Skip to content

fix(ds4): propagate sampler config in layer split#530

Closed
Graffioh wants to merge 1 commit into
Luce-Org:mainfrom
Graffioh:codex/fix-ds4-layer-split-sampling-main
Closed

fix(ds4): propagate sampler config in layer split#530
Graffioh wants to merge 1 commit into
Luce-Org:mainfrom
Graffioh:codex/fix-ds4-layer-split-sampling-main

Conversation

@Graffioh

@Graffioh Graffioh commented Jul 16, 2026

Copy link
Copy Markdown
Contributor

Summary

  • propagate request sampler settings into the DeepSeek4 layer-split adapter
  • reseed the layer-split sampler RNG when sampling with an explicit seed
  • add unit coverage for sampler propagation and deterministic seeding

Scope

This PR covers the layer-split path. #527 remains focused on single-device decode correctness.

Known limitation

run_layer_split_ar_decode() currently supplies generated tokens to sample_logits(), but not the prompt or restored-prefix history. Temperature, top-k, and top-p propagation are covered here; repetition, frequency, and presence penalties remain incomplete until both decode paths share a proper request-history contract.

Validation

  • git diff --check passed
  • unit coverage verifies sampler propagation and deterministic seeding
  • runtime GPU validation pending

@Graffioh Graffioh changed the title fix(deepseek4): propagate sampler config in layer split fix(ds4): propagate sampler config in layer split Jul 16, 2026
@Graffioh
Graffioh force-pushed the codex/fix-ds4-layer-split-sampling-main branch from 45a819e to b87a397 Compare July 16, 2026 13:01
@Graffioh

Copy link
Copy Markdown
Contributor Author

Superseded by #531. The sampler propagation change from this PR has been integrated into #531, where it is combined with the DeepSeek 4 GPU argmax work, mutually exclusive logits/argmax output paths, and CPU/CUDA/HIP parity coverage. Closing this PR so the combined change can be reviewed and merged in one place.

@Graffioh Graffioh closed this Jul 16, 2026
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