update

.ai/AGENTS.md

@@ -10,34 +10,24 @@ Strive to write code as simple and explicit as possible.
 
 ---
 
-## Code formatting
+### Dependencies
+- No new mandatory dependency without discussion (e.g. `einops`)
+- Optional deps guarded with `is_X_available()` and a dummy in `utils/dummy_*.py`
 
+## Code formatting
 - `make style` and `make fix-copies` should be run as the final step before opening a PR
 
 ### Copied Code
-
 - Many classes are kept in sync with a source via a `# Copied from ...` header comment
 - Do not edit a `# Copied from` block directly — run `make fix-copies` to propagate changes from the source
 - Remove the header to intentionally break the link
 
 ### Models
-
-- See [models.md](models.md) for model conventions, attention pattern, implementation rules, dependencies, and gotchas.
-- See the [model-integration](./skills/model-integration/SKILL.md) skill for the full integration workflow, file structure, test setup, and other details.
-
-### Pipelines & Schedulers
-
-- Pipelines inherit from `DiffusionPipeline`
-- Schedulers use `SchedulerMixin` with `ConfigMixin`
-- Use `@torch.no_grad()` on pipeline `__call__`
-- Support `output_type="latent"` for skipping VAE decode
-- Support `generator` parameter for reproducibility
-- Use `self.progress_bar(timesteps)` for progress tracking
-- Don't subclass an existing pipeline for a variant — DO NOT use an existing pipeline class (e.g., `FluxPipeline`) to override another pipeline (e.g., `FluxImg2ImgPipeline`) which will be a part of the core codebase (`src`)
+- All layer calls should be visible directly in `forward` — avoid helper functions that hide `nn.Module` calls.
+- Avoid graph breaks for `torch.compile` compatibility — do not insert NumPy operations in forward implementations and any other patterns that can break `torch.compile` compatibility with `fullgraph=True`.
+- See the **model-integration** skill for the attention pattern, pipeline rules, test setup instructions, and other important details.
 
 ## Skills
 
-Task-specific guides live in `.ai/skills/` and are loaded on demand by AI agents. Available skills include:
-
-- [model-integration](./skills/model-integration/SKILL.md) (adding/converting pipelines)
-- [parity-testing](./skills/parity-testing/SKILL.md) (debugging numerical parity).
+Task-specific guides live in `.ai/skills/` and are loaded on demand by AI agents.
+Available skills: **model-integration** (adding/converting pipelines), **parity-testing** (debugging numerical parity).

.ai/models.md

@@ -1,76 +0,0 @@
-# Model conventions and rules
-
-Shared reference for model-related conventions, patterns, and gotchas.
-Linked from `AGENTS.md`, `skills/model-integration/SKILL.md`, and `review-rules.md`.
-
-## Coding style
-
-- All layer calls should be visible directly in `forward` — avoid helper functions that hide `nn.Module` calls.
-- Avoid graph breaks for `torch.compile` compatibility — do not insert NumPy operations in forward implementations and any other patterns that can break `torch.compile` compatibility with `fullgraph=True`.
-- No new mandatory dependency without discussion (e.g. `einops`). Optional deps guarded with `is_X_available()` and a dummy in `utils/dummy_*.py`.
-
-## Common model conventions
-
-- Models use `ModelMixin` with `register_to_config` for config serialization
-
-## Attention pattern
-
-Attention must follow the diffusers pattern: both the `Attention` class and its processor are defined in the model file. The processor's `__call__` handles the actual compute and must use `dispatch_attention_fn` rather than calling `F.scaled_dot_product_attention` directly. The attention class inherits `AttentionModuleMixin` and declares `_default_processor_cls` and `_available_processors`.
-
-```python
-# transformer_mymodel.py
-
-class MyModelAttnProcessor:
-    _attention_backend = None
-    _parallel_config = None
-
-    def __call__(self, attn, hidden_states, attention_mask=None, ...):
-        query = attn.to_q(hidden_states)
-        key = attn.to_k(hidden_states)
-        value = attn.to_v(hidden_states)
-        # reshape, apply rope, etc.
-        hidden_states = dispatch_attention_fn(
-            query, key, value,
-            attn_mask=attention_mask,
-            backend=self._attention_backend,
-            parallel_config=self._parallel_config,
-        )
-        hidden_states = hidden_states.flatten(2, 3)
-        return attn.to_out[0](hidden_states)
-
-
-class MyModelAttention(nn.Module, AttentionModuleMixin):
-    _default_processor_cls = MyModelAttnProcessor
-    _available_processors = [MyModelAttnProcessor]
-
-    def __init__(self, query_dim, heads=8, dim_head=64, ...):
-        super().__init__()
-        self.to_q = nn.Linear(query_dim, heads * dim_head, bias=False)
-        self.to_k = nn.Linear(query_dim, heads * dim_head, bias=False)
-        self.to_v = nn.Linear(query_dim, heads * dim_head, bias=False)
-        self.to_out = nn.ModuleList([nn.Linear(heads * dim_head, query_dim), nn.Dropout(0.0)])
-        self.set_processor(MyModelAttnProcessor())
-
-    def forward(self, hidden_states, attention_mask=None, **kwargs):
-        return self.processor(self, hidden_states, attention_mask, **kwargs)
-```
-
-Consult the implementations in `src/diffusers/models/transformers/` if you need further references.
-
-## Gotchas
-
-1. **Forgetting `__init__.py` lazy imports.** Every new class must be registered in the appropriate `__init__.py` with lazy imports. Missing this causes `ImportError` that only shows up when users try `from diffusers import YourNewClass`.
-
-2. **Using `einops` or other non-PyTorch deps.** Reference implementations often use `einops.rearrange`. Always rewrite with native PyTorch (`reshape`, `permute`, `unflatten`). Don't add the dependency. If a dependency is truly unavoidable, guard its import: `if is_my_dependency_available(): import my_dependency`.
-
-3. **Missing `make fix-copies` after `# Copied from`.** If you add `# Copied from` annotations, you must run `make fix-copies` to propagate them. CI will fail otherwise.
-
-4. **Wrong `_supports_cache_class` / `_no_split_modules`.** These class attributes control KV cache and device placement. Copy from a similar model and verify -- wrong values cause silent correctness bugs or OOM errors.
-
-5. **Missing `@torch.no_grad()` on pipeline `__call__`.** Forgetting this causes GPU OOM from gradient accumulation during inference.
-
-6. **Config serialization gaps.** Every `__init__` parameter in a `ModelMixin` subclass must be captured by `register_to_config`. If you add a new param but forget to register it, `from_pretrained` will silently use the default instead of the saved value.
-
-7. **Forgetting to update `_import_structure` and `_lazy_modules`.** The top-level `src/diffusers/__init__.py` has both -- missing either one causes partial import failures.
-
-8. **Hardcoded dtype in model forward.** Don't hardcode `torch.float32` or `torch.bfloat16` in the model's forward pass. Use the dtype of the input tensors or `self.dtype` so the model works with any precision.

