Merge branch 'main' into overhaul-release-workflow

* remove str option for quantization config in torchao

* Apply style fixes

* minor fixes

* Added AOBaseConfig docs to torchao.md

* minor fixes for removing str option torchao

* minor change to add back int and uint check

* minor fixes

* minor fixes to tests

* Update tests/quantization/torchao/test_torchao.py

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update docs/source/en/quantization/torchao.md

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update tests/quantization/torchao/test_torchao.py

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* version=2 update to test_torchao.py

---------

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

.github/workflows/pypi_publish.yaml

@@ -1,73 +1,45 @@
-# Adapted from https://blog.deepjyoti30.dev/pypi-release-github-action
-
 name: PyPI release
 
 on:
   workflow_dispatch:
   push:
     tags:
-      - "*"
+      - "v*"
 
 jobs:
-  find-and-checkout-latest-branch:
+  build-and-test:
     runs-on: ubuntu-22.04
-    outputs:
-      latest_branch: ${{ steps.set_latest_branch.outputs.latest_branch }}
     steps:
-      - name: Checkout Repo
+      - name: Checkout repo
         uses: actions/checkout@v6
 
       - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: '3.10'
+          python-version: "3.10"
 
-      - name: Fetch latest branch
-        id: fetch_latest_branch
+      - name: Fetch and checkout latest release branch
         run: |
           pip install -U requests packaging
           LATEST_BRANCH=$(python utils/fetch_latest_release_branch.py)
           echo "Latest branch: $LATEST_BRANCH"
-          echo "latest_branch=$LATEST_BRANCH" >> $GITHUB_ENV
+          git fetch origin "$LATEST_BRANCH"
+          git checkout "$LATEST_BRANCH"
 
-      - name: Set latest branch output
-        id: set_latest_branch
-        run: echo "::set-output name=latest_branch::${{ env.latest_branch }}"
-
-  release:
-    needs: find-and-checkout-latest-branch
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Checkout Repo
-        uses: actions/checkout@v6
-        with:
-          ref: ${{ needs.find-and-checkout-latest-branch.outputs.latest_branch }}
-
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          python-version: "3.10"
-
-      - name: Install dependencies
+      - name: Install build dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install -U setuptools wheel twine
+          pip install -U build
           pip install -U torch --index-url https://download.pytorch.org/whl/cpu
 
       - name: Build the dist files
-        run: python setup.py bdist_wheel && python setup.py sdist
+        run: python -m build
 
-      - name: Publish to the test PyPI
-        env:
-          TWINE_USERNAME: ${{ secrets.TEST_PYPI_USERNAME }}
-          TWINE_PASSWORD: ${{ secrets.TEST_PYPI_PASSWORD }}
-        run: twine upload dist/* -r pypitest --repository-url=https://test.pypi.org/legacy/
+      - name: Install from built wheel
+        run: pip install dist/*.whl
 
       - name: Test installing diffusers and importing
         run: |
-          pip install diffusers && pip uninstall diffusers -y
-          pip install -i https://test.pypi.org/simple/ diffusers
           pip install -U transformers
           python utils/print_env.py
           python -c "from diffusers import __version__; print(__version__)"
@@ -75,8 +47,26 @@ jobs:
           python -c "from diffusers import DiffusionPipeline; pipe = DiffusionPipeline.from_pretrained('hf-internal-testing/tiny-stable-diffusion-pipe', safety_checker=None); pipe('ah suh du')"
           python -c "from diffusers import *"
 
+      - name: Upload build artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: python-dist
+          path: dist/
+
+  publish-to-pypi:
+    needs: build-and-test
+    if: startsWith(github.ref, 'refs/tags/')
+    runs-on: ubuntu-22.04
+    environment: pypi-release
+    permissions:
+      id-token: write
+
+    steps:
+      - name: Download build artifacts
+        uses: actions/download-artifact@v4
+        with:
+          name: python-dist
+          path: dist/
+
       - name: Publish to PyPI
-        env:
-          TWINE_USERNAME: ${{ secrets.PYPI_USERNAME }}
-          TWINE_PASSWORD: ${{ secrets.PYPI_PASSWORD }}
-        run: twine upload dist/* -r pypi
+        uses: pypa/gh-action-pypi-publish@release/v1

docs/source/en/api/pipelines/cogvideox.md

@@ -41,16 +41,15 @@ The quantized CogVideoX 5B model below requires ~16GB of VRAM.
 
 ```py
 import torch
-from diffusers import CogVideoXPipeline, AutoModel
+from diffusers import CogVideoXPipeline, AutoModel, TorchAoConfig
 from diffusers.quantizers import PipelineQuantizationConfig
 from diffusers.hooks import apply_group_offloading
 from diffusers.utils import export_to_video
+from torchao.quantization import Int8WeightOnlyConfig
 
 # quantize weights to int8 with torchao
 pipeline_quant_config = PipelineQuantizationConfig(
-  quant_backend="torchao",
-  quant_kwargs={"quant_type": "int8wo"},
-  components_to_quantize="transformer"
+  quant_mapping={"transformer": TorchAoConfig(Int8WeightOnlyConfig())}
 )
 
 # fp8 layerwise weight-casting

docs/source/en/api/pipelines/ltx2.md

@@ -18,7 +18,7 @@
   <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
 </div>
 
-LTX-2 is a DiT-based audio-video foundation model designed to generate synchronized video and audio within a single model. It brings together the core building blocks of modern video generation, with open weights and a focus on practical, local execution.
+[LTX-2](https://hf.co/papers/2601.03233) is a DiT-based foundation model designed to generate synchronized video and audio within a single model. It brings together the core building blocks of modern video generation, with open weights and a focus on practical, local execution.
 
 You can find all the original LTX-Video checkpoints under the [Lightricks](https://huggingface.co/Lightricks) organization.
 
@@ -293,6 +293,7 @@ import torch
 from diffusers import LTX2ConditionPipeline
 from diffusers.pipelines.ltx2.pipeline_ltx2_condition import LTX2VideoCondition
 from diffusers.pipelines.ltx2.export_utils import encode_video
+from diffusers.pipelines.ltx2.utils import DEFAULT_NEGATIVE_PROMPT
 from diffusers.utils import load_image, load_video
 
 device = "cuda"
@@ -315,19 +316,6 @@ prompt = (
     "landscape is characterized by rugged terrain and a river visible in the distance. The scene captures the "
     "solitude and beauty of a winter drive through a mountainous region."
 )
-negative_prompt = (
-    "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, "
-    "grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, "
-    "deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, "
-    "wrong hand count, artifacts around text, inconsistent perspective, camera shake, incorrect depth of "
-    "field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent "
-    "lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny "
-    "valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, wrong gaze direction, "
-    "mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, "
-    "off-sync audio, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward "
-    "pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, flat lighting, "
-    "inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts."
-)
 
 cond_video = load_video(
     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cosmos/cosmos-video2world-input-vid.mp4"
@@ -343,7 +331,7 @@ frame_rate = 24.0
 video, audio = pipe(
     conditions=conditions,
     prompt=prompt,
-    negative_prompt=negative_prompt,
+    negative_prompt=DEFAULT_NEGATIVE_PROMPT,
     width=width,
     height=height,
     num_frames=121,
@@ -366,6 +354,154 @@ encode_video(
 
 Because the conditioning is done via latent frames, the 8 data space frames corresponding to the specified latent frame for an image condition will tend to be static.
 
+## Multimodal Guidance
+
+LTX-2.X pipelines support multimodal guidance. It is composed of three terms, all using a CFG-style update rule:
+
+1. Classifier-Free Guidance (CFG): standard [CFG](https://huggingface.co/papers/2207.12598) where the perturbed ("weaker") output is generated using the negative prompt.
+2. Spatio-Temporal Guidance (STG): [STG](https://huggingface.co/papers/2411.18664) moves away from a perturbed output created from short-cutting self-attention operations and substitutes in the attention values instead. The idea is that this creates sharper videos and better spatiotemporal consistency.
+3. Modality Isolation Guidance: moves away from a perturbed output created from disabling cross-modality (audio-to-video and video-to-audio) cross attention. This guidance is more specific to [LTX-2.X](https://huggingface.co/papers/2601.03233) models, with the idea that this produces better consistency between the generated audio and video.
+
+These are controlled by the `guidance_scale`, `stg_scale`, and `modality_scale` arguments and can be set separately for video and audio. Additionally, for STG the transformer block indices where self-attention is skipped needs to be specified via the `spatio_temporal_guidance_blocks` argument. The LTX-2.X pipelines also support [guidance rescaling](https://huggingface.co/papers/2305.08891) to help reduce over-exposure, which can be a problem when the guidance scales are set to high values.
+
+```py
+import torch
+from diffusers import LTX2ImageToVideoPipeline
+from diffusers.pipelines.ltx2.export_utils import encode_video
+from diffusers.pipelines.ltx2.utils import DEFAULT_NEGATIVE_PROMPT
+from diffusers.utils import load_image
+
+device = "cuda"
+width = 768
+height = 512
+random_seed = 42
+frame_rate = 24.0
+generator = torch.Generator(device).manual_seed(random_seed)
+model_path = "dg845/LTX-2.3-Diffusers"
+
+pipe = LTX2ImageToVideoPipeline.from_pretrained(model_path, torch_dtype=torch.bfloat16)
+pipe.enable_sequential_cpu_offload(device=device)
+pipe.vae.enable_tiling()
+
+prompt = (
+    "An astronaut hatches from a fragile egg on the surface of the Moon, the shell cracking and peeling apart in "
+    "gentle low-gravity motion. Fine lunar dust lifts and drifts outward with each movement, floating in slow arcs "
+    "before settling back onto the ground. The astronaut pushes free in a deliberate, weightless motion, small "
+    "fragments of the egg tumbling and spinning through the air. In the background, the deep darkness of space subtly "
+    "shifts as stars glide with the camera's movement, emphasizing vast depth and scale. The camera performs a "
+    "smooth, cinematic slow push-in, with natural parallax between the foreground dust, the astronaut, and the "
+    "distant starfield. Ultra-realistic detail, physically accurate low-gravity motion, cinematic lighting, and a "
+    "breath-taking, movie-like shot."
+)
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg",
+)
+
+video, audio = pipe(
+    image=image,
+    prompt=prompt,
+    negative_prompt=DEFAULT_NEGATIVE_PROMPT,
+    width=width,
+    height=height,
+    num_frames=121,
+    frame_rate=frame_rate,
+    num_inference_steps=30,
+    guidance_scale=3.0,  # Recommended LTX-2.3 guidance parameters
+    stg_scale=1.0,  # Note that 0.0 (not 1.0) means that STG is disabled (all other guidance is disabled at 1.0)
+    modality_scale=3.0,
+    guidance_rescale=0.7,
+    audio_guidance_scale=7.0,  # Note that a higher CFG guidance scale is recommended for audio
+    audio_stg_scale=1.0,
+    audio_modality_scale=3.0,
+    audio_guidance_rescale=0.7,
+    spatio_temporal_guidance_blocks=[28],
+    use_cross_timestep=True,
+    generator=generator,
+    output_type="np",
+    return_dict=False,
+)
+
+encode_video(
+    video[0],
+    fps=frame_rate,
+    audio=audio[0].float().cpu(),
+    audio_sample_rate=pipe.vocoder.config.output_sampling_rate,
+    output_path="ltx2_3_i2v_stage_1.mp4",
+)
+```
+
+## Prompt Enhancement
+
+The LTX-2.X models are sensitive to prompting style. Refer to the [official prompting guide](https://ltx.io/model/model-blog/prompting-guide-for-ltx-2) for recommendations on how to write a good prompt. Using prompt enhancement, where the supplied prompts are enhanced using the pipeline's text encoder (by default a [Gemma 3](https://huggingface.co/google/gemma-3-12b-it-qat-q4_0-unquantized) model) given a system prompt, can also improve sample quality. The optional `processor` pipeline component needs to be present to use prompt enhancement. Enable prompt enhancement by supplying a `system_prompt` argument:
+
+
+```py
+import torch
+from transformers import Gemma3Processor
+from diffusers import LTX2Pipeline
+from diffusers.pipelines.ltx2.export_utils import encode_video
+from diffusers.pipelines.ltx2.utils import DEFAULT_NEGATIVE_PROMPT, T2V_DEFAULT_SYSTEM_PROMPT
+
+device = "cuda"
+width = 768
+height = 512
+random_seed = 42
+frame_rate = 24.0
+generator = torch.Generator(device).manual_seed(random_seed)
+model_path = "dg845/LTX-2.3-Diffusers"
+
+pipe = LTX2Pipeline.from_pretrained(model_path, torch_dtype=torch.bfloat16)
+pipe.enable_model_cpu_offload(device=device)
+pipe.vae.enable_tiling()
+if getattr(pipe, "processor", None) is None:
+    processor = Gemma3Processor.from_pretrained("google/gemma-3-12b-it-qat-q4_0-unquantized")
+    pipe.processor = processor
+
+prompt = (
+    "An astronaut hatches from a fragile egg on the surface of the Moon, the shell cracking and peeling apart in "
+    "gentle low-gravity motion. Fine lunar dust lifts and drifts outward with each movement, floating in slow arcs "
+    "before settling back onto the ground. The astronaut pushes free in a deliberate, weightless motion, small "
+    "fragments of the egg tumbling and spinning through the air. In the background, the deep darkness of space subtly "
+    "shifts as stars glide with the camera's movement, emphasizing vast depth and scale. The camera performs a "
+    "smooth, cinematic slow push-in, with natural parallax between the foreground dust, the astronaut, and the "
+    "distant starfield. Ultra-realistic detail, physically accurate low-gravity motion, cinematic lighting, and a "
+    "breath-taking, movie-like shot."
+)
+
+video, audio = pipe(
+    prompt=prompt,
+    negative_prompt=DEFAULT_NEGATIVE_PROMPT,
+    width=width,
+    height=height,
+    num_frames=121,
+    frame_rate=frame_rate,
+    num_inference_steps=30,
+    guidance_scale=3.0,
+    stg_scale=1.0,
+    modality_scale=3.0,
+    guidance_rescale=0.7,
+    audio_guidance_scale=7.0,
+    audio_stg_scale=1.0,
+    audio_modality_scale=3.0,
+    audio_guidance_rescale=0.7,
+    spatio_temporal_guidance_blocks=[28],
+    use_cross_timestep=True,
+    system_prompt=T2V_DEFAULT_SYSTEM_PROMPT,
+    generator=generator,
+    output_type="np",
+    return_dict=False,
+)
+
+encode_video(
+    video[0],
+    fps=frame_rate,
+    audio=audio[0].float().cpu(),
+    audio_sample_rate=pipe.vocoder.config.output_sampling_rate,
+    output_path="ltx2_3_t2v_stage_1.mp4",
+)
+```
+
 ## LTX2Pipeline
 
 [[autodoc]] LTX2Pipeline

docs/source/en/quantization/torchao.md

@@ -29,24 +29,7 @@ from diffusers import DiffusionPipeline, PipelineQuantizationConfig, TorchAoConf
 from torchao.quantization import Int8WeightOnlyConfig
 
 pipeline_quant_config = PipelineQuantizationConfig(
-    quant_mapping={"transformer": TorchAoConfig(Int8WeightOnlyConfig(group_size=128)))}
-)
-pipeline = DiffusionPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-dev",
-    quantization_config=pipeline_quant_config,
-    torch_dtype=torch.bfloat16,
-    device_map="cuda"
-)
-```
-
-For simple use cases, you could also provide a string identifier in [`TorchAo`] as shown below.
-
-```py
-import torch
-from diffusers import DiffusionPipeline, PipelineQuantizationConfig, TorchAoConfig
-
-pipeline_quant_config = PipelineQuantizationConfig(
-    quant_mapping={"transformer": TorchAoConfig("int8wo")}
+    quant_mapping={"transformer": TorchAoConfig(Int8WeightOnlyConfig(group_size=128, version=2))}
 )
 pipeline = DiffusionPipeline.from_pretrained(
     "black-forest-labs/FLUX.1-dev",
@@ -91,18 +74,15 @@ Weight-only quantization stores the model weights in a specific low-bit data typ
 
 Dynamic activation quantization stores the model weights in a low-bit dtype, while also quantizing the activations on-the-fly to save additional memory. This lowers the memory requirements from model weights, while also lowering the memory overhead from activation computations. However, this may come at a quality tradeoff at times, so it is recommended to test different models thoroughly.
 
-The quantization methods supported are as follows:
+Refer to the [official torchao documentation](https://docs.pytorch.org/ao/stable/index.html) for a better understanding of the available quantization methods. An exhaustive list of configuration options are available [here](https://docs.pytorch.org/ao/main/workflows/inference.html#inference-workflows).
 
-| **Category** | **Full Function Names** | **Shorthands** |
-|--------------|-------------------------|----------------|
-| **Integer quantization** | `int4_weight_only`, `int8_dynamic_activation_int4_weight`, `int8_weight_only`, `int8_dynamic_activation_int8_weight` | `int4wo`, `int4dq`, `int8wo`, `int8dq` |
-| **Floating point 8-bit quantization** | `float8_weight_only`, `float8_dynamic_activation_float8_weight`, `float8_static_activation_float8_weight` | `float8wo`, `float8wo_e5m2`, `float8wo_e4m3`, `float8dq`, `float8dq_e4m3`, `float8dq_e4m3_tensor`, `float8dq_e4m3_row` |
-| **Floating point X-bit quantization** | `fpx_weight_only` | `fpX_eAwB` where `X` is the number of bits (1-7), `A` is exponent bits, and `B` is mantissa bits. Constraint: `X == A + B + 1` |
-| **Unsigned Integer quantization** | `uintx_weight_only` | `uint1wo`, `uint2wo`, `uint3wo`, `uint4wo`, `uint5wo`, `uint6wo`, `uint7wo` |
+Some example popular quantization configurations are as follows:
 
-Some quantization methods are aliases (for example, `int8wo` is the commonly used shorthand for `int8_weight_only`). This allows using the quantization methods described in the torchao docs as-is, while also making it convenient to remember their shorthand notations.
-
-Refer to the [official torchao documentation](https://docs.pytorch.org/ao/stable/index.html) for a better understanding of the available quantization methods and the exhaustive list of configuration options available.
+| **Category** | **Configuration Classes** |
+|---|---|
+| **Integer quantization** | [`Int4WeightOnlyConfig`](https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int4WeightOnlyConfig.html), [`Int8WeightOnlyConfig`](https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int8WeightOnlyConfig.html), [`Int8DynamicActivationInt8WeightConfig`](https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Int8DynamicActivationInt8WeightConfig.html) |
+| **Floating point 8-bit quantization** | [`Float8WeightOnlyConfig`](https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Float8WeightOnlyConfig.html), [`Float8DynamicActivationFloat8WeightConfig`](https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.Float8DynamicActivationFloat8WeightConfig.html) |
+| **Unsigned integer quantization** | [`IntxWeightOnlyConfig`](https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.IntxWeightOnlyConfig.html) |
 
 ## Serializing and Deserializing quantized models
 
@@ -111,8 +91,9 @@ To serialize a quantized model in a given dtype, first load the model with the d
 ```python
 import torch
 from diffusers import AutoModel, TorchAoConfig
+from torchao.quantization import Int8WeightOnlyConfig
 
