Add support for generating ONNX Runtime GenAI files only in model bui…

…lder (#186)

### Description

This PR adds support for generating the config and tokenizer files that
are needed for ONNX Runtime GenAI.

### Motivation and Context

This is for when users already have their optimized and/or quantized
ONNX models. By generating the config and tokenizer files only, users
can run their own models with ONNX Runtime GenAI.

src/python/py/models/README.md

@@ -11,6 +11,7 @@ This folder contains the model builder tool, which greatly accelerates creating
    - [Customized or Finetuned PyTorch Model](#customized-or-finetuned-pytorch-model)
    - [GGUF Model](#gguf-model)
    - [Extra Options](#extra-options)
+   - [Config Only](#config-only)
    - [Unit Testing Models](#unit-testing-models)
      - [Option 1: Use the model builder tool directly](#option-1-use-the-model-builder-tool-directly)
      - [Option 2: Edit the config.json file](#option-2-edit-the-configjson-file-on-disk-and-then-run-the-model-builder-tool)
@@ -86,6 +87,18 @@ python3 builder.py -m model_name -o path_to_output_folder -p precision -e execut
 ```
 To see all available options through `--extra_options`, please use the `help` commands in the `Full Usage` section above.
 
+### Config Only
+This scenario is for when you already have your optimized and/or quantized ONNX model and you need to create the config files to run with ONNX Runtime GenAI.
+```
+# From wheel:
+python3 -m onnxruntime_genai.models.builder -m model_name -o path_to_output_folder -p precision -e execution_provider -c cache_dir_for_hf_files --extra_options config_only=true
+
+# From source:
+python3 builder.py -m model_name -o path_to_output_folder -p precision -e execution_provider -c cache_dir_for_hf_files --extra_options config_only=true
+```
+
+Afterwards, please open the `genai_config.json` file in the output folder and modify the fields as needed for your model. You should store your ONNX model in the output folder as well.
+
 ### Unit Testing Models
 This scenario is where your PyTorch model is already downloaded locally (either in the default Hugging Face cache directory or in a local folder on disk). If it is not already downloaded locally, here is an example of how you can download it.
 

src/python/py/models/builder.py

@@ -25,7 +25,7 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
         self.window_size = config.sliding_window if hasattr(config, "sliding_window") else -1  # default is -1 in GroupQueryAttention kernel
         self.intermediate_size = config.intermediate_size
         self.hidden_size = config.hidden_size
-        self.num_kv_heads = config.num_key_value_heads
+        self.num_kv_heads = config.num_key_value_heads if hasattr(config, "num_key_value_heads") else config.num_attention_heads
         self.num_attn_heads = config.num_attention_heads
         self.head_size = config.head_dim if hasattr(config, "head_dim") else config.hidden_size // config.num_attention_heads
         self.num_layers = int(extra_options["num_hidden_layers"]) if "num_hidden_layers" in extra_options else config.num_hidden_layers
@@ -102,12 +102,13 @@ def __init__(self, config, io_dtype, onnx_dtype, ep, cache_dir, extra_options):
 
         # RotaryEmbedding-specific variables
         partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
+        rope_theta = config.rope_theta if hasattr(config, "rope_theta") else 10000
         self.rotemb_attrs = {
             "create_rotary_embedding_caches": True,          # Create cos/sin caches for rotary embeddings
             "partial_rotary_factor": partial_rotary_factor,  # Factor for partial rotary embeddings
             "num_heads": 0,                                  # For partial rotary embeddings (RotaryEmbedding kernel expects a default value of 0)
             "rotary_embedding_dim": 0,                       # For partial rotary embeddings (RotaryEmbedding kernel expects a default value of 0)
-            "theta": config.rope_theta,                      # Base value if calculating cos/sin caches from scratch
+            "theta": rope_theta,                             # Base value if calculating cos/sin caches from scratch
         }
 
         # Attention-specific variables (MHA, GQA, GQA + Rot.Emb., etc.)
@@ -139,8 +140,8 @@ def make_genai_config(self, model_name_or_path, extra_kwargs, out_dir):
                 "context_length": self.context_length,
                 "decoder": {
                     "session_options" : {
-                        "provider_options" : [
-                        ]
+                        "log_id": "onnxruntime-genai",
+                        "provider_options" : []
                     },
                     "filename": self.filename,
                     "head_size": self.head_size,
@@ -1400,23 +1401,26 @@ def create_model(model_name, input_path, output_dir, precision, execution_provid
     # Set input/output precision of ONNX model
     io_dtype = TensorProto.FLOAT if precision in {"int8", "fp32"} or (precision == "int4" and execution_provider == "cpu") else TensorProto.FLOAT16
 
-    # List architecture options in alphabetical order
-    if config.architectures[0] == "GemmaForCausalLM":
-        onnx_model = GemmaModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
-    elif config.architectures[0] == "LlamaForCausalLM":
-        onnx_model = LlamaModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
-    elif config.architectures[0] == "MistralForCausalLM":
-        onnx_model = MistralModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
-    elif config.architectures[0] == "PhiForCausalLM":
-        onnx_model = PhiModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
-    else:
-        raise NotImplementedError(f"The {hf_name} model is not currently supported.")
+    if "config_only" not in extra_options:
+        # List architecture options in alphabetical order
+        if config.architectures[0] == "GemmaForCausalLM":
+            onnx_model = GemmaModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
+        elif config.architectures[0] == "LlamaForCausalLM":
+            onnx_model = LlamaModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
+        elif config.architectures[0] == "MistralForCausalLM":
+            onnx_model = MistralModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
+        elif config.architectures[0] == "PhiForCausalLM":
+            onnx_model = PhiModel(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
+        else:
+            raise NotImplementedError(f"The {hf_name} model is not currently supported.")
 
-    # Make ONNX model
-    onnx_model.make_model(input_path)
+        # Make ONNX model
+        onnx_model.make_model(input_path)
 
-    # Save ONNX model
-    onnx_model.save_model(output_dir)
+        # Save ONNX model
+        onnx_model.save_model(output_dir)
+    else:
+        onnx_model = Model(config, io_dtype, precision, execution_provider, cache_dir, extra_options)
 
     # Make GenAI config
     onnx_model.make_genai_config(hf_name, extra_kwargs, output_dir)
@@ -1497,6 +1501,8 @@ def get_args():
                 filename = Filename for ONNX model (default is 'model.onnx').
                     For models with multiple components, each component is exported to its own ONNX model.
                     The filename for each component will be '<filename>_<component-name>.onnx' (ex: '<filename>_encoder.onnx', '<filename>_decoder.onnx').
+                config_only = Generate config and pre/post processing files only.
+                    Use this option when you already have your optimized and/or quantized ONNX model.
             """),
     )