.ai/review-rules.md

@@ -3,8 +3,8 @@
 Review-specific rules for Claude. Focus on correctness — style is handled by ruff.
 
 Before reviewing, read and apply the guidelines in:
-- [AGENTS.md](AGENTS.md) — coding style, copied code
-- [models.md](models.md) — model conventions, attention pattern, implementation rules, dependencies, gotchas
+- [AGENTS.md](AGENTS.md) — coding style, dependencies, copied code, model conventions
+- [skills/model-integration/SKILL.md](skills/model-integration/SKILL.md) — attention pattern, pipeline rules, implementation checklist, gotchas
 - [skills/parity-testing/SKILL.md](skills/parity-testing/SKILL.md) — testing rules, comparison utilities
 - [skills/parity-testing/pitfalls.md](skills/parity-testing/pitfalls.md) — known pitfalls (dtype mismatches, config assumptions, etc.)
 

.ai/skills/model-integration/SKILL.md

@@ -65,19 +65,89 @@ docs/source/en/api/
 - [ ] Run `make style` and `make quality`
 - [ ] Test parity with reference implementation (see `parity-testing` skill)
 
-### Model conventions, attention pattern, and implementation rules
+### Attention pattern
 
-See [../../models.md](../../models.md) for the attention pattern, implementation rules, common conventions, dependencies, and gotchas. These apply to all model work.
+Attention must follow the diffusers pattern: both the `Attention` class and its processor are defined in the model file. The processor's `__call__` handles the actual compute and must use `dispatch_attention_fn` rather than calling `F.scaled_dot_product_attention` directly. The attention class inherits `AttentionModuleMixin` and declares `_default_processor_cls` and `_available_processors`.
 
-### Model integration specific rules
+```python
+# transformer_mymodel.py
 
-**Don't combine structural changes with behavioral changes.** Restructuring code to fit diffusers APIs (ModelMixin, ConfigMixin, etc.) is unavoidable. But don't also "improve" the algorithm, refactor computation order, or rename internal variables for aesthetics. Keep numerical logic as close to the reference as possible, even if it looks unclean. For standard → modular, this is stricter: copy loop logic verbatim and only restructure into blocks. Clean up in a separate commit after parity is confirmed.
+class MyModelAttnProcessor:
+    _attention_backend = None
+    _parallel_config = None
+
+    def __call__(self, attn, hidden_states, attention_mask=None, ...):
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+        # reshape, apply rope, etc.
+        hidden_states = dispatch_attention_fn(
+            query, key, value,
+            attn_mask=attention_mask,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        return attn.to_out[0](hidden_states)
+
+
+class MyModelAttention(nn.Module, AttentionModuleMixin):
+    _default_processor_cls = MyModelAttnProcessor
+    _available_processors = [MyModelAttnProcessor]
+
+    def __init__(self, query_dim, heads=8, dim_head=64, ...):
+        super().__init__()
+        self.to_q = nn.Linear(query_dim, heads * dim_head, bias=False)
+        self.to_k = nn.Linear(query_dim, heads * dim_head, bias=False)
+        self.to_v = nn.Linear(query_dim, heads * dim_head, bias=False)
+        self.to_out = nn.ModuleList([nn.Linear(heads * dim_head, query_dim), nn.Dropout(0.0)])
+        self.set_processor(MyModelAttnProcessor())
+
+    def forward(self, hidden_states, attention_mask=None, **kwargs):
+        return self.processor(self, hidden_states, attention_mask, **kwargs)
+```
+
+Consult the implementations in `src/diffusers/models/transformers/` if you need further references.
+
+### Implementation rules
+
+1. **Don't combine structural changes with behavioral changes.** Restructuring code to fit diffusers APIs (ModelMixin, ConfigMixin, etc.) is unavoidable. But don't also "improve" the algorithm, refactor computation order, or rename internal variables for aesthetics. Keep numerical logic as close to the reference as possible, even if it looks unclean. For standard → modular, this is stricter: copy loop logic verbatim and only restructure into blocks. Clean up in a separate commit after parity is confirmed.
+2. **Pipelines must inherit from `DiffusionPipeline`.** Consult implementations in `src/diffusers/pipelines` in case you need references.
+3. **Don't subclass an existing pipeline for a variant.** DO NOT use an existing pipeline class (e.g., `FluxPipeline`) to override another pipeline (e.g., `FluxImg2ImgPipeline`) which will be a part of the core codebase (`src`).
 
