Upload 76 files

inference/conversation.py

@@ -0,0 +1,224 @@
+import dataclasses
+from enum import auto, Enum
+from typing import List, Tuple
+
+
+class SeparatorStyle(Enum):
+    """Different separator style."""
+    SINGLE = auto()
+    TWO = auto()
+    MPT = auto()
+    PLAIN = auto()
+    LLAMA_2 = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that keeps all conversation history."""
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
+    sep: str = "###"
+    sep2: str = None
+    version: str = "Unknown"
+
+    skip_next: bool = False
+
+    def get_prompt(self):
+        messages = self.messages
+        if len(messages) > 0 and type(messages[0][1]) is tuple:
+            messages = self.messages.copy()
+            init_role, init_msg = messages[0].copy()
+            init_msg = init_msg[0].replace("<image>", "").strip()
+            if 'mmtag' in self.version:
+                messages[0] = (init_role, init_msg)
+                messages.insert(0, (self.roles[0], "<Image><image></Image>"))
+                messages.insert(1, (self.roles[1], "Received."))
+            else:
+                messages[0] = (init_role, "<image>\n" + init_msg)
+
+        if self.sep_style == SeparatorStyle.SINGLE:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + self.sep
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.TWO:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = self.system
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += message + seps[i % 2]
+                else:
+                    ret += ""
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+        return ret
+
+    def append_message(self, role, message):
+        self.messages.append([role, message])
+
+    def get_images(self, return_pil=False):
+        images = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    import base64
+                    from io import BytesIO
+                    from PIL import Image
+                    msg, image, image_process_mode = msg
+                    if image_process_mode == "Pad":
+                        def expand2square(pil_img, background_color=(122, 116, 104)):
+                            width, height = pil_img.size
+                            if width == height:
+                                return pil_img
+                            elif width > height:
+                                result = Image.new(pil_img.mode, (width, width), background_color)
+                                result.paste(pil_img, (0, (width - height) // 2))
+                                return result
+                            else:
+                                result = Image.new(pil_img.mode, (height, height), background_color)
+                                result.paste(pil_img, ((height - width) // 2, 0))
+                                return result
+                        image = expand2square(image)
+                    elif image_process_mode in ["Default", "Crop"]:
+                        pass
+                    elif image_process_mode == "Resize":
+                        image = image.resize((336, 336))
+                    else:
+                        raise ValueError(f"Invalid image_process_mode: {image_process_mode}")
+                    max_hw, min_hw = max(image.size), min(image.size)
+                    aspect_ratio = max_hw / min_hw
+                    max_len, min_len = 800, 400
+                    shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+                    longest_edge = int(shortest_edge * aspect_ratio)
+                    W, H = image.size
+                    if longest_edge != max(image.size):
+                        if H > W:
+                            H, W = longest_edge, shortest_edge
+                        else:
+                            H, W = shortest_edge, longest_edge
+                        image = image.resize((W, H))
+                    if return_pil:
+                        images.append(image)
+                    else:
+                        buffered = BytesIO()
+                        image.save(buffered, format="PNG")
+                        img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+                        images.append(img_b64_str)
+        return images
+
+    def to_gradio_chatbot(self):
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    import base64
+                    from io import BytesIO
+                    msg, image, image_process_mode = msg
+                    max_hw, min_hw = max(image.size), min(image.size)
+                    aspect_ratio = max_hw / min_hw
+                    max_len, min_len = 800, 400
+                    shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+                    longest_edge = int(shortest_edge * aspect_ratio)
+                    W, H = image.size
+                    if H > W:
+                        H, W = longest_edge, shortest_edge
+                    else:
+                        H, W = shortest_edge, longest_edge
+                    image = image.resize((W, H))
+                    buffered = BytesIO()
+                    image.save(buffered, format="JPEG")
+                    img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+                    img_str = f'<img src="data:image/png;base64,{img_b64_str}" alt="user upload image" />'
+                    msg = img_str + msg.replace('<image>', '').strip()
+                    ret.append([msg, None])
+                else:
+                    ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def copy(self):
+        return Conversation(
+            system=self.system,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            version=self.version)
+
+    def dict(self):
+        if len(self.get_images()) > 0:
+            return {
+                "system": self.system,
+                "roles": self.roles,
+                "messages": [[x, y[0] if type(y) is tuple else y] for x, y in self.messages],
+                "offset": self.offset,
+                "sep": self.sep,
+                "sep2": self.sep2,
+            }
+        return {
+            "system": self.system,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+            "sep": self.sep,
+            "sep2": self.sep2,
+        }
+
+
+conv_phi_v0 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v0",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="<|endoftext|>",
+)
+
+conv_llava_plain = Conversation(
+    system="",
+    roles=("", ""),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.PLAIN,
+    sep="\n",
+)
+
+default_conversation = conv_phi_v0
+conv_templates = {
+    "default": conv_phi_v0,
+    "v0": conv_phi_v0,
+    "phi-2_v0": conv_phi_v0,
+
+    "plain": conv_llava_plain,
+}
+
+
+if __name__ == "__main__":
+    print(default_conversation.get_prompt())

