diff --git a/examples/notebooks/OpenHermesFunctionCalling.ipynb b/examples/notebooks/OpenHermesFunctionCalling.ipynb new file mode 100644 index 000000000..c0de3fdc2 --- /dev/null +++ b/examples/notebooks/OpenHermesFunctionCalling.ipynb @@ -0,0 +1,910 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 1, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n" + ] + } + ], + "source": [ + "import json\n", + "import inspect\n", + "from typing import get_type_hints\n", + "\n", + "class Article:\n", + " pass\n", + "\n", + "class Weather:\n", + " pass\n", + "\n", + "class Directions:\n", + " pass\n", + "\n", + "def calculate_mortgage_payment(loan_amount: int, interest_rate: float, loan_term: int) -> float:\n", + " \"\"\"Get the monthly mortgage payment given an interest rate percentage.\"\"\"\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_article_details(title: str, authors: list[str], short_summary: str, date_published: str, tags: list[str]) -> Article:\n", + " '''Get article details from unstructured article text.\n", + "date_published: formatted as \"MM/DD/YYYY\"'''\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_weather(zip_code: str) -> Weather:\n", + " \"\"\"Get the current weather given a zip code.\"\"\"\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_directions(start: str, destination: str) -> Directions:\n", + " \"\"\"Get directions from Google Directions API.\n", + "start: start address as a string including zipcode (if any)\n", + "destination: end address as a string including zipcode (if any)\"\"\"\n", + " \n", + " # TODO: you must implement this to actually call it later\n", + " pass\n", + "\n", + "def get_type_name(t):\n", + " name = str(t)\n", + " if \"list\" in name or \"dict\" in name:\n", + " return name\n", + " else:\n", + " return t.__name__\n", + "\n", + "def serialize_function_to_json(func):\n", + " signature = inspect.signature(func)\n", + " type_hints = get_type_hints(func)\n", + "\n", + " function_info = {\n", + " \"name\": func.__name__,\n", + " \"description\": func.__doc__,\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {}\n", + " },\n", + " \"returns\": type_hints.get('return', 'void').__name__\n", + " }\n", + "\n", + " for name, _ in signature.parameters.items():\n", + " param_type = get_type_name(type_hints.get(name, type(None)))\n", + " function_info[\"parameters\"][\"properties\"][name] = {\"type\": param_type}\n", + "\n", + " return json.dumps(function_info, indent=2)\n", + "\n", + "print(serialize_function_to_json(get_article_details))" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "metadata": {}, + "outputs": [], + "source": [ + "import xml.etree.ElementTree as ET\n", + "import re\n", + "\n", + "def extract_function_calls(completion):\n", + " completion = completion.strip()\n", + " pattern = r\"((.*?))\"\n", + " match = re.search(pattern, completion, re.DOTALL)\n", + " if not match:\n", + " return None\n", + " \n", + " multiplefn = match.group(1)\n", + " root = ET.fromstring(multiplefn)\n", + " functions = root.findall(\"functioncall\")\n", + " return [json.loads(fn.text) for fn in functions]" + ] + }, + { + "cell_type": "code", + "execution_count": 12, + "metadata": {}, + "outputs": [], + "source": [ + "def generate_hermes_prompt(prompt, functions):\n", + " functions = \"\\n\\n\".join([serialize_function_to_json(fn) for fn in functions])\n", + " prompt = f\"\"\"<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{functions}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {{\"name\": \"function_name\", \"arguments\": {{\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}}}} \n", + " {{\"name\": \"function_name\", \"arguments\": {{\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}}}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "{prompt}<|im_end|>\n", + "<|im_start|>assistant\"\"\"\n", + " return prompt" + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "What's the weather in 10001?<|im_end|>\n", + "<|im_start|>assistant\n", + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.<|im_end|>\n", + "<|im_start|>assistant\n", + "<|im_start|>system\n", + "You are a helpful assistant with access to the following functions:\n", + "\n", + "{\n", + " \"name\": \"get_weather\",\n", + " \"description\": \"Get the current weather given a zip code.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"zip_code\": {\n", + " \"type\": \"str\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Weather\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"calculate_mortgage_payment\",\n", + " \"description\": \"Get the monthly mortgage payment given an interest rate percentage.\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"loan_amount\": {\n", + " \"type\": \"int\"\n", + " },\n", + " \"interest_rate\": {\n", + " \"type\": \"float\"\n", + " },\n", + " \"loan_term\": {\n", + " \"type\": \"int\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"float\"\n", + "}\n", + "\n", + "{\n", + " \"name\": \"get_article_details\",\n", + " \"description\": \"Get article details from unstructured article text.\\ndate_published: formatted as \\\"MM/DD/YYYY\\\"\",\n", + " \"parameters\": {\n", + " \"type\": \"object\",\n", + " \"properties\": {\n", + " \"title\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"authors\": {\n", + " \"type\": \"list[str]\"\n", + " },\n", + " \"short_summary\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"date_published\": {\n", + " \"type\": \"str\"\n", + " },\n", + " \"tags\": {\n", + " \"type\": \"list[str]\"\n", + " }\n", + " }\n", + " },\n", + " \"returns\": \"Article\"\n", + "}\n", + "\n", + "To use these functions respond with:\n", + "\n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " {\"name\": \"function_name\", \"arguments\": {\"arg_1\": \"value_1\", \"arg_2\": value_2, ...}} \n", + " ...\n", + "\n", + "\n", + "Edge cases you must handle:\n", + "- If there are no functions that match the user request, you will respond politely that you cannot help.<|im_end|>\n", + "<|im_start|>user\n", + "What's the current exchange rate for USD to EUR?<|im_end|>\n", + "<|im_start|>assistant\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 10001?\",\n", + " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", + " \"What's the current exchange rate for USD to EUR?\"\n", + "]\n", + "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", + "\n", + "for prompt in prompts:\n", + " print(generate_hermes_prompt(prompt, functions))" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "metadata": {}, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "ggml_init_cublas: GGML_CUDA_FORCE_MMQ: no\n", + "ggml_init_cublas: CUDA_USE_TENSOR_CORES: yes\n", + "ggml_init_cublas: found 1 CUDA devices:\n", + " Device 0: NVIDIA GeForce RTX 2060, compute capability 7.5\n", + "llama_model_loader: loaded meta data with 20 key-value pairs and 291 tensors from ../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf (version GGUF V3 (latest))\n", + "llama_model_loader: - tensor 0: token_embd.weight q4_K [ 4096, 32002, 1, 1 ]\n", + "llama_model_loader: - tensor 1: blk.0.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 2: blk.0.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 3: blk.0.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 4: blk.0.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 5: blk.0.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 6: blk.0.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 7: blk.0.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 8: blk.0.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 9: blk.0.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 10: blk.1.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 11: blk.1.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 12: blk.1.