 ### Test setup
 
 - Slow tests gated with `@slow` and `RUN_SLOW=1`
 - All model-level tests must use the `BaseModelTesterConfig`, `ModelTesterMixin`, `MemoryTesterMixin`, `AttentionTesterMixin`, `LoraTesterMixin`, and `TrainingTesterMixin` classes initially to write the tests. Any additional tests should be added after discussions with the maintainers. Use `tests/models/transformers/test_models_transformer_flux.py` as a reference.
 
+### Common diffusers conventions
+
+- Pipelines inherit from `DiffusionPipeline`
+- Models use `ModelMixin` with `register_to_config` for config serialization
+- Schedulers use `SchedulerMixin` with `ConfigMixin`
+- Use `@torch.no_grad()` on pipeline `__call__`
+- Support `output_type="latent"` for skipping VAE decode
+- Support `generator` parameter for reproducibility
+- Use `self.progress_bar(timesteps)` for progress tracking
+
+## Gotchas
+
+1. **Forgetting `__init__.py` lazy imports.** Every new class must be registered in the appropriate `__init__.py` with lazy imports. Missing this causes `ImportError` that only shows up when users try `from diffusers import YourNewClass`.
+
+2. **Using `einops` or other non-PyTorch deps.** Reference implementations often use `einops.rearrange`. Always rewrite with native PyTorch (`reshape`, `permute`, `unflatten`). Don't add the dependency. If a dependency is truly unavoidable, guard its import: `if is_my_dependency_available(): import my_dependency`.
+
+3. **Missing `make fix-copies` after `# Copied from`.** If you add `# Copied from` annotations, you must run `make fix-copies` to propagate them. CI will fail otherwise.
+
+4. **Wrong `_supports_cache_class` / `_no_split_modules`.** These class attributes control KV cache and device placement. Copy from a similar model and verify -- wrong values cause silent correctness bugs or OOM errors.
+
+5. **Missing `@torch.no_grad()` on pipeline `__call__`.** Forgetting this causes GPU OOM from gradient accumulation during inference.
+
+6. **Config serialization gaps.** Every `__init__` parameter in a `ModelMixin` subclass must be captured by `register_to_config`. If you add a new param but forget to register it, `from_pretrained` will silently use the default instead of the saved value.
+
+7. **Forgetting to update `_import_structure` and `_lazy_modules`.** The top-level `src/diffusers/__init__.py` has both -- missing either one causes partial import failures.
+
+8. **Hardcoded dtype in model forward.** Don't hardcode `torch.float32` or `torch.bfloat16` in the model's forward pass. Use the dtype of the input tensors or `self.dtype` so the model works with any precision.
+
 ---
 
 ## Modular Pipeline Conversion

docs/source/en/_toctree.yml

@@ -161,8 +161,6 @@
     - local: training/ddpo
       title: Reinforcement learning training with DDPO
     title: Methods
-  - local: training/nemo_automodel
-    title: NeMo Automodel
   title: Training
 - isExpanded: false
   sections:

