Skip to content

[Docs] Add AMD-specific HiCache config for DeepSeek V4 playground#31122

Open
seungrokj wants to merge 1 commit into
sgl-project:mainfrom
seungrokj:amd_dsv4_hicache
Open

[Docs] Add AMD-specific HiCache config for DeepSeek V4 playground#31122
seungrokj wants to merge 1 commit into
sgl-project:mainfrom
seungrokj:amd_dsv4_hicache

Conversation

@seungrokj

@seungrokj seungrokj commented Jul 14, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Add fc.amdIo config-driven override to the HiCache playground axis in _playground.jsx, allowing per-model AMD ROCm-specific --hicache-mem-layout, --hicache-io-backend, and --hicache-ratio flags
  • DeepSeek V4 opts in via amdIo: { memLayout: "page_first_direct", ioBackend: "direct", ratio: 4 } — AMD path also skips --hicache-size 0
  • NVIDIA path is unchanged (default: page_first_direct + direct, ratio 2)

Test plan

  • Run mintlify dev and verify DeepSeek V4 playground with MI355X selected: HiCache enable should emit --hicache-ratio 4 without --hicache-size 0, and --hicache-mem-layout page_first_direct --hicache-io-backend direct regardless of storage backend selection
  • Verify NVIDIA hardware (e.g., B200) still emits the default --hicache-ratio 2 --hicache-size 0 and only adds layout/io-backend when a storage backend is selected
  • Verify other models with hicache (e.g., GLM-5.2, Laguna M1) are unaffected on AMD (no amdIo declared)

🤖 Generated with Claude Code

Outcome

This will be applied only on AMD devices & DSv4
image


CI States

Latest PR Test (Base): ⏳ Run #29309928661
Latest PR Test (Extra): ⏳ Run #29309928579

The HiCache playground axis now supports per-model AMD ROCm overrides
via `fc.amdIo = { memLayout, ioBackend, ratio }`. DeepSeek V4 opts in
with AMD-appropriate flags (page_first_direct + direct io, ratio 4,
no --hicache-size 0), matching validated MI355X test configurations.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@gemini-code-assist

Copy link
Copy Markdown
Contributor

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

deepseek documentation Improvements or additions to documentation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant