All pull requests

Pull requests

• Move the MI300 docker image to ROCm 10 and update test expectations

  #49460·Abdennacer-Badaoui opened Oct 9, 2026Member·

  ·
  2

• Add force_gqa flag in sdpa_attention_forward

  #49459·jakobknauer opened Oct 9, 2026·1 task done·

  ·
  1

• Keep AutoVideoProcessor class lookup lazy

  #49458·DimensionSTP opened Oct 9, 2026Contributor·3 of 5 tasks·

  ·
  1

  4

• Fix create_chunked_causal_mask() crashing with block_sequence_ids

  #49457·kaixuanliu opened Oct 9, 2026Contributor·

  ·
  2

• Fix a port allocation race in FSDP tests

  #49455·icavan opened Oct 9, 2026·

  ·
  2

• Fix loading legacy remote-code models without all_tied_weights_keys

  #49454·sywangyi opened Oct 9, 2026Contributor·

  ·
  3

• feat: Add/update AGENTS.md for AI agent guidance

  #49450·jaykrishna316 opened Oct 9, 2026·

  ·
  1

• Fix Trainer checkpoint rotation matching backup directories

  #49449·SrijanRoy123-github opened Oct 9, 2026·

  ·
  1

  1

• fix bridgetower cross-attn mask shape bug

  #49447·kaixuanliu opened Oct 9, 2026Contributor·

  ·
  3

• [Model] Add native Ling 3.0 Flash support

  #49446·icavan opened Oct 9, 2026·4 of 5 tasks·

  ·
  6

• Stop treating HunYuan's DynamicNTKAlpha RoPE as dynamic

  #49445·jiqing-feng opened Oct 9, 2026Contributor·

  ·
  4

• Use 32-row flex attention blocks for head_dim > 256

  #49444·qgallouedec opened Oct 9, 2026Member·

  ·
  3

• Use the Hub rotary kernel for inference only on CUDA

  for patch
  #49439·qgallouedec opened Oct 8, 2026Member·

  ·
  2

• Add GLiNER2

  #49438·MattThomas-fastino opened Oct 8, 2026·

  ·
  9

• [vLLM] Add Qwen-Omni Thinker base models and Gemma 3n per-layer inputs

  #49437·harshaljanjani opened Oct 8, 2026Contributor·3 tasks done·

  ·
  21

• [vLLM] Add Omni token counting and text replacement offsets for the Transformers backend

  #49435·harshaljanjani opened Oct 8, 2026Contributor·3 tasks done·

  ·
  4

• Port Llama 4's MoE to the shared experts interface

  #49432·IlyasMoutawwakil opened Oct 8, 2026Member·6 tasks·

  ·
  3

• Restore GPT-OSS's router-masked expert-parallel plan

  #49431·IlyasMoutawwakil opened Oct 8, 2026Member·6 tasks·

  ·
  3

• Add rerun-failed CPU/GPU callers for the CI dashboard

  #49430·tarekziade opened Oct 8, 2026Collaborator·

  ·
  4

• Expert-parallel dispatch runs the experts masked when the EP group shares one batch

  #49429·IlyasMoutawwakil opened Oct 8, 2026Member·6 tasks·

  ·
  2

• Fix DeepGEMM Mega MoE under expert parallelism

  #49428·IlyasMoutawwakil opened Oct 8, 2026Member·6 tasks·

  ·
  3

• [tests] Some fixes for test suite failures on TPU

  #49425·tengomucho opened Oct 8, 2026Member·

  ·
  21

• Deprecate old input names

  #49424·zucchini-nlp opened Oct 8, 2026Member·

  ·
  7

• Fix Gemma 4 tool call parsing for names containing - or .

  #49423·Varun-51 opened Oct 8, 2026·

  ·
  1

  2

• Expert-parallel dispatch hands the experts the received rows, not the all-to-all's pending result

  #49421·IlyasMoutawwakil opened Oct 8, 2026Member·6 tasks·

  ·
  2