-quantization_config = TorchAoConfig("int8wo")
+quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
 transformer = AutoModel.from_pretrained(
     "black-forest-labs/Flux.1-Dev",
     subfolder="transformer",
@@ -137,18 +118,19 @@ image = pipe(prompt, num_inference_steps=30, guidance_scale=7.0).images[0]
 image.save("output.png")
 ```
 
-If you are using `torch<=2.6.0`, some quantization methods, such as `uint4wo`, cannot be loaded directly and may result in an `UnpicklingError` when trying to load the models, but work as expected when saving them. In order to work around this, one can load the state dict manually into the model. Note, however, that this requires using `weights_only=False` in `torch.load`, so it should be run only if the weights were obtained from a trustable source.
+If you are using `torch<=2.6.0`, some quantization methods, such as `uint4` weight-only, cannot be loaded directly and may result in an `UnpicklingError` when trying to load the models, but work as expected when saving them. In order to work around this, one can load the state dict manually into the model. Note, however, that this requires using `weights_only=False` in `torch.load`, so it should be run only if the weights were obtained from a trustable source.
 
 ```python
 import torch
 from accelerate import init_empty_weights
 from diffusers import FluxPipeline, AutoModel, TorchAoConfig
+from torchao.quantization import IntxWeightOnlyConfig
 
 # Serialize the model
 transformer = AutoModel.from_pretrained(
     "black-forest-labs/Flux.1-Dev",
     subfolder="transformer",
-    quantization_config=TorchAoConfig("uint4wo"),
+    quantization_config=TorchAoConfig(IntxWeightOnlyConfig(dtype=torch.uint4)),
     torch_dtype=torch.bfloat16,
 )
 transformer.save_pretrained("/path/to/flux_uint4wo", safe_serialization=False, max_shard_size="50GB")

examples/profiling/README.md

@@ -1,250 +0,0 @@
-# Profiling Plan: Diffusers Pipeline Profiling with torch.profiler
-
-Education materials to strategically profile pipelines to potentially improve their
-runtime with `torch.compile`. To set these pipelines up for success with `torch.compile`,
-we often have to get rid of DtoH syncs, CPU overheads, kernel launch delays, and
-graph breaks. In this context, profiling serves that purpose for us.
-
-Thanks to Claude Code for paircoding! We acknowledge the [Claude of OSS](https://claude.com/contact-sales/claude-for-oss) support provided to us.
-
-## Table of contents
-
-* [Context](#context)
-* [Target pipelines](#target-pipelines)
-* [Approach taken](#approach)
-* [Verification](#verification)
-* [Interpretation](#interpreting-traces-in-perfetto-ui)
-* [Taking profiling-guided steps for improvements](#afterwards)
-
-Jump to the "Verification" section to get started right away.
-
-## Context
-
-We want to uncover CPU overhead, CPU-GPU sync points, and other bottlenecks in popular diffusers pipelines — especially issues that become non-trivial under `torch.compile`. The approach is inspired by [flux-fast's run_benchmark.py](https://github.com/huggingface/flux-fast/blob/0a1dcc91658f0df14cd7fce862a5c8842784c6da/run_benchmark.py#L66-L85) which uses `torch.profiler` with method-level annotations, and motivated by issues like [diffusers#11696](https://github.com/huggingface/diffusers/pull/11696) (DtoH sync from scheduler `.item()` call).
-
-## Target Pipelines
-
-| Pipeline | Type | Checkpoint | Steps |
-|----------|------|-----------|-------|
-| `FluxPipeline` | text-to-image | `black-forest-labs/FLUX.1-dev` | 2 |
-| `Flux2KleinPipeline` | text-to-image | `black-forest-labs/FLUX.2-klein-base-9B` | 2 |
-| `WanPipeline` | text-to-video | `Wan-AI/Wan2.1-T2V-14B-Diffusers` | 2 |
-| `LTX2Pipeline` | text-to-video | `Lightricks/LTX-2` | 2 |
-| `QwenImagePipeline` | text-to-image | `Qwen/Qwen-Image` | 2 |
-
-> [!NOTE]
-> We use realistic inference call hyperparameters that mimic how these pipelines will be actually used. This
-> include using classifier-free guidance (where applicable), reasonable dimensions such 1024x1024, etc.
-> But we keep the overall running time to a bare minimum (hence 2 `num_inference_steps`). 
-
-## Approach
-
-Follow the flux-fast pattern: **annotate key pipeline methods** with `torch.profiler.record_function` wrappers, then run the pipeline under `torch.profiler.profile` and export a Chrome trace.
-
-### New Files
-
-```
-profiling/
-  profiling_utils.py       # Annotation helper + profiler setup
-  profiling_pipelines.py   # CLI entry point with pipeline configs
-```
-
-### Step 1: `profiling_utils.py` — Annotation and Profiler Infrastructure
-
-**A) `annotate(func, name)` helper** (same pattern as flux-fast):
-
-```python
-def annotate(func, name):
-    """Wrap a function with torch.profiler.record_function for trace annotation."""
-    @functools.wraps(func)
-    def wrapper(*args, **kwargs):
-        with torch.profiler.record_function(name):
-            return func(*args, **kwargs)
-    return wrapper
-```
-
-**B) `annotate_pipeline(pipe)` function** — applies annotations to key methods on any pipeline:
-
-- `pipe.transformer.forward` → `"transformer_forward"`
-- `pipe.vae.decode` → `"vae_decode"` (if present)
-- `pipe.vae.encode` → `"vae_encode"` (if present)
-- `pipe.scheduler.step` → `"scheduler_step"`
-- `pipe.encode_prompt` → `"encode_prompt"` (if present, for full-pipeline profiling)
-
-This is non-invasive — it monkey-patches bound methods without modifying source.
-
-**C) `PipelineProfiler` class:**
-
-- `__init__(pipeline_config, output_dir, mode="eager"|"compile")`
-- `setup_pipeline()` → loads from pretrained, optionally compiles transformer, calls `annotate_pipeline()`
-- `run()`:
-  1. Warm up with 1 unannotated run
-  2. Profile 1 run with `torch.profiler.profile`:
-     - `activities=[CPU, CUDA]`
-     - `record_shapes=True`
-     - `profile_memory=True`
-     - `with_stack=True`
-  3. Export Chrome trace JSON
-  4. Print `key_averages()` summary table (sorted by CUDA time) to stdout
-
-### Step 2: `profiling_pipelines.py` — CLI with Pipeline Configs
-
-**Pipeline config registry** — each entry specifies:
-
-- `pipeline_cls`, `pretrained_model_name_or_path`, `torch_dtype`
-- `call_kwargs` with pipeline-specific defaults:
-
-| Pipeline | Resolution | Frames | Steps | Extra |
-|----------|-----------|--------|-------|-------|
-| Flux | 1024x1024 | — | 2 | `guidance_scale=3.5` |
-| Flux2Klein | 1024x1024 | — | 2 | `guidance_scale=3.5` |
-| Wan | 480x832 | 81 | 2 | — |
-| LTX2 | 768x512 | 121 | 2 | `guidance_scale=4.0` |
-| QwenImage | 1024x1024 | — | 2 | `true_cfg_scale=4.0` |
-
-All configs use `output_type="latent"` by default (skip VAE decode for cleaner denoising-loop traces).
-
-**CLI flags:**
-
-- `--pipeline flux|flux2|wan|ltx2|qwenimage|all`
-- `--mode eager|compile|both`
-- `--output_dir profiling_results/`
-- `--num_steps N` (override, default 4)
-- `--full_decode` (switch output_type from `"latent"` to `"pil"` to include VAE)
-- `--compile_mode default|reduce-overhead|max-autotune`
-- `--compile_fullgraph` flag
-
-**Output:** `{output_dir}/{pipeline}_{mode}.json` Chrome trace + stdout summary.
-
-### Step 3: Known Sync Issues to Validate
-
-The profiling should surface these known/suspected issues:
-
-1. **Scheduler DtoH sync via `nonzero().item()`** — For Flux, this was fixed by adding `scheduler.set_begin_index(0)` before the denoising loop ([diffusers#11696](https://github.com/huggingface/diffusers/pull/11696)). Profiling should reveal whether similar sync points exist in other pipelines.
-
-2. **`modulate_index` tensor rebuilt every forward in `transformer_qwenimage.py`** (line 901-905) — Python list comprehension + `torch.tensor()` each step. Minor but visible in trace.
-
-3. **Any other `.item()`, `.cpu()`, `.numpy()` calls** in the denoising loop hot path — the profiler's `with_stack=True` will surface these as CPU stalls with Python stack traces.
-
-## Verification
-
-1. Run: `python profiling/profiling_pipelines.py --pipeline flux --mode eager --num_steps 2`
-2. Verify `profiling_results/flux_eager.json` is produced
-3. Open trace in [Perfetto UI](https://ui.perfetto.dev/) — confirm:
-   - `transformer_forward` and `scheduler_step` annotations visible
-   - CPU and CUDA timelines present
-   - Stack traces visible on CPU events
-4. Run with `--mode compile` and compare trace for fewer/fused CUDA kernels
-
-You can also use the `run_profiling.sh` script to bulk launch runs for different pipelines.
-
-## Interpreting Traces in Perfetto UI
-
-Open the exported `.json` trace at [ui.perfetto.dev](https://ui.perfetto.dev/). The trace has two main rows: **CPU** (top) and **CUDA** (bottom). In Perfetto, the CPU row is typically labeled with the process/thread name (e.g., `python (PID)` or `MainThread`) and appears at the top. The CUDA row is labeled `GPU 0` (or similar) and appears below the CPU rows.
-
-**Navigation:** Use `W` to zoom in, `S` to zoom out, and `A`/`D` to pan left/right. You can also scroll to zoom and click-drag to pan. Use `Shift+scroll` to scroll vertically through rows.
-
-### What to look for
-
-**1. Gaps between CUDA kernels**
-
-Zoom into the CUDA row during the denoising loop. Ideally, GPU kernels should be back-to-back with no gaps. Gaps mean the GPU is idle waiting for the CPU to launch the next kernel. Common causes:
-- Python overhead between ops (visible as CPU slices in the CPU row during the gap)
-- DtoH sync (`.item()`, `.cpu()`) forcing the GPU to drain before the CPU can proceed
-
-**2. CPU stalls (DtoH syncs)**
-
-These appear on the **CPU row** (not the CUDA row) — they are CPU-side blocking calls that wait for the GPU to finish. Look for long slices labeled `cudaStreamSynchronize` or `cudaDeviceSynchronize`. To find them: zoom into the CPU row during a denoising step and look for unusually wide slices, or use Perfetto's search bar (press `/`) and type `cudaStreamSynchronize` to jump directly to matching events. Click on a slice — if `with_stack=True` was enabled, the bottom panel ("Current Selection") shows the Python stack trace pointing to the exact line causing the sync (e.g., a `.item()` call in the scheduler).
-
-**3. Annotated regions**
-
-Our `record_function` annotations (`transformer_forward`, `scheduler_step`, etc.) appear as labeled spans on the CPU row. This lets you quickly:
-- Measure how long each phase takes (click a span to see duration)
-- See if `scheduler_step` is disproportionately expensive relative to `transformer_forward` (it should be negligible)
-- Spot unexpected CPU work between annotated regions
-
-**4. Eager vs compile comparison**
-
-Open both traces side by side (two Perfetto tabs). Key differences to look for:
-- **Fewer, wider CUDA kernels** in compile mode (fused ops) vs many small kernels in eager
-- **Smaller CPU gaps** between kernels in compile mode (less Python dispatch overhead)
-- **CUDA kernel count per step**: to compare, zoom into a single `transformer_forward` span on the CUDA row and count the distinct kernel slices within it. In eager mode you'll typically see many narrow slices (one per op); in compile mode these fuse into fewer, wider slices. A quick way to estimate: select a time range covering one denoising step on the CUDA row — Perfetto shows the number of slices in the selection summary at the bottom. If compile mode shows a similar kernel count to eager, fusion isn't happening effectively (likely due to graph breaks).
-- **Graph breaks**: if compile mode still shows many small kernels in a section, that section likely has a graph break — check `TORCH_LOGS="+dynamo"` output for details
-
-**5. Memory timeline**
-
-In Perfetto, look for the memory counter track (if `profile_memory=True`). Spikes during the denoising loop suggest unexpected allocations per step. Steady-state memory during denoising is expected — growing memory is not.
-
-**6. Kernel launch latency**
-
-Each CUDA kernel is launched from the CPU. The CPU-side launch calls (`cudaLaunchKernel`) appear as small slices on the **CPU row** — zoom in closely to a denoising step to see them. The corresponding GPU-side kernel executions appear on the **CUDA row** directly below. You can also use Perfetto's search bar (`/`) and type `cudaLaunchKernel` to find them. The time between the CPU dispatch and the GPU kernel starting should be minimal (single-digit microseconds). If you see consistent delays > 10-20us between launch and execution:
-- The launch queue may be starved because of excessive Python work between ops
-- There may be implicit syncs forcing serialization
-- `torch.compile` should help here by batching launches — compare eager vs compile to confirm
-
-To inspect this: zoom into a single denoising step, select a CUDA kernel on the GPU row, and look at the corresponding CPU-side launch slice directly above it. The horizontal offset between them is the launch latency. In a healthy trace, CPU launch slices should be well ahead of GPU execution (the CPU is "feeding" the GPU faster than it can consume).
-
-### Quick checklist per pipeline
-
-| Question | Where to look | Healthy | Unhealthy |
-|----------|--------------|---------|-----------|
-| GPU staying busy? | CUDA row gaps | Back-to-back kernels | Frequent gaps > 100us |
-| CPU blocking on GPU? | `cudaStreamSynchronize` slices | Rare/absent during denoise | Present every step |
-| Scheduler overhead? | `scheduler_step` span duration | < 1% of step time | > 5% of step time |
-| Compile effective? | CUDA kernel count per step | Fewer large kernels | Same as eager |
-| Kernel launch latency? | CPU launch → GPU kernel offset | < 10us, CPU ahead of GPU | > 20us or CPU trailing GPU |
-| Memory stable? | Memory counter track | Flat during denoise loop | Growing per step |
-
-## Afterwards
-
-To keep the profiling iterations fast, we always used [regional compilation](https://pytorch.org/tutorials/recipes/regional_compilation.html). As one would expect the trace with compilation should show
-fewer kernel launches than its eager counterpart:
-
-TODO: show traces
-
-_(The traces above were obtained with Flux2.)_
-
-### Spotting gaps between launches
-
-Then a reasonable next step is to spot frequent gaps between kernel executions. In the compiled
-case, we don't spot any on the surface. But if we zone in, some become apparent.
-
-TODO: show gaps in a compile trace
-
-So, we provided the profile trace (with compilation) to Claude, asked it to find the instances of
-"cudaStreamSynchronize" and "cudaDeviceSynchronize", and to come up with some potential fixes.
-Claude came back pretty strong:
-
-```
-Issue 1 — Gap between transformer forwards:
-- Root cause: tqdm progress bar update() calls between steps add CPU overhead (I/O, time calculations)
-- Fix: profiling/profiling_utils.py — added pipe.set_progress_bar_config(disable=True) during profiling setup.
-This eliminates the tqdm overhead from the trace. (The remaining gap from scheduler step + Python dispatch is
-inherent to eager-mode execution and should shrink significantly under torch.compile.)
-
-Issue 2 — cudaStreamSynchronize during last transformer forward:
-- Root cause: _unpack_latents_with_ids() (called right after the denoising loop) computes h = torch.max(h_ids) +
-1 and w = torch.max(w_ids) + 1 on GPU tensors, then uses them as shape args for torch.zeros((h * w, ch), ...).
-This triggers an implicit .item() DtoH sync, blocking the CPU while the GPU is still finishing the last
-transformer forward's kernels.
-- Fix: Added height/width parameters to _unpack_latents_with_ids(), pre-computed from the known pixel dimensions
-at the call site.
-```
-
-It still didn't eliminate the gaps as expected so, we fed that back to Claude and it spotted
-something more crucial. TODO: caching context fix.
-
-With the fix applied, the improvements were visible:
-
-TODO: show before and after trace
-
-Before:  
-
-- `_set_context` total: 21.6ms (8 calls)                                                                          
-- cache_context total: 21.7ms                                                                                   
-- CPU gaps: 5,523us / 8,007us / 5,508us                                                                         
-                                                                                                            
-After:                                                                                                          
-- `_set_context` total: 0.0ms (8 calls)                                                                           
-- cache_context total: 0.1ms                                                                                    
-- CPU gaps: 158us / 2,777us / 136us  

