All pull requests

Pull requests

• fix(generate): tensor-parallel parity for GptNeoxMLP / GemmaMLP (#1663)

  #2357·virg1n opened Oct 14, 2026·1 of 2 tasks·

  ·
  1

• build(deps): bump the gha-updates group across 1 directory with 3 updates

  CI / actions
  #2356·dependabot[bot] opened Oct 13, 2026Bot·

  ·

• Specify dtype for softmax probabilities

  #2355·Harsh-Prajapati54 opened Oct 12, 2026·

  ·
  1

• fix(model): keep sliding-window KV cache correct for multi-token forwards past the window

  #2354·tritsystem opened Oct 10, 2026·

  ·

• ci: pin runner images to ubuntu-24.04, macos-26, windows-2025

  #2353·bhimrazy opened Oct 5, 2026Collaborator·

  ·

• fix(prompts): add Falcon-180B-Chat prompt style

  #2351·Saadanjum0 opened Oct 2, 2026·

  ·

• Fix docstrings that name parameters the functions do not have

  #2350·VenishPaneliya opened Oct 2, 2026·

  ·

• feat(config): add DeepSeek-R1-Distill-Qwen 1.5B/7B/14B/32B configs

  #2349·jkatdare opened Oct 1, 2026·

  ·
  1

• feat(finetune): optional token-level loss normalization across gradient accumulation

  #2347·jkatdare opened Oct 1, 2026·

  ·
  1

• perf(model): vectorize the training-time sliding window attention mask

  #2345·jkatdare opened Oct 1, 2026·

  ·
  1

• Add assistant-only loss masking for multi-turn Llama 3 SFT

  #2343·marcusacosta opened Sep 28, 2026·

  ·

• docs: deploy LitGPT on Nebius Serverless Endpoints

  #2342·SalikovAlex opened Sep 28, 2026·

  ·

• Keep a comma that sits inside a JSON string

  #2340·SashaMIT opened Sep 28, 2026·2 of 3 tasks·

  ·
  1

  3

• fix(tokenizer): keep trailing eos when encode(eos=True) input already ends with it

  #2338·MohammadHijjawi97 opened Sep 27, 2026·

  ·

• Add configurable per-layer LoRA ranks (Related to Issue #2335)

  #2336·kor80 opened Sep 25, 2026·

  ·
  1

• Enable Flash Attention for safe KV cache decoding

  #2334·mukeshr06 opened Sep 23, 2026·

  ·
  1

• Fix pretraining max steps and checkpoint saving

  #2333·mukeshr06 opened Sep 23, 2026·

  ·
  1

• docs: clarify tokenizer considerations for new-language finetuning

  #2332·mukeshr06 opened Sep 22, 2026·

  ·

• fix: prevent deadlock in pretrain validation when rank dataloader is empty

  #2331·mayuriphad opened Sep 21, 2026·

  ·
  1

  2

• fix(chat): make --compile work

  #2330·manduinca opened Sep 15, 2026·

  ·
  1

• docs: add Simplified Chinese translation (README.zh.md)

  #2328·JasonYeYuhe opened Sep 13, 2026·3 tasks done·

  ·
  1

• Fix adapter KV cache reuse during training

  #2327·Excelius-Wang opened Sep 9, 2026·

  ·
  1

• fix(tokenizer): decode_stream drops leading space per-word on the huggingface backend

  #2326·999thBaam opened Sep 9, 2026·

  ·
  1

  1

• fix(generate): preserve sampling support for CPU float16 probabilities

  #2325·Excelius-Wang opened Sep 9, 2026·

  ·
  1

• fix(merge_lora): use CPU-safe precision to avoid false AMP warning

  #2322·ShamikOfficial opened Sep 7, 2026·1 task done·

  ·
  1