New web interface, move llm engine to separate process

tools/old_llm_engines/qwq_llm_engine.ipynb

@@ -0,0 +1,227 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "source /root/venvs/sia/bin/activate\n",
+    "apt-get update && apt-get install -y cuda-toolkit\n",
+    "pip install flash-attn --no-build-isolation"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Unsloth should be imported before transformers to ensure all optimizations are applied.\n",
+    "from unsloth import FastLanguageModel"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from pathlib import Path\n",
+    "from threading import Thread\n",
+    "from transformers import AutoTokenizer, TextIteratorStreamer, pipeline\n",
+    "from xml_schema_validator import XmlLogitsProcessor"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "temperature = 0.6\n",
+    "model_path = \"/root/models/current\"\n",
+    "xml_schema_text = Path(\"/root/sia/action_schema.xsd\").read_text()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Load tokenizer\n",
+    "tokenizer = AutoTokenizer.from_pretrained(\n",
+    "    model_path,\n",
+    "    legacy=False,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Load model\n",
+    "model, _returned_tokenizer = FastLanguageModel.from_pretrained(\n",
+    "    model_path,\n",
+    "    gpu_memory_utilization = 0.5, # Reduce if out of memory\n",
+    "    load_in_4bit=True,\n",
+    "    attn_implementation=\"flash_attention_2\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# enable unsloth optimizations\n",
+    "FastLanguageModel.for_inference(model)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "# Create inference pipeline with memory-efficient settings\n",
+    "pipeline = pipeline(\n",
+    "    \"text-generation\",\n",
+    "    model=model,\n",
+    "    tokenizer=tokenizer,\n",
+    "    return_full_text=False,\n",
+    "    torch_dtype=torch.float16,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "logits_processor = XmlLogitsProcessor(tokenizer, xml_schema_text)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "messages = [\n",
+    "    {\"role\": \"system\", \"content\": \"Always respond in a <write_stdout></write_stdout> xml block.\"},\n",
+    "    {\"role\": \"user\", \"content\": \"Hi, how are you?\"},\n",
+    "    {\"role\": \"assistant\", \"content\": \"\"},\n",
+    "]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "text = tokenizer.apply_chat_template(\n",
+    "    messages,\n",
+    "    tokenize=False,\n",
+    "    add_generation_prompt=False,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "streamer = TextIteratorStreamer(\n",
+    "    tokenizer,\n",
+    "    skip_prompt=True,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generation_kwargs = {\n",
+    "    \"text_inputs\": text,\n",
+    "    \"do_sample\": True,\n",
+    "    \"temperature\": temperature,\n",
+    "    \"streamer\": streamer,\n",
+    "    \"use_cache\": True,\n",
+    "}"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generation_kwargs[\"logits_processor\"] = [logits_processor.copy()]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "generation_thread = Thread(\n",
+    "    target=pipeline,\n",
+    "    kwargs=generation_kwargs\n",
+    ")\n",
+    "\n",
+    "generation_thread.start()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for text in streamer:\n",
+    "    print(text, end=\"\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "\n",
+    "generation_thread.join()"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "sia",
+   "language": "python",
+   "name": "sia"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}