Refactor IO config to yaml in assets (#2311)

This pull request introduces a major refactor of the HuggingFace ONNX
I/O configuration system in Olive. The changes move away from hardcoded
Python class registries and instead adopt a YAML-driven approach for
specifying input/output configurations for both tasks and diffusers
pipelines. This enables easier extensibility and clearer, data-driven
configuration management. Additionally, the update adds default dummy
input shapes and rich configuration templates for tasks and diffusers
components.

The most important changes are:

**YAML-based IO Configuration System:**
- Added new YAML files (`defaults.yaml`, `tasks.yaml`, `diffusers.yaml`)
in `olive/assets/io_configs/` that define default dummy input shapes,
task templates, and diffusers component/pipeline specifications for ONNX
export. These files replace the previous Python-based registries and
provide a more flexible and maintainable configuration system.
[[1]](diffhunk://#diff-6b06016102bbc575e9fa17da2c2aa9e54172dea3fa610de6604c8b388d54d111R1-R32)
[[2]](diffhunk://#diff-2cf88124ed9192111582c2746192e39b1e675d9224addde89f560e3374611e7dR1-R231)
[[3]](diffhunk://#diff-28e9d219224be5365b9a855dd119d92f7571dbd782775d592418259f9e3361a7R1-R206)

**Refactor of IO Config Python Interface:**
- Replaced the contents of `olive/common/hf/io_config/__init__.py` to
remove all hardcoded class registries and mapping logic, and instead
import new YAML-driven utility functions (`generate_dummy_inputs`,
`get_io_config`, `get_diffusers_io_config`, etc.). This greatly
simplifies the interface and shifts responsibility to the new YAML
configuration system.

**Legal and Organizational Improvements:**
- Added copyright and license headers to `olive/assets/__init__.py` and
`olive/assets/io_configs/__init__.py`.## Describe your changes

## Checklist before requesting a review
- [ ] Add unit tests for this change.
- [ ] Make sure all tests can pass.
- [ ] Update documents if necessary.
- [ ] Lint and apply fixes to your code by running `lintrunner -a`
- [ ] Is this a user-facing change? If yes, give a description of this
change to be included in the release notes.

