Skip to content

chore: CI proxy for #2819 - #2820

Closed
HuiyingLi wants to merge 3 commits into
mainfrom
huiyingl/ci/run-pr-2819
Closed

chore: CI proxy for #2819#2820
HuiyingLi wants to merge 3 commits into
mainfrom
huiyingl/ci/run-pr-2819

Conversation

@HuiyingLi

Copy link
Copy Markdown
Contributor

CI-only proxy PR. Do not merge / do not review. Points at the exact head commit of #2819 (0c29737) so internal CI runs under the internal-contributor queue; results post back to #2819 via the shared SHA. Source of truth: #2819. Close once CI completes.

khazic added 3 commits June 28, 2026 15:51
Domino (sgl-project/SpecForge#571) extends the parallel DFlash draft backbone with a lightweight causal correction head: a GRU encodes a causal state from each block's previous tokens, and a low-rank projection of [backbone hidden | GRU state] adds a correction to the parallel base logits. Training jointly supervises the Domino-refined and backbone-only logits with a base-anchor curriculum (loss = (1-lambda_base)*final_loss + lambda_base*base_loss, lambda_base decaying to 0).

The DFlash path is unchanged: the draft model gains an optional Domino head (projector_type='domino'), a new DominoTrainerModule reuses DFlash anchor sampling and the block mask, and TrainDominoRecipe subclasses the DFlash recipe via four behavior-preserving seams. Adds an example YAML and unit tests (domino core + recipe seams).

Signed-off-by: khazic <khazzz1c@gmail.com>
model.to(bfloat16) rounded the rotary inv_freq buffer to bf16, but HF from_pretrained and the SGLang serving runtime keep it in fp32. That train/inference RoPE mismatch grows with absolute position and erodes draft acceptance. Recompute fresh fp32 frequencies from the rotary config after any low-precision cast (a bf16 round-trip cannot be undone by upcasting), covering the training build and reload paths.

Signed-off-by: khazic <khazzz1c@gmail.com>
…nt mask clone

/simplify cleanups (behavior unchanged): _compute_weighted_losses returns only the three losses instead of also re-returning flattened views (the caller recomputes them as views where needed), and the pre-decay eval mask references weight_mask directly since the decay step rebinds it out-of-place, avoiding a per-step clone.

Signed-off-by: khazic <khazzz1c@gmail.com>
@HuiyingLi
HuiyingLi requested a review from a team as a code owner June 28, 2026 09:51
@copy-pr-bot

copy-pr-bot Bot commented Jun 28, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@HuiyingLi

Copy link
Copy Markdown
Contributor Author

/ok to test 0c29737

@HuiyingLi

Copy link
Copy Markdown
Contributor Author

Closing CI-only proxy because source PR #2819 is merged.

@HuiyingLi HuiyingLi closed this Jun 28, 2026
@HuiyingLi
HuiyingLi deleted the huiyingl/ci/run-pr-2819 branch June 28, 2026 14:29
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