Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Keep token fields when DiffusionTrainer converts a config (#13335)
Vineeth Sai Varikuntlac16236b28894
diff --git a/tests/diffusion/test_diffusion_masked.py b/tests/diffusion/test_diffusion_masked.pyindex 313e5f38495..06b8befbdc0 100644--- a/tests/diffusion/test_diffusion_masked.py+++ b/tests/diffusion/test_diffusion_masked.py@@ -394,6 +394,33 @@ def test_mask_builder_signature_exposed_from_wrapper_closure():assert list(inspect.signature(wrapper).parameters)[:2] == ["config", "inputs_embeds"]+def test_converted_config_keeps_token_fields(monkeypatch, tmp_path):+ from transformers import TrainingArguments+ from trl import SFTConfig+ from unsloth.models.diffusion_trainer import DiffusionConfig, DiffusionTrainer++ seen = {}+ monkeypatch.setattr(+ DiffusionTrainer.__bases__[0],+ "__init__",+ lambda self, model, args, *a, **k: seen.update(args = args),+ )+ for args in (+ SFTConfig(output_dir = str(tmp_path), report_to = "none", hub_token = "hf_x"),+ TrainingArguments(output_dir = str(tmp_path), report_to = "none", hub_token = "hf_x"),+ ):+ DiffusionTrainer(model = None, args = args, diffusion_profile = _profile("dream"))+ assert isinstance(seen["args"], DiffusionConfig)+ assert seen["args"].eos_token is None and seen["args"].pad_token is None+ assert seen["args"].hub_token == "hf_x"++ ParallelismConfig = pytest.importorskip("accelerate.parallelism_config").ParallelismConfig+ parallelism = ParallelismConfig()+ args = SFTConfig(output_dir = str(tmp_path), report_to = "none", parallelism_config = parallelism)+ DiffusionTrainer(model = None, args = args, diffusion_profile = _profile("dream"))+ assert seen["args"].parallelism_config is parallelism++_TINY_TRAIN = r"""import math, os, sys, torchimport unslothdiff --git a/unsloth/models/diffusion_trainer.py b/unsloth/models/diffusion_trainer.pyindex e611babe470..4a56423fb87 100644--- a/unsloth/models/diffusion_trainer.py+++ b/unsloth/models/diffusion_trainer.py@@ -4,7 +4,7 @@"""SFT for text-diffusion language models: SFTTrainer data handling, the family's own denoising loss."""from collections import defaultdict-from dataclasses import dataclass, field+from dataclasses import dataclass, field, fieldsfrom typing import Optionalimport torch@@ -78,11 +78,13 @@ class DiffusionTrainer(SFTTrainer):if args is None:args = DiffusionConfig(output_dir = "outputs")elif not isinstance(args, DiffusionConfig):+ # Not to_dict(): it masks every *_token field (eos_token "<EOS_TOKEN>", which SFTTrainer+ # rejects) and serialises objects such as parallelism_config, which then get dropped.args = DiffusionConfig(**{- k: v- for k, v in args.to_dict().items()- if k in DiffusionConfig.__dataclass_fields__+ f.name: getattr(args, f.name)+ for f in fields(args)+ if f.init and f.name in DiffusionConfig.__dataclass_fields__})args.packing = False