Commits

Branch selector

 main

User selector

All users

All time

Commit History

Commits on Feb 7, 2026

• fix: README commands, paper param contradiction, gradient accumulation flush

  Show description for cd7e385
  

  Marcel

  committed
  cd7e385

  Copy full SHA for cd7e385

• fix: re-review findings — val loss_mask, LoRA device safety, zero-row validity, docs/paper

  Show description for 502d8d0
  

  Marcel

  committed
  502d8d0

  Copy full SHA for 502d8d0

• fix: vLLM top-k logprobs, assistant-only loss masking, prompt boundary capping

  Show description for 753fdef
  

  Marcel

  committed
  753fdef

  Copy full SHA for 753fdef

Commits on Feb 6, 2026

• docs: add LoRA and attention explanation

  

  Marcel

  committed
  0683767

  Copy full SHA for 0683767

• docs: update README to reflect current implementation

  

  Marcel

  committed
  aab8fa3

  Copy full SHA for aab8fa3

Commits on Feb 5, 2026

• fix: chunked KL loss to avoid 318GB OOM at 131K context

  Show description for f62eb53
  

  Marcel

  committed
  f62eb53

  Copy full SHA for f62eb53

• refactor: use full 131K trajectory context for training

  Show description for f4ac04c
  

  Marcel

  committed
  f4ac04c

  Copy full SHA for f4ac04c

• feat: add B=128 config and document 8-GPU teacher caching

  Show description for 358e650
  

  Marcel

  committed
  358e650

  Copy full SHA for 358e650

• feat: batched LoRA via einsum for ~4x training speedup

  Show description for 79f4728
  

  Marcel

  committed
  79f4728

  Copy full SHA for 79f4728

• refactor: complete round 4 code review fixes

  Show description for 7b01dcd
  

  Marcel

  committed
  7b01dcd

  Copy full SHA for 7b01dcd

• fix: round 4 critical bug fixes for training

  Show description for 417e916
  

  Marcel

  committed
  417e916

  Copy full SHA for 417e916

• fix: make max_prompt_tokens configurable in ProducerConsumerPipeline

  Show description for 486414c
  

  Marcel

  committed
  486414c

  Copy full SHA for 486414c

• feat: implement shape-grouped LoRA generator for Qwen3-Coder-Next

  Show description for f6f5073
  

  Marcel

  committed
  f6f5073

  Copy full SHA for f6f5073

• fix: address round 3 correctness issues

  Show description for a7729f6
  

  Marcel

  committed
  a7729f6

  Copy full SHA for a7729f6

• fix: address remaining code review issues (round 2)

  Show description for e32dab2
  

  Marcel

  committed
  e32dab2

  Copy full SHA for e32dab2

• fix: address code review findings

  Show description for ab63aba
  

  Marcel

  committed
  ab63aba

  Copy full SHA for ab63aba

• Fix all code review issues (10 fixes)

  Show description for 35b4b35
  

  Marcel

  committed
  35b4b35

  Copy full SHA for 35b4b35

• Fix critical layer path mapping bug

  Show description for 9364bf3
  

  Marcel

  committed
  9364bf3

  Copy full SHA for 9364bf3

• Add vLLM-based caching and torch.compile optimization

  Show description for d88fe2b
  

  Marcel

  committed
  d88fe2b

  Copy full SHA for d88fe2b

• Wire cached teacher logits into training pipeline

  Show description for 2d9dbdd
  

  Marcel

  committed
  2d9dbdd

  Copy full SHA for 2d9dbdd

• Fix critical implementation issues

  Show description for a7b2561
  

  Marcel

  committed
  a7b2561

  Copy full SHA for a7b2561

• Initial commit: Hypernetwork Distillation System

  Show description for c129f82
  

  Marcel

  committed
  c129f82

  Copy full SHA for c129f82