examples/profiling/profiling_pipelines.py

@@ -1,182 +0,0 @@
-"""
-Profile diffusers pipelines with torch.profiler.
-
-Usage:
-    python profiling/profiling_pipelines.py --pipeline flux --mode eager
-    python profiling/profiling_pipelines.py --pipeline flux --mode compile
-    python profiling/profiling_pipelines.py --pipeline flux --mode both
-    python profiling/profiling_pipelines.py --pipeline all --mode eager
-    python profiling/profiling_pipelines.py --pipeline wan --mode eager --full_decode
-    python profiling/profiling_pipelines.py --pipeline flux --mode compile --num_steps 4
-"""
-
-import argparse
-import copy
-import logging
-
-import torch
-
-from profiling_utils import PipelineProfiler, PipelineProfilingConfig
-
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
-logger = logging.getLogger(__name__)
-
-PROMPT = "A cat holding a sign that says hello world"
-
-
-def build_registry():
-    """Build the pipeline config registry. Imports are deferred to avoid loading all pipelines upfront."""
-    from diffusers import FluxPipeline, Flux2KleinPipeline, WanPipeline, LTX2Pipeline, QwenImagePipeline
-
-    return {
-        "flux": PipelineProfilingConfig(
-            name="flux",
-            pipeline_cls=FluxPipeline,
-            pipeline_init_kwargs={
-                "pretrained_model_name_or_path": "black-forest-labs/FLUX.1-dev",
-                "torch_dtype": torch.bfloat16,
-            },
-            pipeline_call_kwargs={
-                "prompt": PROMPT,
-                "height": 1024,
-                "width": 1024,
-                "num_inference_steps": 4,
-                "guidance_scale": 3.5,
-                "output_type": "latent",
-            },
-        ),
-        "flux2": PipelineProfilingConfig(
-            name="flux2",
-            pipeline_cls=Flux2KleinPipeline,
-            pipeline_init_kwargs={
-                "pretrained_model_name_or_path": "black-forest-labs/FLUX.2-klein-base-9B",
-                "torch_dtype": torch.bfloat16,
-            },
-            pipeline_call_kwargs={
-                "prompt": PROMPT,
-                "height": 1024,
-                "width": 1024,
-                "num_inference_steps": 4,
-                "guidance_scale": 3.5,
-                "output_type": "latent",
-            },
-        ),
-        "wan": PipelineProfilingConfig(
-            name="wan",
-            pipeline_cls=WanPipeline,
-            pipeline_init_kwargs={
-                "pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-14B-Diffusers",
-                "torch_dtype": torch.bfloat16,
-            },
-            pipeline_call_kwargs={
-                "prompt": PROMPT,
-                "negative_prompt": "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards",
-                "height": 480,
-                "width": 832,
-                "num_frames": 81,
-                "num_inference_steps": 4,
-                "output_type": "latent",
-            },
-        ),
-        "ltx2": PipelineProfilingConfig(
-            name="ltx2",
-            pipeline_cls=LTX2Pipeline,
-            pipeline_init_kwargs={
-                "pretrained_model_name_or_path": "Lightricks/LTX-2",
-                "torch_dtype": torch.bfloat16,
-            },
-            pipeline_call_kwargs={
-                "prompt": PROMPT,
-                "negative_prompt": "worst quality, inconsistent motion, blurry, jittery, distorted",
-                "height": 512,
-                "width": 768,
-                "num_frames": 121,
-                "num_inference_steps": 4,
-                "guidance_scale": 4.0,
-                "output_type": "latent",
-            },
-        ),
-        "qwenimage": PipelineProfilingConfig(
-            name="qwenimage",
-            pipeline_cls=QwenImagePipeline,
-            pipeline_init_kwargs={
-                "pretrained_model_name_or_path": "Qwen/Qwen-Image",
-                "torch_dtype": torch.bfloat16,
-            },
-            pipeline_call_kwargs={
-                "prompt": PROMPT,
-                "negative_prompt": " ",
-                "height": 1024,
-                "width": 1024,
-                "num_inference_steps": 4,
-                "true_cfg_scale": 4.0,
-                "output_type": "latent",
-            },
-        ),
-    }
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Profile diffusers pipelines with torch.profiler")
-    parser.add_argument(
-        "--pipeline",
-        choices=["flux", "flux2", "wan", "ltx2", "qwenimage", "all"],
-        required=True,
-        help="Which pipeline to profile",
-    )
-    parser.add_argument(
-        "--mode",
-        choices=["eager", "compile", "both"],
-        default="eager",
-        help="Run in eager mode, compile mode, or both",
-    )
-    parser.add_argument("--output_dir", default="profiling_results", help="Directory for trace output")
-    parser.add_argument("--num_steps", type=int, default=None, help="Override num_inference_steps")
-    parser.add_argument("--full_decode", action="store_true", help="Profile including VAE decode (output_type='pil')")
-    parser.add_argument(
-        "--compile_mode",
-        default="default",
-        choices=["default", "reduce-overhead", "max-autotune"],
-        help="torch.compile mode",
-    )
-    parser.add_argument("--compile_fullgraph", action="store_true", help="Use fullgraph=True for torch.compile")
-    parser.add_argument(
-        "--compile_regional",
-        action="store_true",
-        help="Use compile_repeated_blocks() instead of full model compile",
-    )
-    args = parser.parse_args()
-
-    registry = build_registry()
-
-    pipeline_names = list(registry.keys()) if args.pipeline == "all" else [args.pipeline]
-    modes = ["eager", "compile"] if args.mode == "both" else [args.mode]
-
-    for pipeline_name in pipeline_names:
-        for mode in modes:
-            config = copy.deepcopy(registry[pipeline_name])
-
-            # Apply overrides
-            if args.num_steps is not None:
-                config.pipeline_call_kwargs["num_inference_steps"] = args.num_steps
-            if args.full_decode:
-                config.pipeline_call_kwargs["output_type"] = "pil"
-            if mode == "compile":
-                config.compile_kwargs = {
-                    "fullgraph": args.compile_fullgraph,
-                    "mode": args.compile_mode,
-                }
-                config.compile_regional = args.compile_regional
-
-            logger.info(f"Profiling {pipeline_name} in {mode} mode...")
-            profiler = PipelineProfiler(config, args.output_dir)
-            try:
-                trace_file = profiler.run()
-                logger.info(f"Done: {trace_file}")
-            except Exception as e:
-                logger.error(f"Failed to profile {pipeline_name} ({mode}): {e}")
-
-
-if __name__ == "__main__":
-    main()

