[Discrete Diffusion] Add LLaDA2 pipeline#13226
Open
kashif wants to merge 17 commits intohuggingface:mainfrom
Open
[Discrete Diffusion] Add LLaDA2 pipeline#13226kashif wants to merge 17 commits intohuggingface:mainfrom
kashif wants to merge 17 commits intohuggingface:mainfrom
Conversation
…usion Add support for LLaDA2/LLaDA2.1 discrete diffusion text generation: - BlockRefinementPipeline: block-wise iterative refinement with confidence-based token commitment, supporting editing threshold for LLaDA2.1 models - LLaDA2Pipeline: convenience wrapper with LLaDA2-specific defaults - DiscreteDiffusionPipelineMixin: shared SAR sampling utilities (top-k, top-p, temperature) and prompt/prefix helpers - compute_confidence_aware_loss: CAP-style training loss - Examples: sampling scripts for LLaDA2 and block refinement, training scripts with Qwen causal LM - Docs and tests included
Extract the confidence-based token commit logic from BlockRefinementPipeline into a dedicated BlockRefinementScheduler, following diffusers conventions. The scheduler owns: - Transfer schedule computation (get_num_transfer_tokens) - Timestep management (set_timesteps) - Step logic: confidence-based mask-filling and optional token editing The pipeline now delegates scheduling to self.scheduler.step() and accepts a scheduler parameter in __init__.
12 tests covering set_timesteps, get_num_transfer_tokens, step logic (confidence-based commits, threshold behavior, editing, prompt masking, batched inputs, tuple output).
- Add BlockRefinement and LLaDA2 to docs sidebar navigation - Add BlockRefinementScheduler to schedulers sidebar navigation - Move scheduler autodoc to its own page under api/schedulers/
kashif
changed the title
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
- Add --revision argument for loading model revisions from the Hub - Replace deprecated torch_dtype with dtype for transformers 5.x compat
LLaDA2 models expect a boolean-style (1/0) attention mask, not an additive (0/-inf) mask. The model internally converts to additive, so passing 0/-inf caused double-masking and gibberish output.
…ment.py - Remove toy train_block_refinement_cap.py (self-contained demo with tiny model) - Rename train_block_refinement_qwen_cap.py to train_block_refinement.py (already works with any causal LM via AutoModelForCausalLM) - Fix torch_dtype deprecation and update README with correct script names
- Add usage examples with real model IDs and working code - Add recommended parameters table for LLaDA2.1 quality/speed modes - Note that editing is LLaDA2.1-only (not for LLaDA2.0 models) - Remove misleading config defaults section from BlockRefinement docs
- threshold: 0.95 -> 0.7 (quality mode) - max_post_steps: 0 -> 16 (recommended for LLaDA2.1, harmless for 2.0) - eos_early_stop: False -> True (stop at EOS token) block_length=32, steps=32, temperature=0.0 were already correct. editing_threshold remains None (users enable for LLaDA2.1 models).
LLaDA2.1 is the current generation. Users with LLaDA2.0 models can disable editing by passing editing_threshold=None.
- top_p filtering: add shift-right to preserve at least one token above threshold (matches official code line 1210) - temperature ordering: apply scaling before top-k/top-p filtering so filtering operates on scaled logits (matches official code lines 1232-1235) - greedy branch: return argmax directly when temperature=0 without filtering (matches official code lines 1226-1230)
…put_ids LLaDA2Pipeline._prepare_prompt_ids was a near-copy of DiscreteDiffusionPipelineMixin._prepare_input_ids. Remove the duplicate and call the mixin method directly. Also simplify _extract_input_ids since we always pass return_dict=True.
…ings - Update EXAMPLE_DOC_STRING to use dtype= and LLaDA2.1-mini model ID - Fix sample_block_refinement.py to use dtype=
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add support for LLaDA2/LLaDA2.1 discrete diffusion text generation:
What does this PR do?
Fixes # (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.