inference/inference.ipynb

@@ -0,0 +1,369 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "cdad6b21-030a-40d3-9b31-a229e5b6196d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer, AutoConfig, CLIPImageProcessor"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "1f832710-0e8c-42ec-b581-1b15fd2a6acc",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[2024-01-25 14:31:58,511] [INFO] [real_accelerator.py:110:get_accelerator] Setting ds_accelerator to cuda (auto detect)\n"
+     ]
+    }
+   ],
+   "source": [
+    "from model import LlavaPhiForCausalLM"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "9e68f1d4-1ae3-4d45-b818-4600218d2215",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "e5e13e666e3a43d4ad26cc70904abee8",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "model_name = \"RaviNaik/Llava-Phi2\"\n",
+    "model = LlavaPhiForCausalLM.from_pretrained(model_name)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "49edfa0d-e08a-4d3c-a1d6-34068b122419",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\n"
+     ]
+    }
+   ],
+   "source": [
+    "tokenizer = AutoTokenizer.from_pretrained(model_name)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "dcec20cd-d946-42d7-8e10-c198cd49b486",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image_processor = CLIPImageProcessor.from_pretrained(model_name)\n",
+    "mm_use_im_start_end = getattr(model.config, \"mm_use_im_start_end\", False)\n",
+    "mm_use_im_patch_token = getattr(model.config, \"mm_use_im_patch_token\", True)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "443c13c4-b7e6-4bc5-b6c7-c577bd4708f6",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "if mm_use_im_patch_token:\n",
+    "    tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)\n",
+    "if mm_use_im_start_end:\n",
+    "    tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)\n",
+    "    \n",
+    "if hasattr(model.config, \"max_sequence_length\"):\n",
+    "        context_len = model.config.max_sequence_length\n",
+    "else:\n",
+    "    context_len = 2048"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "d8caee43-0d2a-46d4-bdbc-2cfc7dec9e52",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from transformers import WhisperProcessor, WhisperForConditionalGeneration"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "3acea526-d8ae-4eb6-8dfc-4ea72651b547",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class AudioLanguageConnector:\n",
+    "    def __init__(self, projection_dim):\n",
+    "        model_name = \"microsoft/phi-2\"\n",
+    "        self.phi2_tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n",
+    "        self.phi2_tokenizer.pad_token = self.phi2_tokenizer.eos_token\n",
+    "        self.phi2_tokenizer.max_length = projection_dim\n",
+    "\n",
+    "    def __call__(self, text):\n",
+    "        text = f\"<audio_start> {text} <audio_end>\"\n",
+    "        tokens = self.phi2_tokenizer(text, return_tensors=\"pt\", return_attention_mask=False)\n",
+    "        return tokens\n",
+    "        \n",
+    "\n",
+    "class WhisperWithProjection:\n",
+    "    def __init__(self, projection_dim, device):\n",
+    "        self.device = device\n",
+    "        self.processor = WhisperProcessor.from_pretrained(\"openai/whisper-tiny\", device_map=device)\n",
+    "        self.model = WhisperForConditionalGeneration.from_pretrained(\"openai/whisper-tiny\", device_map=device)\n",
+    "        self.model.config.forced_decoder_ids = None\n",
+    "        # self.audio_language_connector = AudioLanguageConnector(projection_dim)\n",
+    "        \n",
+    "    def __call__(self, audio):\n",
+    "        input_features = self.processor(audio[\"array\"],\n",
+    "                                   sampling_rate=audio[\"sampling_rate\"],\n",
+    "                                   return_tensors=\"pt\").input_features\n",
+    "        # generate token ids\n",
+    "        predicted_ids = self.model.generate(input_features.to(self.device))\n",
+    "        # decode token ids to text        \n",
+    "        transcription = self.processor.batch_decode(predicted_ids, skip_special_tokens=True)\n",
+    "\n",
+    "        # audio_embeddings = self.audio_language_connector(transcription)\n",
+    "        return transcription"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "a2757c91-2ec1-4fe7-9216-03740bf80061",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "IGNORE_INDEX = -100\n",
+    "IMAGE_TOKEN_INDEX = -200\n",
+    "DEFAULT_IMAGE_TOKEN = \"<image>\"\n",
+    "DEFAULT_IMAGE_PATCH_TOKEN = \"<im_patch>\"\n",
+    "DEFAULT_IM_START_TOKEN = \"<im_start>\"\n",
+    "DEFAULT_IM_END_TOKEN = \"<im_end>\"\n",
+    "\n",
+    "from conversation import conv_templates, SeparatorStyle\n",
+    "\n",
+    "class MultiModalPhi2:\n",
+    "    def __init__(self, modelname_or_path=\"RaviNaik/Llava-Phi2\",\n",
+    "                temperature=0.2,\n",
+    "                max_new_tokens=1024,\n",
+    "                device=\"cuda:0\"):\n",
+    "        self.model_name = modelname_or_path\n",
+    "        self.temperature = temperature\n",
+    "        self.max_new_tokens = max_new_tokens\n",
+    "        self.device = device\n",
+    "        self.disable_torch_init()\n",
+    "        self.whisper_w_proj = WhisperWithProjection(projection_dim=512, device=device)\n",
+    "        self.load_pretrained_model()\n",
+    "        \n",
+    "    def disable_torch_init(self):\n",
+    "        \"\"\"\n",
+    "        Disable the redundant torch default initialization to accelerate model creation.\n",
+    "        \"\"\"\n",
+    "        setattr(torch.nn.Linear, \"reset_parameters\", lambda self: None)\n",
+    "        setattr(torch.nn.LayerNorm, \"reset_parameters\", lambda self: None)\n",
+    "        \n",
