[bugfix]: Add compatibility handling for Qwen3.5 GatedDeltaNet padding-free training and fix create_causal_mask patch when cache_positions removed in transformers >5.3.0#202

Draft

meichangsu1 wants to merge 6 commits into

modelscope:mainfrom

meichangsu1:padding_free_bufix_ljl

Commits on May 22, 2026

feat: add support for forward methods with incompatible kwargs
qq_30035749
committed
fix: support native padding-free in GatedDeltaNet and improve kwargs handling
qq_30035749
committed
fix(sequence_parallel): restore global query length for no-cache prefill path
qq_30035749
committed

Commits on May 24, 2026

fix: remove native padding-free version check for GDN patch
qq_30035749
committed

Commits on May 25, 2026

fix: conditionally patch chunk_gated_delta_rule for transformers < 5.9.0
qq_30035749
committed
fix gemini code review
qq_30035749
committed