padding-free / packed-sequence support for Qwen3.5.#186

Merged

meichangsu1 merged 7 commits intomodelscope:mainfrom

meichangsu1:qwen3_5_padding_free_ljl

May 7, 2026

Commits on Apr 30, 2026

refactor: replace `is_packed` with `padding_free` flag for clarity
meichangsu1
committed
refactor: simplify logits handling in forward methods
meichangsu1
committed

Commits on May 3, 2026

wip
meichangsu1
committed

Commits on May 6, 2026

fix: add missing blank line in linear_attention_sp.py for PEP8 compliance
meichangsu1
committed
refactor: clean up Qwen3.5 GDN padding-free patch by removing unused helpers and renaming function
meichangsu1
committed
refactor(sequence_parallel): remove automatic packed metadata caching and rely on explicit position_ids
meichangsu1
committed

Commits on May 7, 2026

refactor(processor): rename Qwen35-specific patch to generic GDN patch
meichangsu1
committed