+    "    def load_pretrained_model(self):\n",
+    "        self.model = LlavaPhiForCausalLM.from_pretrained(self.model_name, device_map=self.device)\n",
+    "        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)\n",
+    "        self.image_processor = CLIPImageProcessor.from_pretrained(self.model_name)\n",
+    "        mm_use_im_start_end = getattr(self.model.config, \"mm_use_im_start_end\", False)\n",
+    "        mm_use_im_patch_token = getattr(self.model.config, \"mm_use_im_patch_token\", True)\n",
+    "        if mm_use_im_patch_token:\n",
+    "            self.tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)\n",
+    "        if mm_use_im_start_end:\n",
+    "            self.tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)\n",
+    "    \n",
+    "    def tokenizer_image_token(self, prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):\n",
+    "        prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]\n",
+    "    \n",
+    "        def insert_separator(X, sep):\n",
+    "            return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]\n",
+    "    \n",
+    "        input_ids = []\n",
+    "        offset = 0\n",
+    "        if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:\n",
+    "            offset = 1\n",
+    "            input_ids.append(prompt_chunks[0][0])\n",
+    "        for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):\n",
+    "            input_ids.extend(x[offset:])\n",
+    "    \n",
+    "        if return_tensors is not None:\n",
+    "            if return_tensors == 'pt':\n",
+    "                return torch.tensor(input_ids, dtype=torch.long)\n",
+    "            raise ValueError(f'Unsupported tensor type: {return_tensors}')\n",
+    "        return input_ids\n",
+    "        \n",
+    "    def __call__(self, text, audio, image):\n",
+    "        qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\\n' + text\n",
+    "        conv = conv_templates[\"phi-2_v0\"].copy()\n",
+    "        conv.append_message(conv.roles[0], qs)\n",
+    "        conv.append_message(conv.roles[1], None)\n",
+    "        prompt = conv.get_prompt()\n",
+    "\n",
+    "        image_tensor = self.image_processor.preprocess(image, return_tensors='pt')['pixel_values'].to(self.device)\n",
+    "        \n",
+    "        input_ids = self.tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0)\n",
+    "        if audio is not None:\n",
+    "            audio_transcript = self.whisper_w_proj(audio)\n",
+    "            audio_embed = self.tokenizer(audio_transcript, return_tensors='pt')[\"input_ids\"]\n",
+    "            input_ids = torch.concat([input_ids, audio_embed], dim=1)\n",
+    "        input_ids = input_ids.to(self.device)\n",
+    "            \n",
+    "        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2\n",
+    "\n",
+    "        with torch.inference_mode():\n",
+    "            output_ids = self.model.generate(\n",
+    "                input_ids,\n",
+    "                images=image_tensor,\n",
+    "                do_sample=True,\n",
+    "                temperature=self.temperature,\n",
+    "                max_new_tokens=self.max_new_tokens,\n",
+    "                eos_token_id=self.tokenizer.eos_token_id,  # End of sequence token\n",
+    "                pad_token_id=self.tokenizer.eos_token_id,  # Pad token\n",
+    "                use_cache=True,\n",
+    "            )\n",
+    "\n",
+    "        input_token_len = input_ids.shape[1]\n",
+    "        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()\n",
+    "        if n_diff_input_output > 0:\n",
+    "            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')\n",
+    "        outputs = self.tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]\n",
+    "        outputs = outputs.strip()\n",
+    "        if outputs.endswith(stop_str):\n",
+    "            outputs = outputs[:-len(stop_str)]\n",
+    "        outputs = outputs.strip()\n",
+    "        return outputs"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "cc47e6a0-3544-4a60-930f-ccae87ef945a",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "9ef56077307d4cef907e25b092061611",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\n"
+     ]
+    }
+   ],
+   "source": [
+    "multimodal_phi2 = MultiModalPhi2()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "id": "cb8aca1b-7d75-45e7-b5a4-71d151f792e1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from PIL import Image\n",
+    "import requests\n",
+    "\n",
+    "url = \"https://www.ilankelman.org/stopsigns/australia.jpg\"\n",
+    "image = Image.open(requests.get(url, stream=True).raw)\n",
+    "\n",
+    "from datasets import load_dataset\n",
+    "audio_ds = load_dataset(\"hf-internal-testing/librispeech_asr_dummy\", \"clean\", split=\"validation\")\n",
+    "audio = audio_ds[0][\"audio\"]\n",
+    "\n",
+    "text = \"tell me about the image\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "id": "6767efc6-be4f-44d3-84ff-34db57d9f940",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'In the image, there is a Chinese writing on a pole in a foreign language. This suggests that the image was taken in a foreign country, possibly in a foreign country. The sign is in a foreign language, which might be in a foreign language. The sign is written in Japanese, which is a common language in Japan. The sign is also written in two different languages, which suggests that it is written in a language that is not in the native language.'"
+      ]
+     },
+     "execution_count": 14,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "multimodal_phi2(text, None, image)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0bdd0b8a-709b-4c82-ac1d-dc746d3a0748",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