docs/source/en/training/nemo_automodel.md

@@ -1,378 +0,0 @@
-<!--Copyright 2025 The HuggingFace Team. All rights reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
-the License. You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
-an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
-specific language governing permissions and limitations under the License.
--->
-
-# NeMo Automodel
-
-[NeMo Automodel](https://github.com/NVIDIA-NeMo/Automodel) is a PyTorch DTensor-native training library from NVIDIA for fine-tuning and pretraining diffusion models at scale. It is Hugging Face native — train any Diffusers-format model from the Hub with no checkpoint conversion. The same YAML recipe and hackable training script runs on any scale from 1 GPU to hundreds of nodes, with [FSDP2](https://pytorch.org/docs/stable/fsdp.html) distributed training, multiresolution bucketed dataloading, and pre-encoded latent space training for maximum GPU utilization. It uses [flow matching](https://huggingface.co/papers/2210.02747) for training and is fully open source (Apache 2.0), NVIDIA-supported, and actively maintained.
-
-NeMo Automodel integrates directly with Diffusers. It loads pretrained models from the Hugging Face Hub using Diffusers model classes and generates outputs with the [`DiffusionPipeline`].
-
-The typical workflow is to install NeMo Automodel (pip or Docker), prepare your data by encoding it into `.meta` files, configure a YAML recipe, launch training with `torchrun`, and run inference with the resulting checkpoint.
-
-## Supported models
-
-| Model | Hugging Face ID | Task | Parameters | Use case |
-|-------|----------------|------|------------|----------|
-| Wan 2.1 T2V 1.3B | [Wan-AI/Wan2.1-T2V-1.3B-Diffusers](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers) | Text-to-Video | 1.3B | video generation on limited hardware (fits on single 40GB A100) |
-| FLUX.1-dev | [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) | Text-to-Image | 12B | high-quality image generation |
-| HunyuanVideo 1.5 | [hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v) | Text-to-Video | 13B | high-quality video generation |
-
-## Installation
-
-### Hardware requirements
-
-| Component | Minimum | Recommended |
-|-----------|---------|-------------|
-| GPU | A100 40GB | A100 80GB / H100 |
-| GPUs | 4 | 8+ |
-| RAM | 128 GB | 256 GB+ |
-| Storage | 500 GB SSD | 2 TB NVMe |
-
-Install NeMo Automodel with pip. For the full set of installation methods (including from source), see the [NeMo Automodel installation guide](https://docs.nvidia.com/nemo/automodel/latest/guides/installation.html).
-
-```bash
-pip3 install nemo-automodel
-```
-
-Alternatively, use the pre-built Docker container which includes all dependencies.
-
-```bash
-docker pull nvcr.io/nvidia/nemo-automodel:26.02.00
-docker run --gpus all -it --rm --shm-size=8g nvcr.io/nvidia/nemo-automodel:26.02.00
-```
-
-> [!WARNING]
-> Checkpoints are lost when the container exits unless you bind-mount the checkpoint directory to the host. For example, add `-v /host/path/checkpoints:/workspace/checkpoints` to the `docker run` command.
-
-
-## Data preparation
-
-NeMo Automodel trains diffusion models in latent space. Raw images or videos must be preprocessed into `.meta` files containing VAE latents and text embeddings before training. This avoids re-encoding on every training step.
-
-Use the built-in preprocessing tool to encode your data. The tool automatically distributes work across all available GPUs.
-
-<hfoptions id="data-prep">
-<hfoption id="video preprocessing">
-
-The video preprocessing command is the same for both Wan 2.1 and HunyuanVideo, but the flags differ. Wan 2.1 uses `--processor wan` with `--resolution_preset` and `--caption_format sidecar`, while HunyuanVideo uses `--processor hunyuan` with `--target_frames` to set the frame count and `--caption_format meta_json`.
-
-**Wan 2.1:**
-
-```bash
-python -m tools.diffusion.preprocessing_multiprocess video \
-    --video_dir /data/videos \
-    --output_dir /cache \
-    --processor wan \
-    --resolution_preset 512p \
-    --caption_format sidecar
-```
-
-**HunyuanVideo:**
-
-```bash
-python -m tools.diffusion.preprocessing_multiprocess video \
-    --video_dir /data/videos \
-    --output_dir /cache \
-    --processor hunyuan \
-    --target_frames 121 \
-    --caption_format meta_json
-```
-
-</hfoption>
-<hfoption id="image preprocessing">
-
-```bash
-python -m tools.diffusion.preprocessing_multiprocess image \
-    --image_dir /data/images \
-    --output_dir /cache \
-    --processor flux \
-    --resolution_preset 512p
-```
-
-</hfoption>
-</hfoptions>
-
-### Output format
-
-Preprocessing produces a cache directory organized by resolution bucket. NeMo Automodel supports multi-resolution training through bucketed sampling. Samples are grouped by spatial resolution so each batch contains same-size samples, avoiding padding waste.
-
-```
-/cache/
-├── 512x512/                      # Resolution bucket
-│   ├── <hash1>.meta              # VAE latents + text embeddings
-│   ├── <hash2>.meta
-│   └── ...
-├── 832x480/                      # Another resolution bucket
-│   └── ...
-├── metadata.json                 # Global config (processor, model, total items)
-└── metadata_shard_0000.json      # Per-sample metadata (paths, resolutions, captions)
-```
-
-> [!TIP]
-> See the [Diffusion Dataset Preparation](https://docs.nvidia.com/nemo/automodel/latest/guides/diffusion/dataset.html) guide for caption formats, input data requirements, and all available preprocessing arguments.
-
-## Training configuration
-
-Fine-tuning is driven by two components:
-
-1. A recipe script ([finetune.py](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/diffusion/finetune/finetune.py)) is a Python entry point that contains the training loop: loading the model, building the dataloader, running forward/backward passes, computing the flow matching loss, checkpointing, and logging.
-2. A YAML configuration file specifies all settings the recipe uses: which model to fine-tune, where the data lives, optimizer hyperparameters, parallelism strategy, and more. You customize training by editing this file rather than modifying code, allowing you to scale from 1 to hundreds of GPUs.
-
-Any YAML field can also be overridden from the CLI:
-
-```bash
-torchrun --nproc-per-node=8 examples/diffusion/finetune/finetune.py \
-  -c examples/diffusion/finetune/wan2_1_t2v_flow.yaml \
-  --optim.learning_rate 1e-5 \
-  --step_scheduler.num_epochs 50
-```
-
-Below is the annotated config for fine-tuning Wan 2.1 T2V 1.3B, with each section explained.
-
-```yaml
-seed: 42
-
-# ── Experiment tracking (optional) ──────────────────────────────────────────
-# Weights & Biases integration for logging metrics, losses, and learning rates.
-# Set mode: "disabled" to turn off.
-wandb:
-  project: wan-t2v-flow-matching
-  mode: online
-  name: wan2_1_t2v_fm
-
-# ── Model ───────────────────────────────────────────────────────────────────
-# pretrained_model_name_or_path: any Hugging Face model ID or local path.
-# mode: "finetune" loads pretrained weights; "pretrain" trains from scratch.
-model:
-  pretrained_model_name_or_path: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
-  mode: finetune
-
-# ── Training schedule ───────────────────────────────────────────────────────
-# global_batch_size: effective batch across all GPUs.
-# Gradient accumulation is computed automatically: global / (local × num_gpus).
-step_scheduler:
-  global_batch_size: 8
-  local_batch_size: 1
-  ckpt_every_steps: 1000     # Save a checkpoint every N steps
-  num_epochs: 100
-  log_every: 2               # Log metrics every N steps
-
-# ── Data ────────────────────────────────────────────────────────────────────
-# _target_: the dataloader factory function.
-#   Use build_video_multiresolution_dataloader for video models (Wan, HunyuanVideo).
-#   Use build_text_to_image_multiresolution_dataloader for image models (FLUX).
-# model_type: "wan" or "hunyuan" (selects the correct latent format).
-# base_resolution: target resolution for multiresolution bucketing.
-data:
-  dataloader:
-    _target_: nemo_automodel.components.datasets.diffusion.build_video_multiresolution_dataloader
-    cache_dir: PATH_TO_YOUR_DATA
-    model_type: wan
-    base_resolution: [512, 512]