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 13: blk.1.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 14: blk.1.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 15: blk.1.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 16: blk.1.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 17: blk.1.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 18: blk.1.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 19: blk.2.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 20: blk.2.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 21: blk.2.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 22: blk.2.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 23: blk.2.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 24: blk.2.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 25: blk.2.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 26: blk.2.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 27: blk.2.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 28: blk.3.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 29: blk.3.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 30: blk.3.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 31: blk.3.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 32: blk.3.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 33: blk.3.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 34: blk.3.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 35: blk.3.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 36: blk.3.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 37: blk.4.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 38: blk.4.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 39: blk.4.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 40: blk.4.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 41: blk.4.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 42: blk.4.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 43: blk.4.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 44: blk.4.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 45: blk.4.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 46: blk.5.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 47: blk.5.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 48: blk.5.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 49: blk.5.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 50: blk.5.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 51: blk.5.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 52: blk.5.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 53: blk.5.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 54: blk.5.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 55: blk.6.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 56: blk.6.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 57: blk.6.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 58: blk.6.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 59: blk.6.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 60: blk.6.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 61: blk.6.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 62: blk.6.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 63: blk.6.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 64: blk.7.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 65: blk.7.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 66: blk.7.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 67: blk.7.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 68: blk.7.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 69: blk.7.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 70: blk.7.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 71: blk.7.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 72: blk.7.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 73: blk.8.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 74: blk.8.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 75: blk.8.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 76: blk.8.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 77: blk.8.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 78: blk.8.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 79: blk.8.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 80: blk.8.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 81: blk.8.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 82: blk.9.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 83: blk.9.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 84: blk.9.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 85: blk.9.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 86: blk.9.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 87: blk.9.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 88: blk.9.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 89: blk.9.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 90: blk.9.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 91: blk.10.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 92: blk.10.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 93: blk.10.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 94: blk.10.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 95: blk.10.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 96: blk.10.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 97: blk.10.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 98: blk.10.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 99: blk.10.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 100: blk.11.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 101: blk.11.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 102: blk.11.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 103: blk.11.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 104: blk.11.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 105: blk.11.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 106: blk.11.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 107: blk.11.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 108: blk.11.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 109: blk.12.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 110: blk.12.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 111: blk.12.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 112: blk.12.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 113: blk.12.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 114: blk.12.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 115: blk.12.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 116: blk.12.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 117: blk.12.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 118: blk.13.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 119: blk.13.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 120: blk.13.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 121: blk.13.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 122: blk.13.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 123: blk.13.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 124: blk.13.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 125: blk.13.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 126: blk.13.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 127: blk.14.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 128: blk.14.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 129: blk.14.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 130: blk.14.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 131: blk.14.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 132: blk.14.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 133: blk.14.