inference/main.py

@@ -0,0 +1,226 @@
+import soundfile as sf
+import librosa
+import torch
+from transformers import (
+    AutoTokenizer,
+    CLIPImageProcessor,
+    WhisperProcessor,
+    WhisperForConditionalGeneration,
+)
+
+from .model import LlavaPhiForCausalLM
+from .conversation import conv_templates, SeparatorStyle
+
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+
+
+class AudioLanguageConnector:
+    def __init__(self, projection_dim):
+        model_name = "microsoft/phi-2"
+        self.phi2_tokenizer = AutoTokenizer.from_pretrained(
+            model_name, trust_remote_code=True
+        )
+        self.phi2_tokenizer.pad_token = self.phi2_tokenizer.eos_token
+        self.phi2_tokenizer.max_length = projection_dim
+
+    def __call__(self, text):
+        text = f"<audio_start> {text} <audio_end>"
+        tokens = self.phi2_tokenizer(
+            text, return_tensors="pt", return_attention_mask=False
+        )
+        return tokens
+
+
+class WhisperWithProjection:
+    def __init__(self, projection_dim, device):
+        self.device = device
+        self.processor = WhisperProcessor.from_pretrained(
+            "openai/whisper-tiny", device_map=device
+        )
+        self.model = WhisperForConditionalGeneration.from_pretrained(
+            "openai/whisper-tiny", device_map=device
+        )
+        self.model.config.forced_decoder_ids = None
+        # self.audio_language_connector = AudioLanguageConnector(projection_dim)
+
+    def __call__(self, audio):
+        array, sampling_rate = sf.read(audio)
+        resampled_array = librosa.resample(
+            array,
+            orig_sr=sampling_rate,
+            target_sr=16000,
+        )
+        input_features = self.processor(
+            resampled_array, sampling_rate=16000, return_tensors="pt"
+        ).input_features
+        # generate token ids
+        predicted_ids = self.model.generate(input_features.to(self.device))
+        # decode token ids to text
+        transcription = self.processor.batch_decode(
+            predicted_ids, skip_special_tokens=True
+        )
+
+        # audio_embeddings = self.audio_language_connector(transcription)
+        return transcription
+
+
+class MultiModalPhi2:
+    def __init__(
+        self,
+        modelname_or_path="RaviNaik/Llava-Phi2",
+        temperature=0.2,
+        max_new_tokens=1024,
+        device="cuda:0",
+    ):
+        self.model_name = modelname_or_path
+        self.temperature = temperature
+        self.max_new_tokens = max_new_tokens
+        self.device = device
+        self.disable_torch_init()
+        self.whisper_w_proj = WhisperWithProjection(projection_dim=512, device=device)
+        self.load_pretrained_model()
+
+    def disable_torch_init(self):
+        """
+        Disable the redundant torch default initialization to accelerate model creation.
+        """
+        setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
+        setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
+
+    def load_pretrained_model(self):
+        self.model = LlavaPhiForCausalLM.from_pretrained(
+            self.model_name, device_map=self.device
+        )
+        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
+        self.image_processor = CLIPImageProcessor.from_pretrained(self.model_name)
+        mm_use_im_start_end = getattr(self.model.config, "mm_use_im_start_end", False)
+        mm_use_im_patch_token = getattr(
+            self.model.config, "mm_use_im_patch_token", True
+        )
+        if mm_use_im_patch_token:
+            self.tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        if mm_use_im_start_end:
+            self.tokenizer.add_tokens(
+                [DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True
+            )
+
+    def tokenizer_image_token(
+        self,
+        prompt,
+        tokenizer,
+        image_token_index=IMAGE_TOKEN_INDEX,
+        return_tensors=None,
+    ):
+        prompt_chunks = [
+            tokenizer(chunk).input_ids for chunk in prompt.split("<image>")
+        ]
+
+        def insert_separator(X, sep):
+            return [ele for sublist in zip(X, [sep] * len(X)) for ele in sublist][:-1]
+
+        input_ids = []
+        offset = 0
+        if (
+            len(prompt_chunks) > 0
+            and len(prompt_chunks[0]) > 0
+            and prompt_chunks[0][0] == tokenizer.bos_token_id
+        ):
+            offset = 1
+            input_ids.append(prompt_chunks[0][0])
+        for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+            input_ids.extend(x[offset:])
+
+        if return_tensors is not None:
+            if return_tensors == "pt":
+                return torch.tensor(input_ids, dtype=torch.long)
+            raise ValueError(f"Unsupported tensor type: {return_tensors}")
+        return input_ids
+
+    def __call__(self, text, audio, image):
+        if text is None:
+            text = ""
+        if image is not None:
+            qs = (
+                DEFAULT_IM_START_TOKEN
+                + DEFAULT_IMAGE_TOKEN
+                + DEFAULT_IM_END_TOKEN
+                + "\n"
+                + text
+            )
+            conv = conv_templates["phi-2_v0"].copy()
+            conv.append_message(conv.roles[0], qs)
+            conv.append_message(conv.roles[1], None)
+            prompt = conv.get_prompt()
+
+            input_ids = self.tokenizer_image_token(
+                prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt"
+            ).unsqueeze(0)
+
+            image_tensor = self.image_processor.preprocess(image, return_tensors="pt")[
+                "pixel_values"
+            ].to(self.device)
+        else:
+            qs = text
+            conv = conv_templates["phi-2_v0"].copy()
+            conv.append_message(conv.roles[0], qs)
+            conv.append_message(conv.roles[1], None)
+            prompt = conv.get_prompt()
+
+            input_ids = self.tokenizer(prompt, return_tensors="pt")["input_ids"]
+
+            image_tensor = None
+
+        if audio is not None:
+            audio_transcript = self.whisper_w_proj(audio)
+            audio_embed = self.tokenizer(audio_transcript, return_tensors="pt")[
+                "input_ids"
+            ]
+            input_ids = torch.concat([input_ids, audio_embed], dim=1)
+        input_ids = input_ids.to(self.device)
+
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+
+        with torch.inference_mode():
+            if image is not None:
+                output_ids = self.model.generate(
+                    input_ids,
+                    images=image_tensor,
+                    do_sample=True,
+                    temperature=self.temperature,
+                    max_new_tokens=self.max_new_tokens,
+                    eos_token_id=self.tokenizer.eos_token_id,  # End of sequence token
+                    pad_token_id=self.tokenizer.eos_token_id,  # Pad token
+                    use_cache=True,
+                )
+            else:
+                output_ids = self.model.generate(
+                    input_ids,
+                    do_sample=True,
+                    temperature=self.temperature,
+                    max_new_tokens=self.max_new_tokens,
+                    eos_token_id=self.tokenizer.eos_token_id,  # End of sequence token
+                    pad_token_id=self.tokenizer.eos_token_id,  # Pad token
+                    use_cache=True,
+                )
+
+        input_token_len = input_ids.shape[1]
+        n_diff_input_output = (
+            (input_ids != output_ids[:, :input_token_len]).sum().item()
+        )
+        if n_diff_input_output > 0:
+            print(
+                f"[Warning] {n_diff_input_output} output_ids are not the same as the input_ids"
+            )
+        outputs = self.tokenizer.batch_decode(
+            output_ids[:, input_token_len:], skip_special_tokens=True
+        )[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[: -len(stop_str)]
+        outputs = outputs.strip()
+        return outputs

