File size: 11,189 Bytes
f35cc9b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 |
{
"nbformat": 4,
"nbformat_minor": 0,
"metadata": {
"colab": {
"provenance": []
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
},
"language_info": {
"name": "python"
}
},
"cells": [
{
"cell_type": "markdown",
"source": [
"#Install dependencies"
],
"metadata": {
"id": "39AMoCOa1ckc"
}
},
{
"cell_type": "code",
"source": [
"!pip install ai-edge-litert-nightly"
],
"metadata": {
"id": "43tAeO0AZ7zp"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"source": [
"from ai_edge_litert import interpreter as interpreter_lib\n",
"from transformers import AutoTokenizer\n",
"import numpy as np\n",
"from collections.abc import Sequence\n",
"import sys"
],
"metadata": {
"id": "i6PMkMVBPr1p"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"# Download model files"
],
"metadata": {
"id": "K5okZCTgYpUd"
}
},
{
"cell_type": "code",
"source": [
"from huggingface_hub import hf_hub_download\n",
"\n",
"model_path = hf_hub_download(repo_id=\"litert-community/DeepSeek-R1-Distill-Qwen-1.5B\", filename=\"deepseek_q8_seq128_ekv1280.tflite\")"
],
"metadata": {
"id": "3t47HAG2tvc3"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"# Create LiteRT interpreter and tokenizer"
],
"metadata": {
"id": "n5Xa4s6XhWqk"
}
},
{
"cell_type": "code",
"source": [
"interpreter = interpreter_lib.InterpreterWithCustomOps(\n",
" custom_op_registerers=[\"pywrap_genai_ops.GenAIOpsRegisterer\"],\n",
" model_path=model_path,\n",
" num_threads=2,\n",
" experimental_default_delegate_latest_features=True)\n",
"tokenizer = AutoTokenizer.from_pretrained(\"deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B\")"
],
"metadata": {
"id": "Rvdn3EIZhaQn"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"# Create pipeline with LiteRT models"
],
"metadata": {
"id": "AM6rDABTXt2F"
}
},
{
"cell_type": "code",
"source": [
"\n",
"class LiteRTLlmPipeline:\n",
"\n",
" def __init__(self, interpreter, tokenizer):\n",
" \"\"\"Initializes the pipeline.\"\"\"\n",
" self._interpreter = interpreter\n",
" self._tokenizer = tokenizer\n",
"\n",
" self._prefill_runner = None\n",
" self._decode_runner = self._interpreter.get_signature_runner(\"decode\")\n",
"\n",
"\n",
" def _init_prefill_runner(self, num_input_tokens: int):\n",
" \"\"\"Initializes all the variables related to the prefill runner.\n",
"\n",
" This method initializes the following variables:\n",
" - self._prefill_runner: The prefill runner based on the input size.\n",
" - self._max_seq_len: The maximum sequence length supported by the model.\n",
" - self._max_kv_cache_seq_len: The maximum sequence length supported by the\n",
" KV cache.\n",
" - self._num_layers: The number of layers in the model.\n",
"\n",
" Args:\n",
" num_input_tokens: The number of input tokens.\n",
" \"\"\"\n",
"\n",
" self._prefill_runner = self._get_prefill_runner(num_input_tokens)\n",
" # input_token_shape has shape (batch, max_seq_len)\n",
" input_token_shape = self._prefill_runner.get_input_details()[\"tokens\"][\n",
" \"shape\"\n",
" ]\n",
" if len(input_token_shape) == 1:\n",
" self._max_seq_len = input_token_shape[0]\n",
" else:\n",
" self._max_seq_len = input_token_shape[1]\n",
"\n",
" # kv cache input has shape [batch=1, seq_len, num_heads, dim].\n",
" kv_cache_shape = self._prefill_runner.get_input_details()[\"kv_cache_k_0\"][\n",
" \"shape\"\n",
" ]\n",
" self._max_kv_cache_seq_len = kv_cache_shape[1]\n",
"\n",
" # The two arguments excluded are `tokens` and `input_pos`. Dividing by 2\n",
" # because each layer has key and value caches.\n",
" self._num_layers = (\n",
" len(self._prefill_runner.get_input_details().keys()) - 2\n",
" ) // 2\n",
"\n",
"\n",
" def _init_kv_cache(self) -> dict[str, np.ndarray]:\n",
" if self._prefill_runner is None:\n",
" raise ValueError(\"Prefill runner is not initialized.\")\n",
" kv_cache = {}\n",
" for i in range(self._num_layers):\n",
" kv_cache[f\"kv_cache_k_{i}\"] = np.zeros(\n",
" self._prefill_runner.get_input_details()[f\"kv_cache_k_{i}\"][\"shape\"],\n",
" dtype=np.float32,\n",
" )\n",
" kv_cache[f\"kv_cache_v_{i}\"] = np.zeros(\n",
" self._prefill_runner.get_input_details()[f\"kv_cache_v_{i}\"][\"shape\"],\n",
" dtype=np.float32,\n",
" )\n",
