Upload LMF_ONNX_test.ipynb

LMF_ONNX_test.ipynb

@@ -0,0 +1,515 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "3c7864eb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "import onnxruntime as ort\n",
+    "from huggingface_hub import hf_hub_download\n",
+    "from transformers import AutoTokenizer\n",
+    "\n",
+    "# Download model (Q4 recommended)\n",
+    "model_id = \"LiquidAI/LFM2.5-1.2B-JP-ONNX\"\n",
+    "model_path = hf_hub_download(model_id, \"onnx/model_q4.onnx\")\n",
+    "\n",
+    "# Download all data files (handles multiple splits for large models)\n",
+    "from huggingface_hub import list_repo_files\n",
+    "for f in list_repo_files(model_id):\n",
+    "    if f.startswith(\"onnx/model_q4.onnx_data\"):\n",
+    "        hf_hub_download(model_id, f)\n",
+    "\n",
+    "# Load model and tokenizer\n",
+    "session = ort.InferenceSession(model_path)\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2275f454",
+   "metadata": {},
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "5788adf8",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "日本の総理大臣は、総理大臣（総理大臣）または首相（首相）とも呼ばれる。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣の長である。総理大臣は、内閣\n"
+     ]
+    }
+   ],
+   "source": [
+    "\n",
+    "# Prepare text completion input (Japanese)\n",
+    "prompt = \"日本の総理大臣は\"\n",
+    "input_ids = np.array([tokenizer.encode(prompt, add_special_tokens=True)], dtype=np.int64)\n",
+    "\n",
+    "# Initialize KV cache\n",
+    "ONNX_DTYPE = {\"tensor(float)\": np.float32, \"tensor(float16)\": np.float16, \"tensor(int64)\": np.int64}\n",
+    "cache = {}\n",
+    "for inp in session.get_inputs():\n",
+    "    if inp.name in {\"input_ids\", \"attention_mask\", \"position_ids\"}:\n",
+    "        continue\n",
+    "    shape = [d if isinstance(d, int) else 1 for d in inp.shape]\n",
+    "    for i, d in enumerate(inp.shape):\n",
+    "        if isinstance(d, str) and \"sequence\" in d.lower():\n",
+    "            shape[i] = 0\n",
+    "    cache[inp.name] = np.zeros(shape, dtype=ONNX_DTYPE.get(inp.type, np.float32))\n",
+    "\n",
+    "# Check if model uses position_ids\n",
+    "input_names = {inp.name for inp in session.get_inputs()}\n",
+    "use_position_ids = \"position_ids\" in input_names\n",
+    "\n",
+    "# Generate tokens\n",
+    "seq_len = input_ids.shape[1]\n",
+    "generated_tokens = []\n",
+    "\n",
+    "for step in range(300):  # max tokens\n",
+    "    if step == 0:\n",
+    "        ids = input_ids\n",
+    "        pos = np.arange(seq_len, dtype=np.int64).reshape(1, -1)\n",
+    "    else:\n",
+    "        ids = np.array([[generated_tokens[-1]]], dtype=np.int64)\n",
+    "        pos = np.array([[seq_len + len(generated_tokens) - 1]], dtype=np.int64)\n",
+    "\n",
+    "    attn_mask = np.ones((1, seq_len + len(generated_tokens)), dtype=np.int64)\n",
+    "    feed = {\"input_ids\": ids, \"attention_mask\": attn_mask, **cache}\n",
+    "    if use_position_ids:\n",
+    "        feed[\"position_ids\"] = pos\n",
+    "\n",
+    "    outputs = session.run(None, feed)\n",
+    "    next_token = int(np.argmax(outputs[0][0, -1]))\n",
+    "    generated_tokens.append(next_token)\n",
+    "\n",
+    "    # Update cache\n",
+    "    for i, out in enumerate(session.get_outputs()[1:], 1):\n",
+    "        name = out.name.replace(\"present_conv\", \"past_conv\").replace(\"present.\", \"past_key_values.\")\n",
+    "        if name in cache:\n",
+    "            cache[name] = outputs[i]\n",
+    "\n",
+    "    if next_token == tokenizer.eos_token_id:\n",
+    "        break\n",
+    "\n",
+    "\n",
+    "print(prompt + tokenizer.decode(generated_tokens, skip_special_tokens=True))\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "85caf150",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "c:\\Users\\showe\\Desktop\\仕事関連\\venv\\lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
+      "  from .autonotebook import tqdm as notebook_tqdm\n"