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 134: blk.14.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 135: blk.14.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 136: blk.15.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 137: blk.15.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 138: blk.15.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 139: blk.15.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 140: blk.15.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 141: blk.15.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 142: blk.15.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 143: blk.15.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 144: blk.15.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 145: blk.16.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 146: blk.16.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 147: blk.16.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 148: blk.16.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 149: blk.16.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 150: blk.16.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 151: blk.16.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 152: blk.16.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 153: blk.16.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 154: blk.17.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 155: blk.17.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 156: blk.17.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 157: blk.17.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 158: blk.17.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 159: blk.17.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 160: blk.17.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 161: blk.17.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 162: blk.17.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 163: blk.18.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 164: blk.18.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 165: blk.18.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 166: blk.18.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 167: blk.18.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 168: blk.18.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 169: blk.18.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 170: blk.18.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 171: blk.18.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 172: blk.19.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 173: blk.19.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 174: blk.19.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 175: blk.19.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 176: blk.19.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 177: blk.19.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 178: blk.19.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 179: blk.19.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 180: blk.19.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 181: blk.20.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 182: blk.20.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 183: blk.20.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 184: blk.20.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 185: blk.20.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 186: blk.20.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 187: blk.20.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 188: blk.20.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 189: blk.20.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 190: blk.21.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 191: blk.21.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 192: blk.21.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 193: blk.21.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 194: blk.21.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 195: blk.21.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 196: blk.21.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 197: blk.21.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 198: blk.21.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 199: blk.22.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 200: blk.22.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 201: blk.22.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 202: blk.22.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 203: blk.22.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 204: blk.22.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 205: blk.22.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 206: blk.22.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 207: blk.22.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 208: blk.23.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 209: blk.23.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 210: blk.23.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 211: blk.23.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 212: blk.23.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 213: blk.23.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 214: blk.23.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 215: blk.23.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 216: blk.23.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 217: blk.24.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 218: blk.24.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 219: blk.24.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 220: blk.24.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 221: blk.24.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 222: blk.24.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 223: blk.24.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 224: blk.24.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 225: blk.24.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 226: blk.25.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 227: blk.25.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 228: blk.25.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 229: blk.25.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 230: blk.25.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 231: blk.25.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 232: blk.25.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 233: blk.25.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 234: blk.25.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 235: blk.26.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 236: blk.26.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 237: blk.26.attn_v.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 238: blk.26.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 239: blk.26.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 240: blk.26.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 241: blk.26.ffn_down.weight q4_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 242: blk.26.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 243: blk.26.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 244: blk.27.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 245: blk.27.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 246: blk.27.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 247: blk.27.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 248: blk.27.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 249: blk.27.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 250: blk.27.