Would cuTile ports of Liger's chunked preference losses (DPO, ORPO, SimPO, KTO) be accepted into the liger suite?
Maintainers usually reply within 8 days
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 35/100
- Issue type
- Feature
- Clarity
- Needs clarification
- Activity status
- Active
- Tech stack
- python
- Domain
- machine-learning, performance
Research direction
Start by reviewing src/tilegym/suites/liger/ and the grpo_loss and fused_linear_cross_entropy patterns, then compare the upstream Liger chunked preference-loss implementations. Resolve whether external ports are accepted and whether the first DPO port covers all loss_type variants. Done requires an agreed scope, matching semantics, and benchmark validation on Ampere or Blackwell.
Written by the indexing model from the issue text.
Description
I would like to contribute cuTile ports of Liger-Kernel's chunked preference losses to src/tilegym/suites/liger/, following the grpo_loss fwd+bwd pattern and the chunked backward-in-forward structure of fused_linear_cross_entropy. Plan: one PR per loss, starting with DPO (LigerFusedLinearDPOLoss semantics: fused linear over chosen/rejected halves, per-sequence logps, reference-model support, the upstream loss_type variants).
Intended split, mirroring the FLCE port: torch/cuBLAS keeps the GEMMs, cuTile owns the two row-wise kernels (fused per-row logsumexp plus target log-prob, and the in-place d_logits write), and the O(n_pairs) preference formulas stay in device-side torch so every variant matches upstream exactly. Upstream implements these in chunked torch.compile rather than a handwritten kernel, so the benchmark baseline would be Liger's compiled path plus a naive full-logits torch reference, validated on Ampere or Blackwell.
Before I start: is the liger suite open to external ports? It is NVIDIA-authored so far, and I do not want to collide with internal plans. Two scope questions if yes:
- Full loss_type coverage in the first DPO PR, or sigmoid first with variants as follow-ups?
- Perf table in the PR description only, like the existing liger suite ops, or an in-repo benchmark file like
tests/benchmark/suites/unsloth/?
- Dominant language
- Python
- Stars
- 830
- Forks
- 92
- Avg merge
- 8d 11h
- Merged PRs (30d)
- 8
Getting set up
- No Dockerfile or Docker Compose file
- Has a pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from NVIDIA/TileGym
-
Difficulty 5/5 Over a week Newbie friendliness 12/100
Maintainers usually reply within 8 days
-
Difficulty 5/5 Over a week Newbie friendliness 35/100
NVIDIA/TileGym#210 · 1 comment ·
Maintainers usually reply within 8 days
-
Difficulty 5/5 Over a week Newbie friendliness 35/100
NVIDIA/TileGym#195 · 2 comments ·
Maintainers usually reply within 8 days
-
Difficulty 5/5 Over a week Newbie friendliness 25/100
NVIDIA/TileGym#191 · 1 comment ·
Maintainers usually reply within 8 days
-
Difficulty 4/5 3-5 days Newbie friendliness 35/100
NVIDIA/TileGym#169 · 5 comments · 1 reaction ·
Maintainers usually reply within 8 days
Similar issues
-
Difficulty 2/5 1-3 hours Newbie friendliness 82/100
Nixtla/mlforecast#823 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 65/100
microsoft/ML-For-Beginners#1061 ·
-
bot-found documentation priority: P3
Difficulty 2/5 Under an hour Newbie friendliness 82/100
madenvel/KalinkaPlayer#503 ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 63/100
EleutherAI/lm-evaluation-harness#4508 ·
Maintainers usually reply within 1 day