfer sgl-project#3701) MnnvlMemory exports the workspace with POSIX
file-descriptor handles and exchanges them over AF_UNIX SCM_RIGHTS, which needs
no CAP_SYS_PTRACE. sglang pins FlashInfer 0.6.18, so the only remaining
requirement is that the DCP group is contained in one node.
nvpohanh added a commit to nvpohanh/sglang that referenced this pull request Oct 9, 2026
…fresh GLM-5.2, DeepSeek-V4, DeepSeek-V4.1

Mirror every NVIDIA sglang / dynamo-sglang AgentX lane on SemiAnalysis
InferenceX main (ceb50ba2) into the Agentic Long-Context Deployment
explorer added in sgl-project#42726, with these open InferenceX PRs overlaid: sgl-project#3816
(Qwen3.8 GB300 SGLang), sgl-project#3840 (Qwen3.5 FP8 GB300 Dynamo+SGLang), sgl-project#2978
(Qwen3.5 B300 Dynamo+SGLang), sgl-project#3817 (DSV4 B300 Dynamo+SGLang), sgl-project#3606
(DSV4 B300 SGLang), sgl-project#3701 (DSV4 GB200 Dynamo+SGLang), sgl-project#3782 (DSV4.1 GB300
SGLang), sgl-project#3744 (DSV4.1 B300 SGLang) and sgl-project#3836 (GLM-5.2 FP8 B300 SGLang).
Lanes of PRs without an authorized /use run are shown as in progress.

- Qwen3.5: new "3.3 Agentic Long-Context Deployment" subsection (the page
  keeps its legacy generator); the hand-written 5.2.3 AgentX HiCache recipe
  is removed.
- Qwen3.8-Flash-Next: new "Agentic Long-Context Deployment" section after
  the Playground with the Deployment panel's agenticLink footer. The static
  "2.3 GB300 long-context serving with HiCache and MTP" section added in
  sgl-project#43159 is removed (its InferenceX sgl-project#3816 configuration is now in the
  explorer), and the Deployment panel warning that linked to it now links
  to the explorer section.
- GLM-5.2 / DeepSeek-V4 / DeepSeek-V4.1: regenerated data; commands change
  only for the overlaid lanes and one derived HiCache variant.
- _agentx.jsx: notes for --mamba-scheduler-strategy and
  SGLANG_DISAGGREGATION_NIXL_BACKEND=LIBFABRIC, and wording for derived
  Dynamo variants whose image does not bundle ai-dynamo.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.

2 participants