+     ]
+    }
+   ],
+   "source": [
+    "import numpy as np\n",
+    "import onnxruntime as ort\n",
+    "from huggingface_hub import hf_hub_download\n",
+    "from transformers import AutoProcessor\n",
+    "from PIL import Image\n",
+    "\n",
+    "# Download model files (fp16 encoder + q4 decoder recommended)\n",
+    "model_id = \"LiquidAI/LFM2.5-VL-1.6B-ONNX\"\n",
+    "embed_tokens_path = hf_hub_download(model_id, \"onnx/embed_tokens_fp16.onnx\")\n",
+    "embed_images_path = hf_hub_download(model_id, \"onnx/embed_images_fp16.onnx\")\n",
+    "decoder_path = hf_hub_download(model_id, \"onnx/decoder_q4.onnx\")\n",
+    "\n",
+    "# Download all data files (handles multiple splits for large models)\n",
+    "from huggingface_hub import list_repo_files\n",
+    "for f in list_repo_files(model_id):\n",
+    "    if any(f.startswith(f\"onnx/{name}\") for name in [\n",
+    "        \"embed_tokens_fp16.onnx_data\",\n",
+    "        \"embed_images_fp16.onnx_data\",\n",
+    "        \"decoder_q4.onnx_data\"\n",
+    "    ]):\n",
+    "        hf_hub_download(model_id, f)\n",
+    "\n",
+    "# Load ONNX sessions\n",
+    "embed_tokens = ort.InferenceSession(embed_tokens_path)\n",
+    "embed_images = ort.InferenceSession(embed_images_path)\n",
+    "decoder = ort.InferenceSession(decoder_path)\n",
+    "\n",
+    "# Load processor\n",
+    "processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "73769fbf",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "この\n",
+      "画像\n",
+      "は\n",
+      "、\n",
+      "雪\n",
+      "に\n",
+      "覆\n",
+      "われた\n",
+      "�\n",
+      "�\n",
+      "面\n",
+      "で\n",
+      "、\n",
+      "晴\n",
+      "れた\n",
+      "日\n",
+      "を\n",
+      "楽し\n",
+      "んでいる\n",
+      "人\n",
+      "物を\n",
+      "描\n",
+      "いて\n",
+      "います\n",
+      "。\n",
+      "太\n",
+      "陽\n",
+      "が\n",
+      "空\n",
+      "を\n",
+      "�\n",
+      "�\n",
+      "か\n",
+      "せ\n",
+      "、\n",
+      "雪\n",
+      "に\n",
+      "長\n",
+      "い\n",
+      "影\n",
+      "を\n",
+      "落\n",
+      "として\n",
+      "います\n",
+      "。\n",
+      "背景\n",
+      "には\n",
+      "、\n",
+      "木\n",
+      "々\n",
+      "や\n",
+      "山\n",
+      "々\n",
+      "が\n",
+      "広\n",
+      "が\n",
+      "り\n",
+      "、\n",
+      "冬\n",
+      "の\n",
+      "風\n",
+      "景\n",
+      "を\n",
+      "完成\n",
+      "させて\n",
+      "います\n",
+      "。\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Prepare input\n",
+    "path = r\"C:\\Users\\showe\\Desktop\\tiikawa.png\"\n",
+    "path = r\"C:\\Users\\showe\\Pictures\\1487423322121.jpg\"\n",
+    "image = Image.open(path)\n",
+    "messages = [{\"role\": \"user\", \"content\": [\n",
+    "    {\"type\": \"image\"},\n",
+    "    {\"type\": \"text\", \"text\": \"この画像について説明してください。\"}\n",
+    "]}]\n",
+    "\n",
+    "# Process inputs\n",
+    "prompt = processor.apply_chat_template(messages, add_generation_prompt=True)\n",
+    "inputs = processor(images=[image], text=prompt, return_tensors=\"pt\")\n",
+    "\n",
+    "# Convert to numpy with correct dtypes\n",
+    "pixel_values = inputs[\"pixel_values\"].numpy().astype(np.float32)\n",
+    "pixel_attention_mask = inputs[\"pixel_attention_mask\"].numpy().astype(np.int64)\n",
+    "spatial_shapes = inputs[\"spatial_shapes\"].numpy().astype(np.int64)\n",
+    "input_ids = inputs[\"input_ids\"].numpy().astype(np.int64)\n",
+    "\n",
+    "# Get image embeddings\n",
+    "image_outputs = embed_images.run(None, {\n",
+    "    \"pixel_values\": pixel_values,\n",
+    "    \"pixel_attention_mask\": pixel_attention_mask,\n",
+    "    \"spatial_shapes\": spatial_shapes,\n",
+    "})\n",
+    "image_embeds = image_outputs[0]\n",
+    "\n",
+    "# Get token embeddings\n",
+    "token_outputs = embed_tokens.run(None, {\"input_ids\": input_ids})\n",
+    "token_embeds = token_outputs[0]\n",
+    "\n",