-    dynamic_batch_size: false  # When true, adjusts batch per bucket to maintain constant memory
-    shuffle: true
-    drop_last: false
-    num_workers: 0
-
-# ── Optimizer ───────────────────────────────────────────────────────────────
-# learning_rate: 5e-6 is a good starting point for fine-tuning.
-# Adjust weight_decay and betas for your dataset.
-optim:
-  learning_rate: 5e-6
-  optimizer:
-    weight_decay: 0.01
-    betas: [0.9, 0.999]
-
-# ── Learning rate scheduler ─────────────────────────────────────────────────
-# Supports cosine, linear, and constant schedules.
-lr_scheduler:
-  lr_decay_style: cosine
-  lr_warmup_steps: 0
-  min_lr: 1e-6
-
-# ── Flow matching ───────────────────────────────────────────────────────────
-# adapter_type: model-specific adapter — must match the model:
-#   "simple" for Wan 2.1, "flux" for FLUX.1-dev, "hunyuan" for HunyuanVideo.
-# timestep_sampling: "uniform" for Wan, "logit_normal" for FLUX and HunyuanVideo.
-# flow_shift: shifts the flow schedule (model-dependent).
-# i2v_prob: probability of image-to-video conditioning during training (video models).
-flow_matching:
-  adapter_type: "simple"
-  adapter_kwargs: {}
-  timestep_sampling: "uniform"
-  logit_mean: 0.0
-  logit_std: 1.0
-  flow_shift: 3.0
-  num_train_timesteps: 1000
-  i2v_prob: 0.3
-  use_loss_weighting: true
-
-# ── FSDP2 distributed training ──────────────────────────────────────────────
-# dp_size: number of GPUs for data parallelism (typically = total GPUs on node).
-# tp_size, cp_size, pp_size: tensor, context, and pipeline parallelism.
-#   For most fine-tuning, dp_size is all you need; leave others at 1.
-fsdp:
-  tp_size: 1
-  cp_size: 1
-  pp_size: 1
-  dp_replicate_size: 1
-  dp_size: 8
-
-# ── Checkpointing ──────────────────────────────────────────────────────────
-# checkpoint_dir: where to save checkpoints (use a persistent path with Docker).
-# restore_from: path to resume training from a previous checkpoint.
-checkpoint:
-  enabled: true
-  checkpoint_dir: PATH_TO_YOUR_CKPT_DIR
-  model_save_format: torch_save
-  save_consolidated: false
-  restore_from: null
-```
-
-### Config field reference
-
-The table below lists the minimal required configs. See the [NeMo Automodel examples](https://github.com/NVIDIA-NeMo/Automodel/tree/main/examples/diffusion/finetune) have full example configs for all models.
-
-| Section | Required? | What to Change |
-|---------|-----------|----------------|
-| `model` | Yes | Set `pretrained_model_name_or_path` to the Hugging Face model ID. Set `mode: finetune` or `mode: pretrain`. |
-| `step_scheduler` | Yes | `global_batch_size` is the effective batch size across all GPUs. `ckpt_every_steps` controls checkpoint frequency. Gradient accumulation is computed automatically. |
-| `data` | Yes | Set `cache_dir` to the path containing your preprocessed `.meta` files. Change `_target_` and `model_type` for different models. |
-| `optim` | Yes | `learning_rate: 5e-6` is a good default for fine-tuning. Adjust for your dataset and model. |
-| `lr_scheduler` | Yes | Choose `cosine`, `linear`, or `constant` for `lr_decay_style`. Set `lr_warmup_steps` for gradual warmup. |
-| `flow_matching` | Yes | `adapter_type` must match the model (`simple` for Wan, `flux` for FLUX, `hunyuan` for HunyuanVideo). See model-specific configs for `adapter_kwargs`. |
-| `fsdp` | Yes | Set `dp_size` to the number of GPUs. For multi-node, set to total GPUs across all nodes. |
-| `checkpoint` | Recommended | Set `checkpoint_dir` to a persistent path, especially in Docker. Use `restore_from` to resume from a previous checkpoint. |
-| `wandb` | Optional | Configure to enable Weights & Biases experiment tracking. Set `mode: disabled` to turn off. |
-
-
-
-## Launch training
-
-<hfoptions id="launch-training">
-<hfoption id="single-node">
-
-```bash
-torchrun --nproc-per-node=8 \
-  examples/diffusion/finetune/finetune.py \
-  -c examples/diffusion/finetune/wan2_1_t2v_flow.yaml
-```
-
-</hfoption>
-<hfoption id="multi-node">
-
-Run the following on each node, setting `NODE_RANK` accordingly:
-
-```bash
-export MASTER_ADDR=node0.hostname
-export MASTER_PORT=29500
-export NODE_RANK=0  # 0 on master, 1 on second node, etc.
-
-torchrun \
-  --nnodes=2 \
-  --nproc-per-node=8 \
-  --node_rank=${NODE_RANK} \
-  --rdzv_backend=c10d \
-  --rdzv_endpoint=${MASTER_ADDR}:${MASTER_PORT} \
-  examples/diffusion/finetune/finetune.py \
-  -c examples/diffusion/finetune/wan2_1_t2v_flow_multinode.yaml
-```
-
-> [!NOTE]
-> For multi-node training, set `fsdp.dp_size` in the YAML to the **total** number of GPUs across all nodes (e.g., 16 for 2 nodes with 8 GPUs each).
-
-</hfoption>
-</hfoptions>
-
-## Generation
-
-After training, generate videos or images from text prompts using the fine-tuned checkpoint.
-
-<hfoptions id="generation">
-<hfoption id="Wan 2.1">
-
-```bash
-python examples/diffusion/generate/generate.py \
-  -c examples/diffusion/generate/configs/generate_wan.yaml
-```
-
-With a fine-tuned checkpoint:
-
-```bash
-python examples/diffusion/generate/generate.py \
-  -c examples/diffusion/generate/configs/generate_wan.yaml \
-  --model.checkpoint ./checkpoints/step_1000 \
-  --inference.prompts '["A dog running on a beach"]'
-```
-
-</hfoption>
-<hfoption id="FLUX">
-
-```bash
-python examples/diffusion/generate/generate.py \
-  -c examples/diffusion/generate/configs/generate_flux.yaml
-```
-
-With a fine-tuned checkpoint:
-
-```bash
-python examples/diffusion/generate/generate.py \
-  -c examples/diffusion/generate/configs/generate_flux.yaml \
-  --model.checkpoint ./checkpoints/step_1000 \
-  --inference.prompts '["A dog running on a beach"]'
-```
-
-</hfoption>
-<hfoption id="HunyuanVideo">
-
-```bash
-python examples/diffusion/generate/generate.py \
-  -c examples/diffusion/generate/configs/generate_hunyuan.yaml
-```
-
-With a fine-tuned checkpoint:
-
-```bash
-python examples/diffusion/generate/generate.py \
-  -c examples/diffusion/generate/configs/generate_hunyuan.yaml \
-  --model.checkpoint ./checkpoints/step_1000 \
-  --inference.prompts '["A dog running on a beach"]'
-```
-
-</hfoption>
-</hfoptions>
-
-## Diffusers integration
-
-NeMo Automodel is built on top of Diffusers and uses it as the backbone for model loading and inference. It loads models directly from the Hugging Face Hub using Diffusers model classes such as [`WanTransformer3DModel`], [`FluxTransformer2DModel`], and [`HunyuanVideoTransformer3DModel`], and generates outputs via Diffusers pipelines like [`WanPipeline`] and [`FluxPipeline`].
-
-This integration provides several benefits for Diffusers users:
-
-- **No checkpoint conversion**: pretrained weights from the Hub work out of the box. Point `pretrained_model_name_or_path` at any Diffusers-format model ID and start training immediately.
-- **Day-0 model support**: when a new diffusion model is added to Diffusers and uploaded to the Hub, it can be fine-tuned with NeMo Automodel without waiting for a dedicated training script.
-- **Pipeline-compatible outputs**: fine-tuned checkpoints are saved in a format that can be loaded directly back into Diffusers pipelines for inference, sharing on the Hub, or further optimization with tools like quantization and compilation.
-- **Scalable training for Diffusers models**: NeMo Automodel adds distributed training capabilities (FSDP2, multi-node, multiresolution bucketing) that go beyond what the built-in Diffusers training scripts provide, while keeping the same model and pipeline interfaces.
-- **Shared ecosystem**: any model, LoRA adapter, or pipeline component from the Diffusers ecosystem remains compatible throughout the training and inference workflow.
-
-## NVIDIA Team
-
-- Pranav Prashant Thombre, pthombre@nvidia.com
-- Linnan Wang, linnanw@nvidia.com
-- Alexandros Koumparoulis, akoumparouli@nvidia.com
-
-## Resources
-
-- [NeMo Automodel GitHub](https://github.com/NVIDIA-NeMo/Automodel)
-- [Diffusion Fine-Tuning Guide](https://docs.nvidia.com/nemo/automodel/latest/guides/diffusion/finetune.html)
-- [Diffusion Dataset Preparation](https://docs.nvidia.com/nemo/automodel/latest/guides/diffusion/dataset.html)
-- [Diffusion Model Coverage](https://docs.nvidia.com/nemo/automodel/latest/model-coverage/diffusion.html)
-- [NeMo Automodel for Transformers (LLM/VLM fine-tuning)](https://huggingface.co/docs/transformers/en/community_integrations/nemo_automodel_finetuning)

