433 lines
32 KiB
Plaintext
433 lines
32 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"I'm pretty happy with my model's accuracy relative to GPT-4. How does it compare cost-wise?\n",
|
|
"\n",
|
|
"I'll really push this to its limits -- let's see how quickly our poor model can classify the [full 2-million-recipe dataset](https://huggingface.co/datasets/corbt/all-recipes) 😈."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Requirement already satisfied: datasets==2.14.4 in /usr/local/lib/python3.10/dist-packages (2.14.4)\n",
|
|
"Requirement already satisfied: vllm==0.1.3 in /usr/local/lib/python3.10/dist-packages (0.1.3)\n",
|
|
"Requirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (1.24.4)\n",
|
|
"Requirement already satisfied: pyarrow>=8.0.0 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (12.0.1)\n",
|
|
"Requirement already satisfied: dill<0.3.8,>=0.3.0 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (0.3.7)\n",
|
|
"Requirement already satisfied: pandas in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (2.0.3)\n",
|
|
"Requirement already satisfied: requests>=2.19.0 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (2.28.1)\n",
|
|
"Requirement already satisfied: tqdm>=4.62.1 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (4.66.1)\n",
|
|
"Requirement already satisfied: xxhash in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (3.3.0)\n",
|
|
"Requirement already satisfied: multiprocess in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (0.70.15)\n",
|
|
"Requirement already satisfied: fsspec[http]>=2021.11.1 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (2023.6.0)\n",
|
|
"Requirement already satisfied: aiohttp in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (3.8.5)\n",
|
|
"Requirement already satisfied: huggingface-hub<1.0.0,>=0.14.0 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (0.16.4)\n",
|
|
"Requirement already satisfied: packaging in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (23.1)\n",
|
|
"Requirement already satisfied: pyyaml>=5.1 in /usr/local/lib/python3.10/dist-packages (from datasets==2.14.4) (6.0)\n",
|
|
"Requirement already satisfied: ninja in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (1.11.1)\n",
|
|
"Requirement already satisfied: psutil in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (5.9.5)\n",
|
|
"Requirement already satisfied: ray>=2.5.1 in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (2.6.3)\n",
|
|
"Requirement already satisfied: sentencepiece in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (0.1.99)\n",
|
|
"Requirement already satisfied: torch>=2.0.0 in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (2.0.1+cu118)\n",
|
|
"Requirement already satisfied: transformers>=4.31.0 in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (4.33.0.dev0)\n",
|
|
"Requirement already satisfied: xformers>=0.0.19 in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (0.0.21)\n",
|
|
"Requirement already satisfied: fastapi in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (0.101.1)\n",
|
|
"Requirement already satisfied: uvicorn in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (0.23.2)\n",
|
|
"Requirement already satisfied: pydantic<2 in /usr/local/lib/python3.10/dist-packages (from vllm==0.1.3) (1.10.12)\n",
|
|
"Requirement already satisfied: attrs>=17.3.0 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (23.1.0)\n",
|
|
"Requirement already satisfied: charset-normalizer<4.0,>=2.0 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (2.1.1)\n",