examples/profiling/profiling_utils.py

@@ -1,146 +0,0 @@
-import functools
-import gc
-import logging
-import os
-from dataclasses import dataclass, field
-from typing import Any
-
-import torch
-import torch.profiler
-
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
-logger = logging.getLogger(__name__)
-
-
-def annotate(func, name):
-    """Wrap a function with torch.profiler.record_function for trace annotation."""
-
-    @functools.wraps(func)
-    def wrapper(*args, **kwargs):
-        with torch.profiler.record_function(name):
-            return func(*args, **kwargs)
-
-    return wrapper
-
-
-def annotate_pipeline(pipe):
-    """Apply profiler annotations to key pipeline methods.
-
-    Monkey-patches bound methods so they appear as named spans in the trace.
-    Non-invasive — no source modifications required.
-    """
-    annotations = [
-        ("transformer", "forward", "transformer_forward"),
-        ("vae", "decode", "vae_decode"),
-        ("vae", "encode", "vae_encode"),
-        ("scheduler", "step", "scheduler_step"),
-    ]
-
-    # Annotate sub-component methods
-    for component_name, method_name, label in annotations:
-        component = getattr(pipe, component_name, None)
-        if component is None:
-            continue
-        method = getattr(component, method_name, None)
-        if method is None:
-            continue
-        setattr(component, method_name, annotate(method, label))
-
-    # Annotate pipeline-level methods
-    if hasattr(pipe, "encode_prompt"):
-        pipe.encode_prompt = annotate(pipe.encode_prompt, "encode_prompt")
-
-
-def flush():
-    gc.collect()
-    torch.cuda.empty_cache()
-    torch.cuda.reset_max_memory_allocated()
-    torch.cuda.reset_peak_memory_stats()
-
-
-@dataclass
-class PipelineProfilingConfig:
-    name: str
-    pipeline_cls: Any
-    pipeline_init_kwargs: dict[str, Any]
-    pipeline_call_kwargs: dict[str, Any]
-    compile_kwargs: dict[str, Any] | None = field(default=None)
-    compile_regional: bool = False
-
-
-class PipelineProfiler:
-    def __init__(self, config: PipelineProfilingConfig, output_dir: str = "profiling_results"):
-        self.config = config
-        self.output_dir = output_dir
-        os.makedirs(output_dir, exist_ok=True)
-
-    def setup_pipeline(self):
-        """Load the pipeline from pretrained, optionally compile, and annotate."""
-        logger.info(f"Loading pipeline: {self.config.name}")
-        pipe = self.config.pipeline_cls.from_pretrained(**self.config.pipeline_init_kwargs)
-        pipe.to("cuda")
-
-        if self.config.compile_kwargs:
-            if self.config.compile_regional:
-                logger.info(f"Regional compilation (compile_repeated_blocks) with kwargs: {self.config.compile_kwargs}")
-                pipe.transformer.compile_repeated_blocks(**self.config.compile_kwargs)
-            else:
-                logger.info(f"Full compilation with kwargs: {self.config.compile_kwargs}")
-                pipe.transformer.compile(**self.config.compile_kwargs)
-
-        # Disable tqdm progress bar to avoid CPU overhead / IO between steps
-        pipe.set_progress_bar_config(disable=True)
-
-        annotate_pipeline(pipe)
-        return pipe
-
-    def run(self):
-        """Execute the profiling run: warmup, then profile one pipeline call."""
-        pipe = self.setup_pipeline()
-        flush()
-
-        mode = "compile" if self.config.compile_kwargs else "eager"
-        trace_file = os.path.join(self.output_dir, f"{self.config.name}_{mode}.json")
-
-        # Warmup (pipeline __call__ is already decorated with @torch.no_grad())
-        logger.info("Running warmup...")
-        pipe(**self.config.pipeline_call_kwargs)
-        flush()
-
-        # Profile
-        logger.info("Running profiled iteration...")
-        activities = [
-            torch.profiler.ProfilerActivity.CPU,
-            torch.profiler.ProfilerActivity.CUDA,
-        ]
-        with torch.profiler.profile(
-            activities=activities,
-            record_shapes=True,
-            profile_memory=True,
-            with_stack=True,
-        ) as prof:
-            with torch.profiler.record_function("pipeline_call"):
-                pipe(**self.config.pipeline_call_kwargs)
-
-        # Export trace
-        prof.export_chrome_trace(trace_file)
-        logger.info(f"Chrome trace saved to: {trace_file}")
-
-        # Print summary
-        print("\n" + "=" * 80)
-        print(f"Profile summary: {self.config.name} ({mode})")
-        print("=" * 80)
-        print(
-            prof.key_averages().table(
-                sort_by="cuda_time_total",
-                row_limit=20,
-            )
-        )
-
-        # Cleanup
-        pipe.to("cpu")
-        del pipe
-        flush()
-
-        return trace_file

examples/profiling/run_profiling.sh

@@ -1,46 +0,0 @@
-#!/bin/bash
-# Run profiling across all pipelines in eager and compile (regional) modes.
-#
-# Usage:
-#   bash profiling/run_profiling.sh
-#   bash profiling/run_profiling.sh --output_dir my_results
-
-set -euo pipefail
-
-OUTPUT_DIR="profiling_results"
-while [[ $# -gt 0 ]]; do
-    case "$1" in
-        --output_dir) OUTPUT_DIR="$2"; shift 2 ;;
-        *) echo "Unknown arg: $1"; exit 1 ;;
-    esac
-done
-NUM_STEPS=2
-# PIPELINES=("flux" "flux2" "wan" "ltx2" "qwenimage")
-PIPELINES=("flux2")
-MODES=("eager" "compile")
-
-for pipeline in "${PIPELINES[@]}"; do
-    for mode in "${MODES[@]}"; do
-        echo "============================================================"
-        echo "Profiling: ${pipeline} | mode: ${mode}"
-        echo "============================================================"
-
-        COMPILE_ARGS=""
-        if [ "$mode" = "compile" ]; then
-            COMPILE_ARGS="--compile_regional --compile_fullgraph --compile_mode default"
-        fi
-
-        python profiling/profiling_pipelines.py \
-            --pipeline "$pipeline" \
-            --mode "$mode" \
-            --output_dir "$OUTPUT_DIR" \
-            --num_steps "$NUM_STEPS" \
-            $COMPILE_ARGS
-
-        echo ""
-    done
-done
-
-echo "============================================================"
-echo "All traces saved to: ${OUTPUT_DIR}/"
-echo "============================================================"

src/diffusers/hooks/hooks.py

@@ -271,31 +271,12 @@ class HookRegistry:
             if hook._is_stateful:
                 hook._set_context(self._module_ref, name)
 
-        for registry in self._get_child_registries():
-            registry._set_context(name)
-
-    def _get_child_registries(self) -> list["HookRegistry"]:
-        """Return registries of child modules, using a cached list when available.
-
-        The cache is built on first call and reused for subsequent calls. This avoids the cost of walking the full
-        module tree via named_modules() on every _set_context call, which is significant for large models (e.g. ~2.7ms
-        per call on Flux2).
-        """
-        if not hasattr(self, "_child_registries_cache"):
-            self._child_registries_cache = None
-
-        if self._child_registries_cache is not None:
-            return self._child_registries_cache
-
-        registries = []
         for module_name, module in unwrap_module(self._module_ref).named_modules():
             if module_name == "":
                 continue
             module = unwrap_module(module)
             if hasattr(module, "_diffusers_hook"):
-                registries.append(module._diffusers_hook)
-        self._child_registries_cache = registries
-        return registries
+                module._diffusers_hook._set_context(name)
 
     def __repr__(self) -> str:
         registry_repr = ""

src/diffusers/pipelines/flux2/pipeline_flux2_klein.py

@@ -397,9 +397,7 @@ class Flux2KleinPipeline(DiffusionPipeline, Flux2LoraLoaderMixin):
 
     @staticmethod
     # Copied from diffusers.pipelines.flux2.pipeline_flux2.Flux2Pipeline._unpack_latents_with_ids
-    def _unpack_latents_with_ids(
-        x: torch.Tensor, x_ids: torch.Tensor, height: int | None = None, width: int | None = None
-    ) -> list[torch.Tensor]:
+    def _unpack_latents_with_ids(x: torch.Tensor, x_ids: torch.Tensor) -> list[torch.Tensor]:
         """
         using position ids to scatter tokens into place
         """
@@ -409,9 +407,8 @@ class Flux2KleinPipeline(DiffusionPipeline, Flux2LoraLoaderMixin):
             h_ids = pos[:, 1].to(torch.int64)
             w_ids = pos[:, 2].to(torch.int64)
 
-            # Use provided height/width to avoid DtoH sync from torch.max().item()
-            h = height if height is not None else torch.max(h_ids) + 1
-            w = width if width is not None else torch.max(w_ids) + 1
+            h = torch.max(h_ids) + 1
+            w = torch.max(w_ids) + 1
 
             flat_ids = h_ids * w + w_ids
 
@@ -898,10 +895,7 @@ class Flux2KleinPipeline(DiffusionPipeline, Flux2LoraLoaderMixin):
 
         self._current_timestep = None
 
-        # Pass pre-computed latent height/width to avoid DtoH sync from torch.max().item()
-        latent_height = 2 * (int(height) // (self.vae_scale_factor * 2))
-        latent_width = 2 * (int(width) // (self.vae_scale_factor * 2))
-        latents = self._unpack_latents_with_ids(latents, latent_ids, latent_height // 2, latent_width // 2)
+        latents = self._unpack_latents_with_ids(latents, latent_ids)
 
         latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype)
         latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(

src/diffusers/pipelines/ltx2/utils.py

@@ -1,6 +1,155 @@
+# Copyright 2026 Lightricks and 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.
+
 # Pre-trained sigma values for distilled model are taken from
 # https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py
 DISTILLED_SIGMA_VALUES = [1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875]
 
 # Reduced schedule for super-resolution stage 2 (subset of distilled values)
 STAGE_2_DISTILLED_SIGMA_VALUES = [0.909375, 0.725, 0.421875]
+
+
+# Default negative prompt from
+# https://github.com/Lightricks/LTX-2/blob/ae855f8538843825f9015a419cf4ba5edaf5eec2/packages/ltx-pipelines/src/ltx_pipelines/utils/constants.py#L131-L143
+DEFAULT_NEGATIVE_PROMPT = (
+    "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, "
+    "grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, "
+    "deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, "
+    "wrong hand count, artifacts around text, inconsistent perspective, camera shake, incorrect depth of "
+    "field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent "
+    "lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny "
+    "valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, wrong gaze direction, "
+    "mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, "
+    "off-sync audio, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward "
+    "pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, flat lighting, "
+    "inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts."
+)
+
+
+# System prompts for prompt enhancement
+# https://github.com/Lightricks/LTX-2/blob/ae855f8538843825f9015a419cf4ba5edaf5eec2/packages/ltx-core/src/ltx_core/text_encoders/gemma/encoders/prompts/gemma_t2v_system_prompt.txt#L1
+# Disable line-too-long rule in ruff to keep the prompts exactly the same (e.g. in terms of newlines)
+# Supported in ruff>=0.15.0
+# ruff: disable[E501]
+T2V_DEFAULT_SYSTEM_PROMPT = """
+You are a Creative Assistant. Given a user's raw input prompt describing a scene or concept, expand it into a detailed
+video generation prompt with specific visuals and integrated audio to guide a text-to-video model.
+
+#### Guidelines
+- Strictly follow all aspects of the user's raw input: include every element requested (style, visuals, motions,
+  actions, camera movement, audio).
+    - If the input is vague, invent concrete details: lighting, textures, materials, scene settings, etc.
+        - For characters: describe gender, clothing, hair, expressions. DO NOT invent unrequested characters.
+- Use active language: present-progressive verbs ("is walking," "speaking"). If no action specified, describe natural
+  movements.
+- Maintain chronological flow: use temporal connectors ("as," "then," "while").
+- Audio layer: Describe complete soundscape (background audio, ambient sounds, SFX, speech/music when requested).
+  Integrate sounds chronologically alongside actions. Be specific (e.g., "soft footsteps on tile"), not vague (e.g.,
+  "ambient sound is present").
+- Speech (only when requested):
+    - For ANY speech-related input (talking, conversation, singing, etc.), ALWAYS include exact words in quotes with
+      voice characteristics (e.g., "The man says in an excited voice: 'You won't believe what I just saw!'").
+    - Specify language if not English and accent if relevant.
+- Style: Include visual style at the beginning: "Style: <style>, <rest of prompt>." Default to cinematic-realistic if
+  unspecified. Omit if unclear.
+- Visual and audio only: NO non-visual/auditory senses (smell, taste, touch).
+- Restrained language: Avoid dramatic/exaggerated terms. Use mild, natural phrasing.
+    - Colors: Use plain terms ("red dress"), not intensified ("vibrant blue," "bright red").
+    - Lighting: Use neutral descriptions ("soft overhead light"), not harsh ("blinding light").
+    - Facial features: Use delicate modifiers for subtle features (i.e., "subtle freckles").
+
+#### Important notes:
+- Analyze the user's raw input carefully. In cases of FPV or POV, exclude the description of the subject whose POV is
+  requested.
+- Camera motion: DO NOT invent camera motion unless requested by the user.
+- Speech: DO NOT modify user-provided character dialogue unless it's a typo.
+- No timestamps or cuts: DO NOT use timestamps or describe scene cuts unless explicitly requested.
+- Format: DO NOT use phrases like "The scene opens with...". Start directly with Style (optional) and chronological
+  scene description.
+- Format: DO NOT start your response with special characters.
+- DO NOT invent dialogue unless the user mentions speech/talking/singing/conversation.
+- If the user's raw input prompt is highly detailed, chronological and in the requested format: DO NOT make major edits
+  or introduce new elements. Add/enhance audio descriptions if missing.
+
+#### Output Format (Strict):
+- Single continuous paragraph in natural language (English).
+- NO titles, headings, prefaces, code fences, or Markdown.
+- If unsafe/invalid, return original user prompt. Never ask questions or clarifications.
+
+Your output quality is CRITICAL. Generate visually rich, dynamic prompts with integrated audio for high-quality video
+generation.
+
+#### Example Input: "A woman at a coffee shop talking on the phone" Output: Style: realistic with cinematic lighting.
+In a medium close-up, a woman in her early 30s with shoulder-length brown hair sits at a small wooden table by the
+window. She wears a cream-colored turtleneck sweater, holding a white ceramic coffee cup in one hand and a smartphone
+to her ear with the other. Ambient cafe sounds fill the space—espresso machine hiss, quiet conversations, gentle
+clinking of cups. The woman listens intently, nodding slightly, then takes a sip of her coffee and sets it down with a
+soft clink. Her face brightens into a warm smile as she speaks in a clear, friendly voice, 'That sounds perfect! I'd
+love to meet up this weekend. How about Saturday afternoon?' She laughs softly—a genuine chuckle—and shifts in her
+chair. Behind her, other patrons move subtly in and out of focus. 'Great, I'll see you then,' she concludes cheerfully,
+lowering the phone.
+"""
+# ruff: enable[E501]
+
+# ruff: disable[E501]
+I2V_DEFAULT_SYSTEM_PROMPT = """
+You are a Creative Assistant writing concise, action-focused image-to-video prompts. Given an image (first frame) and
+user Raw Input Prompt, generate a prompt to guide video generation from that image.
+
+#### Guidelines:
+- Analyze the Image: Identify Subject, Setting, Elements, Style and Mood.
+- Follow user Raw Input Prompt: Include all requested motion, actions, camera movements, audio, and details. If in
+  conflict with the image, prioritize user request while maintaining visual consistency (describe transition from image
+  to user's scene).
+- Describe only changes from the image: Don't reiterate established visual details. Inaccurate descriptions may cause
+  scene cuts.
+- Active language: Use present-progressive verbs ("is walking," "speaking"). If no action specified, describe natural
+  movements.
+- Chronological flow: Use temporal connectors ("as," "then," "while").
+- Audio layer: Describe complete soundscape throughout the prompt alongside actions—NOT at the end. Align audio
+  intensity with action tempo. Include natural background audio, ambient sounds, effects, speech or music (when
+  requested). Be specific (e.g., "soft footsteps on tile") not vague (e.g., "ambient sound").
+- Speech (only when requested): Provide exact words in quotes with character's visual/voice characteristics (e.g., "The
+  tall man speaks in a low, gravelly voice"), language if not English and accent if relevant. If general conversation
+  mentioned without text, generate contextual quoted dialogue. (i.e., "The man is talking" input -> the output should
+  include exact spoken words, like: "The man is talking in an excited voice saying: 'You won't believe what I just
+  saw!' His hands gesture expressively as he speaks, eyebrows raised with enthusiasm. The ambient sound of a quiet room
+  underscores his animated speech.")
+- Style: Include visual style at beginning: "Style: <style>, <rest of prompt>." If unclear, omit to avoid conflicts.
+- Visual and audio only: Describe only what is seen and heard. NO smell, taste, or tactile sensations.
+- Restrained language: Avoid dramatic terms. Use mild, natural, understated phrasing.
+
+#### Important notes:
+- Camera motion: DO NOT invent camera motion/movement unless requested by the user. Make sure to include camera motion
+  only if specified in the input.
+- Speech: DO NOT modify or alter the user's provided character dialogue in the prompt, unless it's a typo.
+- No timestamps or cuts: DO NOT use timestamps or describe scene cuts unless explicitly requested.
+- Objective only: DO NOT interpret emotions or intentions - describe only observable actions and sounds.
+- Format: DO NOT use phrases like "The scene opens with..." / "The video starts...". Start directly with Style
+  (optional) and chronological scene description.
+- Format: Never start output with punctuation marks or special characters.
+- DO NOT invent dialogue unless the user mentions speech/talking/singing/conversation.
+- Your performance is CRITICAL. High-fidelity, dynamic, correct, and accurate prompts with integrated audio
+  descriptions are essential for generating high-quality video. Your goal is flawless execution of these rules.
+
+#### Output Format (Strict):
+- Single concise paragraph in natural English. NO titles, headings, prefaces, sections, code fences, or Markdown.
+- If unsafe/invalid, return original user prompt. Never ask questions or clarifications.
+
+#### Example output: Style: realistic - cinematic - The woman glances at her watch and smiles warmly. She speaks in a
+cheerful, friendly voice, "I think we're right on time!" In the background, a café barista prepares drinks at the
+counter. The barista calls out in a clear, upbeat tone, "Two cappuccinos ready!" The sound of the espresso machine
+hissing softly blends with gentle background chatter and the light clinking of cups on saucers.
+"""
+# ruff: enable[E501]