inference/model/__init__.py

@@ -0,0 +1,2 @@
+from .language_model.llava_phi import LlavaPhiForCausalLM
+from .language_model.configuration_llava_phi import LlavaPhiConfig, LlavaPhiVisionConfig, ProjectorConfig

inference/model/builder.py

@@ -0,0 +1,180 @@
+import os
+import warnings
+
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    AutoConfig,
+    BitsAndBytesConfig,
+    CLIPImageProcessor,
+)
+import torch
+from .language_model.llava_phi import LlavaPhiForCausalLM
+from .language_model.configuration_llava_phi import LlavaPhiConfig
+
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+
+
+def load_pretrained_model(
+    model_path,
+    model_base,
+    model_name,
+    load_8bit=False,
+    load_4bit=False,
+    device_map="cuda",
+    device="cuda",
+):
+    kwargs = {"device_map": device_map}
+    if load_8bit:
+        kwargs["load_in_8bit"] = True
+    elif load_4bit:
+        kwargs["load_in_4bit"] = True
+        kwargs["quantization_config"] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type="nf4",
+        )
+    # else:  # TODO: after fine-tuning LLava-Phi, load the model weights with fp16 will pose nan
+    #     kwargs['torch_dtype'] = torch.float16
+
+    if "phi" in model_name.lower():
+        # Load LLaVA-Phi model
+        if "lora" in model_name.lower() and model_base is None:
+            warnings.warn(
+                "There is `lora` in model name but no `model_base` is provided. If you are loading a LoRA model, please provide the `model_base` argument."
+            )
+        if "lora" in model_name.lower() and model_base is not None:
+            lora_cfg_pretrained = AutoConfig.from_pretrained(model_path)
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            print("Loading LLaVA-Phi from base model...")
+            model = LlavaPhiForCausalLM.from_pretrained(
+                model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs
+            )
+            token_num, tokem_dim = model.lm_head.out_features, model.lm_head.in_features
+            if model.lm_head.weight.shape[0] != token_num:
+                model.lm_head.weight = torch.nn.Parameter(
+                    torch.empty(
+                        token_num, tokem_dim, device=model.device, dtype=model.dtype
+                    )
+                )
+                model.model.embed_tokens.weight = torch.nn.Parameter(
+                    torch.empty(
+                        token_num, tokem_dim, device=model.device, dtype=model.dtype
+                    )
+                )
+
+            print("Loading additional LLaVA-Phi weights...")
+            if os.path.exists(os.path.join(model_path, "non_lora_trainables.bin")):
+                non_lora_trainables = torch.load(
+                    os.path.join(model_path, "non_lora_trainables.bin"),
+                    map_location="cpu",
+                )
+            else:
+                # this is probably from HF Hub
+                from huggingface_hub import hf_hub_download
+
+                def load_from_hf(repo_id, filename, subfolder=None):
+                    cache_file = hf_hub_download(
+                        repo_id=repo_id, filename=filename, subfolder=subfolder
+                    )
+                    return torch.load(cache_file, map_location="cpu")
+
+                non_lora_trainables = load_from_hf(
+                    model_path, "non_lora_trainables.bin"
+                )
+            non_lora_trainables = {
+                (k[11:] if k.startswith("base_model.") else k): v
+                for k, v in non_lora_trainables.items()
+            }
+            if any(k.startswith("model.model.") for k in non_lora_trainables):
+                non_lora_trainables = {
+                    (k[6:] if k.startswith("model.") else k): v
+                    for k, v in non_lora_trainables.items()
+                }
+            model.load_state_dict(non_lora_trainables, strict=False)
+
+            from peft import PeftModel
+
+            print("Loading LoRA weights...")
+            model = PeftModel.from_pretrained(model, model_path)
+            print("Merging LoRA weights...")
+            model = model.merge_and_unload()
+            print("Model is loaded...")
+        elif model_base is not None:
+            # this may be mm projector only
+            print("Loading LLaVA-Phi from base model...")
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            cfg_pretrained = AutoConfig.from_pretrained(model_path)
+            model = LlavaPhiForCausalLM.from_pretrained(
+                model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs
+            )
+
+            mm_projector_weights = torch.load(
+                os.path.join(model_path, "mm_projector.bin"), map_location="cpu"
+            )
+            mm_projector_weights = {
+                k: v.to(torch.float16) for k, v in mm_projector_weights.items()
+            }
+            model.load_state_dict(mm_projector_weights, strict=False)
+        else:
+            print("load llaVA-Phi MLLM!!!")
+            config = LlavaPhiConfig.from_pretrained(model_path, trust_remote_code=True)
+            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
+            model = LlavaPhiForCausalLM.from_pretrained(
+                model_path, config=config, use_safetensors=True, **kwargs
+            ).to("cuda")
+    else:
+        # Load language model
+        if model_base is not None:
+            # PEFT model
+            from peft import PeftModel
+
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(
+                model_base,
+                torch_dtype=torch.float16,
+                low_cpu_mem_usage=True,
+                device_map="auto",
+            )
+            print(f"Loading LoRA weights from {model_path}")
+            model = PeftModel.from_pretrained(model, model_path)
+            print(f"Merging weights")
+            model = model.merge_and_unload()
+            print("Convert to FP16...")
+            model.to(torch.float16)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(
+                model_path, low_cpu_mem_usage=True, **kwargs
+            )
+
+    image_processor = CLIPImageProcessor.from_pretrained(model_path)
+
+    if "phi" in model_name.lower():
+        mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
+        mm_use_im_patch_token = getattr(model.config, "mm_use_im_patch_token", True)
+
+        # TODO: the tokenizer length of phi-2 is 50295, but the output class of lm_head is 51200
+        if mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        if mm_use_im_start_end:
+            tokenizer.add_tokens(
+                [DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True
+            )
+            # model.resize_token_embeddings(len(tokenizer))
+    else:
+        raise ValueError(f"Unsupported model name: {model_name}")
+
+    if hasattr(model.config, "max_sequence_length"):
+        context_len = model.config.max_sequence_length
+    else:
+        context_len = 2048
+    model.to(device="cuda")
+    print(kwargs)
+    return tokenizer, model, image_processor, context_len

