Change8

v5.11.0

📦 transformersView on GitHub →
6 features🐛 20 fixes🔧 16 symbols

Summary

This release introduces two major new models, DiffusionGemma and DeepSeek-V3.2, alongside significant enhancements to the custom kernel API for better fusion capabilities. Numerous bug fixes address issues across various models, CI stability, and dependency compatibility.

Migration Steps

  1. If using FalconMamba fast-path, ensure you install kernels via 'pip install kernels' instead of relying on older instructions.
  2. If encountering issues with Mamba2 CUDA kernel path for NemotronH/Zamba2, be aware of potential dtype mismatches in 'out_proj'.

✨ New Features

  • Added DiffusionGemma model, which uses an encoder-decoder architecture and multi-canvas sampling for faster inference.
  • Added DeepSeek-V3.2-Exp model featuring DeepSeek Sparse Attention (DSA) for efficient long-context handling.
  • Extended the KernelConfig API to support n-to-1 module fusion and parameter transformation for custom kernel integration.
  • Added fine-grained fp8/fp4 Triton kernel support.
  • Added 'max_requests_per_batch' parameter for Continuous Batching (CB).
  • Added cosmos3 model.

🐛 Bug Fixes

  • Resolved a dtype mismatch in the Mamba2 CUDA kernel path for NemotronH/Zamba2 (specifically 'out_proj').
  • Corrected the FalconMamba fast-path warning to recommend 'pip install kernels' instead of 'mamba-ssm'.
  • Fixed model parallel beam search bugs in Qwen2-VL, Qwen2.5-VL, and Qwen3-VL MoE model families.
  • Fixed offsets in processing.
  • Fixed buggy action sha pin.
  • Fixed trailing comma bug in DataCollatorForLanguageModeling example.
  • Fixed missing Gemma4Processor._compute_audio_num_tokens.
  • Fixed InternVL models.
  • Reduced tokens in test_compile_static_cache to avoid flaky bfloat16 drift for afmoe.
  • Fixed a regression introduced by PR #45534.
  • Used torchvision's native LANCZOS interpolation instead of PIL fallback.
  • Fixed tests: 'Cohere2MoeModel' object has no attribute 'hf_device_map'.
  • Fixed noisy logging from image_processing module aliases issue.
  • Fixed conversion script and config bugs for gemma4_unified.
  • Fixed crashes when fork is not possible in Continuous Batching (CB).
  • Fixed missing f-string prefixes in error messages.
  • Fixed deepspeed docker issues.
  • Fixed conversion for clip models.
  • Fixed flip_back graph break.
  • Fixed vLLM x v5 compatibility issue by setting default untied embeddings in AudioFlamingo3 and VibeVoice.

Affected Symbols