src/diffusers/quantizers/quantization_config.py

@@ -23,20 +23,17 @@ https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e
 from __future__ import annotations
 
 import copy
-import dataclasses
 import importlib.metadata
-import inspect
 import json
 import os
 import warnings
-from dataclasses import dataclass, is_dataclass
+from dataclasses import dataclass
 from enum import Enum
-from functools import partial
 from typing import Any, Callable
 
 from packaging import version
 
-from ..utils import deprecate, is_torch_available, is_torchao_available, is_torchao_version, logging
+from ..utils import deprecate, is_torch_available, is_torchao_version, logging
 
 
 if is_torch_available():
@@ -53,16 +50,6 @@ class QuantizationMethod(str, Enum):
     MODELOPT = "modelopt"
 
 
-if is_torchao_available():
-    from torchao.quantization.quant_primitives import MappingType
-
-    class TorchAoJSONEncoder(json.JSONEncoder):
-        def default(self, obj):
-            if isinstance(obj, MappingType):
-                return obj.name
-            return super().default(obj)
-
-
 @dataclass
 class QuantizationConfigMixin:
     """
@@ -446,49 +433,21 @@ class TorchAoConfig(QuantizationConfigMixin):
     """This is a config class for torchao quantization/sparsity techniques.
 
     Args:
-        quant_type (`str` | AOBaseConfig):
-            The type of quantization we want to use, currently supporting:
-                - **Integer quantization:**
-                    - Full function names: `int4_weight_only`, `int8_dynamic_activation_int4_weight`,
-                      `int8_weight_only`, `int8_dynamic_activation_int8_weight`
-                    - Shorthands: `int4wo`, `int4dq`, `int8wo`, `int8dq`
-
-                - **Floating point 8-bit quantization:**
-                    - Full function names: `float8_weight_only`, `float8_dynamic_activation_float8_weight`,
-                      `float8_static_activation_float8_weight`
-                    - Shorthands: `float8wo`, `float8wo_e5m2`, `float8wo_e4m3`, `float8dq`, `float8dq_e4m3`,
-                      `float8_e4m3_tensor`, `float8_e4m3_row`,
-
-                - **Floating point X-bit quantization:** (in torchao <= 0.14.1, not supported in torchao >= 0.15.0)
-                    - Full function names: `fpx_weight_only`
-                    - Shorthands: `fpX_eAwB`, where `X` is the number of bits (between `1` to `7`), `A` is the number
-                      of exponent bits and `B` is the number of mantissa bits. The constraint of `X == A + B + 1` must
-                      be satisfied for a given shorthand notation.
-
-                - **Unsigned Integer quantization:**
-                    - Full function names: `uintx_weight_only`
-                    - Shorthands: `uint1wo`, `uint2wo`, `uint3wo`, `uint4wo`, `uint5wo`, `uint6wo`, `uint7wo`
-                - An AOBaseConfig instance: for more advanced configuration options.
+        quant_type (`AOBaseConfig`):
+            An `AOBaseConfig` subclass instance specifying the quantization type. See the [torchao
+            documentation](https://docs.pytorch.org/ao/main/api_ref_quantization.html#inference-apis-for-quantize) for
+            available config classes (e.g. `Int4WeightOnlyConfig`, `Int8WeightOnlyConfig`, `Float8WeightOnlyConfig`,
+            `Float8DynamicActivationFloat8WeightConfig`, etc.).
         modules_to_not_convert (`list[str]`, *optional*, default to `None`):
             The list of modules to not quantize, useful for quantizing models that explicitly require to have some
             modules left in their original precision.
-        kwargs (`dict[str, Any]`, *optional*):
-            The keyword arguments for the chosen type of quantization, for example, int4_weight_only quantization
-            supports two keyword arguments `group_size` and `inner_k_tiles` currently. More API examples and
-            documentation of arguments can be found in
-            https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques
 
     Example:
         ```python
         from diffusers import FluxTransformer2DModel, TorchAoConfig
-
-        # AOBaseConfig-based configuration
         from torchao.quantization import Int8WeightOnlyConfig
 
         quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
-
-        # String-based config
-        quantization_config = TorchAoConfig("int8wo")
         transformer = FluxTransformer2DModel.from_pretrained(
             "black-forest-labs/Flux.1-Dev",
             subfolder="transformer",
@@ -500,7 +459,7 @@ class TorchAoConfig(QuantizationConfigMixin):
 
     def __init__(
         self,
-        quant_type: str | "AOBaseConfig",  # noqa: F821
+        quant_type: "AOBaseConfig",  # noqa: F821
         modules_to_not_convert: list[str] | None = None,
         **kwargs,
     ) -> None:
@@ -508,89 +467,28 @@ class TorchAoConfig(QuantizationConfigMixin):
         self.quant_type = quant_type
         self.modules_to_not_convert = modules_to_not_convert
 
-        # When we load from serialized config, "quant_type_kwargs" will be the key
-        if "quant_type_kwargs" in kwargs:
-            self.quant_type_kwargs = kwargs["quant_type_kwargs"]
-        else:
-            self.quant_type_kwargs = kwargs
-
         self.post_init()
 
     def post_init(self):
-        if not isinstance(self.quant_type, str):
-            if is_torchao_version("<=", "0.9.0"):
-                raise ValueError(
-                    f"torchao <= 0.9.0 only supports string quant_type, got {type(self.quant_type).__name__}. "
-                    f"Upgrade to torchao > 0.9.0 to use AOBaseConfig."
-                )
+        if is_torchao_version("<=", "0.9.0"):
+            raise ValueError("TorchAoConfig requires torchao > 0.9.0. Please upgrade with `pip install -U torchao`.")
 
-            from torchao.quantization.quant_api import AOBaseConfig
+        from torchao.quantization.quant_api import AOBaseConfig
 
-            if not isinstance(self.quant_type, AOBaseConfig):
-                raise TypeError(f"quant_type must be a AOBaseConfig instance, got {type(self.quant_type).__name__}")
-
-        elif isinstance(self.quant_type, str):
-            TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()
-
-            if self.quant_type not in TORCHAO_QUANT_TYPE_METHODS.keys():
-                is_floatx_quant_type = self.quant_type.startswith("fp")
-                is_float_quant_type = self.quant_type.startswith("float") or is_floatx_quant_type
-                if is_float_quant_type and not self._is_xpu_or_cuda_capability_atleast_8_9():
-                    raise ValueError(
-                        f"Requested quantization type: {self.quant_type} is not supported on GPUs with CUDA capability <= 8.9. You "
-                        f"can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`."
-                    )
-                elif is_floatx_quant_type and not is_torchao_version("<=", "0.14.1"):
-                    raise ValueError(
-                        f"Requested quantization type: {self.quant_type} is only supported in torchao <= 0.14.1. "
-                        f"Please downgrade to torchao <= 0.14.1 to use this quantization type."
-                    )
-
-                raise ValueError(
-                    f"Requested quantization type: {self.quant_type} is not supported or is an incorrect `quant_type` name. If you think the "
-                    f"provided quantization type should be supported, please open an issue at https://github.com/huggingface/diffusers/issues."
-                )
-
-            method = TORCHAO_QUANT_TYPE_METHODS[self.quant_type]
-            signature = inspect.signature(method)
-            all_kwargs = {
-                param.name
-                for param in signature.parameters.values()
-                if param.kind in [inspect.Parameter.KEYWORD_ONLY, inspect.Parameter.POSITIONAL_OR_KEYWORD]
-            }
-            unsupported_kwargs = list(self.quant_type_kwargs.keys() - all_kwargs)
-
-            if len(unsupported_kwargs) > 0:
-                raise ValueError(
-                    f'The quantization method "{self.quant_type}" does not support the following keyword arguments: '
-                    f"{unsupported_kwargs}. The following keywords arguments are supported: {all_kwargs}."
-                )
+        if not isinstance(self.quant_type, AOBaseConfig):
+            raise TypeError(f"quant_type must be an AOBaseConfig instance, got {type(self.quant_type).__name__}")
 
     def to_dict(self):
         """Convert configuration to a dictionary."""
         d = super().to_dict()
 
-        if isinstance(self.quant_type, str):
-            # Handle layout serialization if present
-            if "quant_type_kwargs" in d and "layout" in d["quant_type_kwargs"]:
-                if is_dataclass(d["quant_type_kwargs"]["layout"]):
-                    d["quant_type_kwargs"]["layout"] = [
-                        d["quant_type_kwargs"]["layout"].__class__.__name__,
-                        dataclasses.asdict(d["quant_type_kwargs"]["layout"]),
-                    ]
-                if isinstance(d["quant_type_kwargs"]["layout"], list):
-                    assert len(d["quant_type_kwargs"]["layout"]) == 2, "layout saves layout name and layout kwargs"
-                    assert isinstance(d["quant_type_kwargs"]["layout"][0], str), "layout name must be a string"
-                    assert isinstance(d["quant_type_kwargs"]["layout"][1], dict), "layout kwargs must be a dict"
-                else:
-                    raise ValueError("layout must be a list")
-        else:
-            # Handle AOBaseConfig serialization
-            from torchao.core.config import config_to_dict
+        # Handle AOBaseConfig serialization
+        from torchao.core.config import config_to_dict
 
-            # For now we assume there is 1 config per Transformer, however in the future
-            # We may want to support a config per fqn.
-            d["quant_type"] = {"default": config_to_dict(self.quant_type)}
+        # For now we assume there is 1 config per Transformer, however in the future
+        # we may want to support a config per fqn.
+        # See: https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.quantize_.html
+        d["quant_type"] = {"default": config_to_dict(self.quant_type)}
 
         return d
 
@@ -602,8 +500,6 @@ class TorchAoConfig(QuantizationConfigMixin):
         config_dict = config_dict.copy()
         quant_type = config_dict.pop("quant_type")
 
-        if isinstance(quant_type, str):
-            return cls(quant_type=quant_type, **config_dict)
         # Check if we only have one key which is "default"
         # In the future we may update this
         assert len(quant_type) == 1 and "default" in quant_type, (
@@ -618,210 +514,13 @@ class TorchAoConfig(QuantizationConfigMixin):
 
         return cls(quant_type=quant_type, **config_dict)
 