|
|
"Requirement already satisfied: multidict<7.0,>=4.5 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (6.0.4)\n",
|
|
"Requirement already satisfied: async-timeout<5.0,>=4.0.0a3 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (4.0.3)\n",
|
|
"Requirement already satisfied: yarl<2.0,>=1.0 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (1.9.2)\n",
|
|
"Requirement already satisfied: frozenlist>=1.1.1 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (1.4.0)\n",
|
|
"Requirement already satisfied: aiosignal>=1.1.2 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets==2.14.4) (1.3.1)\n",
|
|
"Requirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from huggingface-hub<1.0.0,>=0.14.0->datasets==2.14.4) (3.9.0)\n",
|
|
"Requirement already satisfied: typing-extensions>=3.7.4.3 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub<1.0.0,>=0.14.0->datasets==2.14.4) (4.7.1)\n",
|
|
"Requirement already satisfied: click>=7.0 in /usr/local/lib/python3.10/dist-packages (from ray>=2.5.1->vllm==0.1.3) (8.1.7)\n",
|
|
"Requirement already satisfied: jsonschema in /usr/local/lib/python3.10/dist-packages (from ray>=2.5.1->vllm==0.1.3) (4.18.0)\n",
|
|
"Requirement already satisfied: msgpack<2.0.0,>=1.0.0 in /usr/local/lib/python3.10/dist-packages (from ray>=2.5.1->vllm==0.1.3) (1.0.5)\n",
|
|
"Requirement already satisfied: protobuf!=3.19.5,>=3.15.3 in /usr/local/lib/python3.10/dist-packages (from ray>=2.5.1->vllm==0.1.3) (4.24.1)\n",
|
|
"Requirement already satisfied: grpcio>=1.42.0 in /usr/local/lib/python3.10/dist-packages (from ray>=2.5.1->vllm==0.1.3) (1.57.0)\n",
|
|
"Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets==2.14.4) (3.4)\n",
|
|
"Requirement already satisfied: urllib3<1.27,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets==2.14.4) (1.26.13)\n",
|
|
"Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets==2.14.4) (2022.12.7)\n",
|
|
"Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from torch>=2.0.0->vllm==0.1.3) (1.11.1)\n",
|
|
"Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from torch>=2.0.0->vllm==0.1.3) (3.0)\n",
|
|
"Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from torch>=2.0.0->vllm==0.1.3) (3.1.2)\n",
|
|
"Requirement already satisfied: triton==2.0.0 in /usr/local/lib/python3.10/dist-packages (from torch>=2.0.0->vllm==0.1.3) (2.0.0)\n",
|
|
"Requirement already satisfied: cmake in /usr/local/lib/python3.10/dist-packages (from triton==2.0.0->torch>=2.0.0->vllm==0.1.3) (3.25.0)\n",
|
|
"Requirement already satisfied: lit in /usr/local/lib/python3.10/dist-packages (from triton==2.0.0->torch>=2.0.0->vllm==0.1.3) (15.0.7)\n",
|
|
"Requirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.10/dist-packages (from transformers>=4.31.0->vllm==0.1.3) (2023.8.8)\n",
|
|
"Requirement already satisfied: tokenizers!=0.11.3,<0.14,>=0.11.1 in /usr/local/lib/python3.10/dist-packages (from transformers>=4.31.0->vllm==0.1.3) (0.13.3)\n",
|
|
"Requirement already satisfied: safetensors>=0.3.1 in /usr/local/lib/python3.10/dist-packages (from transformers>=4.31.0->vllm==0.1.3) (0.3.2)\n",
|
|
"Requirement already satisfied: starlette<0.28.0,>=0.27.0 in /usr/local/lib/python3.10/dist-packages (from fastapi->vllm==0.1.3) (0.27.0)\n",
|
|
"Requirement already satisfied: python-dateutil>=2.8.2 in /usr/local/lib/python3.10/dist-packages (from pandas->datasets==2.14.4) (2.8.2)\n",
|
|
"Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.10/dist-packages (from pandas->datasets==2.14.4) (2023.3)\n",
