move to contained directory

README.md

-# ollama
+# Ollama
 
-🙊
+- Run models, fast
+- Download, manage and import models
 
-## Running
-
-Install dependencies:
+## Install
 
 ```
-pip install -r requirements.txt
+pip install ollama
 ```
 
-Put your model in `models/` and run:
+## Example quickstart
 
-```
-python3 ollama.py serve
+```python
+import ollama
+model_name = "huggingface.co/thebloke/llama-7b-ggml"
+model = ollama.pull(model_name)
+ollama.load(model)
+ollama.generate(model_name, "hi")
 ```
 
-To run the app:
+## Reference
 
-```
-cd desktop
-npm install
-npm start
+### `ollama.load`
+
+Load a model from a path or a docker image
+
+```python
+ollama.load("model name")
 ```
 
-## Building
+### `ollama.generate("message")`
 
-If using Apple silicon, you need a Python version that supports arm64:
+Generate a completion
 
-```bash
-wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh
-bash Miniforge3-MacOSX-arm64.sh
+```python
+ollama.generate(model, "hi")
 ```
 
-Get the dependencies:
+### `ollama.models`
 
-```bash
-pip install -r requirements.txt
+List models
+
+```
+models = ollama.models()
 ```
 
-Then build a binary for your current platform:
+### `ollama.serve`
 
-```bash
-python3 build.py
-```
+Serve the ollama http server
 
-### Building the app
+## Cooing Soon
 
+### `ollama.pull`
+
+Examples:
+
+```python
+ollama.pull("huggingface.co/thebloke/llama-7b-ggml")
 ```
-cd desktop
-npm run package
-```
 
-## API
+### `ollama.import`
+
+Import an existing model into the model store
+
+```python
+ollama.import("./path/to/model")
+```
 
-### `GET /models`
+### `ollama.search`
 
-Returns a list of available models
+Search for compatible models that Ollama can run
 
-### `POST /generate`
+```python
+ollama.search("llama-7b")
+```
 
-Generates completions as a series of JSON objects
+## Future CLI
 
-model: `string` - The name of the model to use in the `models` folder.
-prompt: `string` - The prompt to use.
+```
+ollama run huggingface.co/thebloke/llama-7b-ggml
+```

ollama.py → ollama/ollama.py

@@ -2,6 +2,9 @@ import json
 import os
 import threading
 import click
+from transformers import AutoModel
+from tqdm import tqdm
+from pathlib import Path
 from llama_cpp import Llama
 from flask import Flask, Response, stream_with_context, request
 from flask_cors import CORS
@@ -15,33 +18,81 @@ llms = {}
 lock = threading.Lock()
 
 
-def load(model):
+def models_directory():
+    home_dir = Path.home()
+    models_dir = home_dir / ".ollama/models"
+
+    if not models_dir.exists():
+        models_dir.mkdir(parents=True)
+
+    return models_dir
+
+
+def load(model=None, path=None):
+    """
+    Load a model.
+
+    The model can be specified by providing either the path or the model name,
+    but not both. If both are provided, this function will raise a ValueError.
+    If the model does not exist or could not be loaded, this function returns an error.
+
+    Args:
+    model (str, optional): The name of the model to load.
+    path (str, optional): The path to the model file.
+
+    Returns:
+    dict or None: If the model cannot be loaded, a dictionary with an 'error' key is returned.
+                  If the model is successfully loaded, None is returned.
+    """
+
     with lock:
-        if not os.path.exists(f"{model}"):
-            return {"error": "The model does not exist."}
-        if model not in llms:
-            llms[model] = Llama(model_path=f"{model}")
+        if path is not None and model is not None:
+            raise ValueError(
+                "Both path and model are specified. Please provide only one of them."
+            )
+        elif path is not None:
+            name = os.path.basename(path)
+            load_from = path
+        elif model is not None:
+            name = model
+            dir = models_directory()
+            load_from = str(dir / f"{model}.bin")
+        else:
+            raise ValueError("Either path or model must be specified.")
+
+        if not os.path.exists(load_from):
+            return {"error": f"The model at {load_from} does not exist."}
+
+        if name not in llms:
+            # TODO: download model from a repository if it does not exist
+            llms[name] = Llama(model_path=load_from)
+
+    # TODO: this should start a persistent instance of ollama with the model loaded
     return None
 
 
 def unload(model):
-    with lock:
-        if not os.path.exists(f"{model}"):
-            return {"error": "The model does not exist."}
-        llms.pop(model, None)
-    return None
+    """
+    Unload a model.
+
+    Remove a model from the list of loaded models. If the model is not loaded, this is a no-op.
 
+    Args:
+    model (str): The name of the model to unload.
+    """
+    llms.pop(model, None)
 
-def query(model, prompt):
+
+def generate(model, prompt):
     # auto load
     error = load(model)
+    print(error)
     if error is not None:
         return error
     generated = llms[model](
         str(prompt),  # TODO: optimize prompt based on model
         max_tokens=4096,
         stop=["Q:", "\n"],
-        echo=True,
         stream=True,
     )
     for output in generated:
@@ -49,7 +100,8 @@ def query(model, prompt):
 
 
 def models():
-    all_files = os.listdir("./models")
+    dir = models_directory()
+    all_files = os.listdir(dir)
     bin_files = [
         file.replace(".bin", "") for file in all_files if file.endswith(".bin")
     ]
@@ -74,9 +126,7 @@ def unload_route_handler():
     model = data.get("model")
     if not model:
         return Response("Model is required", status=400)
-    error = unload(model)
-    if error is not None:
-        return error
+    unload(model)
     return Response(status=204)
 
 
@@ -92,7 +142,7 @@ def generate_route_handler():
     if not os.path.exists(f"{model}"):
         return {"error": "The model does not exist."}, 400
     return Response(
-        stream_with_context(query(model, prompt)), mimetype="text/event-stream"
+        stream_with_context(generate(model, prompt)), mimetype="text/event-stream"
     )
 
 
@@ -118,20 +168,56 @@ def serve(port, debug):
     app.run(host="0.0.0.0", port=port, debug=debug)
 
 
-@cli.command()
-@click.option("--model", default="vicuna-7b-v1.3.ggmlv3.q8_0", help="The model to use")
+@cli.command(name="load")
+@click.argument("model")
+@click.option("--file", default=False, help="Indicates that a file path is provided")
+def load_cli(model, file):
+    if file:
+        error = load(path=model)
+    else:
+        error = load(model)
+    if error is not None:
+        print(error)
+        return
+    print("Model loaded")
+
+
+@cli.command(name="generate")
+@click.argument("model")
 @click.option("--prompt", default="", help="The prompt for the model")
-def generate(model, prompt):
+def generate_cli(model, prompt):
     if prompt == "":
         prompt = input("Prompt: ")
     output = ""
     prompt = template(model, prompt)
-    for generated in query(model, prompt):
+    for generated in generate(model, prompt):
         generated_json = json.loads(generated)
         text = generated_json["choices"][0]["text"]
         output += text
         print(f"\r{output}", end="", flush=True)
 
 
+def download_model(model_name):
+    dir = models_directory()
+    AutoModel.from_pretrained(model_name, cache_dir=dir)
+
+
+@cli.command(name="models")
+def models_cli():
+    print(models())
+
+
+@cli.command(name="pull")
+@click.argument("model")
+def pull_cli(model):
+    print("not implemented")
+
+
+@cli.command(name="import")
+@click.argument("model")
+def import_cli(model):
+    print("not implemented")
+
+
 if __name__ == "__main__":
     cli()