-    @classmethod
-    def _get_torchao_quant_type_to_method(cls):
-        r"""
-        Returns supported torchao quantization types with all commonly used notations.
-        """
-
-        if is_torchao_available():
-            # TODO(aryan): Support sparsify
-            from torchao.quantization import (
-                float8_dynamic_activation_float8_weight,
-                float8_static_activation_float8_weight,
-                float8_weight_only,
-                int4_weight_only,
-                int8_dynamic_activation_int4_weight,
-                int8_dynamic_activation_int8_weight,
-                int8_weight_only,
-                uintx_weight_only,
-            )
-
-            if is_torchao_version("<=", "0.14.1"):
-                from torchao.quantization import fpx_weight_only
-            # TODO(aryan): Add a note on how to use PerAxis and PerGroup observers
-            from torchao.quantization.observer import PerRow, PerTensor
-
-            def generate_float8dq_types(dtype: torch.dtype):
-                name = "e5m2" if dtype == torch.float8_e5m2 else "e4m3"
-                types = {}
-
-                for granularity_cls in [PerTensor, PerRow]:
-                    # Note: Activation and Weights cannot have different granularities
-                    granularity_name = "tensor" if granularity_cls is PerTensor else "row"
-                    types[f"float8dq_{name}_{granularity_name}"] = partial(
-                        float8_dynamic_activation_float8_weight,
-                        activation_dtype=dtype,
-                        weight_dtype=dtype,
-                        granularity=(granularity_cls(), granularity_cls()),
-                    )
-
-                return types
-
-            def generate_fpx_quantization_types(bits: int):
-                if is_torchao_version("<=", "0.14.1"):
-                    types = {}
-
-                    for ebits in range(1, bits):
-                        mbits = bits - ebits - 1
-                        types[f"fp{bits}_e{ebits}m{mbits}"] = partial(fpx_weight_only, ebits=ebits, mbits=mbits)
-
-                    non_sign_bits = bits - 1
-                    default_ebits = (non_sign_bits + 1) // 2
-                    default_mbits = non_sign_bits - default_ebits
-                    types[f"fp{bits}"] = partial(fpx_weight_only, ebits=default_ebits, mbits=default_mbits)
-
-                    return types
-                else:
-                    raise ValueError("Floating point X-bit quantization is not supported in torchao >= 0.15.0")
-
-            INT4_QUANTIZATION_TYPES = {
-                # int4 weight + bfloat16/float16 activation
-                "int4wo": int4_weight_only,
-                "int4_weight_only": int4_weight_only,
-                # int4 weight + int8 activation
-                "int4dq": int8_dynamic_activation_int4_weight,
-                "int8_dynamic_activation_int4_weight": int8_dynamic_activation_int4_weight,
-            }
-
-            INT8_QUANTIZATION_TYPES = {
-                # int8 weight + bfloat16/float16 activation
-                "int8wo": int8_weight_only,
-                "int8_weight_only": int8_weight_only,
-                # int8 weight + int8 activation
-                "int8dq": int8_dynamic_activation_int8_weight,
-                "int8_dynamic_activation_int8_weight": int8_dynamic_activation_int8_weight,
-            }
-
-            # TODO(aryan): handle torch 2.2/2.3
-            FLOATX_QUANTIZATION_TYPES = {
-                # float8_e5m2 weight + bfloat16/float16 activation
-                "float8wo": partial(float8_weight_only, weight_dtype=torch.float8_e5m2),
-                "float8_weight_only": float8_weight_only,
-                "float8wo_e5m2": partial(float8_weight_only, weight_dtype=torch.float8_e5m2),
-                # float8_e4m3 weight + bfloat16/float16 activation
-                "float8wo_e4m3": partial(float8_weight_only, weight_dtype=torch.float8_e4m3fn),
-                # float8_e5m2 weight + float8 activation (dynamic)
-                "float8dq": float8_dynamic_activation_float8_weight,
-                "float8_dynamic_activation_float8_weight": float8_dynamic_activation_float8_weight,
-                # ===== Matrix multiplication is not supported in float8_e5m2 so the following errors out.
-                # However, changing activation_dtype=torch.float8_e4m3 might work here =====
-                # "float8dq_e5m2": partial(
-                #     float8_dynamic_activation_float8_weight,
-                #     activation_dtype=torch.float8_e5m2,
-                #     weight_dtype=torch.float8_e5m2,
-                # ),
-                # **generate_float8dq_types(torch.float8_e5m2),
-                # ===== =====
-                # float8_e4m3 weight + float8 activation (dynamic)
-                "float8dq_e4m3": partial(
-                    float8_dynamic_activation_float8_weight,
-                    activation_dtype=torch.float8_e4m3fn,
-                    weight_dtype=torch.float8_e4m3fn,
-                ),
-                **generate_float8dq_types(torch.float8_e4m3fn),
-                # float8 weight + float8 activation (static)
-                "float8_static_activation_float8_weight": float8_static_activation_float8_weight,
-            }
-
-            if is_torchao_version("<=", "0.14.1"):
-                FLOATX_QUANTIZATION_TYPES.update(generate_fpx_quantization_types(3))
-                FLOATX_QUANTIZATION_TYPES.update(generate_fpx_quantization_types(4))
-                FLOATX_QUANTIZATION_TYPES.update(generate_fpx_quantization_types(5))
-                FLOATX_QUANTIZATION_TYPES.update(generate_fpx_quantization_types(6))
-                FLOATX_QUANTIZATION_TYPES.update(generate_fpx_quantization_types(7))
-
-            UINTX_QUANTIZATION_DTYPES = {
-                "uintx_weight_only": uintx_weight_only,
-                "uint1wo": partial(uintx_weight_only, dtype=torch.uint1),
-                "uint2wo": partial(uintx_weight_only, dtype=torch.uint2),
-                "uint3wo": partial(uintx_weight_only, dtype=torch.uint3),
-                "uint4wo": partial(uintx_weight_only, dtype=torch.uint4),
-                "uint5wo": partial(uintx_weight_only, dtype=torch.uint5),
-                "uint6wo": partial(uintx_weight_only, dtype=torch.uint6),
-                "uint7wo": partial(uintx_weight_only, dtype=torch.uint7),
-                # "uint8wo": partial(uintx_weight_only, dtype=torch.uint8),  # uint8 quantization is not supported
-            }
-
-            QUANTIZATION_TYPES = {}
-            QUANTIZATION_TYPES.update(INT4_QUANTIZATION_TYPES)
-            QUANTIZATION_TYPES.update(INT8_QUANTIZATION_TYPES)
-            QUANTIZATION_TYPES.update(UINTX_QUANTIZATION_DTYPES)
-
-            if cls._is_xpu_or_cuda_capability_atleast_8_9():
-                QUANTIZATION_TYPES.update(FLOATX_QUANTIZATION_TYPES)
-
-            return QUANTIZATION_TYPES
-        else:
-            raise ValueError(
-                "TorchAoConfig requires torchao to be installed, please install with `pip install torchao`"
-            )
-
-    @staticmethod
-    def _is_xpu_or_cuda_capability_atleast_8_9() -> bool:
-        if torch.cuda.is_available():
-            major, minor = torch.cuda.get_device_capability()
-            if major == 8:
-                return minor >= 9
-            return major >= 9
-        elif torch.xpu.is_available():
-            return True
-        else:
-            raise RuntimeError("TorchAO requires a CUDA compatible GPU or Intel XPU and installation of PyTorch.")
-
     def get_apply_tensor_subclass(self):
         """Create the appropriate quantization method based on configuration."""
-        if not isinstance(self.quant_type, str):
-            return self.quant_type
-        else:
-            methods = self._get_torchao_quant_type_to_method()
-            quant_type_kwargs = self.quant_type_kwargs.copy()
-            if (
-                not torch.cuda.is_available()
-                and is_torchao_available()
-                and self.quant_type == "int4_weight_only"
-                and version.parse(importlib.metadata.version("torchao")) >= version.parse("0.8.0")
-                and quant_type_kwargs.get("layout", None) is None
-            ):
-                if torch.xpu.is_available():
-                    if version.parse(importlib.metadata.version("torchao")) >= version.parse(
-                        "0.11.0"
-                    ) and version.parse(importlib.metadata.version("torch")) > version.parse("2.7.9"):
-                        from torchao.dtypes import Int4XPULayout
-                        from torchao.quantization.quant_primitives import ZeroPointDomain
-
-                        quant_type_kwargs["layout"] = Int4XPULayout()
-                        quant_type_kwargs["zero_point_domain"] = ZeroPointDomain.INT
-                    else:
-                        raise ValueError(
-                            "TorchAoConfig requires torchao >= 0.11.0 and torch >= 2.8.0 for XPU support. Please upgrade the version or use run on CPU with the cpu version pytorch."
-                        )
-                else:
-                    from torchao.dtypes import Int4CPULayout
-
-                    quant_type_kwargs["layout"] = Int4CPULayout()
-
-            return methods[self.quant_type](**quant_type_kwargs)
+        return self.quant_type
 
     def __repr__(self):
-        r"""
-        Example of how this looks for `TorchAoConfig("uint4wo", group_size=32)`:
-
-        ```
-        TorchAoConfig {
-            "modules_to_not_convert": null,
-            "quant_method": "torchao",
-            "quant_type": "uint4wo",
-            "quant_type_kwargs": {
-                "group_size": 32
-            }
-        }
-        ```
-        """
         config_dict = self.to_dict()
-        return (
-            f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True, cls=TorchAoJSONEncoder)}\n"
-        )
+        return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n"
 
 
 @dataclass

src/diffusers/quantizers/torchao/torchao_quantizer.py

@@ -20,7 +20,6 @@ https://github.com/huggingface/transformers/blob/3a8eb74668e9c2cc563b2f5c62fac17
 import importlib
 import re
 import types
-from fnmatch import fnmatch
 from typing import TYPE_CHECKING, Any
 
 from packaging import version
@@ -199,13 +198,13 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
                     )
 
     def update_torch_dtype(self, torch_dtype):
-        quant_type = self.quantization_config.quant_type
-        if isinstance(quant_type, str) and (quant_type.startswith("int") or quant_type.startswith("uint")):
-            if torch_dtype is not None and torch_dtype != torch.bfloat16:
-                logger.warning(
-                    f"You are trying to set torch_dtype to {torch_dtype} for int4/int8/uintx quantization, but "
-                    f"only bfloat16 is supported right now. Please set `torch_dtype=torch.bfloat16`."
-                )
+        config_name = self.quantization_config.quant_type.__class__.__name__
+        is_int_quant = config_name.startswith("Int") or config_name.startswith("Uint")
+        if is_int_quant and torch_dtype is not None and torch_dtype != torch.bfloat16:
+            logger.warning(
+                f"You are trying to set torch_dtype to {torch_dtype} for integer quantization, but "
+                f"only bfloat16 is supported right now. Please set `torch_dtype=torch.bfloat16`."
+            )
 
         if torch_dtype is None:
             # We need to set the torch_dtype, otherwise we have dtype mismatch when performing the quantized linear op
@@ -219,45 +218,16 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
         return torch_dtype
 
     def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype":
-        quant_type = self.quantization_config.quant_type
         from accelerate.utils import CustomDtype
 
-        if isinstance(quant_type, str):
-            if quant_type.startswith("int8"):
-                # Note that int4 weights are created by packing into torch.int8, but since there is no torch.int4, we use torch.int8
-                return torch.int8
-            elif quant_type.startswith("int4"):
-                return CustomDtype.INT4
-            elif quant_type == "uintx_weight_only":
-                return self.quantization_config.quant_type_kwargs.get("dtype", torch.uint8)
-            elif quant_type.startswith("uint"):
-                return {
-                    1: torch.uint1,
-                    2: torch.uint2,
-                    3: torch.uint3,
-                    4: torch.uint4,
-                    5: torch.uint5,
-                    6: torch.uint6,
-                    7: torch.uint7,
-                }[int(quant_type[4])]
-            elif quant_type.startswith("float") or quant_type.startswith("fp"):
-                return torch.bfloat16
+        quant_type = self.quantization_config.quant_type
+        config_name = quant_type.__class__.__name__
+        size_digit = fuzzy_match_size(config_name)
 
-        elif is_torchao_version(">", "0.9.0"):
-            from torchao.core.config import AOBaseConfig
-
-            quant_type = self.quantization_config.quant_type
-            if isinstance(quant_type, AOBaseConfig):
-                # Extract size digit using fuzzy match on the class name
-                config_name = quant_type.__class__.__name__
-                size_digit = fuzzy_match_size(config_name)
-
-                # Map the extracted digit to appropriate dtype
-                if size_digit == "4":
-                    return CustomDtype.INT4
-                else:
-                    # Default to int8
-                    return torch.int8
+        if size_digit == "4":
+            return CustomDtype.INT4
+        else:
+            return torch.int8
 
         if isinstance(target_dtype, SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION):
             return target_dtype
@@ -337,29 +307,14 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
         - Use a division factor of 8 for int4 weights
         - Use a division factor of 4 for int8 weights
         """
-        # Original mapping for non-AOBaseConfig types
-        # For the uint types, this is a best guess. Once these types become more used
-        # we can look into their nuances.
-        if is_torchao_version(">", "0.9.0"):
-            from torchao.core.config import AOBaseConfig
-
-            quant_type = self.quantization_config.quant_type
-            if isinstance(quant_type, AOBaseConfig):
-                # Extract size digit using fuzzy match on the class name
-                config_name = quant_type.__class__.__name__
-                size_digit = fuzzy_match_size(config_name)
-
-                if size_digit == "4":
-                    return 8
-                else:
-                    return 4
-
-        map_to_target_dtype = {"int4_*": 8, "int8_*": 4, "uint*": 8, "float8*": 4}
         quant_type = self.quantization_config.quant_type
-        for pattern, target_dtype in map_to_target_dtype.items():
-            if fnmatch(quant_type, pattern):
-                return target_dtype
-        raise ValueError(f"Unsupported quant_type: {quant_type!r}")
+        config_name = quant_type.__class__.__name__
+        size_digit = fuzzy_match_size(config_name)
+
+        if size_digit == "4":
+            return 8
+        else:
+            return 4
 
     def _process_model_before_weight_loading(
         self,
@@ -415,9 +370,17 @@ class TorchAoHfQuantizer(DiffusersQuantizer):
 
         return _is_torchao_serializable
 
+    _TRAINABLE_QUANTIZATION_CONFIGS = (
+        "Int8WeightOnlyConfig",
+        "Int8DynamicActivationInt8WeightConfig",
+        "Int8StaticActivationInt8WeightConfig",
+        "Float8WeightOnlyConfig",
+        "Float8DynamicActivationFloat8WeightConfig",
+    )
+
     @property
     def is_trainable(self):
-        return self.quantization_config.quant_type.startswith("int8")
+        return self.quantization_config.quant_type.__class__.__name__ in self._TRAINABLE_QUANTIZATION_CONFIGS
 
     @property
     def is_compileable(self) -> bool:

tests/models/testing_utils/quantization.py

@@ -25,7 +25,6 @@ from diffusers.utils.import_utils import (
     is_nvidia_modelopt_available,
     is_optimum_quanto_available,
     is_torchao_available,
-    is_torchao_version,
 )
 
 from ...testing_utils import (
@@ -63,8 +62,7 @@ if is_gguf_available():
     pass
 
 if is_torchao_available():
-    if is_torchao_version(">=", "0.9.0"):
-        pass
+    import torchao.quantization as _torchao_quantization
 
 
 class LoRALayer(torch.nn.Module):
@@ -806,9 +804,9 @@ class TorchAoConfigMixin:
     """
 
     TORCHAO_QUANT_TYPES = {
-        "int4wo": {"quant_type": "int4_weight_only"},
-        "int8wo": {"quant_type": "int8_weight_only"},
-        "int8dq": {"quant_type": "int8_dynamic_activation_int8_weight"},
+        "int4wo": "Int4WeightOnlyConfig",
+        "int8wo": "Int8WeightOnlyConfig",
+        "int8dq": "Int8DynamicActivationInt8WeightConfig",
     }
 
     TORCHAO_EXPECTED_MEMORY_REDUCTIONS = {
@@ -817,8 +815,13 @@ class TorchAoConfigMixin:
         "int8dq": 1.5,
     }
 