+    "# Replace <image> tokens with image embeddings\n",
+    "image_token_id = processor.tokenizer.convert_tokens_to_ids(\"<image>\")\n",
+    "image_positions = np.where(input_ids[0] == image_token_id)[0]\n",
+    "for i, pos in enumerate(image_positions):\n",
+    "    if i < len(image_embeds):\n",
+    "        token_embeds[0, pos] = image_embeds[i]\n",
+    "\n",
+    "# Initialize KV cache for stateful decoding\n",
+    "ONNX_DTYPE = {\"tensor(float)\": np.float32, \"tensor(float16)\": np.float16, \"tensor(int64)\": np.int64}\n",
+    "cache = {}\n",
+    "for inp in decoder.get_inputs():\n",
+    "    if inp.name in {\"inputs_embeds\", \"attention_mask\", \"position_ids\"}:\n",
+    "        continue\n",
+    "    shape = [d if isinstance(d, int) else 1 for d in inp.shape]\n",
+    "    for i, d in enumerate(inp.shape):\n",
+    "        if isinstance(d, str) and \"sequence\" in d.lower():\n",
+    "            shape[i] = 0\n",
+    "    cache[inp.name] = np.zeros(shape, dtype=ONNX_DTYPE.get(inp.type, np.float32))\n",
+    "\n",
+    "# Generate tokens\n",
+    "seq_len = token_embeds.shape[1]\n",
+    "generated_tokens = []\n",
+    "\n",
+    "for step in range(100):  # max tokens\n",
+    "    if step == 0:\n",
+    "        embeds = token_embeds.astype(np.float32)\n",
+    "    else:\n",
+    "        last_token = np.array([[generated_tokens[-1]]], dtype=np.int64)\n",
+    "        embeds = embed_tokens.run(None, {\"input_ids\": last_token})[0].astype(np.float32)\n",
+    "\n",
+    "    attn_mask = np.ones((1, seq_len + len(generated_tokens)), dtype=np.int64)\n",
+    "    feed = {\"inputs_embeds\": embeds, \"attention_mask\": attn_mask, **cache}\n",
+    "\n",
+    "    outputs = decoder.run(None, feed)\n",
+    "    next_token = int(np.argmax(outputs[0][0, -1]))\n",
+    "    generated_tokens.append(next_token)\n",
+    "\n",
+    "    # Update cache\n",
+    "    for i, out in enumerate(decoder.get_outputs()[1:], 1):\n",
+    "        name = out.name.replace(\"present_conv\", \"past_conv\").replace(\"present.\", \"past_key_values.\")\n",
+    "        if name in cache:\n",
+    "            cache[name] = outputs[i]\n",
+    "\n",
+    "    if next_token == processor.tokenizer.eos_token_id:\n",
+    "        break\n",
+    "\n",
+    "    print(processor.tokenizer.decode(next_token, skip_special_tokens=True))\n",
+    "# print(processor.tokenizer.decode(generated_tokens, skip_special_tokens=True))\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "9144e5cb",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "この画像は、雪に覆われた斜面で、晴れた日を楽しんでいる人物を描いています。太陽が空を輝かせ、雪に長い影を落としています。背景には、木々や山々が広がり、冬の風景を完成させています。\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(processor.tokenizer.decode(generated_tokens, skip_special_tokens=True))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c3864c06",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "np.array(image)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "id": "1f9373fd",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "import onnxruntime as ort\n",
+    "from huggingface_hub import hf_hub_download\n",
+    "from transformers import AutoTokenizer\n",
+    "from IPython.display import display, HTML, clear_output, Markdown\n",
+    "\n",
+    "# Download model (Q4 recommended)\n",
+    "model_id = \"LiquidAI/LFM2.5-1.2B-Instruct-ONNX\"\n",
+    "model_path = hf_hub_download(model_id, \"onnx/model_q4.onnx\")\n",
+    "\n",
+    "# Download all data files (handles multiple splits for large models)\n",
+    "from huggingface_hub import list_repo_files\n",
+    "for f in list_repo_files(model_id):\n",
+    "    if f.startswith(\"onnx/model_q4.onnx_data\"):\n",
+    "        hf_hub_download(model_id, f)\n",
+    "\n",
+    "# Load model and tokenizer\n",
+    "session = ort.InferenceSession(model_path)\n",
+    "tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 28,