inference/model/language_model/configuration_llava_phi.py

@@ -0,0 +1,191 @@
+import os
+from typing import Union
+from transformers import PretrainedConfig, PhiConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class LlavaPhiVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPVisionModel`]. It is used to instantiate a
+    CLIP vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+        mm_vision_select_feature (`str`, *optional*, defaults to `"patch"`):
+            The feature to select from the vision encoder output. Can be one of `"patch"` or `"cls_patch"`.
+        mm_vision_select_layer (`int`, *optional*, defaults to `-2`):
+            The layer to select from the vision encoder output.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPVisionConfig, CLIPVisionModel
+
+    >>> # Initializing a CLIPVisionConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPVisionConfig()
+
+    >>> # Initializing a CLIPVisionModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "llava_phi_clip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        projection_dim=512,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=32,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        mm_vision_select_feature="patch",
+        mm_vision_select_layer=-2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.mm_vision_select_feature = mm_vision_select_feature
+        self.mm_vision_select_layer = mm_vision_select_layer
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+    ) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "llava_phi-phi":
+            config_dict = config_dict["vision_config"]
+
+        if (
+            "model_type" in config_dict
+            and hasattr(cls, "model_type")
+            and config_dict["model_type"] != cls.model_type
+        ):
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class ProjectorConfig(PretrainedConfig):
+    model_type = "llava_phi_projector"
+
+    def __init__(
+        self, mm_projector_type="linear", mm_hidden_size=768, hidden_size=2560, **kwargs
+    ):
+        self.mm_projector_type = mm_projector_type
+        self.mm_hidden_size = mm_hidden_size
+        self.hidden_size = hidden_size
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+    ) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "llava_phi-phi":
+            config_dict = config_dict["projector_config"]
+
+        if (
+            "model_type" in config_dict
+            and hasattr(cls, "model_type")
+            and config_dict["model_type"] != cls.model_type
+        ):
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+DEFAULT_VISUAL_CONFIG = {
+    "vision_tower": LlavaPhiVisionConfig().to_dict(),
+    "mm_projector": ProjectorConfig().to_dict(),
+}
+
+
+class LlavaPhiConfig(PhiConfig):
+    model_type = "llava_phi"
+
+    def __init__(self, vision_config=None, **kwargs):
+        if vision_config is None:
+            self.vision_config = DEFAULT_VISUAL_CONFIG
+        else:
+            self.vision_config = vision_config
+
+        super().__init__(**kwargs)
+
+
+if __name__ == "__main__":
+    print(LlavaPhiVisionConfig())

inference/model/language_model/llava_phi.py

@@ -0,0 +1,126 @@
+import os
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+    PhiModel, PhiPreTrainedModel
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from ..llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
+from transformers.utils import logging
+from .configuration_llava_phi import LlavaPhiConfig
+
+logger = logging.get_logger(__name__)
+
+
+class LLavaPhiModel(LlavaMetaModel, PhiModel):
+    config_class = LlavaPhiConfig
+
+    def __init__(self, config):
+        super(LLavaPhiModel, self).__init__(config)
+
+
+class LlavaPhiForCausalLM(PhiPreTrainedModel, LlavaMetaForCausalLM):
+    config_class = LlavaPhiConfig
+
+    def __init__(self, config):
+        super(PhiPreTrainedModel, self).__init__(config)
+        self.model = LLavaPhiModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=True)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            labels: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            images: Optional[torch.FloatTensor] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        input_ids, attention_mask, past_key_values, inputs_embeds, labels = self.prepare_inputs_labels_for_multimodal(
+            input_ids, attention_mask, past_key_values, labels, images)
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model/pipeline parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+        
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+            self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "images": kwargs.get("images", None),
+            }
+        )
+        return model_inputs
+
+
+AutoConfig.register("llava_phi", LlavaPhiConfig)
+AutoModelForCausalLM.register(LlavaPhiConfig, LlavaPhiForCausalLM)