-    def _create_quantized_model(self, config_kwargs, **extra_kwargs):
-        config = TorchAoConfig(**config_kwargs)
+    @staticmethod
+    def _get_quant_config(config_name):
+        config_cls = getattr(_torchao_quantization, config_name)
+        return TorchAoConfig(config_cls())
+
+    def _create_quantized_model(self, config_name, **extra_kwargs):
+        config = self._get_quant_config(config_name)
         kwargs = getattr(self, "pretrained_model_kwargs", {}).copy()
         kwargs["quantization_config"] = config
         kwargs["device_map"] = str(torch_device)

tests/quantization/torchao/test_torchao.py

@@ -55,6 +55,20 @@ from ..test_torch_compile_utils import QuantCompileTests
 enable_full_determinism()
 
 
+def _is_xpu_or_cuda_capability_atleast_8_9() -> bool:
+    if is_torch_available():
+        import torch
+
+        if torch.cuda.is_available():
+            major, minor = torch.cuda.get_device_capability()
+            if major == 8:
+                return minor >= 9
+            return major >= 9
+        elif torch.xpu.is_available():
+            return True
+    return False
+
+
 if is_torch_available():
     import torch
     import torch.nn as nn
@@ -64,12 +78,17 @@ if is_torch_available():
 
 if is_torchao_available():
     from torchao.dtypes import AffineQuantizedTensor
+    from torchao.quantization import (
+        Float8WeightOnlyConfig,
+        Int4WeightOnlyConfig,
+        Int8DynamicActivationInt8WeightConfig,
+        Int8WeightOnlyConfig,
+    )
     from torchao.quantization.linear_activation_quantized_tensor import LinearActivationQuantizedTensor
-    from torchao.quantization.quant_primitives import MappingType
     from torchao.utils import get_model_size_in_bytes
 
-    if version.parse(importlib.metadata.version("torchao")) >= version.Version("0.9.0"):
-        from torchao.quantization import Int8WeightOnlyConfig
+    if version.parse(importlib.metadata.version("torchao")) >= version.Version("0.10.0"):
+        from torchao.quantization import Int8DynamicActivationIntxWeightConfig, IntxWeightOnlyConfig
 
 
 @require_torch
@@ -80,53 +99,30 @@ class TorchAoConfigTest(unittest.TestCase):
         """
         Makes sure the config format is properly set
         """
-        quantization_config = TorchAoConfig("int4_weight_only")
+        quantization_config = TorchAoConfig(Int4WeightOnlyConfig(version=2))
         torchao_orig_config = quantization_config.to_dict()
-
-        for key in torchao_orig_config:
-            self.assertEqual(getattr(quantization_config, key), torchao_orig_config[key])
+        self.assertIn("quant_type", torchao_orig_config)
+        self.assertIn("quant_method", torchao_orig_config)
 
     def test_post_init_check(self):
         """
-        Test kwargs validations in TorchAoConfig
+        Test that non-AOBaseConfig types are rejected
         """
-        _ = TorchAoConfig("int4_weight_only")
-        with self.assertRaisesRegex(ValueError, "is not supported"):
-            _ = TorchAoConfig("uint8")
+        _ = TorchAoConfig(Int4WeightOnlyConfig())
+        with self.assertRaises(TypeError):
+            _ = TorchAoConfig("int4_weight_only")
 
-        with self.assertRaisesRegex(ValueError, "does not support the following keyword arguments"):
-            _ = TorchAoConfig("int4_weight_only", group_size1=32)
+        with self.assertRaises(TypeError):
+            _ = TorchAoConfig(42)
 
     def test_repr(self):
         """
         Check that there is no error in the repr
         """
-        quantization_config = TorchAoConfig("int4_weight_only", modules_to_not_convert=["conv"], group_size=8)
-        expected_repr = """TorchAoConfig {
-            "modules_to_not_convert": [
-                "conv"
-            ],
-            "quant_method": "torchao",
-            "quant_type": "int4_weight_only",
-            "quant_type_kwargs": {
-                "group_size": 8
-            }
-        }""".replace(" ", "").replace("\n", "")
-        quantization_repr = repr(quantization_config).replace(" ", "").replace("\n", "")
-        self.assertEqual(quantization_repr, expected_repr)
-
-        quantization_config = TorchAoConfig("int4dq", group_size=64, act_mapping_type=MappingType.SYMMETRIC)
-        expected_repr = """TorchAoConfig {
-            "modules_to_not_convert": null,
-            "quant_method": "torchao",
-            "quant_type": "int4dq",
-            "quant_type_kwargs": {
-                "act_mapping_type": "SYMMETRIC",
-                "group_size": 64
-            }
-        }""".replace(" ", "").replace("\n", "")
-        quantization_repr = repr(quantization_config).replace(" ", "").replace("\n", "")
-        self.assertEqual(quantization_repr, expected_repr)
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig(version=2), modules_to_not_convert=["conv"])
+        quantization_repr = repr(quantization_config)
+        self.assertIn("TorchAoConfig", quantization_repr)
+        self.assertIn("torchao", quantization_repr)
 
 
 # Slices for these tests have been obtained on our aws-g6e-xlarge-plus runners
@@ -234,79 +230,30 @@ class TorchAoTest(unittest.TestCase):
         for model_id in ["hf-internal-testing/tiny-flux-pipe", "hf-internal-testing/tiny-flux-sharded"]:
             # fmt: off
             QUANTIZATION_TYPES_TO_TEST = [
-                ("int4wo", np.array([0.4648, 0.5234, 0.5547, 0.4219, 0.4414, 0.6445, 0.4336, 0.4531, 0.5625])),
-                ("int4dq", np.array([0.4688, 0.5195, 0.5547, 0.418, 0.4414, 0.6406, 0.4336, 0.4531, 0.5625])),
-                ("int8wo", np.array([0.4648, 0.5195, 0.5547, 0.4199, 0.4414, 0.6445, 0.4316, 0.4531, 0.5625])),
-                ("int8dq", np.array([0.4648, 0.5195, 0.5547, 0.4199, 0.4414, 0.6445, 0.4316, 0.4531, 0.5625])),
-                ("uint4wo", np.array([0.4609, 0.5234, 0.5508, 0.4199, 0.4336, 0.6406, 0.4316, 0.4531, 0.5625])),
-                ("uint7wo", np.array([0.4648, 0.5195, 0.5547, 0.4219, 0.4414, 0.6445, 0.4316, 0.4531, 0.5625])),
+                (Int4WeightOnlyConfig(version=2), np.array([0.4648, 0.5234, 0.5547, 0.4219, 0.4414, 0.6445, 0.4336, 0.4531, 0.5625])),
+                (Int8DynamicActivationIntxWeightConfig(version=2), np.array([0.4688, 0.5195, 0.5547, 0.418, 0.4414, 0.6406, 0.4336, 0.4531, 0.5625])),
+                (Int8WeightOnlyConfig(version=2), np.array([0.4648, 0.5195, 0.5547, 0.4199, 0.4414, 0.6445, 0.4316, 0.4531, 0.5625])),
+                (Int8DynamicActivationInt8WeightConfig(version=2), np.array([0.4648, 0.5195, 0.5547, 0.4199, 0.4414, 0.6445, 0.4316, 0.4531, 0.5625])),
+                (IntxWeightOnlyConfig(dtype=torch.uint4, group_size=16, version=2), np.array([0.4609, 0.5234, 0.5508, 0.4199, 0.4336, 0.6406, 0.4316, 0.4531, 0.5625])),
+                (IntxWeightOnlyConfig(dtype=torch.uint7, group_size=16, version=2), np.array([0.4648, 0.5195, 0.5547, 0.4219, 0.4414, 0.6445, 0.4316, 0.4531, 0.5625])),
             ]
 
-            if TorchAoConfig._is_xpu_or_cuda_capability_atleast_8_9():
+            if _is_xpu_or_cuda_capability_atleast_8_9():
                 QUANTIZATION_TYPES_TO_TEST.extend([
-                    ("float8wo_e5m2", np.array([0.4590, 0.5273, 0.5547, 0.4219, 0.4375, 0.6406, 0.4316, 0.4512, 0.5625])),
-                    ("float8wo_e4m3", np.array([0.4648, 0.5234, 0.5547, 0.4219, 0.4414, 0.6406, 0.4316, 0.4531, 0.5625])),
-                    # =====
-                    # The following lead to an internal torch error:
-                    #    RuntimeError: mat2 shape (32x4 must be divisible by 16
-                    # Skip these for now; TODO(aryan): investigate later
-                    # ("float8dq_e4m3", np.array([0, 0, 0, 0, 0, 0, 0, 0, 0])),
-                    # ("float8dq_e4m3_tensor", np.array([0, 0, 0, 0, 0, 0, 0, 0, 0])),
-                    # =====
-                    # Cutlass fails to initialize for below
-                    # ("float8dq_e4m3_row", np.array([0, 0, 0, 0, 0, 0, 0, 0, 0])),
-                    # =====
+                    (Float8WeightOnlyConfig(weight_dtype=torch.float8_e5m2), np.array([0.4590, 0.5273, 0.5547, 0.4219, 0.4375, 0.6406, 0.4316, 0.4512, 0.5625])),
+                    (Float8WeightOnlyConfig(weight_dtype=torch.float8_e4m3fn), np.array([0.4648, 0.5234, 0.5547, 0.4219, 0.4414, 0.6406, 0.4316, 0.4531, 0.5625])),
                 ])
-                if version.parse(importlib.metadata.version("torchao")) <= version.Version("0.14.1"):
-                    QUANTIZATION_TYPES_TO_TEST.extend([
-                        ("fp4", np.array([0.4668, 0.5195, 0.5547, 0.4199, 0.4434, 0.6445, 0.4316, 0.4531, 0.5625])),
-                        ("fp6", np.array([0.4668, 0.5195, 0.5547, 0.4199, 0.4434, 0.6445, 0.4316, 0.4531, 0.5625])),
-                    ])
             # fmt: on
 
-            for quantization_name, expected_slice in QUANTIZATION_TYPES_TO_TEST:
-                quant_kwargs = {}
-                if quantization_name in ["uint4wo", "uint7wo"]:
-                    # The dummy flux model that we use has smaller dimensions. This imposes some restrictions on group_size here
-                    quant_kwargs.update({"group_size": 16})
-                quantization_config = TorchAoConfig(
-                    quant_type=quantization_name, modules_to_not_convert=["x_embedder"], **quant_kwargs
-                )
+            for quant_config, expected_slice in QUANTIZATION_TYPES_TO_TEST:
+                quantization_config = TorchAoConfig(quant_type=quant_config, modules_to_not_convert=["x_embedder"])
                 self._test_quant_type(quantization_config, expected_slice, model_id)
 
-    @unittest.skip("Skipping floatx quantization tests")
-    def test_floatx_quantization(self):
-        for model_id in ["hf-internal-testing/tiny-flux-pipe", "hf-internal-testing/tiny-flux-sharded"]:
-            if TorchAoConfig._is_xpu_or_cuda_capability_atleast_8_9():
-                if version.parse(importlib.metadata.version("torchao")) <= version.Version("0.14.1"):
-                    quantization_config = TorchAoConfig(quant_type="fp4", modules_to_not_convert=["x_embedder"])
-                    self._test_quant_type(
-                        quantization_config,
-                        np.array(
-                            [
-                                0.4648,
-                                0.5195,
-                                0.5547,
-                                0.4180,
-                                0.4434,
-                                0.6445,
-                                0.4316,
-                                0.4531,
-                                0.5625,
-                            ]
-                        ),
-                        model_id,
-                    )
-                else:
-                    # Make sure the correct error is thrown
-                    with self.assertRaisesRegex(ValueError, "Please downgrade"):
-                        quantization_config = TorchAoConfig(quant_type="fp4", modules_to_not_convert=["x_embedder"])
-
     def test_int4wo_quant_bfloat16_conversion(self):
         """
         Tests whether the dtype of model will be modified to bfloat16 for int4 weight-only quantization.
         """
-        quantization_config = TorchAoConfig("int4_weight_only", group_size=64)
+        quantization_config = TorchAoConfig(Int4WeightOnlyConfig(group_size=64))
         quantized_model = FluxTransformer2DModel.from_pretrained(
             "hf-internal-testing/tiny-flux-pipe",
             subfolder="transformer",
@@ -361,7 +308,7 @@ class TorchAoTest(unittest.TestCase):
             else:
                 expected_slice = expected_slice_offload
             with tempfile.TemporaryDirectory() as offload_folder:
-                quantization_config = TorchAoConfig("int4_weight_only", group_size=64)
+                quantization_config = TorchAoConfig(Int4WeightOnlyConfig(group_size=64))
                 quantized_model = FluxTransformer2DModel.from_pretrained(
                     "hf-internal-testing/tiny-flux-pipe",
                     subfolder="transformer",
@@ -385,7 +332,7 @@ class TorchAoTest(unittest.TestCase):
                 self.assertTrue(numpy_cosine_similarity_distance(output_slice, expected_slice) < 2e-3)
 
             with tempfile.TemporaryDirectory() as offload_folder:
-                quantization_config = TorchAoConfig("int4_weight_only", group_size=64)
+                quantization_config = TorchAoConfig(Int4WeightOnlyConfig(group_size=64))
                 quantized_model = FluxTransformer2DModel.from_pretrained(
                     "hf-internal-testing/tiny-flux-sharded",
                     subfolder="transformer",
@@ -406,7 +353,7 @@ class TorchAoTest(unittest.TestCase):
                 self.assertTrue(numpy_cosine_similarity_distance(output_slice, expected_slice) < 2e-3)
 
     def test_modules_to_not_convert(self):
-        quantization_config = TorchAoConfig("int8_weight_only", modules_to_not_convert=["transformer_blocks.0"])
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig(), modules_to_not_convert=["transformer_blocks.0"])
         quantized_model_with_not_convert = FluxTransformer2DModel.from_pretrained(
             "hf-internal-testing/tiny-flux-pipe",
             subfolder="transformer",
@@ -422,7 +369,7 @@ class TorchAoTest(unittest.TestCase):
         quantized_layer = quantized_model_with_not_convert.proj_out
         self.assertTrue(isinstance(quantized_layer.weight, AffineQuantizedTensor))
 