|
|
"Requirement already satisfied: tzdata>=2022.1 in /usr/local/lib/python3.10/dist-packages (from pandas->datasets==2.14.4) (2023.3)\n",
|
|
"Requirement already satisfied: h11>=0.8 in /usr/local/lib/python3.10/dist-packages (from uvicorn->vllm==0.1.3) (0.14.0)\n",
|
|
"Requirement already satisfied: six>=1.5 in /usr/lib/python3/dist-packages (from python-dateutil>=2.8.2->pandas->datasets==2.14.4) (1.16.0)\n",
|
|
"Requirement already satisfied: anyio<5,>=3.4.0 in /usr/local/lib/python3.10/dist-packages (from starlette<0.28.0,>=0.27.0->fastapi->vllm==0.1.3) (3.7.1)\n",
|
|
"Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->torch>=2.0.0->vllm==0.1.3) (2.1.2)\n",
|
|
"Requirement already satisfied: jsonschema-specifications>=2023.03.6 in /usr/local/lib/python3.10/dist-packages (from jsonschema->ray>=2.5.1->vllm==0.1.3) (2023.6.1)\n",
|
|
"Requirement already satisfied: referencing>=0.28.4 in /usr/local/lib/python3.10/dist-packages (from jsonschema->ray>=2.5.1->vllm==0.1.3) (0.29.1)\n",
|
|
"Requirement already satisfied: rpds-py>=0.7.1 in /usr/local/lib/python3.10/dist-packages (from jsonschema->ray>=2.5.1->vllm==0.1.3) (0.8.10)\n",
|
|
"Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->torch>=2.0.0->vllm==0.1.3) (1.2.1)\n",
|
|
"Requirement already satisfied: sniffio>=1.1 in /usr/local/lib/python3.10/dist-packages (from anyio<5,>=3.4.0->starlette<0.28.0,>=0.27.0->fastapi->vllm==0.1.3) (1.3.0)\n",
|
|
"Requirement already satisfied: exceptiongroup in /usr/local/lib/python3.10/dist-packages (from anyio<5,>=3.4.0->starlette<0.28.0,>=0.27.0->fastapi->vllm==0.1.3) (1.1.2)\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
|
|
"\u001b[0m\n",
|
|
"\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m A new release of pip is available: \u001b[0m\u001b[31;49m23.1.2\u001b[0m\u001b[39;49m -> \u001b[0m\u001b[32;49m23.2.1\u001b[0m\n",
|
|
"\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m To update, run: \u001b[0m\u001b[32;49mpython -m pip install --upgrade pip\u001b[0m\n",
|
|
"Note: you may need to restart the kernel to use updated packages.\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"%pip install datasets==2.14.4 vllm==0.1.3"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Number of recipes: 2,147,248\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"from datasets import load_dataset\n",
|
|
"\n",
|
|
"all_recipes = load_dataset(\"corbt/all-recipes\")[\"train\"][\"input\"]\n",
|
|
"\n",
|
|
"print(f\"Number of recipes: {len(all_recipes):,}\")\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 5,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"INFO 08-24 19:38:29 llm_engine.py:70] Initializing an LLM engine with config: model='./models/run1/merged', tokenizer='./models/run1/merged', tokenizer_mode=auto, trust_remote_code=False, dtype=torch.float16, use_dummy_weights=False, download_dir=None, use_np_weights=False, tensor_parallel_size=1, seed=0)\n",
|
|
"INFO 08-24 19:39:48 llm_engine.py:196] # GPU blocks: 3419, # CPU blocks: 512\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"from vllm import LLM, SamplingParams\n",
|
|
"\n",
|
|
"llm = LLM(model=\"./models/run1/merged\", max_num_batched_tokens=4096)\n",
|
|
"\n",
|
|
"sampling_params = SamplingParams(\n",
|
|
" # 120 should be fine for the work we're doing here.\n",
|
|
" max_tokens=120,\n",
|
|
" # This is a deterministic task so temperature=0 is best.\n",
|
|
" temperature=0,\n",
|
|