examples/dreambooth/README_flux2.md

@@ -347,17 +347,16 @@ When LoRA was first adapted from language models to diffusion models, it was app
 More recently, SOTA text-to-image diffusion models replaced the Unet with a diffusion Transformer(DiT). With this change, we may also want to explore 
 applying LoRA training onto different types of layers and blocks. To allow more flexibility and control over the targeted modules we added `--lora_layers`- in which you can specify in a comma separated string
 the exact modules for LoRA training. Here are some examples of target modules you can provide: 
-- for attention only layers: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.to_qkv_mlp_proj"`
-- to train the same modules as in the fal trainer: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.to_qkv_mlp_proj,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.linear_in,ff.linear_out,ff_context.linear_in,ff_context.linear_out"`
-- to train the same modules as in ostris ai-toolkit / replicate trainer: `--lora_blocks="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.to_qkv_mlp_proj,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.linear_in,ff.linear_out,ff_context.linear_in,ff_context.linear_out,norm_out.linear,norm_out.proj_out"`
+- for attention only layers: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0"`
+- to train the same modules as in the fal trainer: `--lora_layers="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2"`
+- to train the same modules as in ostris ai-toolkit / replicate trainer: `--lora_blocks="attn.to_k,attn.to_q,attn.to_v,attn.to_out.0,attn.add_k_proj,attn.add_q_proj,attn.add_v_proj,attn.to_add_out,ff.net.0.proj,ff.net.2,ff_context.net.0.proj,ff_context.net.2,norm1_context.linear, norm1.linear,norm.linear,proj_mlp,proj_out"`
 > [!NOTE]
 > `--lora_layers` can also be used to specify which **blocks** to apply LoRA training to. To do so, simply add a block prefix to each layer in the comma separated string:
 > **single DiT blocks**: to target the ith single transformer block, add the prefix `single_transformer_blocks.i`, e.g. - `single_transformer_blocks.i.attn.to_k`
-> **MMDiT blocks**: to target the ith MMDiT block, add the prefix `transformer_blocks.i`, e.g. - `transformer_blocks.i.attn.to_k`
+> **MMDiT blocks**: to target the ith MMDiT block, add the prefix `transformer_blocks.i`, e.g. - `transformer_blocks.i.attn.to_k` 
 > [!NOTE]
 > keep in mind that while training more layers can improve quality and expressiveness, it also increases the size of the output LoRA weights.
-> [!NOTE]
-In FLUX2, the q, k, and v projections are fused into a single linear layer named attn.to_qkv_mlp_proj within the single transformer block. Also, the attention output is just attn.to_out, not attn.to_out.0 — it’s no longer a ModuleList like in transformer block.
+
 