inference/model/llava_arch.py

@@ -0,0 +1,330 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import torch
+
+from .multimodal_encoder.clip_encoder import CLIPVisionTower
+from .multimodal_projector.builder import build_vision_projector
+from .language_model.configuration_llava_phi import (
+    LlavaPhiConfig,
+    LlavaPhiVisionConfig,
+    ProjectorConfig,
+)
+
+# from llava_phi.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+
+
+class LlavaMetaModel:
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+        self.vision_tower = CLIPVisionTower(
+            LlavaPhiVisionConfig(**config.vision_config["vision_tower"])
+        )
+        self.mm_projector = build_vision_projector(
+            ProjectorConfig(**config.vision_config["mm_projector"])
+        )
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, "vision_tower", None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+
+class LlavaMetaForCausalLM(ABC):
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def encode_images(self, images):
+        image_features = self.get_model().get_vision_tower()(images)
+        image_features = self.get_model().mm_projector(image_features)
+        return image_features
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, attention_mask, past_key_values, labels, images
+    ):
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            if (
+                past_key_values is not None
+                and vision_tower is not None
+                and images is not None
+                and input_ids.shape[1] == 1
+            ):
+                attention_mask = torch.ones(
+                    (attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1),
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+            return input_ids, attention_mask, past_key_values, None, labels
+
+        if type(images) is list or images.ndim == 5:
+            concat_images = torch.cat([image for image in images], dim=0)
+            image_features = self.encode_images(concat_images)
+            split_sizes = [image.shape[0] for image in images]
+            image_features = torch.split(image_features, split_sizes, dim=0)
+            image_features = [x.flatten(0, 1) for x in image_features]
+        else:
+            image_features = self.encode_images(images)
+
+        new_input_embeds = []
+        new_labels = [] if labels is not None else None
+        cur_image_idx = 0
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
+                # multimodal LLM, but the current sample is not multimodal
+                # FIXME: this is a hacky fix, for deepspeed zero3 to work
+                half_len = cur_input_ids.shape[0] // 2
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(
+                    cur_input_ids[:half_len]
+                )
+                cur_input_embeds_2 = self.get_model().embed_tokens(
+                    cur_input_ids[half_len:]
+                )
+                cur_input_embeds = torch.cat(
+                    [cur_input_embeds_1, cur_image_features[0:0], cur_input_embeds_2],
+                    dim=0,
+                )
+                new_input_embeds.append(cur_input_embeds)
+                if labels is not None:
+                    new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+            image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
+            cur_new_input_embeds = []
+            if labels is not None:
+                cur_labels = labels[batch_idx]
+                cur_new_labels = []
+                assert cur_labels.shape == cur_input_ids.shape
+            while image_token_indices.numel() > 0:
+                cur_image_features = image_features[cur_image_idx]
+                image_token_start = image_token_indices[0]
+                if getattr(self.config, "tune_mm_mlp_adapter", False) and getattr(
+                    self.config, "mm_use_im_start_end", False
+                ):
+                    cur_new_input_embeds.append(
+                        self.get_model()
+                        .embed_tokens(cur_input_ids[: image_token_start - 1])
+                        .detach()
+                    )
+                    cur_new_input_embeds.append(
+                        self.get_model().embed_tokens(
+                            cur_input_ids[image_token_start - 1 : image_token_start]
+                        )
+                    )
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_input_embeds.append(
+                        self.get_model().embed_tokens(
+                            cur_input_ids[image_token_start + 1 : image_token_start + 2]
+                        )
+                    )
+                    if labels is not None:
+                        cur_new_labels.append(cur_labels[:image_token_start])
+                        cur_new_labels.append(
+                            torch.full(
+                                (cur_image_features.shape[0],),
+                                IGNORE_INDEX,
+                                device=labels.device,
+                                dtype=labels.dtype,
+                            )
+                        )
+                        cur_new_labels.append(
+                            cur_labels[image_token_start : image_token_start + 1]
+                        )
+                        cur_labels = cur_labels[image_token_start + 2 :]
+                else:
+                    cur_new_input_embeds.append(
+                        self.get_model().embed_tokens(cur_input_ids[:image_token_start])
+                    )
+                    cur_new_input_embeds.append(cur_image_features)
+                    if labels is not None:
+                        cur_new_labels.append(cur_labels[:image_token_start])
+                        cur_new_labels.append(
+                            torch.full(
+                                (cur_image_features.shape[0],),
+                                IGNORE_INDEX,
+                                device=labels.device,
+                                dtype=labels.dtype,
+                            )
+                        )
+                        cur_labels = cur_labels[image_token_start + 1 :]
+                cur_image_idx += 1
+                if getattr(self.config, "tune_mm_mlp_adapter", False) and getattr(
+                    self.config, "mm_use_im_start_end", False
+                ):
+                    cur_input_ids = cur_input_ids[image_token_start + 2 :]
+                else:
+                    cur_input_ids = cur_input_ids[image_token_start + 1 :]
+                image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
+            if cur_input_ids.numel() > 0:
+                if getattr(self.config, "tune_mm_mlp_adapter", False) and getattr(
+                    self.config, "mm_use_im_start_end", False
+                ):
+                    cur_new_input_embeds.append(
+                        self.get_model().embed_tokens(cur_input_ids).detach()
+                    )
+                else:
+                    cur_new_input_embeds.append(
+                        self.get_model().embed_tokens(cur_input_ids)
+                    )
+                if labels is not None:
+                    cur_new_labels.append(cur_labels)
+            cur_new_input_embeds = [
+                x.to(device=self.device) for x in cur_new_input_embeds
+            ]
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
+            new_input_embeds.append(cur_new_input_embeds)
+            if labels is not None:
+                cur_new_labels = torch.cat(cur_new_labels, dim=0)
+                new_labels.append(cur_new_labels)
+
+        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
+            max_len = max(x.shape[0] for x in new_input_embeds)
+
+            new_input_embeds_align = []
+            for cur_new_embed in new_input_embeds:
+                cur_new_embed = torch.cat(
+                    (
+                        cur_new_embed,
+                        torch.zeros(
+                            (max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]),
+                            dtype=cur_new_embed.dtype,
+                            device=cur_new_embed.device,
+                        ),
+                    ),
+                    dim=0,
+                )
+                new_input_embeds_align.append(cur_new_embed)
+            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)
+
+            if labels is not None:
+                new_labels_align = []
+                _new_labels = new_labels
+                for cur_new_label in new_labels:
+                    cur_new_label = torch.cat(
+                        (
+                            cur_new_label,
+                            torch.full(
+                                (max_len - cur_new_label.shape[0],),
+                                IGNORE_INDEX,
+                                dtype=cur_new_label.dtype,
+                                device=cur_new_label.device,
+                            ),
+                        ),
+                        dim=0,
+                    )
+                    new_labels_align.append(cur_new_label)
+                new_labels = torch.stack(new_labels_align, dim=0)
+
+            if attention_mask is not None:
+                new_attention_mask = []
+                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(
+                    attention_mask, _new_labels, new_labels
+                ):
+                    new_attn_mask_pad_left = torch.full(
+                        (cur_new_labels.shape[0] - labels.shape[1],),
+                        True,
+                        dtype=attention_mask.dtype,
+                        device=attention_mask.device,
+                    )
+                    new_attn_mask_pad_right = torch.full(
+                        (cur_new_labels_align.shape[0] - cur_new_labels.shape[0],),
+                        False,
+                        dtype=attention_mask.dtype,
+                        device=attention_mask.device,
+                    )
+                    cur_new_attention_mask = torch.cat(
+                        (
+                            new_attn_mask_pad_left,
+                            cur_attention_mask,
+                            new_attn_mask_pad_right,
+                        ),
+                        dim=0,
+                    )
+                    new_attention_mask.append(cur_new_attention_mask)
+                attention_mask = torch.stack(new_attention_mask, dim=0)
+                assert attention_mask.shape == new_labels.shape
+        else:
+            new_input_embeds = torch.stack(new_input_embeds, dim=0)
+            if labels is not None:
+                new_labels = torch.stack(new_labels, dim=0)
+
+            if attention_mask is not None:
+                new_attn_mask_pad_left = torch.full(
+                    (
+                        attention_mask.shape[0],
+                        new_input_embeds.shape[1] - input_ids.shape[1],
+                    ),
+                    True,
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+                attention_mask = torch.cat(
+                    (new_attn_mask_pad_left, attention_mask), dim=1
+                )
+                assert attention_mask.shape == new_input_embeds.shape[:2]
+
+        return None, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens(
+                [DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True
+            )
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True
+                )
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True
+                )
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False