")\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 6,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Start time: 1692906050.3340027\n",
|
|
"Processing recipes 0 to 10,000...\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processed prompts: 100%|██████████| 10000/10000 [04:51<00:00, 34.30it/s]\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processing recipes 10,000 to 20,000...\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processed prompts: 100%|██████████| 10000/10000 [04:54<00:00, 33.98it/s]\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processing recipes 20,000 to 30,000...\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processed prompts: 100%|██████████| 10000/10000 [04:53<00:00, 34.11it/s]\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processing recipes 30,000 to 40,000...\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processed prompts: 100%|██████████| 10000/10000 [04:53<00:00, 34.11it/s]\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processing recipes 40,000 to 50,000...\n"
|
|
]
|
|
},
|
|
{
|
|
"name": "stderr",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"Processed prompts: 48%|████▊ | 4796/10000 [02:21<03:18, 26.22it/s]"
|
|
]
|
|
},
|
|
{
|
|
"ename": "KeyboardInterrupt",
|
|
"evalue": "",
|
|
"output_type": "error",
|
|
"traceback": [
|
|
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
|
|
"\u001b[0;31mKeyboardInterrupt\u001b[0m Traceback (most recent call last)",
|
|
"Cell \u001b[0;32mIn[6], line 12\u001b[0m\n\u001b[1;32m 10\u001b[0m \u001b[39mfor\u001b[39;00m i \u001b[39min\u001b[39;00m \u001b[39mrange\u001b[39m(\u001b[39m0\u001b[39m, \u001b[39mlen\u001b[39m(all_recipes), BATCH_SIZE):\n\u001b[1;32m 11\u001b[0m \u001b[39mprint\u001b[39m(\u001b[39mf\u001b[39m\u001b[39m\"\u001b[39m\u001b[39mProcessing recipes \u001b[39m\u001b[39m{\u001b[39;00mi\u001b[39m:\u001b[39;00m\u001b[39m,\u001b[39m\u001b[39m}\u001b[39;00m\u001b[39m to \u001b[39m\u001b[39m{\u001b[39;00mi\u001b[39m+\u001b[39mBATCH_SIZE\u001b[39m:\u001b[39;00m\u001b[39m,\u001b[39m\u001b[39m}\u001b[39;00m\u001b[39m...\u001b[39m\u001b[39m\"\u001b[39m)\n\u001b[0;32m---> 12\u001b[0m outputs \u001b[39m=\u001b[39m llm\u001b[39m.\u001b[39;49mgenerate(all_recipes[i:i\u001b[39m+\u001b[39;49mBATCH_SIZE], sampling_params\u001b[39m=\u001b[39;49msampling_params)\n\u001b[1;32m 14\u001b[0m all_outputs\u001b[39m.\u001b[39mextend([o\u001b[39m.\u001b[39moutputs[\u001b[39m0\u001b[39m]\u001b[39m.\u001b[39mtext \u001b[39mfor\u001b[39;00m o \u001b[39min\u001b[39;00m outputs])\n\u001b[1;32m 16\u001b[0m end_time \u001b[39m=\u001b[39m time\u001b[39m.\u001b[39mtime()\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/entrypoints/llm.py:130\u001b[0m, in \u001b[0;36mLLM.generate\u001b[0;34m(self, prompts, sampling_params, prompt_token_ids, use_tqdm)\u001b[0m\n\u001b[1;32m 128\u001b[0m token_ids \u001b[39m=\u001b[39m prompt_token_ids[i]\n\u001b[1;32m 129\u001b[0m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_add_request(prompt, sampling_params, token_ids)\n\u001b[0;32m--> 130\u001b[0m \u001b[39mreturn\u001b[39;00m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49m_run_engine(use_tqdm)\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/entrypoints/llm.py:150\u001b[0m, in \u001b[0;36mLLM._run_engine\u001b[0;34m(self, use_tqdm)\u001b[0m\n\u001b[1;32m 148\u001b[0m outputs: List[RequestOutput] \u001b[39m=\u001b[39m []\n\u001b[1;32m 149\u001b[0m \u001b[39mwhile\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mllm_engine\u001b[39m.