## (Optional) Issue link

---------

Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>

Author: xiaoyu-work

docs/source/how-to/extending/how-to-add-new-task.md

@@ -0,0 +1,308 @@
+# How to Add a New Task or Diffusers Component for ONNX Export
+
+This guide explains how to add IO configurations for a new HuggingFace task or diffusers component to enable ONNX model export.
+
+Olive uses YAML-based IO configurations to define input/output specifications for ONNX export. These configurations specify tensor shapes, data types, and dynamic axes for each model input and output.
+
+There are two types of configurations:
+- **Task configs** (`tasks.yaml`): For HuggingFace transformers tasks like text-generation, text-classification, etc.
+- **Diffusers component configs** (`diffusers.yaml`): For Stable Diffusion and similar diffusion model components like UNet, VAE, text encoders, etc.
+
+## File Locations
+
+IO config files are located in `olive/assets/io_configs/`:
+
+```
+olive/assets/io_configs/
+├── tasks.yaml      # Task-based configurations
+├── diffusers.yaml  # Diffusers component configurations
+└── defaults.yaml   # Default dimension values and aliases
+```
+
+## Task-based IO Configs (`tasks.yaml`)
+
+### Format
+
+Each task defines its input/output specifications:
+
+```yaml
+task-name:
+  inputs:
+    input_name:
+      shape: [dim1, dim2, ...]      # Shape template for dummy input generation
+      axes: {0: axis_name, 1: ...}  # Dynamic axes for ONNX export
+      dtype: int64 | float          # Data type (default: int64)
+      max_value: vocab_size         # Optional: max value for random input
+      optional: true                # Optional: skip if not in model.forward()
+  outputs:
+    output_name:
+      axes: {0: axis_name, ...}     # Dynamic axes for ONNX export
+  with_past:                        # Optional: overrides for KV cache scenarios
+    input_name:
+      shape: [...]
+      axes: {...}
+```
+
+### Field Descriptions
+
+| Field | Description |
+|-------|-------------|
+| `shape` | List of dimension names or integers. Used to generate dummy inputs for ONNX export. Dimension names are resolved from model config or defaults. |
+| `axes` | Dict mapping axis index to axis name. Defines which dimensions are dynamic in the exported ONNX model. |
+| `dtype` | Data type: `int64`, `int32`, or `float`. Defaults to `int64` for inputs. |
+| `optional` | If `true`, the input is only included if it exists in `model.forward()` signature. |
+| `max_value` | Maximum value for random input generation (e.g., `vocab_size` for input_ids). |
+| `with_past` | Alternative shapes/axes when using KV cache (`use_past_in_inputs=True`). |
+
+### Example: Adding a New Task
+
+To add support for a new task, add an entry to `tasks.yaml`:
+
+```yaml
+# Custom task for a new model type
+my-custom-task:
+  inputs:
+    input_ids:
+      shape: [batch_size, sequence_length]
+      axes: {0: batch_size, 1: sequence_length}
+      dtype: int64
+      max_value: vocab_size
+    attention_mask:
+      shape: [batch_size, sequence_length]
+      axes: {0: batch_size, 1: sequence_length}
+      dtype: int64
+    custom_input:
+      shape: [batch_size, custom_dim]
+      axes: {0: batch_size, 1: custom_dim}
+      dtype: float
+      optional: true
+  outputs:
+    logits:
+      axes: {0: batch_size, 1: sequence_length, 2: vocab_size}
+    custom_output:
+      axes: {0: batch_size, 1: hidden_size}
+```
+
+### Supported Tasks
+
+Currently supported tasks include:
+- `text-generation`
+- `text-classification`
+- `feature-extraction`
+- `fill-mask`
+- `token-classification`
+- `question-answering`
+- `multiple-choice`
+- `text2text-generation`
+- `image-classification`
+- `object-detection`
+- `semantic-segmentation`
+- `audio-classification`
+- `automatic-speech-recognition`
+- `zero-shot-image-classification`
+
+## Diffusers Component Configs (`diffusers.yaml`)
+
+### Format
+
+Diffusers configurations define components and pipelines:
+
+```yaml
+components:
+  component_name:
+    inputs:
+      input_name:
+        shape: [dim1, dim2, ...]
+        axes: {0: axis_name, ...}
+        dtype: int64 | float
+    outputs:
+      output_name:
+        axes: {0: axis_name, ...}
+    sdxl_inputs:           # Optional: additional inputs for SDXL
+      extra_input:
+        shape: [...]
+        axes: {...}
+    optional_inputs:       # Optional: conditional inputs
+      optional_input:
+        shape: [...]
+        axes: {...}
+        condition: config_attr  # Only include if config.config_attr is True
+
+pipelines:
+  pipeline_name:
+    - component_name
+    - component_config:alias_name  # Use component_config with alias
+```
+
+### Example: Adding a New Diffusers Component
+
+```yaml
+components:
+  my_custom_transformer:
+    inputs:
+      hidden_states:
+        shape: [batch_size, in_channels, height, width]
+        axes: {0: batch_size, 1: in_channels, 2: height, 3: width}
+        dtype: float
+      encoder_hidden_states:
+        shape: [batch_size, sequence_length, hidden_size]
+        axes: {0: batch_size, 1: sequence_length, 2: hidden_size}
+        dtype: float
+      timestep:
+        shape: [batch_size]
+        axes: {0: batch_size}
+        dtype: float
+    outputs:
+      out_sample:
+        axes: {0: batch_size, 1: in_channels, 2: height, 3: width}
+    optional_inputs:
+      guidance:
+        shape: [batch_size]
+        axes: {0: batch_size}
+        dtype: float
+        condition: guidance_embeds  # Only if config.guidance_embeds is True
+
+pipelines:
+  my_custom_pipeline:
+    - text_encoder
+    - my_custom_transformer:transformer
+    - vae_encoder
+    - vae_decoder
+```
+
+### Supported Diffusers Components
+
+Currently supported components include:
+- `text_encoder`, `text_encoder_with_projection`, `t5_encoder`, `gemma2_text_encoder`
+- `unet`, `sd3_transformer`, `flux_transformer`, `sana_transformer`
+- `vae_encoder`, `vae_decoder`, `dcae_encoder`, `dcae_decoder`
+
+Supported pipelines: `sd`, `sdxl`, `sd3`, `flux`, `sana`
+
+## Default Values (`defaults.yaml`)
+
+The `defaults.yaml` file defines:
+1. **Aliases**: Alternative attribute names for the same concept across different models
+2. **Default dimensions**: Fallback values when dimensions can't be resolved from model config
+
+### Aliases
+
+Aliases help resolve config attributes that have different names across models:
+
+```yaml
+aliases:
+  num_layers: [num_hidden_layers, n_layer, n_layers]
+  hidden_size: [dim, d_model, n_embd]
+  num_attention_heads: [num_heads, n_head, n_heads, encoder_attention_heads]
+  num_kv_heads: [num_key_value_heads]
+  height: [sample_size, image_size, vision_config.image_size]
+  width: [sample_size, image_size, vision_config.image_size]
+  num_channels: [in_channels, vision_config.num_channels]
+```
+
+### Default Dimensions
+
+Default values used when dimensions can't be resolved from model config:
+
+```yaml
+batch_size: 2
+sequence_length: 16
+past_sequence_length: 16
+vocab_size: 32000
+height: 64
+width: 64
+num_channels: 3
+```
+
+### Adding New Defaults
+
+If your model uses a dimension not already defined, add it to `defaults.yaml`:
+
+```yaml
+# Add new dimension for your model
+my_custom_dim: 128
+
+# Add aliases if the same concept has different names
+aliases:
+  my_custom_dim: [custom_dim, my_dim]
+```
+
+## Dimension Resolution
+
+When generating dummy inputs, dimensions in `shape` are resolved in this order:
+
+1. **Model config with aliases**: Check `config.attr_name` for each alias
+2. **Computed dimensions**: Special dimensions like `height_latent = height // 8`
+3. **Default values**: Fall back to values in `defaults.yaml`
+
+## Usage in Olive Workflows
+
+Once you've added your IO config, Olive will automatically use it during ONNX conversion.
+
+### Task-based Models
+
+For HuggingFace transformers models, specify the task in `HfModel`:
+
+```yaml
+# olive_config.yaml
+input_model:
+  type: HfModel
+  model_path: my-model
+  task: my-custom-task  # Uses the task config you defined
+
+passes:
+  conversion:
+    type: OnnxConversion
+```
+
+### Diffusers Models
+
+For diffusion models, use `DiffusersModel`. Olive automatically detects the pipeline type and exports all components using the IO configs defined in `diffusers.yaml`:
+
+```yaml
+# olive_config.yaml
+input_model:
+  type: DiffusersModel
+  model_path: stabilityai/stable-diffusion-xl-base-1.0
+
+passes:
+  conversion:
+    type: OnnxConversion
+```
+
+Olive will automatically:
+1. Detect the pipeline type (e.g., `sdxl`)
+2. Identify exportable components (text_encoder, text_encoder_2, unet, vae_encoder, vae_decoder)
+3. Use the corresponding IO configs from `diffusers.yaml` for each component
+
+## Testing Your Config
+
+After adding a new IO config, verify it works:
+
+```python
+from olive.common.hf.io_config import get_io_config, generate_dummy_inputs
+
+# Test task config
+io_config = get_io_config("my-model-path", task="my-custom-task")
+print(io_config["input_names"])
+print(io_config["output_names"])
+print(io_config["dynamic_axes"])
+
+# Generate dummy inputs
+dummy_inputs = generate_dummy_inputs("my-model-path", task="my-custom-task")
+for name, tensor in dummy_inputs.items():
+    print(f"{name}: {tensor.shape}")
+```
+
+For diffusers:
+
+```python
+from olive.common.hf.io_config import get_diffusers_io_config, generate_diffusers_dummy_inputs
+
+# Test diffusers config
+io_config = get_diffusers_io_config("my_custom_transformer", config)
+print(io_config["input_names"])
+
+# Generate dummy inputs
+dummy_inputs = generate_diffusers_dummy_inputs("my_custom_transformer", config)
+```