" return kv_cache\n",
"\n",
" def _get_prefill_runner(self, num_input_tokens: int) :\n",
" \"\"\"Gets the prefill runner with the best suitable input size.\n",
"\n",
" Args:\n",
" num_input_tokens: The number of input tokens.\n",
"\n",
" Returns:\n",
" The prefill runner with the smallest input size.\n",
" \"\"\"\n",
" best_signature = None\n",
" delta = sys.maxsize\n",
" max_prefill_len = -1\n",
" for key in self._interpreter.get_signature_list().keys():\n",
" if \"prefill\" not in key:\n",
" continue\n",
" input_pos = self._interpreter.get_signature_runner(key).get_input_details()[\n",
" \"input_pos\"\n",
" ]\n",
" # input_pos[\"shape\"] has shape (max_seq_len, )\n",
" seq_size = input_pos[\"shape\"][0]\n",
" max_prefill_len = max(max_prefill_len, seq_size)\n",
" if num_input_tokens <= seq_size and seq_size - num_input_tokens < delta:\n",
" delta = seq_size - num_input_tokens\n",
" best_signature = key\n",
" if best_signature is None:\n",
" raise ValueError(\n",
" \"The largest prefill length supported is %d, but we have %d number of input tokens\"\n",
" %(max_prefill_len, num_input_tokens)\n",
" )\n",
" return self._interpreter.get_signature_runner(best_signature)\n",
"\n",
" def _greedy_sampler(self, logits: np.ndarray) -> int:\n",
" return int(np.argmax(logits))\n",
"\n",
" def generate(self, prompt: str, max_decode_steps: int | None = None) -> str:\n",
" messages=[{ 'role': 'user', 'content': prompt}]\n",
" token_ids = self._tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True)\n",
" # Initialize the prefill runner with the suitable input size.\n",
" self._init_prefill_runner(len(token_ids))\n",
"\n",
" actual_max_decode_steps = self._max_kv_cache_seq_len - len(token_ids)\n",
" if max_decode_steps is not None:\n",
" actual_max_decode_steps = min(actual_max_decode_steps, max_decode_steps)\n",
"\n",
" input_token_ids = [0] * self._max_seq_len\n",
" input_token_ids[:len(token_ids)] = token_ids\n",
" model_inputs = self._init_kv_cache()\n",
" model_inputs.update({\n",
" \"tokens\": np.asarray([input_token_ids], dtype=np.int32),\n",
" \"input_pos\": np.arange(self._max_seq_len, dtype=np.int32),\n",
" })\n",
" decode_text = []\n",
" decode_step = 0\n",
" print('Running prefill')\n",
" for step in range(actual_max_decode_steps+1):\n",
" signature_runner = self._prefill_runner if step == 0 else self._decode_runner\n",
" model_outputs = signature_runner(**model_inputs)\n",
" # At prefill stage, output logits has shape (batch=1, seq_size, vocab_size)\n",
" # At decode stage, output logits has shape (batch=1, 1, vocab_size).\n",
" selected_logit = len(token_ids)-1 if step == 0 else 0\n",
" logits = model_outputs.pop(\"logits\")[0][selected_logit]\n",
"\n",
" if step == 0:\n",
" print('Running decode')\n",
"\n",
" # Decode text output.\n",
" next_token = self._greedy_sampler(logits)\n",
" if next_token == self._tokenizer.eos_token_id:\n",
" break\n",
" decode_text.append(self._tokenizer.decode(next_token, skip_special_tokens=False))\n",
" print(decode_text[-1], end='', flush=True)\n",
" # The rest of the outputs is the updated kv cache.\n",
" model_inputs = model_outputs\n",
" model_inputs.update({\n",
" \"tokens\": np.array([[next_token]], dtype=np.int32),\n",
" \"input_pos\": np.array([decode_step + len(token_ids)], dtype=np.int32),})\n",
" decode_step += 1\n",
"\n",
"\n",
"\n",
" print() # print a new line at the end.\n",
" return ''.join(decode_text)\n"
],
"metadata": {
"id": "UBSGrHrM4ANm"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "markdown",
"source": [
"# Generate text from model"
],
"metadata": {
"id": "dASKx_JtYXwe"
}
},
{
"cell_type": "code",
"source": [
"# Disclaimer: Model performance demonstrated with the Python API in this notebook is not representative of performance on a local device.\n",
"pipeline = LiteRTLlmPipeline(interpreter, tokenizer)"
],
"metadata": {
"id": "AZhlDQWg61AL"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"source": [
"prompt = \"what is 8 mod 5\"\n",
"output = pipeline.generate(prompt, max_decode_steps = None)"
],
"metadata": {
"id": "wT9BIiATkjzL"
},
"execution_count": null,
"outputs": []
},
{
"cell_type": "code",
"source": [],
"metadata": {
"id": "GNzDBxDFEuAJ"
},
"execution_count": null,
"outputs": []
}
]
} |