 ## Training Image-to-Image
 

examples/dreambooth/train_dreambooth_lora_flux2.py

@@ -1256,13 +1256,7 @@ def main(args):
     if args.lora_layers is not None:
         target_modules = [layer.strip() for layer in args.lora_layers.split(",")]
     else:
-        # target_modules = ["to_k", "to_q", "to_v", "to_out.0"] # just train transformer_blocks
-
-        # train transformer_blocks and single_transformer_blocks
-        target_modules = ["to_k", "to_q", "to_v", "to_out.0"] + [
-            "to_qkv_mlp_proj",
-            *[f"single_transformer_blocks.{i}.attn.to_out" for i in range(48)],
-        ]
+        target_modules = ["to_k", "to_q", "to_v", "to_out.0"]
 
     # now we will add new LoRA weights the transformer layers
     transformer_lora_config = LoraConfig(

examples/dreambooth/train_dreambooth_lora_flux2_img2img.py

@@ -1206,13 +1206,7 @@ def main(args):
     if args.lora_layers is not None:
         target_modules = [layer.strip() for layer in args.lora_layers.split(",")]
     else:
-        # target_modules = ["to_k", "to_q", "to_v", "to_out.0"] # just train transformer_blocks
-
-        # train transformer_blocks and single_transformer_blocks
-        target_modules = ["to_k", "to_q", "to_v", "to_out.0"] + [
-            "to_qkv_mlp_proj",
-            *[f"single_transformer_blocks.{i}.attn.to_out" for i in range(48)],
-        ]
+        target_modules = ["to_k", "to_q", "to_v", "to_out.0"]
 
     # now we will add new LoRA weights the transformer layers
     transformer_lora_config = LoraConfig(

examples/dreambooth/train_dreambooth_lora_flux2_klein.py

@@ -1249,13 +1249,7 @@ def main(args):
     if args.lora_layers is not None:
         target_modules = [layer.strip() for layer in args.lora_layers.split(",")]
     else:
-        # target_modules = ["to_k", "to_q", "to_v", "to_out.0"] # just train transformer_blocks
-
-        # train transformer_blocks and single_transformer_blocks
-        target_modules = ["to_k", "to_q", "to_v", "to_out.0"] + [
-            "to_qkv_mlp_proj",
-            *[f"single_transformer_blocks.{i}.attn.to_out" for i in range(24)],
-        ]
+        target_modules = ["to_k", "to_q", "to_v", "to_out.0"]
 
     # now we will add new LoRA weights the transformer layers
     transformer_lora_config = LoraConfig(

examples/dreambooth/train_dreambooth_lora_flux2_klein_img2img.py

@@ -1200,13 +1200,7 @@ def main(args):
     if args.lora_layers is not None:
         target_modules = [layer.strip() for layer in args.lora_layers.split(",")]
     else:
-        # target_modules = ["to_k", "to_q", "to_v", "to_out.0"] # just train transformer_blocks
-
-        # train transformer_blocks and single_transformer_blocks
-        target_modules = ["to_k", "to_q", "to_v", "to_out.0"] + [
-            "to_qkv_mlp_proj",
-            *[f"single_transformer_blocks.{i}.attn.to_out" for i in range(24)],
-        ]
+        target_modules = ["to_k", "to_q", "to_v", "to_out.0"]
 
     # now we will add new LoRA weights the transformer layers
     transformer_lora_config = LoraConfig(

src/diffusers/models/attention_dispatch.py

@@ -862,23 +862,23 @@ def _native_attention_backward_op(
     key.requires_grad_(True)
     value.requires_grad_(True)
 
-    with torch.enable_grad():
-        query_t, key_t, value_t = (x.permute(0, 2, 1, 3) for x in (query, key, value))
-        out = torch.nn.functional.scaled_dot_product_attention(
-            query=query_t,
-            key=key_t,
-            value=value_t,
-            attn_mask=ctx.attn_mask,
-            dropout_p=ctx.dropout_p,
-            is_causal=ctx.is_causal,
-            scale=ctx.scale,
-            enable_gqa=ctx.enable_gqa,
-        )
-        out = out.permute(0, 2, 1, 3)
+    query_t, key_t, value_t = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+    out = torch.nn.functional.scaled_dot_product_attention(
+        query=query_t,
+        key=key_t,
+        value=value_t,
+        attn_mask=ctx.attn_mask,
+        dropout_p=ctx.dropout_p,
+        is_causal=ctx.is_causal,
+        scale=ctx.scale,
+        enable_gqa=ctx.enable_gqa,
+    )
+    out = out.permute(0, 2, 1, 3)
 
-        grad_query_t, grad_key_t, grad_value_t = torch.autograd.grad(
-            outputs=out, inputs=[query_t, key_t, value_t], grad_outputs=grad_out, retain_graph=False
-        )
+    grad_out_t = grad_out.permute(0, 2, 1, 3)
+    grad_query_t, grad_key_t, grad_value_t = torch.autograd.grad(
+        outputs=out, inputs=[query_t, key_t, value_t], grad_outputs=grad_out_t, retain_graph=False
+    )
 
     grad_query = grad_query_t.permute(0, 2, 1, 3)
     grad_key = grad_key_t.permute(0, 2, 1, 3)

src/diffusers/pipelines/pipeline_utils.py

@@ -764,7 +764,7 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
         token = kwargs.pop("token", None)
         revision = kwargs.pop("revision", None)
         from_flax = kwargs.pop("from_flax", False)
-        torch_dtype = kwargs.pop("torch_dtype", None)
+        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
         custom_pipeline = kwargs.pop("custom_pipeline", None)
         custom_revision = kwargs.pop("custom_revision", None)
         provider = kwargs.pop("provider", None)

tests/models/autoencoders/testing_utils.py

@@ -44,9 +44,9 @@ class AutoencoderTesterMixin:
             if isinstance(output, dict):
                 output = output.to_tuple()[0]
 
