refactor

ollama/__init__.py

@@ -0,0 +1,9 @@
+from ollama.model import models
+from ollama.engine import generate, load, unload
+
+__all__ = [
+    'models',
+    'generate',
+    'load',
+    'unload',
+]

ollama/cmd/cli.py

@@ -0,0 +1,43 @@
+import json
+from pathlib import Path
+from argparse import ArgumentParser
+
+from ollama import model, engine
+from ollama.cmd import server
+
+
+def main():
+    parser = ArgumentParser()
+    parser.add_argument('--models-home', default=Path.home() / '.ollama' / 'models')
+
+    subparsers = parser.add_subparsers()
+
+    server.set_parser(subparsers.add_parser('serve'))
+
+    list_parser = subparsers.add_parser('list')
+    list_parser.set_defaults(fn=list)
+
+    generate_parser = subparsers.add_parser('generate')
+    generate_parser.add_argument('model')
+    generate_parser.add_argument('prompt')
+    generate_parser.set_defaults(fn=generate)
+
+    args = parser.parse_args()
+    args = vars(args)
+
+    fn = args.pop('fn')
+    fn(**args)
+
+
+def list(*args, **kwargs):
+    for m in model.models(*args, **kwargs):
+        print(m)
+
+
+def generate(*args, **kwargs):
+    for output in engine.generate(*args, **kwargs):
+        output = json.loads(output)
+
+        choices = output.get('choices', [])
+        if len(choices) > 0:
+            print(choices[0].get('text', ''), end='')

ollama/cmd/server.py

@@ -0,0 +1,75 @@
+from aiohttp import web
+
+from ollama import engine
+
+
+def set_parser(parser):
+    parser.add_argument('--host', default='127.0.0.1')
+    parser.add_argument('--port', default=7734)
+    parser.set_defaults(fn=serve)
+
+
+def serve(models_home='.', *args, **kwargs):
+    app = web.Application()
+    app.add_routes([
+        web.post('/load', load),
+        web.post('/unload', unload),
+        web.post('/generate', generate),
+    ])
+
+    app.update({
+        'llms': {},
+        'models_home': models_home,
+    })
+
+    web.run_app(app, **kwargs)
+
+
+async def load(request):
+    body = await request.json()
+    model = body.get('model')
+    if not model:
+        raise web.HTTPBadRequest()
+
+    kwargs = {
+        'llms': request.app.get('llms'),
+        'models_home': request.app.get('models_home'),
+    }
+
+    engine.load(model, **kwargs)
+    return web.Response()
+
+
+async def unload(request):
+    body = await request.json()
+    model = body.get('model')
+    if not model:
+        raise web.HTTPBadRequest()
+
+    engine.unload(model, llms=request.app.get('llms'))
+    return web.Response()
+
+
+async def generate(request):
+    body = await request.json()
+    model = body.get('model')
+    if not model:
+        raise web.HTTPBadRequest()
+
+    prompt = body.get('prompt')
+    if not prompt:
+        raise web.HTTPBadRequest()
+
+    response = web.StreamResponse()
+    await response.prepare(request)
+
+    kwargs = {
+        'llms': request.app.get('llms'),
+        'models_home': request.app.get('models_home'),
+    }
+
+    for output in engine.generate(model, prompt, **kwargs):
+        await response.write(output.encode('utf-8'))
+        await response.write(b'\n')
+
+    return response

ollama/engine.py

@@ -0,0 +1,57 @@
+import os
+import json
+import sys
+from contextlib import contextmanager
+from llama_cpp import Llama as LLM
+
+import ollama.model
+
+
+@contextmanager
+def suppress_stderr():
+    stderr = os.dup(sys.stderr.fileno())
+    with open(os.devnull, 'w') as devnull:
+        os.dup2(devnull.fileno(), sys.stderr.fileno())
+        yield
+
+    os.dup2(stderr, sys.stderr.fileno())
+
+
+def generate(model, prompt, models_home='.', llms={}, *args, **kwargs):
+    llm = load(model, models_home=models_home, llms=llms)
+
+    if 'max_tokens' not in kwargs:
+        kwargs.update({'max_tokens': 16384})
+
+    if 'stop' not in kwargs:
+        kwargs.update({'stop': ['Q:', '\n']})
+
+    if 'stream' not in kwargs:
+        kwargs.update({'stream': True})
+
+    for output in llm(prompt, *args, **kwargs):
+        yield json.dumps(output)
+
+
+def load(model, models_home='.', llms={}):
+    llm = llms.get(model, None)
+    if not llm:
+        model_path = {
+            name: path
+            for name, path in ollama.model.models(models_home)
+        }.get(model, None)
+
+        if model_path is None:
+            raise ValueError('Model not found')
+
+        # suppress LLM's output
+        with suppress_stderr():
+            llm = LLM(model_path, verbose=False)
+            llms.update({model: llm})
+
+    return llm
+
+
+def unload(model, llms={}):
+    if model in llms:
+        llms.pop(model)

ollama/model.py

@@ -0,0 +1,9 @@
+from os import walk, path
+
+
+def models(models_home='.', *args, **kwargs):
+    for root, _, files in walk(models_home):
+        for file in files:
+            base, ext = path.splitext(file)
+            if ext == '.bin':
+                yield base, path.join(root, file)