docs/source/how-to/index.md

@@ -32,6 +32,7 @@ The Olive CLI provides a set of primitives such as `quantize`, `finetune`, `onnx
 
 - [Olive design overview](extending/design)
 - [How to add a new Pass](extending/how-to-add-optimization-pass)
+- [How to add a new task for ONNX export](extending/how-to-add-new-task.md)
 - [How to add custom model evaluator](extending/custom-model-evaluator)
 - [How to add custom scripts to load datasets](extending/custom-scripts)
 
@@ -57,6 +58,7 @@ configure-workflows/systems
 configure-workflows/engine-configuration
 extending/design
 extending/how-to-add-optimization-pass
+extending/how-to-add-new-task
 extending/custom-model-evaluator
 extending/custom-scripts
 ```

olive/assets/__init__.py

@@ -0,0 +1,4 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# -------------------------------------------------------------------------

assets/cost_models/Llama-2-13B.csv olive/assets/cost_models/Llama-2-13B.csvassets/cost_models/Llama-2-13B.csv renamed to olive/assets/cost_models/Llama-2-13B.csv

@@ -0,0 +1,4 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# -------------------------------------------------------------------------

assets/cost_models/Llama-2-7B.csv olive/assets/cost_models/Llama-2-7B.csvassets/cost_models/Llama-2-7B.csv renamed to olive/assets/cost_models/Llama-2-7B.csv

@@ -0,0 +1,40 @@
+# Default dimension values for dummy input generation
+# These values are used when dimensions cannot be resolved from model config
+
+# Attribute aliases (same concept, different naming across models)
+aliases:
+  # Layer count
+  num_layers: [num_hidden_layers, n_layer, n_layers]
+  # Hidden dimensions
+  hidden_size: [dim, d_model, n_embd]
+  num_attention_heads: [num_heads, n_head, n_heads, encoder_attention_heads]
+  num_kv_heads: [num_key_value_heads]
+  # Image dimensions
+  height: [sample_size, image_size, vision_config.image_size]
+  width: [sample_size, image_size, vision_config.image_size]
+  num_channels: [in_channels, vision_config.num_channels]
+
+# Common
+batch_size: 2
+sequence_length: 16
+past_sequence_length: 16
+num_choices: 4
+vocab_size: 32000
+
+# Image
+width: 64
+height: 64
+num_channels: 3
+point_batch_size: 3
+nb_points_per_image: 2
+visual_seq_length: 16
+
+# Multimodal (CLIP, etc.)
+text_batch_size: 2
+image_batch_size: 2
+projection_dim: 512
+
+# Audio
+feature_size: 80
+nb_max_frames: 3000
+audio_sequence_length: 16000