padding-free / packed-sequence support for Qwen3.5.

[meichangsu1]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

This PR focuses on three sequence-parallel / padding-free fixes for Qwen3.5 in the Transformers backend.

Main changes:

• Add Qwen3.5 support for padding-free / packed inputs.
  • Introduce a dedicated Qwen3.5 （Qwen3.6） GatedDeltaNet padding-free patch.
  • Pass explicit packed-sequence metadata (cu_seq_lens_q, cu_seq_lens_k, max_length_q, max_length_k) for packed Qwen3.5 inputs.
  • Make Qwen3.5 linear attention use flash-linear-attention kernels with packed cu_seqlens when padding-free is enabled.
• Fix gather_loss_tensors to remove sequence-parallel padding before loss computation.
  • Trim SP/RP-added padding from gathered logps and labels.
  • Ensure packed/padding-free loss computation uses only real tokens after gather.
• Add a non-padding-free fallback path for Qwen3.5 GatedDeltaNet when flash-linear-attention is unavailable.
  • If padding-free is not enabled, fall back to torch-native GatedDeltaNet computation.
  • This keeps non-packed Qwen3.5 sequence-parallel training usable without requiring FLA.

Experiment results

[gemini-code-assist]

Code Review

This pull request introduces support for padding-free and packed sequence inputs for Qwen 3.5 models, specifically targeting GatedDeltaNet and linear attention within a sequence parallel context. Key changes include a new patching mechanism for Qwen 3.5, refactored attention logic to handle variable sequence lengths without padding, and fallback implementations for linear attention kernels when specialized libraries are missing. Feedback highlights a regression in how sequence boundaries are determined in the attention strategy and identifies inconsistencies in the return types and activation handling within the new torch-based fallback for causal convolution.