Skip to content

Support W&B 0.28 and vLLM 0.24 DSV4 LoRA#754

Merged
bradhilton merged 2 commits into
mainfrom
agent/wandb-0-28-main
Jul 14, 2026
Merged

Support W&B 0.28 and vLLM 0.24 DSV4 LoRA#754
bradhilton merged 2 commits into
mainfrom
agent/wandb-0-28-main

Conversation

@bradhilton

Copy link
Copy Markdown
Collaborator

Summary

  • upgrade ART's backend W&B pin from 0.25.0 to 0.28.0
  • declare the lora_manager attribute required by vLLM 0.24's runtime-checkable LoRA protocol for DeepSeek V4
  • add a focused regression for the DSV4 protocol contract

Why

Caladan's live DeepSeek-V4-Flash qualification exposed two integration gaps. ART's backend W&B pin conflicted with Caladan's validated W&B 0.28 dependency, and vLLM 0.24 rejected the patched DSV4 model after weight loading because the structural LoRA protocol now requires lora_manager.

The Click 8.2 override is required because W&B 0.28 requires Click 8.2 or newer while ART's current SkyPilot development dependency still declares an older upper bound. ART's existing CLI smoke test remains functional; this follows the same override already used by Caladan.

Validation

  • 94 selected W&B/backend/trainer/DSV4 tests passed
  • 2 focused DSV4 runtime patch tests passed
  • uv lock --check passed
  • focused Ruff and format checks passed
  • Caladan live qualification passed base and rank-8 LoRA inference on 2xB200
  • Caladan full suite: 426 passed, 3 skipped, 19 subtests; full prek passed

Full ART type checking still reports existing CPU-environment Megatron/Torch and missing optional einops diagnostics; none originate in this diff.

@bradhilton bradhilton marked this pull request as ready for review July 14, 2026 00:26
@bradhilton bradhilton merged commit 88fad83 into main Jul 14, 2026
9 checks passed
@bradhilton bradhilton deleted the agent/wandb-0-28-main branch July 14, 2026 00:50
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