All pull requests

Pull requests

• FP8 scaled_mm: keep native fp16 output on torch >= 2.14

  #16063·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• MXFP4 torch decode fallback: run ldexp on the operands' GPU

  #16062·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• gpt-oss: accept an absolute transformers_weights inside the local model dir

  #16061·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• Scope WavLM's 4-bit attention skip to WavLM's own layers

  #16060·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• Cross entropy: trace under torch.compile(dynamic=True)

  #16059·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• Packed mRoPE flash attention: cover a padded tail, same decision when compiled

  #16058·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• GRPO: keep masking completions cut by the vLLM context limit

  #16057·danielhanchen opened Oct 30, 2026Member·

  ·
  3

  

• Studio Images: enable GGUF LoRAs on MLX and diffusers, and explain when unavailable

  #16053·oobabooga opened Oct 30, 2026Member·

  ·
  2

• fix(datasets): preserve ShareGPT message metadata

  #16052·hsusul opened Oct 30, 2026·

  ·
  1

• Flex attention: small shared-memory tiles for short no-grad calls so one short prompt does not leave the process on eager attention

  #16049·danielhanchen opened Oct 30, 2026Member·

  ·

• Studio: stop offering the VibeVoice q4-ios GGUF that audio.cpp cannot load

  #16048·oobabooga opened Oct 30, 2026Member·

  ·
  2

• Studio: download the commit the manifest was planned from

  #16047·danielhanchen opened Oct 29, 2026Member·

  ·

• Studio: warn when a Windows AMD GGUF load leaves the system short of commit

  #16046·debojyoti-dev opened Oct 29, 2026Contributor·

  ·
  1

• Studio: let managed accounts use local MCP servers when allowed

  #16044·NilayYadav opened Oct 29, 2026Collaborator·

  ·
  3

  

• feat(distributed): Zero-Disk in-flight token streaming & in-memory GGUF v3 Hub exporter

  #16043·thealidev opened Oct 29, 2026·

  ·

• Don't warn about loss_type = "nll" on a TRL whose SFTConfig predates it

  #16042·debojyoti-dev opened Oct 29, 2026Contributor·

  ·
  1

• Gemma 3: generate on the non-flex attention path instead of flex_attention's fallback

  #16041·danielhanchen opened Oct 29, 2026Member·

  ·

• Desktop: switch the theme without a view transition on Linux

  #16040·oobabooga opened Oct 29, 2026Member·

  ·
  1

  2

• Studio: include Schema transform output in recipe dataset downloads

  #16039·NilayYadav opened Oct 29, 2026Collaborator·

  ·
  1

  

• Studio: scheduled tasks use Memory and project instructions

  #16038·NilayYadav opened Oct 29, 2026Collaborator·

  ·
  1

  

• Studio: stop skipping scheduled runs right after Studio starts

  #16037·NilayYadav opened Oct 29, 2026Collaborator·

  ·
  1

  

• Studio: let conversation search return every earlier turn that fits

  #16036·NilayYadav opened Oct 29, 2026Collaborator·

  ·
  

• Studio: do not hand agents a tool call cut off at the token limit

  #16035·NilayYadav opened Oct 29, 2026Collaborator·

  ·
  2

  

• Opt-in fused batch-1 NF4 decode for Llama (UNSLOTH_DECODE_FUSED_NF4=1): 1.6-2.3x per token in eager generate

  #16032·danielhanchen opened Oct 29, 2026Member·

  ·
  17

• Studio diffusion: test-launch sd.cpp binaries without the Windows missing-DLL box

  #16031·oobabooga opened Oct 29, 2026Member·

  ·
  2