-        quantization_config = TorchAoConfig("int8_weight_only")
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
         quantized_model = FluxTransformer2DModel.from_pretrained(
             "hf-internal-testing/tiny-flux-pipe",
             subfolder="transformer",
@@ -436,7 +383,7 @@ class TorchAoTest(unittest.TestCase):
         self.assertTrue(size_quantized < size_quantized_with_not_convert)
 
     def test_training(self):
-        quantization_config = TorchAoConfig("int8_weight_only")
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
         quantized_model = FluxTransformer2DModel.from_pretrained(
             "hf-internal-testing/tiny-flux-pipe",
             subfolder="transformer",
@@ -470,7 +417,7 @@ class TorchAoTest(unittest.TestCase):
     def test_torch_compile(self):
         r"""Test that verifies if torch.compile works with torchao quantization."""
         for model_id in ["hf-internal-testing/tiny-flux-pipe", "hf-internal-testing/tiny-flux-sharded"]:
-            quantization_config = TorchAoConfig("int8_weight_only")
+            quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
             components = self.get_dummy_components(quantization_config, model_id=model_id)
             pipe = FluxPipeline(**components)
             pipe.to(device=torch_device)
@@ -491,11 +438,15 @@ class TorchAoTest(unittest.TestCase):
         memory footprint of the converted model and the class type of the linear layers of the converted models
         """
         for model_id in ["hf-internal-testing/tiny-flux-pipe", "hf-internal-testing/tiny-flux-sharded"]:
-            transformer_int4wo = self.get_dummy_components(TorchAoConfig("int4wo"), model_id=model_id)["transformer"]
+            transformer_int4wo = self.get_dummy_components(TorchAoConfig(Int4WeightOnlyConfig()), model_id=model_id)[
+                "transformer"
+            ]
             transformer_int4wo_gs32 = self.get_dummy_components(
-                TorchAoConfig("int4wo", group_size=32), model_id=model_id
+                TorchAoConfig(Int4WeightOnlyConfig(group_size=32)), model_id=model_id
             )["transformer"]
-            transformer_int8wo = self.get_dummy_components(TorchAoConfig("int8wo"), model_id=model_id)["transformer"]
+            transformer_int8wo = self.get_dummy_components(TorchAoConfig(Int8WeightOnlyConfig()), model_id=model_id)[
+                "transformer"
+            ]
             transformer_bf16 = self.get_dummy_components(None, model_id=model_id)["transformer"]
 
             # Will not quantized all the layers by default due to the model weights shapes not being divisible by group_size=64
@@ -553,20 +504,22 @@ class TorchAoTest(unittest.TestCase):
         unquantized_model_memory = get_memory_consumption_stat(transformer_bf16, inputs)
         del transformer_bf16
 
-        transformer_int8wo = self.get_dummy_components(TorchAoConfig("int8wo"), model_id=model_id)["transformer"]
+        transformer_int8wo = self.get_dummy_components(TorchAoConfig(Int8WeightOnlyConfig()), model_id=model_id)[
+            "transformer"
+        ]
         transformer_int8wo.to(torch_device)
         quantized_model_memory = get_memory_consumption_stat(transformer_int8wo, inputs)
         assert unquantized_model_memory / quantized_model_memory >= expected_memory_saving_ratio
 
     def test_wrong_config(self):
-        with self.assertRaises(ValueError):
+        with self.assertRaises(TypeError):
             self.get_dummy_components(TorchAoConfig("int42"))
 
     def test_sequential_cpu_offload(self):
         r"""
         A test that checks if inference runs as expected when sequential cpu offloading is enabled.
         """
-        quantization_config = TorchAoConfig("int8wo")
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
         components = self.get_dummy_components(quantization_config)
         pipe = FluxPipeline(**components)
         pipe.enable_sequential_cpu_offload()
@@ -595,8 +548,8 @@ class TorchAoSerializationTest(unittest.TestCase):
         gc.collect()
         backend_empty_cache(torch_device)
 
-    def get_dummy_model(self, quant_method, quant_method_kwargs, device=None):
-        quantization_config = TorchAoConfig(quant_method, **quant_method_kwargs)
+    def get_dummy_model(self, quant_type, device=None):
+        quantization_config = TorchAoConfig(quant_type)
         quantized_model = FluxTransformer2DModel.from_pretrained(
             self.model_name,
             subfolder="transformer",
@@ -632,8 +585,8 @@ class TorchAoSerializationTest(unittest.TestCase):
             "timestep": timestep,
         }
 
-    def _test_original_model_expected_slice(self, quant_method, quant_method_kwargs, expected_slice):
-        quantized_model = self.get_dummy_model(quant_method, quant_method_kwargs, torch_device)
+    def _test_original_model_expected_slice(self, quant_type, expected_slice):
+        quantized_model = self.get_dummy_model(quant_type, torch_device)
         inputs = self.get_dummy_tensor_inputs(torch_device)
         output = quantized_model(**inputs)[0]
         output_slice = output.flatten()[-9:].detach().float().cpu().numpy()
@@ -641,8 +594,8 @@ class TorchAoSerializationTest(unittest.TestCase):
         self.assertTrue(isinstance(weight, (AffineQuantizedTensor, LinearActivationQuantizedTensor)))
         self.assertTrue(numpy_cosine_similarity_distance(output_slice, expected_slice) < 1e-3)
 
-    def _check_serialization_expected_slice(self, quant_method, quant_method_kwargs, expected_slice, device):
-        quantized_model = self.get_dummy_model(quant_method, quant_method_kwargs, device)
+    def _check_serialization_expected_slice(self, quant_type, expected_slice, device):
+        quantized_model = self.get_dummy_model(quant_type, device)
 
         with tempfile.TemporaryDirectory() as tmp_dir:
             quantized_model.save_pretrained(tmp_dir, safe_serialization=False)
@@ -662,40 +615,39 @@ class TorchAoSerializationTest(unittest.TestCase):
         self.assertTrue(numpy_cosine_similarity_distance(output_slice, expected_slice) < 1e-3)
 
     def test_int_a8w8_accelerator(self):
-        quant_method, quant_method_kwargs = "int8_dynamic_activation_int8_weight", {}
+        quant_type = Int8DynamicActivationInt8WeightConfig()
         expected_slice = np.array([0.3633, -0.1357, -0.0188, -0.249, -0.4688, 0.5078, -0.1289, -0.6914, 0.4551])
         device = torch_device
-        self._test_original_model_expected_slice(quant_method, quant_method_kwargs, expected_slice)
-        self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device)
+        self._test_original_model_expected_slice(quant_type, expected_slice)
+        self._check_serialization_expected_slice(quant_type, expected_slice, device)
 
     def test_int_a16w8_accelerator(self):
-        quant_method, quant_method_kwargs = "int8_weight_only", {}
+        quant_type = Int8WeightOnlyConfig()
         expected_slice = np.array([0.3613, -0.127, -0.0223, -0.2539, -0.459, 0.4961, -0.1357, -0.6992, 0.4551])
         device = torch_device
-        self._test_original_model_expected_slice(quant_method, quant_method_kwargs, expected_slice)
-        self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device)
+        self._test_original_model_expected_slice(quant_type, expected_slice)
+        self._check_serialization_expected_slice(quant_type, expected_slice, device)
 
     def test_int_a8w8_cpu(self):
-        quant_method, quant_method_kwargs = "int8_dynamic_activation_int8_weight", {}
+        quant_type = Int8DynamicActivationInt8WeightConfig()
         expected_slice = np.array([0.3633, -0.1357, -0.0188, -0.249, -0.4688, 0.5078, -0.1289, -0.6914, 0.4551])
         device = "cpu"
-        self._test_original_model_expected_slice(quant_method, quant_method_kwargs, expected_slice)
-        self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device)
+        self._test_original_model_expected_slice(quant_type, expected_slice)
+        self._check_serialization_expected_slice(quant_type, expected_slice, device)
 
     def test_int_a16w8_cpu(self):
-        quant_method, quant_method_kwargs = "int8_weight_only", {}
+        quant_type = Int8WeightOnlyConfig()
         expected_slice = np.array([0.3613, -0.127, -0.0223, -0.2539, -0.459, 0.4961, -0.1357, -0.6992, 0.4551])
         device = "cpu"
-        self._test_original_model_expected_slice(quant_method, quant_method_kwargs, expected_slice)
-        self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device)
+        self._test_original_model_expected_slice(quant_type, expected_slice)
+        self._check_serialization_expected_slice(quant_type, expected_slice, device)
 
-    @require_torchao_version_greater_or_equal("0.9.0")
     def test_aobase_config(self):
-        quant_method, quant_method_kwargs = Int8WeightOnlyConfig(), {}
+        quant_type = Int8WeightOnlyConfig()
         expected_slice = np.array([0.3613, -0.127, -0.0223, -0.2539, -0.459, 0.4961, -0.1357, -0.6992, 0.4551])
         device = torch_device
-        self._test_original_model_expected_slice(quant_method, quant_method_kwargs, expected_slice)
-        self._check_serialization_expected_slice(quant_method, quant_method_kwargs, expected_slice, device)
+        self._test_original_model_expected_slice(quant_type, expected_slice)
+        self._check_serialization_expected_slice(quant_type, expected_slice, device)
 
 
 @require_torchao_version_greater_or_equal("0.14.0")
@@ -817,29 +769,25 @@ class SlowTorchAoTests(unittest.TestCase):
     def test_quantization(self):
         # fmt: off
         QUANTIZATION_TYPES_TO_TEST = [
-            ("int8wo", np.array([0.0505, 0.0742, 0.1367, 0.0429, 0.0585, 0.1386, 0.0585, 0.0703, 0.1367, 0.0566, 0.0703, 0.1464, 0.0546, 0.0703, 0.1425, 0.0546, 0.3535, 0.7578, 0.5000, 0.4062, 0.7656, 0.5117, 0.4121, 0.7656, 0.5117, 0.3984, 0.7578, 0.5234, 0.4023, 0.7382, 0.5390, 0.4570])),
-            ("int8dq", np.array([0.0546, 0.0761, 0.1386, 0.0488, 0.0644, 0.1425, 0.0605, 0.0742, 0.1406, 0.0625, 0.0722, 0.1523, 0.0625, 0.0742, 0.1503, 0.0605, 0.3886, 0.7968, 0.5507, 0.4492, 0.7890, 0.5351, 0.4316, 0.8007, 0.5390, 0.4179, 0.8281, 0.5820, 0.4531, 0.7812, 0.5703, 0.4921])),
+            (Int8WeightOnlyConfig(), np.array([0.0505, 0.0742, 0.1367, 0.0429, 0.0585, 0.1386, 0.0585, 0.0703, 0.1367, 0.0566, 0.0703, 0.1464, 0.0546, 0.0703, 0.1425, 0.0546, 0.3535, 0.7578, 0.5000, 0.4062, 0.7656, 0.5117, 0.4121, 0.7656, 0.5117, 0.3984, 0.7578, 0.5234, 0.4023, 0.7382, 0.5390, 0.4570])),
+            (Int8DynamicActivationInt8WeightConfig(), np.array([0.0546, 0.0761, 0.1386, 0.0488, 0.0644, 0.1425, 0.0605, 0.0742, 0.1406, 0.0625, 0.0722, 0.1523, 0.0625, 0.0742, 0.1503, 0.0605, 0.3886, 0.7968, 0.5507, 0.4492, 0.7890, 0.5351, 0.4316, 0.8007, 0.5390, 0.4179, 0.8281, 0.5820, 0.4531, 0.7812, 0.5703, 0.4921])),
         ]
 
-        if TorchAoConfig._is_xpu_or_cuda_capability_atleast_8_9():
+        if _is_xpu_or_cuda_capability_atleast_8_9():
             QUANTIZATION_TYPES_TO_TEST.extend([
-                ("float8wo_e4m3", np.array([0.0546, 0.0722, 0.1328, 0.0468, 0.0585, 0.1367, 0.0605, 0.0703, 0.1328, 0.0625, 0.0703, 0.1445, 0.0585, 0.0703, 0.1406, 0.0605, 0.3496, 0.7109, 0.4843, 0.4042, 0.7226, 0.5000, 0.4160, 0.7031, 0.4824, 0.3886, 0.6757, 0.4667, 0.3710, 0.6679, 0.4902, 0.4238])),
+                (Float8WeightOnlyConfig(weight_dtype=torch.float8_e4m3fn), np.array([0.0546, 0.0722, 0.1328, 0.0468, 0.0585, 0.1367, 0.0605, 0.0703, 0.1328, 0.0625, 0.0703, 0.1445, 0.0585, 0.0703, 0.1406, 0.0605, 0.3496, 0.7109, 0.4843, 0.4042, 0.7226, 0.5000, 0.4160, 0.7031, 0.4824, 0.3886, 0.6757, 0.4667, 0.3710, 0.6679, 0.4902, 0.4238])),
             ])
-            if version.parse(importlib.metadata.version("torchao")) <= version.Version("0.14.1"):
-                QUANTIZATION_TYPES_TO_TEST.extend([
-                    ("fp5_e3m1", np.array([0.0527, 0.0762, 0.1309, 0.0449, 0.0645, 0.1328, 0.0566, 0.0723, 0.125, 0.0566, 0.0703, 0.1328, 0.0566, 0.0742, 0.1348, 0.0566, 0.3633, 0.7617, 0.5273, 0.4277, 0.7891, 0.5469, 0.4375, 0.8008, 0.5586, 0.4336, 0.7383, 0.5156, 0.3906, 0.6992, 0.5156, 0.4375])),
-                ])
         # fmt: on
 
-        for quantization_name, expected_slice in QUANTIZATION_TYPES_TO_TEST:
-            quantization_config = TorchAoConfig(quant_type=quantization_name, modules_to_not_convert=["x_embedder"])
+        for quant_config, expected_slice in QUANTIZATION_TYPES_TO_TEST:
+            quantization_config = TorchAoConfig(quant_type=quant_config, modules_to_not_convert=["x_embedder"])
             self._test_quant_type(quantization_config, expected_slice)
             gc.collect()
             backend_empty_cache(torch_device)
             backend_synchronize(torch_device)
 
     def test_serialization_int8wo(self):
-        quantization_config = TorchAoConfig("int8wo")
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
         components = self.get_dummy_components(quantization_config)
         pipe = FluxPipeline(**components)
         pipe.enable_model_cpu_offload()
@@ -876,7 +824,7 @@ class SlowTorchAoTests(unittest.TestCase):
     def test_memory_footprint_int4wo(self):
         # The original checkpoints are in bf16 and about 24 GB
         expected_memory_in_gb = 6.0
-        quantization_config = TorchAoConfig("int4wo")
+        quantization_config = TorchAoConfig(Int4WeightOnlyConfig())
         cache_dir = None
         transformer = FluxTransformer2DModel.from_pretrained(
             "black-forest-labs/FLUX.1-dev",
@@ -891,7 +839,7 @@ class SlowTorchAoTests(unittest.TestCase):
     def test_memory_footprint_int8wo(self):
         # The original checkpoints are in bf16 and about 24 GB
         expected_memory_in_gb = 12.0
-        quantization_config = TorchAoConfig("int8wo")
+        quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
         cache_dir = None
         transformer = FluxTransformer2DModel.from_pretrained(
             "black-forest-labs/FLUX.1-dev",