Repository navigation

Monolithic fused head-loss kernel (torch.compile) (#507) - #549

Merged

jlamypoirier merged 37 commits into

mainfrom

jlp_monolithic_head_loss

Jul 20, 2026

Commits on Jun 26, 2026
- Monolithic head-loss kernel: scaffolding + cross-entropy (#507)
  
  jlamypoirier
  and
  claude
  committed
  b6c71a8
- Monolithic head-loss kernel: add z-loss (#507)
  
  jlamypoirier
  and
  claude
  committed
  3b4070a
- Monolithic head-loss kernel: add from-distribution KL losses (#507)
  
  jlamypoirier
  and
  claude
  committed
  f5ca0d7
Commits on Jun 27, 2026
- Monolithic head-loss kernel: wire distillation + fix dropped temperature (#507)
  
  jlamypoirier
  and
  claude
  committed
  e4d5845
- Monolithic head-loss kernel: GRPO objective + new_logprobs (#507)
  
  jlamypoirier
  and
  claude
  committed
  c872b11
Commits on Jun 29, 2026
- Monolithic head-loss kernel: GRPO metrics + entropy, kill the second softmax (#507)
  
  jlamypoirier
  and
  claude
  committed
  c8fd86f
- Monolithic head-loss kernel: GSPO three-phase eager-seam path (#507)
  
  jlamypoirier
  and
  claude
  committed
  ea6f569
- Monolithic head-loss kernel: multi-loss combo coverage (#507)
  
  jlamypoirier
  and
  claude
  committed
  e5147c5
Commits on Jun 30, 2026
- Monolithic head-loss kernel: GSPO as a kernel kind (#507)
  
  jlamypoirier
  and
  claude
  committed
  ecea37c
- Monolithic head-loss kernel: make shared loss cores public (#507)
  
  jlamypoirier
  and
  claude
  committed
  3340da4
- Address review (#549): gate GRPO metrics off-log, drop triton enum, share CE/z-loss cores
  
  jlamypoirier
  and
  claude
  committed
  40a1f37
Commits on Jul 2, 2026
- Rework monolithic head loss as a `monolithic` loss type (#507)
  
  jlamypoirier
  and
  claude
  committed
  c9c95c8
- Fold combinable loss boilerplate into a CombinableLoss base (#507)
  
  jlamypoirier
  and
  claude
  committed
  70250d7
- Move standalone combinable forward-backward onto the loss object (#507)
  
  jlamypoirier
  and
  claude
  committed
  86b6816
- Route entropy losses through combinable_forward_backward too (#507)
  
  jlamypoirier
  and
  claude
  committed
  1bc212c
- Fix entropy/z-loss test threshold comparing the wrong symbol (#507)
  
  jlamypoirier
  and
  claude
  committed
  dcfa978
- Rename combinable_extract -> get_inputs, combinable_core -> fused_core (#507)
  
  jlamypoirier
  and
  claude
  committed
  5295e01
- Extract single-loss template into SingleLoss intermediate (#507)
  
  jlamypoirier
  and
  claude
  committed
  735f4fd
- Untrack stray benchmark; trim loss docstrings/comments (#507)
  
  jlamypoirier
  and
  claude
  committed
  869db68
- Untrack bench_monolithic.py (#507)
  
  jlamypoirier
  and
  claude
  committed
  cb5be2f
Commits on Jul 3, 2026
- Guard label reverse-KL / monolithic use_triton; test monolithic under TP (#507)
  
  jlamypoirier
  and
  claude
  committed
  c8fb19f
- Fine-review nits: type hints, strict zip, docstring fix (#507)
  
  jlamypoirier
  and
  claude
  committed
  71421fb
- Skip the gradient term for zero-weight losses (#507)
  
  jlamypoirier
  and
  claude
  committed
  b7a7227
- Add CombinableLossConfig base, drop the combinable ClassVar (#507)
  
  jlamypoirier
  and
  claude
  committed
  9254905
- Stack the composite loss's logits_scale_factor onto its children (#507)
  
  jlamypoirier
  and
  claude
  committed
  7cf65dd
- Fine-review cleanups: CombinableLoss stubs, drop redundant guard (#507)
  
  jlamypoirier
  and
  claude
  committed
  66fa29e
Commits on Jul 7, 2026
- Fuse GSPO into the monolithic compiled loss (#507) (#552)
  
  jlamypoirier
  and
  claude
  authored
  96a6991
Commits on Jul 8, 2026
- Fuse the GSPO segment seam into compiled blocks (#507) (#554)
  
  jlamypoirier
  and
  claude
  authored
  9e7dd32
Commits on Jul 10, 2026
- Merge main into monolithic head-loss (compiled)
  
  jlamypoirier
  and
  claude
  committed
  15421c1
- Merge remote-tracking branch 'origin/main' into jlp_monolithic_head_loss
  jlamypoirier
  committed
  8c4204a
Commits on Jul 13, 2026
- Policy metrics in the compiled monolithic loss (#507) (#565)
  
  jlamypoirier
  and
  claude
  authored
  18c4662
Commits on Jul 14, 2026
- Triton monolithic head-loss kernel (#507) (#558)
  
  jlamypoirier
  and
  claude
  authored
  52fabf0
Commits on Jul 15, 2026
- Fuse distillation + z-loss in the triton monolithic path (#507) (#566)
  
  jlamypoirier
  and
  claude
  authored
  74994b4
Commits on Jul 16, 2026
- Merge main into monolithic head-loss; integrate #562 data metrics
  
  jlamypoirier
  and
  claude
  committed
  1a8eb00
- Address review consistency findings on the monolithic head-loss kernel (#507)
  
  jlamypoirier
  and
  claude
  committed
  fdde689
Commits on Jul 20, 2026
- Auto-select the monolithic head-loss implementation (#507) (#568)
  
  jlamypoirier
  and
  claude
  authored
  f0c337b
- Review fixes: widen data-metrics type hint, name regularized log-sum-exp (#507)
  
  jlamypoirier
  and
  claude
  committed
  82f95bc

Commits

Commits on Jun 26, 2026

Commits on Jun 27, 2026

Commits on Jun 29, 2026

Commits on Jun 30, 2026

Commits on Jul 2, 2026

Commits on Jul 3, 2026

Commits on Jul 7, 2026

Commits on Jul 8, 2026

Commits on Jul 10, 2026

Commits on Jul 13, 2026

Commits on Jul 14, 2026

Commits on Jul 15, 2026

Commits on Jul 16, 2026

Commits on Jul 20, 2026