All pull requests

Pull requests

• fix: do not double the curly braces in the Llama2FunctionCalling prompt

  #2419·Yupsecous opened Nov 14, 2026·

  ·
  1

• pretrain: add --compile and skip torch.compile on CUDA without Triton (fails on Windows)

  #2417·Zephyr4772 opened Nov 13, 2026·

  ·

• generate_full/adapter/adapter_v2: decode only the generated tokens (IndexError for non-Alpaca prompt styles)

  #2416·Zephyr4772 opened Nov 13, 2026·

  ·

• Bound SFT collator storage by truncating before padding

  #2415·Afloat16 opened Nov 13, 2026·

  ·
  1

• Keep finetuning dataloader workers alive between epochs and validations (35 s per epoch/validation on Windows)

  #2413·Zephyr4772 opened Nov 13, 2026·

  ·

• Avoid overflow when accumulating chunked half-precision losses

  #2412·Afloat16 opened Nov 12, 2026·

  ·
  1

• Keep optimizer runtime overrides separate from reusable configurations

  #2410·Afloat16 opened Nov 12, 2026·

  ·
  1

• Reuse MoE input activations for shared experts

  #2408·Afloat16 opened Nov 12, 2026·

  ·
  1

• Preserve SFT source samples when applying transforms

  #2406·Afloat16 opened Nov 12, 2026·

  ·
  1

• Train wrapped LoRA biases for lora_only policy

  #2404·Afloat16 opened Nov 12, 2026·

  ·
  1

• Honor explicit adapter constructor device and dtype

  #2402·Afloat16 opened Nov 12, 2026·

  ·
  1

• Preserve training mode when example generation is skipped

  #2400·Afloat16 opened Nov 12, 2026·

  ·
  1

• Keep all-ignored unchunked cross entropy finite

  #2398·Afloat16 opened Nov 12, 2026·

  ·
  1

• Allow fine-tuning with zero learning-rate warmup

  #2396·Afloat16 opened Nov 12, 2026·

  ·
  1

• Fix CodeLlama-34b-Instruct mapping to the CodeLlama prompt style

  #2394·selimfedakar opened Nov 9, 2026·

  ·
  1

• fix(tokenizer): keep an existing trailing EOS when eos=True

  #2392·PratikDhanave opened Nov 7, 2026·

  ·

• fix(get_lr): avoid division by zero when warmup_iters >= max_iters

  #2391·davrens101 opened Nov 5, 2026·

  ·
  1

• Preserve prompt style when saving LLM

  #2390·shivamkumar0502 opened Nov 4, 2026·

  ·

• Keep grouped MoE routing within eligible expert groups

  #2389·davrens101 opened Nov 3, 2026·

  ·
  1

• Return point-mass probabilities for greedy speculative decoding

  #2386·davrens101 opened Nov 2, 2026·

  ·
  1

• build(deps-dev): bump litdata from 0.2.59 to 0.2.76

  dependencies
  #2384·dependabot[bot] opened Nov 1, 2026Bot·

  ·

• build(deps-dev): update transformers requirement from <4.57,>=4.51.3 to >=4.51.3,<5.21

  dependencies
  #2383·dependabot[bot] opened Nov 1, 2026Bot·

  ·

• Handle token ID output in LLM benchmarks

  #2382·remote-controlled-man opened Nov 1, 2026·

  ·
  1

  1

• Preserve the active prompt style when saving LLMs

  #2380·remote-controlled-man opened Nov 1, 2026·

  ·
  1

  1

• fix: explain sequence length limits during finetuning

  #2379·emme1t opened Nov 1, 2026·

  ·
  1