All pull requests

Pull requests

• Studio: keep the whole int8 Qwen-Image-2.1 denoiser resident at 12 GB, released only while the encoders run (298 to 212 ms per step)

  #12536·danielhanchen opened Oct 2, 2026Member·

  ·
  1

• [XPU] Enable FP8 training on XPU

  #12535·JoshuaL3000 opened Oct 2, 2026Contributor·

  ·

• Studio small-host route: prefetch the streamed text encoder and fuse the int8 dequant (T4 FLUX.1-schnell 5.19 to 4.49 s per step, pixel-identical)

  #12532·danielhanchen opened Oct 2, 2026Member·

  ·
  1

  

• Studio: fused fp16 Wan block kernels on T4 (Wan2.2-5B 11.7 to 10.0 s per step, within base's spread)

  #12531·danielhanchen opened Oct 2, 2026Member·

  ·
  1

  

• Count only the lock's own waits in the unlockable-filesystem row

  #12529·lonexreb opened Oct 2, 2026Contributor·

  ·
  

• fix: preserve shared upstream gradients in RoPE backward

  #12528·taking-lying-flat opened Oct 2, 2026Contributor·

  ·
  2

  

• Studio: MiniMax-H3 GGUF uses the BF16 cuBLAS matmul path by default on sm80+ (1.2 to 2.3x per step, closer to an F32 reference)

  #12527·danielhanchen opened Oct 2, 2026Member·

  ·
  3

  

• Studio: move torchao int8 weights back to the GPU after an oversized request streams pinned groups

  #12526·danielhanchen opened Oct 2, 2026Member·

  ·
  3

  

• Studio int8 GEMM: stop the device probe from reserving 0.7 to 1.4 GB of spill memory

  #12525·danielhanchen opened Oct 2, 2026Member·

  ·
  2

  

• Studio: engage the fused int8 GEMM when the placement pins the whole denoiser, and restore torchao weights after an oversized request

  #12524·danielhanchen opened Oct 2, 2026Member·

  ·
  1

  

• Studio: renew the chat-run lease while long prefill is still advancing

  #12523·indrajeetapache opened Oct 2, 2026Contributor·4 tasks·

  ·
  1

  

• Unsloth Studio / Desktop: add a llama-bench tab to Benchmarks and keep llama-bench in the llama.cpp install

  #12522·LeoBorcherding opened Oct 2, 2026Collaborator·

  ·
  

• Unsloth Studio / Desktop: add an Evals tab to Benchmarks that scores a GGUF with lm-eval-harness

  #12521·LeoBorcherding opened Oct 2, 2026Collaborator·

  ·

• Studio: fuse MiniMax-H3's ConvRot rotation into the int8 activation quant

  #12517·danielhanchen opened Oct 2, 2026Member·

  ·
  3

• Studio: use ComfyUI's default settings for FLUX.1, Qwen-Image, Z-Image, Ideogram 4, Wan2.2 and HunyuanVideo-1.5

  #12516·danielhanchen opened Oct 2, 2026Member·

  ·
  2

• Parity: keep the reap-margin check off the edge of its own window

  #12514·lonexreb opened Oct 2, 2026Contributor·

  ·
  1

• Studio: draw the command palette like chat search

  #12511·shimmyshimmer opened Oct 2, 2026Member·

  ·
  4

• Enhance smoke test to validate fp16 matmul execution

  defer#non-studio#
  #12509·MaddipatlaChetan24 opened Oct 2, 2026·

  ·
  1

• Studio: MiniMax-H3 int8 Linears take the fused-dequant int8 GEMM (A100 9% faster per step, ahead of ComfyUI)

  new-feature#
  #12507·danielhanchen opened Oct 2, 2026Member·

  ·
  2

• Studio: keep the composer permission shield still while the plus spins

  affects UI#codex-converged#
  #12506·shimmyshimmer opened Oct 2, 2026Member·

  ·
  3

• Studio: stop rejecting --mmproj-device CUDA1 when gpu_ids are saved

  edge case#
  #12505·1fanwang opened Oct 2, 2026·

  ·
  1

  1

• Studio: center the sidebar row kebab and pin in their hover circle

  affects UI#
  #12502·shimmyshimmer opened Oct 2, 2026Member·

  ·
  3

• Studio: lift a carried model picker row like a sidebar chat, and lighten both drag copies

  affects UI#new-feature#
  #12499·shimmyshimmer opened Oct 2, 2026Member·

  ·

• studio: enable anthropic studio tools

  new-feature#
  #12497·mahiatlinux opened Oct 2, 2026Collaborator·

  ·
  1

• Studio: stop Whisper dropping sentences from clips longer than 30 seconds

  important#
  #12481·NilayYadav opened Oct 2, 2026Collaborator·

  ·
  2