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 251: blk.27.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 252: blk.27.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 253: blk.28.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 254: blk.28.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 255: blk.28.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 256: blk.28.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 257: blk.28.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 258: blk.28.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 259: blk.28.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 260: blk.28.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 261: blk.28.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 262: blk.29.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 263: blk.29.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 264: blk.29.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 265: blk.29.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 266: blk.29.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 267: blk.29.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 268: blk.29.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 269: blk.29.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 270: blk.29.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 271: blk.30.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 272: blk.30.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 273: blk.30.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 274: blk.30.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 275: blk.30.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 276: blk.30.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 277: blk.30.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 278: blk.30.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 279: blk.30.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 280: blk.31.attn_q.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 281: blk.31.attn_k.weight q4_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 282: blk.31.attn_v.weight q6_K [ 4096, 1024, 1, 1 ]\n", + "llama_model_loader: - tensor 283: blk.31.attn_output.weight q4_K [ 4096, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 284: blk.31.ffn_gate.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 285: blk.31.ffn_up.weight q4_K [ 4096, 14336, 1, 1 ]\n", + "llama_model_loader: - tensor 286: blk.31.ffn_down.weight q6_K [ 14336, 4096, 1, 1 ]\n", + "llama_model_loader: - tensor 287: blk.31.attn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 288: blk.31.ffn_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 289: output_norm.weight f32 [ 4096, 1, 1, 1 ]\n", + "llama_model_loader: - tensor 290: output.weight q6_K [ 4096, 32002, 1, 1 ]\n", + "llama_model_loader: - kv 0: general.architecture str = llama\n", + "llama_model_loader: - kv 1: general.name str = teknium_openhermes-2.5-mistral-7b\n", + "llama_model_loader: - kv 2: llama.context_length u32 = 32768\n", + "llama_model_loader: - kv 3: llama.embedding_length u32 = 4096\n", + "llama_model_loader: - kv 4: llama.block_count u32 = 32\n", + "llama_model_loader: - kv 5: llama.feed_forward_length u32 = 14336\n", + "llama_model_loader: - kv 6: llama.rope.dimension_count u32 = 128\n", + "llama_model_loader: - kv 7: llama.attention.head_count u32 = 32\n", + "llama_model_loader: - kv 8: llama.attention.head_count_kv u32 = 8\n", + "llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000010\n", + "llama_model_loader: - kv 10: llama.rope.freq_base f32 = 10000.000000\n", + "llama_model_loader: - kv 11: general.file_type u32 = 15\n", + "llama_model_loader: - kv 12: tokenizer.ggml.model str = llama\n", + "llama_model_loader: - kv 13: tokenizer.ggml.tokens arr[str,32002] = [\"\", \"\", \"\", \"<0x00>\", \"<...\n", + "llama_model_loader: - kv 14: tokenizer.ggml.scores arr[f32,32002] = [0.000000, 0.000000, 0.000000, 0.0000...\n", + "llama_model_loader: - kv 15: tokenizer.ggml.token_type arr[i32,32002] = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...\n", + "llama_model_loader: - kv 16: tokenizer.ggml.bos_token_id u32 = 1\n", + "llama_model_loader: - kv 17: tokenizer.ggml.eos_token_id u32 = 32000\n", + "llama_model_loader: - kv 18: tokenizer.ggml.padding_token_id u32 = 0\n", + "llama_model_loader: - kv 19: general.quantization_version u32 = 2\n", + "llama_model_loader: - type f32: 65 tensors\n", + "llama_model_loader: - type q4_K: 193 tensors\n", + "llama_model_loader: - type q6_K: 33 tensors\n", + "llm_load_vocab: special tokens definition check successful ( 261/32002 ).\n", + "llm_load_print_meta: format = GGUF V3 (latest)\n", + "llm_load_print_meta: arch = llama\n", + "llm_load_print_meta: vocab type = SPM\n", + "llm_load_print_meta: n_vocab = 32002\n", + "llm_load_print_meta: n_merges = 0\n", + "llm_load_print_meta: n_ctx_train = 32768\n", + "llm_load_print_meta: n_embd = 4096\n", + "llm_load_print_meta: n_head = 32\n", + "llm_load_print_meta: n_head_kv = 8\n", + "llm_load_print_meta: n_layer = 32\n", + "llm_load_print_meta: n_rot = 128\n", + "llm_load_print_meta: n_gqa = 4\n", + "llm_load_print_meta: f_norm_eps = 0.0e+00\n", + "llm_load_print_meta: f_norm_rms_eps = 1.0e-05\n", + "llm_load_print_meta: f_clamp_kqv = 0.0e+00\n", + "llm_load_print_meta: f_max_alibi_bias = 0.0e+00\n", + "llm_load_print_meta: n_ff = 14336\n", + "llm_load_print_meta: rope scaling = linear\n", + "llm_load_print_meta: freq_base_train = 10000.0\n", + "llm_load_print_meta: freq_scale_train = 1\n", + "llm_load_print_meta: n_yarn_orig_ctx = 32768\n", + "llm_load_print_meta: rope_finetuned = unknown\n", + "llm_load_print_meta: model type = 7B\n", + "llm_load_print_meta: model ftype = mostly Q4_K - Medium\n", + "llm_load_print_meta: model params = 7.24 B\n", + "llm_load_print_meta: model size = 4.07 GiB (4.83 BPW) \n", + "llm_load_print_meta: general.name = teknium_openhermes-2.5-mistral-7b\n", + "llm_load_print_meta: BOS token = 1 ''\n", + "llm_load_print_meta: EOS token = 32000 '<|im_end|>'\n", + "llm_load_print_meta: UNK token = 0 ''\n", + "llm_load_print_meta: PAD token = 0 ''\n", + "llm_load_print_meta: LF token = 13 '<0x0A>'\n", + "llm_load_tensors: ggml ctx size = 0.11 MiB\n", + "llm_load_tensors: using CUDA for GPU acceleration\n", + "llm_load_tensors: mem required = 70.42 MiB\n", + "llm_load_tensors: offloading 32 repeating layers to GPU\n", + "llm_load_tensors: offloading non-repeating layers to GPU\n", + "llm_load_tensors: offloaded 35/35 layers to GPU\n", + "llm_load_tensors: VRAM used: 4095.06 MiB\n", + "...............................................................................................\n", + "llama_new_context_with_model: n_ctx = 2048\n", + "llama_new_context_with_model: freq_base = 10000.0\n", + "llama_new_context_with_model: freq_scale = 1\n", + "llama_kv_cache_init: offloading v cache to GPU\n", + "llama_kv_cache_init: offloading k cache to GPU\n", + "llama_kv_cache_init: VRAM kv self = 256.00 MiB\n", + "llama_new_context_with_model: kv self size = 256.00 MiB\n", + "llama_build_graph: non-view tensors processed: 740/740\n", + "llama_new_context_with_model: compute buffer total size = 159.07 MiB\n", + "llama_new_context_with_model: VRAM scratch buffer: 156.00 MiB\n", + "llama_new_context_with_model: total VRAM used: 4507.07 MiB (model: 4095.06 MiB, context: 412.00 MiB)\n" + ] + } + ], + "source": [ + "import llama_cpp\n", + "\n", + "llama = llama_cpp.Llama(model_path=\"../../models/OpenHermes-2.5-Mistral-7B-GGUF/openhermes-2.5-mistral-7b.Q4_K_M.gguf\", n_gpu_layers=-1, n_ctx=2048, verbose=False)" + ] + }, + { + "cell_type": "code", + "execution_count": 22, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "[{'name': 'get_weather', 'arguments': {'zip_code': '10001'}}]\n", + "====================================================================================================\n", + "[{'name': 'calculate_mortgage_payment', 'arguments': {'loan_amount': 200000, 'interest_rate': 0.04, 'loan_term': 30}}]\n", + "====================================================================================================\n", + "Unfortunately, I do not have a built-in function to check currency exchange rates. However, you can use third-party APIs or websites like Google Finance or XE to get this information.