\u001b[39mhas_unfinished_requests():\n\u001b[0;32m--> 150\u001b[0m step_outputs \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49mllm_engine\u001b[39m.\u001b[39;49mstep()\n\u001b[1;32m 151\u001b[0m \u001b[39mfor\u001b[39;00m output \u001b[39min\u001b[39;00m step_outputs:\n\u001b[1;32m 152\u001b[0m \u001b[39mif\u001b[39;00m output\u001b[39m.\u001b[39mfinished:\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/engine/llm_engine.py:313\u001b[0m, in \u001b[0;36mLLMEngine.step\u001b[0;34m(self)\u001b[0m\n\u001b[1;32m 307\u001b[0m \u001b[39mreturn\u001b[39;00m [\n\u001b[1;32m 308\u001b[0m RequestOutput\u001b[39m.\u001b[39mfrom_seq_group(seq_group)\n\u001b[1;32m 309\u001b[0m \u001b[39mfor\u001b[39;00m seq_group \u001b[39min\u001b[39;00m scheduler_outputs\u001b[39m.\u001b[39mignored_seq_groups\n\u001b[1;32m 310\u001b[0m ]\n\u001b[1;32m 312\u001b[0m \u001b[39m# Execute the model.\u001b[39;00m\n\u001b[0;32m--> 313\u001b[0m output \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49m_run_workers(\n\u001b[1;32m 314\u001b[0m \u001b[39m\"\u001b[39;49m\u001b[39mexecute_model\u001b[39;49m\u001b[39m\"\u001b[39;49m,\n\u001b[1;32m 315\u001b[0m seq_group_metadata_list\u001b[39m=\u001b[39;49mseq_group_metadata_list,\n\u001b[1;32m 316\u001b[0m blocks_to_swap_in\u001b[39m=\u001b[39;49mscheduler_outputs\u001b[39m.\u001b[39;49mblocks_to_swap_in,\n\u001b[1;32m 317\u001b[0m blocks_to_swap_out\u001b[39m=\u001b[39;49mscheduler_outputs\u001b[39m.\u001b[39;49mblocks_to_swap_out,\n\u001b[1;32m 318\u001b[0m blocks_to_copy\u001b[39m=\u001b[39;49mscheduler_outputs\u001b[39m.\u001b[39;49mblocks_to_copy,\n\u001b[1;32m 319\u001b[0m )\n\u001b[1;32m 320\u001b[0m \u001b[39m# Update the scheduler with the model outputs.\u001b[39;00m\n\u001b[1;32m 321\u001b[0m seq_groups \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mscheduler\u001b[39m.\u001b[39mupdate(output)\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/engine/llm_engine.py:470\u001b[0m, in \u001b[0;36mLLMEngine._run_workers\u001b[0;34m(self, method, get_all_outputs, *args, **kwargs)\u001b[0m\n\u001b[1;32m 467\u001b[0m \u001b[39melse\u001b[39;00m:\n\u001b[1;32m 468\u001b[0m executor \u001b[39m=\u001b[39m \u001b[39mgetattr\u001b[39m(worker, method)\n\u001b[0;32m--> 470\u001b[0m output \u001b[39m=\u001b[39m executor(\u001b[39m*\u001b[39;49margs, \u001b[39m*\u001b[39;49m\u001b[39m*\u001b[39;49mkwargs)\n\u001b[1;32m 471\u001b[0m all_outputs\u001b[39m.\u001b[39mappend(output)\n\u001b[1;32m 473\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mparallel_config\u001b[39m.\u001b[39mworker_use_ray:\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py:115\u001b[0m, in \u001b[0;36mcontext_decorator.<locals>.decorate_context\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m 112\u001b[0m \u001b[39m@functools\u001b[39m\u001b[39m.\u001b[39mwraps(func)\n\u001b[1;32m 113\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39mdecorate_context\u001b[39m(\u001b[39m*\u001b[39margs, \u001b[39m*\u001b[39m\u001b[39m*\u001b[39mkwargs):\n\u001b[1;32m 114\u001b[0m \u001b[39mwith\u001b[39;00m ctx_factory():\n\u001b[0;32m--> 115\u001b[0m \u001b[39mreturn\u001b[39;00m func(\u001b[39m*\u001b[39;49margs, \u001b[39m*\u001b[39;49m\u001b[39m*\u001b[39;49mkwargs)\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/worker/worker.py:293\u001b[0m, in \u001b[0;36mWorker.execute_model\u001b[0;34m(self, seq_group_metadata_list, blocks_to_swap_in, blocks_to_swap_out, blocks_to_copy)\u001b[0m\n\u001b[1;32m 289\u001b[0m input_tokens, input_positions, input_metadata \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_prepare_inputs(\n\u001b[1;32m 290\u001b[0m seq_group_metadata_list)\n\u001b[1;32m 292\u001b[0m \u001b[39m# Execute the model.