diff --git a/miles/backends/fsdp_utils/configs/qwen_image.py b/miles/backends/fsdp_utils/configs/qwen_image.py index 73005559..13240225 100644 --- a/miles/backends/fsdp_utils/configs/qwen_image.py +++ b/miles/backends/fsdp_utils/configs/qwen_image.py @@ -114,7 +114,7 @@ def collate_cond_for_sample_batch( ) -> dict: """Pad+stack per-sample encoder_hidden_states to (M, max_len, D), build the corresponding (M, max_len) bool mask from txt_seq_lens, and - list-concat img_shapes / txt_seq_lens. Mask isn't transmitted from + list-concat img_shapes. Mask isn't transmitted from rollout — it is fully derivable from txt_seq_lens which is. When ``pad_to_len`` is given, pad to ``max(max(seq_lens), pad_to_len)`` @@ -165,7 +165,6 @@ def collate_cond_for_sample_batch( return { "encoder_hidden_states": encoder_hidden_states, "encoder_hidden_states_mask": mask, - "txt_seq_lens": seq_lens, "img_shapes": img_shapes, } diff --git a/requirements.txt b/requirements.txt index 2d8f27ca..fd93a32f 100644 --- a/requirements.txt +++ b/requirements.txt @@ -2,7 +2,10 @@ accelerate==1.12.0 addict==2.4.0 blobfile==3.0.0 datasets==4.4.2 -diffusers==0.38.0 +# MiniMax H3 (PR #14355) is merged into diffusers main but ships in no release yet: +# 0.39.0 predates the merge. Pinned to the merge commit on main; switch to the next +# stable release once it is out. +diffusers @ git+https://github.com/huggingface/diffusers.git@f53d552036a0d1bd5570782a39cd40cfabf112bc fastapi==0.128.8 httpx[http2]==0.28.1 ltx-core @ git+https://github.com/Lightricks/LTX-2.git@780984275fd47128b02bef9b5c085404276866ee#subdirectory=packages/ltx-core