+   "id": "2b426e38",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "<translation>会議やインタビュー、録音を検索可能なテキストに自動変換します。Nottaで効率的に作業できます</translation>"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "execution_count": 28,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "content = '''\n",
+    "以下の文章を流暢な日本語に翻訳してください。\n",
+    "出力は<translate>タグで囲んでください。\n",
+    "タグ以外の出力は絶対やめてください。\n",
+    "\n",
+    "【例】\n",
+    "<translate>this is translated text</translate>\n",
+    "\n",
+    "【翻訳前文章】\n",
+    "Automatically convert your meetings, interviews, or recordings into searchable text with Notta. Transcribe, edit, summarize, and collaborate all in a single workflow to stay productive.\n",
+    "'''\n",
+    "\n",
+    "# Prepare chat input\n",
+    "\n",
+    "messages = [{\"role\": \"user\", \"content\": content}]\n",
+    "prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
+    "input_ids = np.array([tokenizer.encode(prompt, add_special_tokens=False)], dtype=np.int64)\n",
+    "\n",
+    "# Initialize KV cache\n",
+    "ONNX_DTYPE = {\"tensor(float)\": np.float32, \"tensor(float16)\": np.float16, \"tensor(int64)\": np.int64}\n",
+    "cache = {}\n",
+    "for inp in session.get_inputs():\n",
+    "    if inp.name in {\"input_ids\", \"attention_mask\", \"position_ids\"}:\n",
+    "        continue\n",
+    "    shape = [d if isinstance(d, int) else 1 for d in inp.shape]\n",
+    "    for i, d in enumerate(inp.shape):\n",
+    "        if isinstance(d, str) and \"sequence\" in d.lower():\n",
+    "            shape[i] = 0\n",
+    "    cache[inp.name] = np.zeros(shape, dtype=ONNX_DTYPE.get(inp.type, np.float32))\n",
+    "\n",
+    "# Check if model uses position_ids\n",
+    "input_names = {inp.name for inp in session.get_inputs()}\n",
+    "use_position_ids = \"position_ids\" in input_names\n",
+    "\n",
+    "# Generate tokens\n",
+    "seq_len = input_ids.shape[1]\n",
+    "generated_tokens = []\n",
+    "\n",
+    "for step in range(512):  # max tokens\n",
+    "    if step == 0:\n",
+    "        ids = input_ids\n",
+    "        pos = np.arange(seq_len, dtype=np.int64).reshape(1, -1)\n",
+    "    else:\n",
+    "        ids = np.array([[generated_tokens[-1]]], dtype=np.int64)\n",
+    "        pos = np.array([[seq_len + len(generated_tokens) - 1]], dtype=np.int64)\n",
+    "\n",
+    "    attn_mask = np.ones((1, seq_len + len(generated_tokens)), dtype=np.int64)\n",
+    "    feed = {\"input_ids\": ids, \"attention_mask\": attn_mask, **cache}\n",
+    "    if use_position_ids:\n",
+    "        feed[\"position_ids\"] = pos\n",
+    "\n",
+    "    outputs = session.run(None, feed)\n",
+    "    next_token = int(np.argmax(outputs[0][0, -1]))\n",
+    "    generated_tokens.append(next_token)\n",
+    "\n",
+    "    # Update cache\n",
+    "    for i, out in enumerate(session.get_outputs()[1:], 1):\n",
+    "        name = out.name.replace(\"present_conv\", \"past_conv\").replace(\"present.\", \"past_key_values.\")\n",
+    "        if name in cache:\n",
+    "            cache[name] = outputs[i]\n",
+    "\n",
+    "    token_str = tokenizer.decode([next_token], skip_special_tokens=True)\n",
+    "    print(token_str, end=\"\", flush=True)\n",
+    "\n",
+    "    if next_token == tokenizer.eos_token_id:\n",
+    "        break\n",
+    "\n",
+    "clear_output()\n",
+    "Markdown(tokenizer.decode(generated_tokens, skip_special_tokens=True))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 25,
+   "id": "1186c8f5",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'<translate>LFM2.5は、乗算ゲートと短い畳み込みを組み合わせたハイブリッドアーキテクチャで、CPU、GPU、NPUハードウェアでの高速推論に最適化されています</translate>'"
+      ]
+     },
+     "execution_count": 25,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "tokenizer.decode(generated_tokens, skip_special_tokens=True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ddc378d4",
+   "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": "3.10.11"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}