\u001b[39;00m\n\u001b[0;32m--> 293\u001b[0m output \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49mmodel(\n\u001b[1;32m 294\u001b[0m input_ids\u001b[39m=\u001b[39;49minput_tokens,\n\u001b[1;32m 295\u001b[0m positions\u001b[39m=\u001b[39;49minput_positions,\n\u001b[1;32m 296\u001b[0m kv_caches\u001b[39m=\u001b[39;49m\u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49mgpu_cache,\n\u001b[1;32m 297\u001b[0m input_metadata\u001b[39m=\u001b[39;49minput_metadata,\n\u001b[1;32m 298\u001b[0m cache_events\u001b[39m=\u001b[39;49mcache_events,\n\u001b[1;32m 299\u001b[0m )\n\u001b[1;32m 300\u001b[0m \u001b[39mreturn\u001b[39;00m output\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py:1501\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1496\u001b[0m \u001b[39m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[1;32m 1497\u001b[0m \u001b[39m# this function, and just call forward.\u001b[39;00m\n\u001b[1;32m 1498\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39mnot\u001b[39;00m (\u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_backward_hooks \u001b[39mor\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_backward_pre_hooks \u001b[39mor\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_forward_hooks \u001b[39mor\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_forward_pre_hooks\n\u001b[1;32m 1499\u001b[0m \u001b[39mor\u001b[39;00m _global_backward_pre_hooks \u001b[39mor\u001b[39;00m _global_backward_hooks\n\u001b[1;32m 1500\u001b[0m \u001b[39mor\u001b[39;00m _global_forward_hooks \u001b[39mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[0;32m-> 1501\u001b[0m \u001b[39mreturn\u001b[39;00m forward_call(\u001b[39m*\u001b[39;49margs, \u001b[39m*\u001b[39;49m\u001b[39m*\u001b[39;49mkwargs)\n\u001b[1;32m 1502\u001b[0m \u001b[39m# Do not call functions when jit is used\u001b[39;00m\n\u001b[1;32m 1503\u001b[0m full_backward_hooks, non_full_backward_hooks \u001b[39m=\u001b[39m [], []\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/model_executor/models/llama.py:255\u001b[0m, in \u001b[0;36mLlamaForCausalLM.forward\u001b[0;34m(self, input_ids, positions, kv_caches, input_metadata, cache_events)\u001b[0m\n\u001b[1;32m 245\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39mforward\u001b[39m(\n\u001b[1;32m 246\u001b[0m \u001b[39mself\u001b[39m,\n\u001b[1;32m 247\u001b[0m input_ids: torch\u001b[39m.\u001b[39mTensor,\n\u001b[0;32m (...)\u001b[0m\n\u001b[1;32m 251\u001b[0m cache_events: Optional[List[torch\u001b[39m.\u001b[39mcuda\u001b[39m.\u001b[39mEvent]],\n\u001b[1;32m 252\u001b[0m ) \u001b[39m-\u001b[39m\u001b[39m>\u001b[39m Dict[\u001b[39mint\u001b[39m, SequenceOutputs]:\n\u001b[1;32m 253\u001b[0m hidden_states \u001b[39m=\u001b[39m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39mmodel(input_ids, positions, kv_caches,\n\u001b[1;32m 254\u001b[0m input_metadata, cache_events)\n\u001b[0;32m--> 255\u001b[0m next_tokens \u001b[39m=\u001b[39m \u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49msampler(\u001b[39mself\u001b[39;49m\u001b[39m.\u001b[39;49mlm_head\u001b[39m.\u001b[39;49mweight, hidden_states,\n\u001b[1;32m 256\u001b[0m input_metadata)\n\u001b[1;32m 257\u001b[0m \u001b[39mreturn\u001b[39;00m next_tokens\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py:1501\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1496\u001b[0m \u001b[39m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[1;32m 1497\u001b[0m \u001b[39m# this function, and just call forward.