-        assert output is not None
+        self.assertIsNotNone(output)
         expected_shape = inputs_dict["sample"].shape
-        assert output.shape == expected_shape, "Input and output shapes do not match"
+        self.assertEqual(output.shape, expected_shape, "Input and output shapes do not match")
 
     def test_enable_disable_tiling(self):
         if not hasattr(self.model_class, "enable_tiling"):

tests/models/testing_utils/parallelism.py

@@ -98,64 +98,6 @@ def _context_parallel_worker(rank, world_size, master_port, model_class, init_di
             dist.destroy_process_group()
 
 
-def _context_parallel_backward_worker(
-    rank, world_size, master_port, model_class, init_dict, cp_dict, inputs_dict, return_dict
-):
-    """Worker function for context parallel backward pass testing."""
-    try:
-        # Set up distributed environment
-        os.environ["MASTER_ADDR"] = "localhost"
-        os.environ["MASTER_PORT"] = str(master_port)
-        os.environ["RANK"] = str(rank)
-        os.environ["WORLD_SIZE"] = str(world_size)
-
-        # Get device configuration
-        device_config = DEVICE_CONFIG.get(torch_device, DEVICE_CONFIG["cuda"])
-        backend = device_config["backend"]
-        device_module = device_config["module"]
-
-        # Initialize process group
-        dist.init_process_group(backend=backend, rank=rank, world_size=world_size)
-
-        # Set device for this process
-        device_module.set_device(rank)
-        device = torch.device(f"{torch_device}:{rank}")
-
-        # Create model in training mode
-        model = model_class(**init_dict)
-        model.to(device)
-        model.train()
-
-        # Move inputs to device
-        inputs_on_device = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in inputs_dict.items()}
-
-        # Enable context parallelism
-        cp_config = ContextParallelConfig(**cp_dict)
-        model.enable_parallelism(config=cp_config)
-
-        # Run forward and backward pass
-        output = model(**inputs_on_device, return_dict=False)[0]
-        loss = output.sum()
-        loss.backward()
-
-        # Check that backward actually produced at least one valid gradient
-        grads = [p.grad for p in model.parameters() if p.requires_grad and p.grad is not None]
-        has_valid_grads = len(grads) > 0 and all(torch.isfinite(g).all() for g in grads)
-
-        # Only rank 0 reports results
-        if rank == 0:
-            return_dict["status"] = "success"
-            return_dict["has_valid_grads"] = bool(has_valid_grads)
-
-    except Exception as e:
-        if rank == 0:
-            return_dict["status"] = "error"
-            return_dict["error"] = str(e)
-    finally:
-        if dist.is_initialized():
-            dist.destroy_process_group()
-
-
 def _custom_mesh_worker(
     rank,
     world_size,
@@ -262,51 +204,6 @@ class ContextParallelTesterMixin:
     def test_context_parallel_batch_inputs(self, cp_type):
         self.test_context_parallel_inference(cp_type, batch_size=2)
 
-    @pytest.mark.parametrize("cp_type", ["ulysses_degree", "ring_degree"], ids=["ulysses", "ring"])
-    def test_context_parallel_backward(self, cp_type, batch_size: int = 1):
-        if not torch.distributed.is_available():
-            pytest.skip("torch.distributed is not available.")
-
-        if not hasattr(self.model_class, "_cp_plan") or self.model_class._cp_plan is None:
-            pytest.skip("Model does not have a _cp_plan defined for context parallel inference.")
-
-        if cp_type == "ring_degree":
-            active_backend, _ = _AttentionBackendRegistry.get_active_backend()
-            if active_backend == AttentionBackendName.NATIVE:
-                pytest.skip("Ring attention is not supported with the native attention backend.")
-
-        world_size = 2
-        init_dict = self.get_init_dict()
-        inputs_dict = self.get_dummy_inputs(batch_size=batch_size)
-
-        # Move all tensors to CPU for multiprocessing
-        inputs_dict = {k: v.cpu() if isinstance(v, torch.Tensor) else v for k, v in inputs_dict.items()}
-        cp_dict = {cp_type: world_size}
-
-        # Find a free port for distributed communication
-        master_port = _find_free_port()
-
-        # Use multiprocessing manager for cross-process communication
-        manager = mp.Manager()
-        return_dict = manager.dict()
-
-        # Spawn worker processes
-        mp.spawn(
-            _context_parallel_backward_worker,
-            args=(world_size, master_port, self.model_class, init_dict, cp_dict, inputs_dict, return_dict),
-            nprocs=world_size,
-            join=True,
-        )
-
-        assert return_dict.get("status") == "success", (
-            f"Context parallel backward pass failed: {return_dict.get('error', 'Unknown error')}"
-        )
-        assert return_dict.get("has_valid_grads"), "Context parallel backward pass did not produce valid gradients."
-
-    @pytest.mark.parametrize("cp_type", ["ulysses_degree", "ring_degree"], ids=["ulysses", "ring"])
-    def test_context_parallel_backward_batch_inputs(self, cp_type):
-        self.test_context_parallel_backward(cp_type, batch_size=2)
-
     @pytest.mark.parametrize(
         "cp_type,mesh_shape,mesh_dim_names",
         [