[Bugfix][HunyuanImage3] Fix offline AR garbage output by switching to Instruct chat template#3243

Merged

Gaohan123 merged 17 commits into

vllm-project:mainfrom

TaffyOfficial:feature/hunyuan-t2t-sdpa-fa

May 5, 2026

Commits on Apr 30, 2026

fix(hunyuan_image3): handle list pixel_values from Siglip2 in transformers>=5.x
zuiho-kai
committed
fix(hunyuan_image3): use instruct chat format for T2T prompt
zuiho-kai
committed
fix(hunyuan_image3): use instruct chat template for all chat tasks
zuiho-kai
committed
fix(hunyuan_image3): use Assistant: prefix to match HF tokenizer
zuiho-kai
committed
fix(hunyuan_image3): segment-tokenize prompt to bypass BPE merges
zuiho-kai
committed
docs(hunyuan_image3): document timestep-slot equivalence with HF
zuiho-kai
committed
docs(hunyuan_image3): document image preprocessing alignment with HF
zuiho-kai
committed
fix(hunyuan_image3): cast VAE pixels at encoder boundary, not in processor
zuiho-kai
committed
fix(hunyuan_image3): route MoE in fp32 to match HF reference
zuiho-kai
committed
fix(hunyuan_image3): stop AR-only output at </think> for i2t/t2t
zuiho-kai
committed
fix(hunyuan_image3): import SharedFusedMoE
zuiho-kai
committed
refactor(hunyuan_image3): extract prompt_utils as shared builder
zuiho-kai
committed
test(hunyuan_image3): regression tests for AR prompt template (PR #3243)
zuiho-kai
committed
test(hunyuan_image3): apply ruff format to prompt_utils regression tests
zuiho-kai
committed

Commits on May 5, 2026