\u001b[39;00m\n\u001b[1;32m 1498\u001b[0m \u001b[39mif\u001b[39;00m \u001b[39mnot\u001b[39;00m (\u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_backward_hooks \u001b[39mor\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_backward_pre_hooks \u001b[39mor\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_forward_hooks \u001b[39mor\u001b[39;00m \u001b[39mself\u001b[39m\u001b[39m.\u001b[39m_forward_pre_hooks\n\u001b[1;32m 1499\u001b[0m \u001b[39mor\u001b[39;00m _global_backward_pre_hooks \u001b[39mor\u001b[39;00m _global_backward_hooks\n\u001b[1;32m 1500\u001b[0m \u001b[39mor\u001b[39;00m _global_forward_hooks \u001b[39mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[0;32m-> 1501\u001b[0m \u001b[39mreturn\u001b[39;00m forward_call(\u001b[39m*\u001b[39;49margs, \u001b[39m*\u001b[39;49m\u001b[39m*\u001b[39;49mkwargs)\n\u001b[1;32m 1502\u001b[0m \u001b[39m# Do not call functions when jit is used\u001b[39;00m\n\u001b[1;32m 1503\u001b[0m full_backward_hooks, non_full_backward_hooks \u001b[39m=\u001b[39m [], []\n",
|
|
"File \u001b[0;32m/usr/local/lib/python3.10/dist-packages/vllm/model_executor/layers/sampler.py:44\u001b[0m, in \u001b[0;36mSampler.forward\u001b[0;34m(self, embedding, hidden_states, input_metadata, embedding_bias)\u001b[0m\n\u001b[1;32m 36\u001b[0m \u001b[39mdef\u001b[39;00m \u001b[39mforward\u001b[39m(\n\u001b[1;32m 37\u001b[0m \u001b[39mself\u001b[39m,\n\u001b[1;32m 38\u001b[0m embedding: torch\u001b[39m.\u001b[39mTensor,\n\u001b[0;32m (...)\u001b[0m\n\u001b[1;32m 42\u001b[0m ) \u001b[39m-\u001b[39m\u001b[39m>\u001b[39m Dict[\u001b[39mint\u001b[39m, SequenceOutputs]:\n\u001b[1;32m 43\u001b[0m \u001b[39m# Get the hidden states that we use for sampling.\u001b[39;00m\n\u001b[0;32m---> 44\u001b[0m hidden_states \u001b[39m=\u001b[39m _prune_hidden_states(hidden_states, input_metadata)\n\u001b[1;32m 46\u001b[0m \u001b[39m# Get the logits for the next tokens.\u001b[39;00m\n\u001b[1;32m 47\u001b[0m logits \u001b[39m=\u001b[39m torch\u001b[39m.\u001b[39mmatmul(hidden_states, embedding\u001b[39m.\u001b[39mt())\n",
|
|
"\u001b[0;31mKeyboardInterrupt\u001b[0m: "
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"# We'll process our recipes in batches of 10,000.\n",
|
|
"\n",
|
|
"import time\n",
|
|
"\n",
|
|
"BATCH_SIZE = 10000\n",
|
|
"all_outputs = []\n",
|
|
"\n",
|
|
"start_time = time.time()\n",
|
|
"print(f\"Start time: {start_time}\")\n",
|
|
"for i in range(0, len(all_recipes), BATCH_SIZE):\n",
|
|
" print(f\"Processing recipes {i:,} to {i+BATCH_SIZE:,}...\")\n",
|
|
" outputs = llm.generate(\n",
|
|
" all_recipes[i : i + BATCH_SIZE], sampling_params=sampling_params\n",
|
|
" )\n",
|
|
"\n",
|
|
" all_outputs.extend([o.outputs[0].text for o in outputs])\n",
|
|
"\n",
|
|
"end_time = time.time()\n",
|
|
"print(f\"End time: {end_time}\")\n",
|
|
"print(f\"Total hours: {((end_time - start_time) / 3600):.2f}\")\n",
|
|
"\n",
|
|
"# Ended up running this in a separate script to leave it running in the background.\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"Nice! I've processed all 2,147,248 recipes in under 17 hours. Let's do a cost comparison with GPT-3.5 and GPT-4. I'll use the GPT-4 latency/cost numbers based on the 5000 samples used to generate our model's training data."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 19,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"data": {
|
