All pull requests

Pull requests

• Fix grouped RMSNorm in Mamba2 non-fused path

  #49712·Subodh-17 opened Oct 18, 2026·4 tasks done·

  ·
  1

  2

• Add regression test for void-only segmentation maps in Mask2Former (closes #30064)

  #49711·habib-analyst opened Oct 18, 2026·

  ·
  1

  2

• Fix expert-parallel token dispatch gradients on torch 2.7 to 2.10

  #49710·digit50 opened Oct 18, 2026·3 of 6 tasks·

  ·
  1

  1

• [Atlas Inference] Strix Halo (gfx1151) Mamba-2 + conv1d Hub kernel path

  #49709·AzeezIsh opened Oct 18, 2026Contributor·5 of 6 tasks·

  ·
  1

• Restore SentencePiece precompiled_charsmap normalizer in DebertaV2Tokenizer

  #49708·deep1blue opened Oct 18, 2026·

  ·
  1

  2

• Fix chunked prefill dropping multimodal encoder outputs in VLMs

  #49707·deep1blue opened Oct 18, 2026·

  ·
  1

  1

• Preserve ModelOutput instances in move_to_device during generate()

  #49706·deep1blue opened Oct 18, 2026·

  ·
  1

  1

• [serge] Fix 2 integration tests regressed by commit c48824e84024 (PR #49404)

  #49705·sergereview[bot] opened Oct 17, 2026Bot·

  ·
  3

• fix(mask2former): register input_projections as nn.ModuleList (fixes #49698)

  #49703·Omarinho12345 opened Oct 17, 2026·

  ·
  1

  2

• update speculative decoding docs

  #49697·merveenoyan opened Oct 17, 2026Contributor·

  ·
  2

• Fix encoder logits processors when precomputed encoder_outputs lack token IDs

  #49696·kunalsanga opened Oct 17, 2026Contributor·

  ·
  1

  2

• Qwen4-Exp: skip QSA block scoring when the budget covers every visible key

  #49694·crisostomi opened Oct 17, 2026·2 of 5 tasks·

  ·
  2

• Fix multi-zoom crops in the SAM mask-generation pipeline and related bugs (Fixes #35530)

  #49689·stlukyanenko opened Oct 17, 2026·4 of 6 tasks·

  ·
  1

  2

• Fix zero expert gradients in bfloat16 with expert parallelism and ep_size < fsdp_size

  #49688·qgallouedec opened Oct 17, 2026Member·

  ·
  1

  8

• Add UltraVBERT

  #49685·athrael-soju opened Oct 17, 2026·3 of 5 tasks·

  ·
  4

• Avoid a host-to-device copy in eager_mask

  #49684·hmellor opened Oct 16, 2026Member·

  ·
  4

• Add v1/audio/speech endpoint

  #49681·SunMarc opened Oct 16, 2026Member·3 tasks·

  ·
  5

• 🚨 Separate EP logic into its own file

  #49679·3outeille opened Oct 16, 2026Member·2/2·

  ·
  15

• Fix/gemma4 tool names response template

  #49678·artin-saberi-ai opened Oct 16, 2026·

  ·
  2

• Declare the input modalities of the VLMs that report text only

  #49677·qgallouedec opened Oct 16, 2026Member·

  ·
  3

• Make TextStreamer linear on long lines by decoding a bounded token window

  #49676·deadstrobe5 opened Oct 16, 2026·3 of 5 tasks·

  ·
  1

  1

• [Pipelines] Reject format specs in zero-shot hypothesis_template

  #49673·errmakov opened Oct 16, 2026·4 of 6 tasks·

  ·
  1

  1

• mimi moshi output capturing

  #49670·eustlb opened Oct 16, 2026Collaborator·1/4·

  ·
  3

• Cleaned up tokenizer tests

  #49669·tarekziade opened Oct 16, 2026Collaborator·

  ·
  5

• Show the actual accelerate version for transformers env

  #49668·Sai-Suraj-27 opened Oct 16, 2026Contributor·6 tasks·

  ·
  8