Skip to content

[Refactor] Remove prerelease DeepSeek V4.1 config aliases - #38946

Merged
hnyls2002 merged 2 commits into
sgl-project:dsv4.1from
hnyls2002:codex/dsv41-drop-prerelease-config-aliases
Sep 10, 2026
Merged

hnyls2002 merged 2 commits into
sgl-project:dsv4.1from
hnyls2002:codex/dsv41-drop-prerelease-config-aliases

Conversation

@hnyls2002

@hnyls2002 hnyls2002 commented Sep 10, 2026

Copy link
Copy Markdown
Collaborator
  • Stack on dsv4.1: remaining model and runtime integration #38798 with base dsv4.1. The public DeepSeek-V4.1-Flash config uses the canonical names.
  • Remove the three dotted deepseek_v4.1* registrations, five prerelease field aliases, vision_config.max_num_tokens fallback, and DeepseekV41ForConditionalGeneration alias. Old internal checkpoint configs must be migrated to the official names.
  • Let the canonical vision config inherit construction and serialization from PretrainedConfig after removing its legacy conversion.
  • Preserve canonical nested config flattening, vision field mapping, DeepseekV41ForCausalLM routing, config overrides, and save/reload behavior. Keep the existing V4/V3.2 registrations.
  • Validation: 144 passed, 1 skipped, 25 subtests. The unmodified public config produces identical full config snapshots before and after this cleanup for main/DSpark draft construction, nested overrides, save/reload, text/vision subconfigs, and V4/V3.2 registration. Formatting and lint checks passed. GPU weight loading was not rerun for this cleanup.

CI States

Latest PR Test (Base): ❌ Run #34534472878
Latest PR Test (Extra): ❌ Run #34534472729
Latest PR Test (AMD ROCm 10): ❌ Run #34534472990

@hnyls2002
hnyls2002 merged commit b1ad23f into sgl-project:dsv4.1 Sep 10, 2026
81 of 91 checks passed
wonderisland pushed a commit to shinemo-ai/sglang that referenced this pull request Sep 11, 2026
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