.. _GenAI Model Compilation:

GenAI Model Compilation
=================

Overview
--------

The **ModelSDK container** provides a command-line tool ``llima-compile`` to compile models directly from Hugging Face or GGUF files:

.. code-block:: console

   llima-compile [options] <model_path>

When you run this command, the tool handles the entire compilation pipeline including calibration, quantization, and code generation. The pipeline consists of several stages that differ slightly depending on the input format:

**For HuggingFace Models:**

1. **DEVKIT** - Generate runtime orchestration scripts
2. **ONNX** - Convert model to ONNX intermediate representation
3. **QUANTIZE** - Quantize model weights and calibrate
4. **COMPILE** - Compile to Modalix machine code

**For GGUF Models:**

1. **DEVKIT** - Generate runtime orchestration scripts
2. **MODEL_SDK_DIRECT** - Convert GGUF directly to Model SDK format (quantization already applied)
3. **COMPILE** - Compile to Modalix machine code

You can run individual stages using ``--onnx``, ``--quantize``, ``--model_sdk``, ``--compile``, or ``--devkit`` flags if needed.

The compilation process generates the following directory structure in your output directory:

.. code-block:: text

   output_directory/
   ├── onnx_files/                # ONNX intermediate files (HF models only)
   │   └── ...
   └── sima_files/                # Compiled model files
       ├── devkit/                # Python runtime orchestration files
       │   ├── tokenizer.json
       │   ├── vlm_config.json
       │   └── ...
       ├── mpk/                   # MPK archives with compiled binaries
       │   ├── layer_0.tar.gz
       │   └── ...
       └── ...

Command-Line Arguments
----------------------

The ``llima-compile`` tool accepts various arguments to customize the compilation process. The following tables describe the available options:

.. list-table::
   :widths: 30 70
   :header-rows: 1

   * - Argument
     - Description
   * - ``model_path``
     - Input model path (GGUF file or Hugging Face directory).
   * - ``-o, --output``
     - Output directory for compiled files. Defaults to the model name.
   * - ``-c, --configuration_file``
     - Python script to configure precision per layer (e.g., for mixed-precision).
   * - ``--max_num_tokens``
     - Max context length. Default: 1024.
   * - ``--resume``
     - Resume interrupted builds by skipping existing files.
   * - ``-j, --jobs``
     - Number of parallel compilation jobs. Default: Number of physical CPU cores.
   * - ``--log_level``
     - Logging level (DEBUG, INFO, WARNING, ERROR). Default: WARNING.


.. list-table::
   :widths: 30 70
   :header-rows: 1

   * - Advanced Argument
     - Description
   * - ``--language_group_size``
     - Batch size for parallel token processing during prefill. Larger values (e.g., 256) can improve TTFT for large input prompts, but can decrease TTFT for smaller input prompts. Default: 128.
   * - ``--future_token_mask_size``
     - Mask size for reusing compiled models across token positions. Larger values reduce number of compiled binary files, but may reduce TPS. Default: 128.


Configuration File
------------------

The configuration file allows customizing compilation on a per-layer basis, enabling mixed-precision compilation and selective layer compilation.

LLM inference consists of two distinct phases, and the compiler generates optimized models for each:

- **Prefill (Group models)**: Processes the input prompt in batches using ``language_group_size`` (e.g., 128 tokens at once). This phase determines TTFT (Time To First Token) and is optimized for throughput.
- **Decode (Single-token models)**: Generates output tokens one at a time autoregressively. This phase determines TPS (Tokens Per Second) and is optimized for low-latency generation.

Because these phases have different performance characteristics, you can apply different quantization strategies to each using the ``is_group`` flag in the configuration function.

**Input Parameters**

The ``get_layer_configuration`` function is called for each layer and receives:

- ``model_properties``: Dictionary with ``{"num_hidden_layers": int}``
- ``layer``: Dictionary with:
   - ``"part"``: Layer type - ``"PRE"``, ``"CACHE"``, ``"POST"``, or ``"VISION"``
   - ``"is_group"``: ``True`` for batch processing layers, ``False`` for single-token layers
   - ``"index"``: Layer index (0 to num_hidden_layers-1)

**Return Values**

The function returns a dictionary with:

- ``"precision"``: Quantization level (required)
   - ``"BF16"``: Full precision - best quality, largest size, slowest
   - ``"A_BF16_W_INT8"``: Medium quantization - good quality, moderate size
   - ``"A_BF16_W_INT4"``: High quantization - acceptable quality, smallest size, fastest
- ``"compile"``: Set to ``False`` to skip compiling this layer (optional, default: ``True``)

.. note::
   
   **Best Practice:** Use INT8 (``A_BF16_W_INT8``) for group layers to maintain quality during prefill, INT4 (``A_BF16_W_INT4``) for single-token layers for fast generation, and BF16 for vision encoders to preserve image understanding quality. For most models, this configuration provides the optimal balance between model accuracy, throughput, and memory usage.

Examples
--------

**Example 1: Compiling a Simple LLM**

Compile a Llama model, downloaded from Hugging Face, with default settings:

.. code-block:: console
  
   sima-user@docker-image-id:/home/docker$ hf download meta-llama/Llama-3.2-3B --local-dir Llama-3.2-3B-Instruct
   sima-user@docker-image-id:/home/docker$ llima-compile Llama-3.2-3B-Instruct -o Llama-3.2-3B-Instruct_out

This will:

- Use default BF16 precision for all layers
- Set context length to 1024 tokens
- Output to ``Llama-3.2-3B-Instruct`` directory

**Example 2: Compiling with Custom Context Length**

.. code-block:: console

   sima-user@docker-image-id:/home/docker$ hf download meta-llama/Llama-3.2-3B --local-dir Llama-3.2-3B-Instruct
   sima-user@docker-image-id:/home/docker$ llima-compile --max_num_tokens 4096 Llama-3.2-3B-Instruct -o Llama-3.2-3B-Instruct_out

This will:

- Use default BF16 precision for all layers
- Set context length to 4096 tokens
- Output to ``Llama-3.2-3B-Instruct`` directory


**Example 3: Compiling Gemma 3 VLM with Mixed Precision**

For complex models like Gemma 3 VLM, you may need to specify different precisions for different layers (e.g., keeping the vision encoder in BF16).

1. **Download the model**:

   .. code-block:: console

     sima-user@docker-image-id:/home/docker$ hf download simaai/gemma3-siglip448 --local-dir gemma-3-model

2. **Create a configuration file** (e.g., ``config.py``):

   .. code-block:: python

      def get_layer_configuration(model_properties, layer):
          # Keep vision encoder in full precision
          if layer["part"] == "VISION":
              precision = "BF16"
          # Use INT8 for batch processing layers (better quality)
          elif layer["is_group"]:
              precision = "A_BF16_W_INT8"
          # Use INT4 for single-token layers (smaller size)
          else:
              precision = "A_BF16_W_INT4"
          return {"precision": precision}

3. **Run the compiler**:

   .. code-block:: console

      sima-user@docker-image-id:/home/docker$ llima-compile -c config.py --max_num_tokens 2048 gemma-3-model -o gemma-3-model_out

**Example 4: Advanced Configuration**

Mixed precision with layer-specific control:

.. code-block:: python

   def get_layer_configuration(model_properties, layer):
       # Skip compiling certain cache layers
       if layer["part"] == "CACHE" and layer["index"] > 20:
           return {"compile": False}
       
       # Higher precision for early layers
       if layer["index"] < 4:
           return {"precision": "BF16"}
       
       # Standard quantization for middle layers
       return {"precision": "A_BF16_W_INT8"}