\n", + "====================================================================================================\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 10001?\",\n", + " \"Determine the monthly mortgage payment for a loan amount of $200,000, an interest rate of 4%, and a loan term of 30 years.\",\n", + " \"What's the current exchange rate for USD to EUR?\"\n", + "]\n", + "functions = [get_weather, calculate_mortgage_payment, get_article_details]\n", + "\n", + "for prompt in prompts:\n", + " prompt = generate_hermes_prompt(prompt, functions)\n", + " completion = llama.create_completion(prompt, max_tokens=-1)[\"choices\"][0][\"text\"]\n", + " function_calls = extract_function_calls(completion)\n", + " if function_calls:\n", + " print(function_calls)\n", + " else:\n", + " print(completion.strip())\n", + " print(\"=\"*100)" + ] + }, + { + "cell_type": "code", + "execution_count": 23, + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "get_weather\n", + "{'zip_code': '05751'}\n", + "====================================================================================================\n", + "get_weather\n", + "{'zip_code': '05751'}\n", + "get_weather\n", + "{'zip_code': '07030'}\n", + "calculate_mortgage_payment\n", + "{'loan_amount': 250000, 'interest_rate': 4.18, 'loan_term': 30}\n", + "====================================================================================================\n", + "I don't have a function to get exchange rates, but I can provide some resources where you can find this information. You can check websites like Google Finance, XE.com, or Yahoo Finance for up-to-date currency exchange rates.\n", + "====================================================================================================\n" + ] + } + ], + "source": [ + "prompts = [\n", + " \"What's the weather in 05751?\",\n", + " \"I'm planning a trip to Killington, Vermont (05751) from Hoboken, NJ (07030). Can you get me weather for both locations and directions?\",\n", + " \"What's the current exchange rate for USD to EUR?\"\n", + "]\n", + "\n", + "for prompt in prompts:\n", + " completion = llama.create_completion(generate_hermes_prompt(prompt, functions), max_tokens=-1)[\"choices\"][0][\"text\"]\n", + " function_calls = extract_function_calls(completion)\n", + "\n", + " if function_calls:\n", + " for function in function_calls:\n", + " print(function[\"name\"])\n", + " print(function[\"arguments\"])\n", + " else:\n", + " print(completion.strip())\n", + "\n", + " print(\"=\"*100)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": ".venv", + "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": "3.11.5+" + } + }, + "nbformat": 4, + "nbformat_minor": 2 +} diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 3dcb4b58a..3efd95df5 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -50,6 +50,9 @@ _LlamaSamplingContext, # type: ignore ) from ._logger import set_verbose +from ._utils import ( + suppress_stdout_stderr +) class Llama: @@ -182,7 +185,8 @@ def __init__( self.numa = numa if not Llama.__backend_initialized: - llama_cpp.llama_backend_init(self.numa) + with suppress_stdout_stderr(disable=verbose): + llama_cpp.llama_backend_init(self.numa) Llama.__backend_initialized = True self.model_path = model_path @@ -1567,6 +1571,38 @@ def create_chat_completion( logit_bias=logit_bias, ) + def create_chat_completion_openai_v1( + self, + *args: Any, + **kwargs: Any, + ): + """Generate a chat completion with return type based on the the OpenAI v1 API. + + OpenAI python package is required to use this method. + + You can install it with `pip install openai`. + + Args: + *args: Positional arguments to pass to create_chat_completion. + **kwargs: Keyword arguments to pass to create_chat_completion. + + Returns: + Generated chat completion or a stream of chat completion chunks. + """ + try: + from openai.types.chat import ChatCompletion, ChatCompletionChunk + stream = kwargs.get("stream", False) # type: ignore + assert isinstance(stream, bool) + if stream: + return (ChatCompletionChunk(**chunk) for chunk in self.create_chat_completion(*args, **kwargs)) # type: ignore + else: + return ChatCompletion(**self.create_chat_completion(*args, **kwargs)) # type: ignore + except ImportError: + raise ImportError( + "To use create_chat_completion_openai_v1, you must install the openai package." + "You can install it with `pip install openai`." + ) + def __getstate__(self): return dict( model_path=self.model_path, diff --git a/llama_cpp/llama_chat_format.py b/llama_cpp/llama_chat_format.py index 2e4204121..af60d5f26 100644 --- a/llama_cpp/llama_chat_format.py +++ b/llama_cpp/llama_chat_format.py @@ -31,6 +31,7 @@ ### Chat Completion Handler ### + class LlamaChatCompletionHandler(Protocol): """Base Protocol for a llama chat completion handler. @@ -77,8 +78,7 @@ def __call__( ) -> Union[ llama_types.CreateChatCompletionResponse, Iterator[llama_types.CreateChatCompletionStreamResponse], - ]: - ... + ]: ... class LlamaChatCompletionHandlerNotFoundException(Exception): @@ -134,6 +134,7 @@ def decorator(f: LlamaChatCompletionHandler): ### Chat Formatter ### + @dataclasses.dataclass class ChatFormatterResponse: """Dataclass that stores completion parameters for a given chat format and @@ -157,8 +158,7 @@ def __call__( *, messages: List[llama_types.ChatCompletionRequestMessage], **kwargs: Any, - ) -> ChatFormatterResponse: - ... + ) -> ChatFormatterResponse: ... class Jinja2ChatFormatter(ChatFormatter): @@ -195,7 +195,7 @@ def raise_exception(message: str): eos_token=self.eos_token, bos_token=self.bos_token, raise_exception=raise_exception, - add_generation_prompt=self.add_generation_prompt + add_generation_prompt=self.add_generation_prompt, ) return ChatFormatterResponse(prompt=prompt, stop=[self.eos_token]) @@ -255,11 +255,13 @@ def _convert_text_completion_chunks_to_chat( "choices": [ { "index": 0, - "delta": { - "content": chunk["choices"][0]["text"], - } - if chunk["choices"][0]["finish_reason"] is None - else {}, + "delta": ( + { + "content": chunk["choices"][0]["text"], + } + if chunk["choices"][0]["finish_reason"] is None + else {} + ), "finish_reason": chunk["choices"][0]["finish_reason"], } ], @@ -338,10 +340,12 @@ def chat_completion_handler( # create grammar from json schema if "schema" in response_format: grammar = llama_grammar.LlamaGrammar.from_json_schema( - json.dumps(response_format["schema"]) + json.dumps(response_format["schema"]), verbose=llama.verbose ) except Exception as e: - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) completion_or_chunks = llama.create_completion( prompt=prompt, @@ -452,7 +456,9 @@ def hf_tokenizer_config_to_chat_completion_handler( tokenizer_config: Dict[str, Any], add_generation_prompt: bool = True, ) -> LlamaChatCompletionHandler: - chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config, add_generation_prompt=add_generation_prompt) + chat_formatter = hf_tokenizer_config_to_chat_formatter( + tokenizer_config, add_generation_prompt=add_generation_prompt + ) return chat_formatter_to_chat_completion_handler(chat_formatter) @@ -463,11 +469,12 @@ def guess_chat_format_from_gguf_metadata(metadata: Dict[str, str]) -> Optional[s if metadata["tokenizer.chat_template"] == CHATML_CHAT_TEMPLATE: return "chatml" - if metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE: + if metadata["tokenizer.chat_template"] == MISTRAL_INSTRUCT_CHAT_TEMPLATE: return "mistral-instruct" return None + ### Utility functions for formatting chat prompts ### # TODO: Replace these with jinja2 templates @@ -916,9 +923,17 @@ def format_mistral_instruct( stop = eos prompt = bos for message in messages: - if message["role"] == "user" and message["content"] is not None and isinstance(message["content"], str): + if ( + message["role"] == "user" + and message["content"] is not None + and isinstance(message["content"], str) + ): prompt += "[INST] " + message["content"] - elif message["role"] == "assistant" and message["content"] is not None and isinstance(message["content"], str): + elif ( + message["role"] == "assistant" + and message["content"] is not None + and isinstance(message["content"], str) + ): prompt += " [/INST]" + message["content"] + eos prompt += " [/INST]" return ChatFormatterResponse(prompt=prompt, stop=stop) @@ -958,6 +973,7 @@ def format_openchat( _prompt = _format_chatml(system_message, _messages, _sep) return ChatFormatterResponse(prompt=_prompt, stop=_sep) + # Chat format for Saiga models, see more details and available models: # https://huggingface.co/collections/IlyaGusev/saiga2-saigamistral-6505d4ccc3d1e53166b636cd @register_chat_format("saiga") @@ -979,8 +995,10 @@ def format_saiga( _prompt += "bot" return ChatFormatterResponse(prompt=_prompt.strip()) + # Tricky chat formats that require custom chat handlers + @register_chat_completion_handler("functionary") def functionary_chat_handler( llama: llama.Llama, @@ -1253,7 +1271,8 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): json.dumps(function_body) ) grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.json_schema_to_gbnf(json.dumps(function_body)) + llama_grammar.json_schema_to_gbnf(json.dumps(function_body)), + verbose=llama.verbose, ) print(grammar_text) except Exception as e: @@ -1264,11 +1283,14 @@ def message_to_str(msg: llama_types.ChatCompletionRequestMessage): print(e) with suppress_stdout_stderr(disable=llama.verbose): grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF + llama_grammar.JSON_GBNF, + verbose=llama.verbose, ) else: with suppress_stdout_stderr(disable=llama.verbose): - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) completion: llama_types.Completion = llama.create_completion( prompt=new_prompt, @@ -1365,11 +1387,13 @@ def functionary_v1_v2_chat_handler( **kwargs, # type: ignore ) -> Union[llama_types.ChatCompletion, Iterator[llama_types.ChatCompletionChunk]]: SYSTEM_MESSAGE = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary""" - + tokenizer = llama.tokenizer_ - assert hasattr(tokenizer, "hf_tokenizer"), "Please provide a valid hf_tokenizer_path from https://huggingface.co/meetkai when initializing the Llama class" + assert hasattr( + tokenizer, "hf_tokenizer" + ), "Please provide a valid hf_tokenizer_path from https://huggingface.co/meetkai when initializing the Llama class" from transformers import AutoTokenizer - + if "<|START_OF_FUNCTION_CALL|>" in tokenizer.hf_tokenizer.additional_special_tokens: version = "v1" END_SYSTEM_TOKEN = "<|END_OF_SYSTEM|>" @@ -1513,13 +1537,16 @@ def prepare_messages_for_inference( "name" ] = f"functions.{message['function_call']['name']}" all_messages.append(message) - + if version == "v1": suffix = "assistant:\n" else: suffix = "<|from|>assistant\n<|recipient|>" - - return tokenizer.hf_tokenizer.apply_chat_template(all_messages, tokenize=False) + suffix + + return ( + tokenizer.hf_tokenizer.apply_chat_template(all_messages, tokenize=False) + + suffix + ) if tools is not None: functions = [tool["function"] for tool in tools if tool["type"] == "function"] @@ -1529,8 +1556,10 @@ def prepare_messages_for_inference( tool_choice if isinstance(tool_choice, str) else tool_choice["function"] ) - prompt = prepare_messages_for_inference(messages, tokenizer, version, functions, tools) - + prompt = prepare_messages_for_inference( + messages, tokenizer, version, functions, tools + ) + # If no tools/functions are provided if function_call is None and (functions is None or len(functions) == 0): if version == "v1": @@ -1538,7 +1567,7 @@ def prepare_messages_for_inference( else: stop = STOP_TOKEN prompt += "all\n<|content|>" - + completion_or_completion_chunks = llama.create_completion( prompt=prompt, temperature=temperature, @@ -1561,9 +1590,9 @@ def prepare_messages_for_inference( grammar=grammar, ) return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream) # type: ignore - + assert stream is False # TODO: support stream mode - + def get_grammar(function_call): function_body = None for function in functions or []: @@ -1574,7 +1603,7 @@ def get_grammar(function_call): if tool["type"] == "function" and tool["function"]["name"] == function_call: function_body = tool["function"]["parameters"] break - + try: with suppress_stdout_stderr(disable=llama.verbose): grammar_text = llama_grammar.json_schema_to_gbnf( @@ -1592,11 +1621,11 @@ def get_grammar(function_call): print(e) with suppress_stdout_stderr(disable=llama.verbose): grammar = llama_grammar.LlamaGrammar.from_string( - llama_grammar.JSON_GBNF + llama_grammar.JSON_GBNF, verbose=llama.verbose ) - + return grammar - + def create_completion(stop): completion: llama_types.Completion = llama.create_completion( prompt=prompt, @@ -1619,11 +1648,11 @@ def create_completion(stop): logits_processor=logits_processor, grammar=grammar, ) - + return completion - + function_calls, function_bodies = [], [] - + if version == "v1": # If no or "auto" tool_choice/function_call if function_call is None or ( @@ -1632,7 +1661,9 @@ def create_completion(stop): stops = ["\n", END_ASSISTANT_TOKEN] # If tool_choice/function_call is "none" elif isinstance(function_call, str) and function_call == "none": - prompt = prepare_messages_for_inference(messages, tokenizer, version, [], []) + prompt = prepare_messages_for_inference( + messages, tokenizer, version, [], [] + ) stops = END_ASSISTANT_TOKEN # If tool_choice/function_call is provided elif isinstance(function_call, dict): @@ -1647,14 +1678,27 @@ def create_completion(stop): completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] - + # If the generation does not involve a function call - if START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN not in completion_text: + if ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN not in completion_text + ): return _convert_completion_to_chat(completion, stream=stream) # type: ignore # If the generation involves a function call in completion, generate the parameters - elif START_FUNCTION_CALL_TOKEN not in prompt and START_FUNCTION_CALL_TOKEN in completion_text: - prompt += completion_text.replace(f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN) + "\n" - function_calls.append(completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip()) + elif ( + START_FUNCTION_CALL_TOKEN not in prompt + and START_FUNCTION_CALL_TOKEN in completion_text + ): + prompt += ( + completion_text.replace( + f"{START_FUNCTION_CALL_TOKEN} ", START_FUNCTION_CALL_TOKEN + ) + + "\n" + ) + function_calls.append( + completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip() + ) grammar = get_grammar(function_calls[-1]) completion = create_completion(stop=END_FUNCTION_CALL_TOKEN) function_bodies.append(completion["choices"][0]["text"].strip()) @@ -1672,7 +1716,10 @@ def create_completion(stop): stops = CONTENT_TOKEN # If tool_choice/function_call is "none" elif isinstance(function_call, str) and function_call == "none": - prompt = prepare_messages_for_inference(messages, tokenizer, version, [], []) + "all\n<|content|>" + prompt = ( + prepare_messages_for_inference(messages, tokenizer, version, [], []) + + "all\n<|content|>" + ) stops = STOP_TOKEN # If tool_choice/function_call is provided elif isinstance(function_call, dict): @@ -1684,15 +1731,17 @@ def create_completion(stop): else: prompt = prompt stops = STOP_TOKEN - + completion = create_completion(stop=stops) completion_text = completion["choices"][0]["text"] - + # If the generation does not involve a function call - if prompt.endswith("all\n<|content|>") and not completion_text.startswith("all"): + if prompt.endswith("all\n<|content|>") and not completion_text.startswith( + "all" + ): return _convert_completion_to_chat(completion, stream=stream) # type: ignore # Generate model response if the model decides not to call any function - elif (prompt.endswith(RECIPIENT_TOKEN) and completion_text.startswith("all")): + elif prompt.endswith(RECIPIENT_TOKEN) and completion_text.startswith("all"): prompt += completion_text + CONTENT_TOKEN completion = create_completion(stop=STOP_TOKEN) return _convert_completion_to_chat(completion, stream=stream) # type: ignore @@ -1704,7 +1753,7 @@ def create_completion(stop): function_bodies.append(completion["choices"][0]["text"].strip()) prompt += f"{function_calls[-1]}\n{CONTENT_TOKEN}{function_bodies[-1]}" grammar = None - + # Try to generate the beginning of next turn # If empty completion, break from loop next_turn_completion_text = create_completion( @@ -1718,17 +1767,21 @@ def create_completion(stop): else: function_bodies.append(completion_text.strip()) break - + assert "usage" in completion assert len(function_calls) > 0 assert len(function_calls) == len(function_bodies) - + tool_calls = [] for function_call, function_body in zip(function_calls, function_bodies): tool_calls.append( { - "id": "call_" + "".join( - [random.choice(string.ascii_letters + string.digits) for _ in range(24)] + "id": "call_" + + "".join( + [ + random.choice(string.ascii_letters + string.digits) + for _ in range(24) + ] ), "type": "function", "function": { @@ -1924,7 +1977,9 @@ def __call__( json.dumps(response_format["schema"]) ) except Exception as e: - grammar = llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF + ) return _convert_completion_to_chat( llama.create_completion( @@ -1950,3 +2005,601 @@ def __call__( ), stream=stream, ) + + +@register_chat_completion_handler("chatml-function-calling") +def chatml_function_calling( + llama: llama.Llama, + messages: List[llama_types.ChatCompletionRequestMessage], + functions: Optional[List[llama_types.ChatCompletionFunction]] = None, + function_call: Optional[llama_types.ChatCompletionRequestFunctionCall] = None, + tools: Optional[List[llama_types.ChatCompletionTool]] = None, + tool_choice: Optional[llama_types.ChatCompletionToolChoiceOption] = None, + temperature: float = 0.2, + top_p: float = 0.95, + top_k: int = 40, + min_p: float = 0.05, + typical_p: float = 1.0, + stream: bool = False, + stop: Optional[Union[str, List[str]]] = [], + response_format: Optional[llama_types.ChatCompletionRequestResponseFormat] = None, + max_tokens: Optional[int] = None, + presence_penalty: float = 0.0, + frequency_penalty: float = 0.0, + repeat_penalty: float = 1.1, + tfs_z: float = 1.0, + mirostat_mode: int = 0, + mirostat_tau: float = 5.0, + mirostat_eta: float = 0.1, + model: Optional[str] = None, + logits_processor: Optional[llama.LogitsProcessorList] = None, + grammar: Optional[llama.LlamaGrammar] = None, + **kwargs, # type: ignore +) -> Union[ + llama_types.CreateChatCompletionResponse, + Iterator[llama_types.CreateChatCompletionStreamResponse], +]: + function_calling_template = ( + "{% for message in messages %}" + "<|im_start|>{{ message.role }}\n" + # System message + "{% if message.role == 'system' %}" + "{{ message.content }}" + "{% if tool_calls %}" + "\n\nYou have access to the following functions:\n" + "{% for tool in tools %}" + "\nfunctions.{{ tool.function.name }}:\n" + "{{ tool.function.parameters | tojson }}" + "\n{% endfor %}" + "\n\nYou can respond to users messages with either a single message or one or more function calls." + "\n\nTo respond with a message begin the message with 'message:', use the following format:" + "\n\nmessage:" + "\n" + "\n\nTo respond with one or more function calls begin the message with 'functions.:', use the following format:" + "\n\nfunctions.:" + '\n{ "arg1": "value1", "arg2": "value2" }' + "\nfunctions.:" + '\n{ "arg1": "value1", "arg2": "value2" }' + "{% endif %}" + "\n<|im_end|>\n" + "{% endif %}" + # User message + "{% if message.role == 'user' %}" + "{{ message.content }}" + "\n<|im_end|>\n" + "{% endif %}" + # Assistant message + "{% if message.role == 'assistant' %}" + ## Reglar message + "{% if message.content and message.content | length > 0 %}" + "message:\n" + "{{ message.content }}" + "\n<|im_end|>\n" + "{% endif %}" + ## Function calls + "{% if message.tool_calls %}" + "{% for tool_call in message.tool_calls %}" + "functions.{{ tool_call.function.name }}:\n" + "{{ tool_call.function.arguments }}" + "{% endfor %}" + "\n<|im_end|>\n" + "{% endif %}" + "{% endif %}" + "{% endfor %}" + ) + template_renderer = jinja2.Environment( + loader=jinja2.BaseLoader(), + autoescape=jinja2.select_autoescape(["html", "xml"]), + undefined=jinja2.StrictUndefined, + ).from_string(function_calling_template) + + # Convert legacy functions to tools + if functions is not None: + tools = [ + { + "type": "function", + "function": function, + } + for function in functions + ] + + # Convert legacy function_call to tool_choice + if function_call is not None: + if isinstance(function_call, str) and ( + function_call == "none" or function_call == "auto" + ): + tool_choice = function_call + if isinstance(function_call, dict) and "name" in function_call: + tool_choice = { + "type": "function", + "function": { + "name": function_call["name"], + }, + } + + # Case 1: No tool choice by user + if ( + tool_choice is None + or (isinstance(tool_choice, str) and tool_choice == "none") + or tools is None + or len(tools) == 0 + ): + prompt = template_renderer.render( + messages=messages, + tools=[], + tool_calls=None, + ) + if response_format is not None and response_format["type"] == "json_object": + try: + grammar = ( + llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(response_format["schema"]) + ) + if "schema" in response_format + else None + ) + except Exception as e: + if llama.verbose: + print( + "Failed to parse response format as JSON schema, falling back to default grammar" + ) + print(e) + grammar = ( + llama_grammar.LlamaGrammar.from_string(llama_grammar.JSON_GBNF) + if grammar is None + else grammar + ) + return _convert_completion_to_chat( + llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ), + stream=stream, + ) + + def _convert_completion_to_chat_function( + tool_name: str, + completion_or_chunks: Union[ + llama_types.CreateCompletionResponse, + Iterator[llama_types.CreateCompletionStreamResponse], + ], + stream: bool, + ): + if not stream: + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + assert "usage" in completion + tool_id = "call_" + "_0_" + tool_name + "_" + completion["id"] + # TODO: Fix for legacy function calls + chat_completion: llama_types.CreateChatCompletionResponse = { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "index": 0, + "message": { + "role": "assistant", + "content": None, + "function_call": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + "tool_calls": [ + { + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + ], + }, + "finish_reason": "tool_calls", + } + ], + "usage": completion["usage"], + } + return chat_completion + else: + chunks: Iterator[llama_types.CreateCompletionStreamResponse] = completion_or_chunks # type: ignore + + def _stream_response_to_function_stream( + chunks: Iterator[llama_types.CreateCompletionStreamResponse], + ) -> Iterator[llama_types.CreateChatCompletionStreamResponse]: + # blank first message + first = True + id_ = None + created = None + model = None + tool_id = None + for chunk in chunks: + if first: + id_ = "chat" + chunk["id"] + created = chunk["created"] + model = chunk["model"] + tool_id = "call_" + "_0_" + tool_name + "_" + chunk["id"] + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": "assistant", + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": "", + }, + } + ], + }, + } + ], + } + first = False + continue + assert tool_id is not None + yield { + "id": "chat" + chunk["id"], + "object": "chat.completion.chunk", + "created": chunk["created"], + "model": chunk["model"], + "choices": [ + { + "index": 0, + "finish_reason": None, + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": { + "name": tool_name, + "arguments": chunk["choices"][0]["text"], + }, + "tool_calls": [ + { + "index": 0, + "id": tool_id, + "type": "function", + "function": { + "name": tool_name, + "arguments": chunk["choices"][0][ + "text" + ], + }, + } + ], + }, + } + ], + } + + if id_ is not None and created is not None and model is not None: + yield { + "id": id_, + "object": "chat.completion.chunk", + "created": created, + "model": model, + "choices": [ + { + "index": 0, + "finish_reason": "tool_calls", + "logprobs": None, + "delta": { + "role": None, + "content": None, + "function_call": None, + "tool_calls": None, + }, + } + ], + } + + return _stream_response_to_function_stream(chunks) + + # Case 2: Tool choice by user + if isinstance(tool_choice, dict): + tool_name = tool_choice["function"]["name"] + tool = next( + (tool for tool in tools if tool["function"]["name"] == tool_name), None + ) + if tool is None: + raise ValueError(f"Tool with name '{tool_name}' not found in tools") + prompt = template_renderer.render( + messages=messages, + tools=tools, + tool_calls=True, + ) + prompt += f"functions.{tool_name}:\n" + try: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(tool["function"]["parameters"]), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=stop, + max_tokens=max_tokens, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + return _convert_completion_to_chat_function( + tool_name, completion_or_chunks, stream + ) + + # Case 3: Automatic tool choice + assert isinstance(tool_choice, str) and tool_choice == "auto" + function_names = " | ".join( + [f'''"functions.{tool['function']['name']}:"''' for tool in tools] + ) + initial_gbnf_tool_grammar = ( + """root ::= functions | "message:"\n""" + f"""functions ::= {function_names}\n""" + ) + follow_up_gbnf_tool_grammar = ( + """root ::= functions | "<|im_end|>"\n""" + f"""functions ::= {function_names}\n""" + ) + prompt = template_renderer.render( + messages=messages, + tools=tools, + tool_calls=True, + ) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=0, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=[":"], + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + initial_gbnf_tool_grammar, verbose=llama.verbose + ), + ) + completion: llama_types.CreateCompletionResponse = completion_or_chunks # type: ignore + text = completion["choices"][0]["text"] + if "message" in text: + return _convert_completion_to_chat( + llama.create_completion( + prompt=prompt + "message:\n", + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=stream, + stop=["<|im_end|>"], + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + follow_up_gbnf_tool_grammar, verbose=llama.verbose + ), + ), + stream=stream, + ) + + # One or more function calls + tool_name = text[len("functions.") :] + tool = next((tool for tool in tools if tool["function"]["name"] == tool_name), None) + if not stream: + completions = [] + completions_tool_name = [] + while tool is not None: + prompt += f"functions.{tool_name}:\n" + try: + grammar = llama_grammar.LlamaGrammar.from_json_schema( + json.dumps(tool["function"]["parameters"]), verbose=llama.verbose + ) + except Exception as e: + grammar = llama_grammar.LlamaGrammar.from_string( + llama_grammar.JSON_GBNF, verbose=llama.verbose + ) + if llama.verbose: + print( + "Failed to parse function body as JSON schema, falling back to default grammar" + ) + print(e) + completion_or_chunks = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=stop, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=grammar, + ) + completions.append(completion_or_chunks) + completions_tool_name.append(tool_name) + prompt += completion_or_chunks["choices"][0]["text"] + prompt += "\n" + + response = llama.create_completion( + prompt=prompt, + temperature=temperature, + top_p=top_p, + top_k=top_k, + min_p=min_p, + typical_p=typical_p, + stream=False, + stop=stop, + max_tokens=None, + presence_penalty=presence_penalty, + frequency_penalty=frequency_penalty, + repeat_penalty=repeat_penalty, + tfs_z=tfs_z, + mirostat_mode=mirostat_mode, + mirostat_tau=mirostat_tau, + mirostat_eta=mirostat_eta, + model=model, + logits_processor=logits_processor, + grammar=llama_grammar.LlamaGrammar.from_string( + follow_up_gbnf_tool_grammar, verbose=llama.verbose + ), + ) + + tool_name = response["choices"][0]["text"][len("functions.") :] + tool = next( + (tool for tool in tools if tool["function"]["name"] == tool_name), None + ) + + # Merge completions + function_call = { + "function_call": { + "name": tool_name, + "arguments": completions[0]["choices"][0]["text"], + } + } if len(completions) == 1 else {} + return { + "id": "chat" + completion["id"], + "object": "chat.completion", + "created": completion["created"], + "model": completion["model"], + "choices": [ + { + "finish_reason": "tool_calls", + "index": 0, + "message": { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_" + + f"_{i}_" + + tool_name + + "_" + + completion["id"], + "type": "function", + "function": { + "name": tool_name, + "arguments": completion["choices"][0]["text"], + }, + } + for i, (tool_name, completion) in enumerate( + zip(completions_tool_name, completions) + ) + ], + **function_call + }, + } + ], + "usage": { + "completion_tokens": sum( + completion["usage"]["completion_tokens"] + for completion in completions + ), + "prompt_tokens": sum( + completion["usage"]["prompt_tokens"] for completion in completions + ), + "total_tokens": sum( + completion["usage"]["total_tokens"] for completion in completions + ), + }, + } + + raise ValueError("Automatic streaming tool choice is not supported") diff --git a/llama_cpp/llama_types.py b/llama_cpp/llama_types.py index c3deba87b..1b1befebe 100644 --- a/llama_cpp/llama_types.py +++ b/llama_cpp/llama_types.py @@ -97,7 +97,7 @@ class CreateChatCompletionResponse(TypedDict): class ChatCompletionMessageToolCallChunkFunction(TypedDict): - name: str + name: Optional[str] arguments: str @@ -118,12 +118,12 @@ class ChatCompletionStreamResponseDeltaFunctionCall(TypedDict): class ChatCompletionStreamResponseDelta(TypedDict): - content: NotRequired[str] + content: NotRequired[Optional[str]] function_call: NotRequired[ - ChatCompletionStreamResponseDeltaFunctionCall + Optional[ChatCompletionStreamResponseDeltaFunctionCall] ] # DEPRECATED - tool_calls: NotRequired[List[ChatCompletionMessageToolCallChunk]] - role: NotRequired[Literal["system", "user", "assistant", "tool"]] + tool_calls: NotRequired[Optional[List[ChatCompletionMessageToolCallChunk]]] + role: NotRequired[Optional[Literal["system", "user", "assistant", "tool"]]] class ChatCompletionStreamResponseChoice(TypedDict): @@ -132,6 +132,7 @@ class ChatCompletionStreamResponseChoice(TypedDict): ChatCompletionStreamResponseDelta, ChatCompletionStreamResponseDeltaEmpty ] finish_reason: Optional[Literal["stop", "length", "tool_calls", "function_call"]] + logprobs: NotRequired[Optional[CompletionLogprobs]] class CreateChatCompletionStreamResponse(TypedDict):