|
"text/html": [
|
|
"<div>\n",
|
|
"<style scoped>\n",
|
|
" .dataframe tbody tr th:only-of-type {\n",
|
|
" vertical-align: middle;\n",
|
|
" }\n",
|
|
"\n",
|
|
" .dataframe tbody tr th {\n",
|
|
" vertical-align: top;\n",
|
|
" }\n",
|
|
"\n",
|
|
" .dataframe thead th {\n",
|
|
" text-align: right;\n",
|
|
" }\n",
|
|
"</style>\n",
|
|
"<table border=\"1\" class=\"dataframe\">\n",
|
|
" <thead>\n",
|
|
" <tr style=\"text-align: right;\">\n",
|
|
" <th></th>\n",
|
|
" <th>Model</th>\n",
|
|
" <th>Cost to Classify One Recipe</th>\n",
|
|
" <th>Cost to Classify Entire Dataset</th>\n",
|
|
" </tr>\n",
|
|
" </thead>\n",
|
|
" <tbody>\n",
|
|
" <tr>\n",
|
|
" <th>0</th>\n",
|
|
" <td>Llama 2 7B (finetuned)</td>\n",
|
|
" <td>0.000009</td>\n",
|
|
" <td>18.86</td>\n",
|
|
" </tr>\n",
|
|
" <tr>\n",
|
|
" <th>1</th>\n",
|
|
" <td>GPT-3.5</td>\n",
|
|
" <td>0.000481</td>\n",
|
|
" <td>1,033.26</td>\n",
|
|
" </tr>\n",
|
|
" <tr>\n",
|
|
" <th>2</th>\n",
|
|
" <td>GPT-3.5 (finetuned)</td>\n",
|
|
" <td>0.004044</td>\n",
|
|
" <td>8,683.47</td>\n",
|
|
" </tr>\n",
|
|
" <tr>\n",
|
|
" <th>3</th>\n",
|
|
" <td>GPT-4</td>\n",
|
|
" <td>0.010800</td>\n",
|
|
" <td>23,190.28</td>\n",
|
|
" </tr>\n",
|
|
" </tbody>\n",
|
|
"</table>\n",
|
|
"</div>"
|
|
],
|
|
"text/plain": [
|
|
" Model Cost to Classify One Recipe \\\n",
|
|
"0 Llama 2 7B (finetuned) 0.000009 \n",
|
|
"1 GPT-3.5 0.000481 \n",
|
|
"2 GPT-3.5 (finetuned) 0.004044 \n",
|
|
"3 GPT-4 0.010800 \n",
|
|
"\n",
|
|
" Cost to Classify Entire Dataset \n",
|
|
"0 18.86 \n",
|
|
"1 1,033.26 \n",
|
|
"2 8,683.47 \n",
|
|
"3 23,190.28 "
|
|
]
|
|
},
|
|
"execution_count": 19,
|
|
"metadata": {},
|
|
"output_type": "execute_result"
|
|
}
|
|
],
|
|
"source": [
|
|
"import pandas as pd\n",
|
|
"\n",
|
|
"# I used an on-demand Nvidia L40 on RunPod for this, at an hourly cost of $1.14.\n",
|
|
"finetuned_hourly_cost = 1.14\n",
|
|
"\n",
|
|
"finetuned_total_hours = 16.5\n",
|
|
"\n",
|
|
"finetuned_avg_cost = finetuned_hourly_cost * finetuned_total_hours / len(all_recipes)\n",
|
|
"\n",
|
|
"# The average input and output tokens for OpenAI, based on the 5000 recipes I\n",
|
|
"# sent them when generating training data.\n",
|
|
"avg_input_tokens = 276\n",
|
|
"avg_output_tokens = 42\n",
|
|
"\n",
|
|
"# Token pricing from https://openai.com/pricing\n",
|
|
"gpt_4_avg_cost = avg_input_tokens * 0.03 / 1000 + avg_output_tokens * 0.06 / 1000\n",
|
|
"\n",
|
|
"gpt_35_avg_cost = avg_input_tokens * 0.0015 / 1000 + avg_output_tokens * 0.0016 / 1000\n",
|
|
"\n",
|
|
"gpt_35_finetuned_avg_cost = (\n",
|
|
" avg_input_tokens * 0.012 / 1000 + avg_output_tokens * 0.016 / 1000 + 0.06 / 1000\n",
|
|
")\n",
|
|
"\n",
|
|
"costs = pd.DataFrame(\n",
|
|
" {\n",
|
|
" \"Model\": [\n",
|
|
" \"Llama 2 7B (finetuned)\",\n",
|
|
" \"GPT-3.5\",\n",
|
|
" \"GPT-3.5 (finetuned)\",\n",
|
|
" \"GPT-4\",\n",
|
|
" ],\n",
|
|
" \"Cost to Classify One Recipe\": [\n",
|
|
" finetuned_avg_cost,\n",
|
|
" gpt_35_avg_cost,\n",
|
|
" gpt_35_finetuned_avg_cost,\n",
|
|
" gpt_4_avg_cost,\n",
|
|
" ],\n",
|
|
" }\n",
|
|
")\n",
|
|
"\n",
|
|
"costs[\"Cost to Classify Entire Dataset\"] = (\n",
|
|
" costs[\"Cost to Classify One Recipe\"] * len(all_recipes)\n",
|
|
").map(lambda x: f\"{x:,.2f}\")\n",
|
|
"\n",
|
|
"\n",
|
|
"costs\n"
|
|
]
|
|
}
|
|
],
|
|
"metadata": {
|
|
"kernelspec": {
|
|
"display_name": "Python 3 (ipykernel)",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.10.6"
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 2
|
|
}
|