inference/model/multimodal_encoder/clip_encoder.py

@@ -0,0 +1,89 @@
+from abc import ABC
+
+import torch
+import torch.nn as nn
+
+from transformers import CLIPPreTrainedModel, CLIPVisionConfig
+from transformers.models.clip.modeling_clip import CLIPVisionTransformer
+from inference.model.language_model.configuration_llava_phi import LlavaPhiVisionConfig
+
+
+class CLIPVisionTower(CLIPPreTrainedModel):
+    config_class = LlavaPhiVisionConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.vision_model = CLIPVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[
+            self.config.mm_vision_select_layer
+        ]
+        if self.config.mm_vision_select_feature == "patch":
+            image_features = image_features[:, 1:]
+        elif self.config.mm_vision_select_feature == "cls_patch":
+            image_features = image_features
+        else:
+            raise ValueError(
+                f"Unexpected select feature: {self.config.mm_vision_select_feature}"
+            )
+        return image_features
+
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_model(
+                    image.to(device=self.device, dtype=self.dtype).unsqueeze(0),
+                    output_hidden_states=True,
+                )
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_model(
+                images.to(device=self.device, dtype=self.dtype),
+                output_hidden_states=True,
+            )
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return list(self.vision_model.parameters())[0].dtype
+
+    @property
+    def device(self):
+        return list(self.vision_model.parameters())[0].device
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+
+if __name__ == "__main__":
+    clip_config = CLIPVisionConfig.from_pretrained(
+        "/data/private/zhumj/GPTcode/mm-phi/openai/clip-vit-large-patch14-336"
+    )
+    print("################ clip_config ##############")
+    print(clip_config)
+    phi_vis_config = LlavaPhiVisionConfig(**clip_config.to_dict())
+    print("################ phi_vis_config ##############")
+    print(phi_vis_config)
+
+    model = CLIPVisionTower(clip_config)
+    # print(list(model.vision_model.parameters())[0].dtype)

inference/model/multimodal_projector/builder.py

@@ -0,0 +1,50 @@
+import torch
+import torch.nn as nn
+import re
+
+
+class IdentityMap(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+    @property
+    def config(self):
+        return {"mm_projector_type": "identity"}
+
+
+class SimpleResBlock(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.pre_norm = nn.LayerNorm(channels)
+
+        self.proj = nn.Sequential(
+            nn.Linear(channels, channels), nn.GELU(), nn.Linear(channels, channels)
+        )
+
+    def forward(self, x):
+        x = self.pre_norm(x)
+        return x + self.proj(x)
+
+
+def build_vision_projector(config):
+    projector_type = getattr(config, "mm_projector_type", "linear")
+
+    if projector_type == "linear":
+        return nn.Linear(config.mm_hidden_size, config.hidden_size)
+
+    mlp_gelu_match = re.match(r"^mlp(\d+)x_gelu$", projector_type)
+    if mlp_gelu_match:
+        mlp_depth = int(mlp_gelu_match.group(1))
+        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
+        for _ in range(1, mlp_depth):
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
+        return nn.Sequential(*modules)
+
+    if projector_type == "identity":
+        return IdentityMap()
+
+    raise ValueError(f"Unknown projector type: {projector_type}")

llava-phi/llava_phi/__init__.py

@@ -0,0 +1 @@
+from .model import LlavaPhiForCausalLM

llava-phi/llava_phi/constants.py

@@ -0,0 +1,12 @@
+CONTROLLER_HEART_BEAT_EXPIRATION = 30
+WORKER_HEART_BEAT_INTERVAL = 15
+
+LOGDIR = "."
+
+# Model Constants
+IGNORE_INDEX = -100
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"

llava-phi/llava_phi/conversation.py

@@ -0,0 +1,224 @@
+import dataclasses
+from enum import auto, Enum
+from typing import List, Tuple
+
+
+class SeparatorStyle(Enum):
+    """Different separator style."""
+    SINGLE = auto()
+    TWO = auto()
+    MPT = auto()
+    PLAIN = auto()
+    LLAMA_2 = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that keeps all conversation history."""
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    sep_style: SeparatorStyle = SeparatorStyle.SINGLE
+    sep: str = "###"
+    sep2: str = None
+    version: str = "Unknown"
+
+    skip_next: bool = False
+
+    def get_prompt(self):
+        messages = self.messages
+        if len(messages) > 0 and type(messages[0][1]) is tuple:
+            messages = self.messages.copy()
+            init_role, init_msg = messages[0].copy()
+            init_msg = init_msg[0].replace("<image>", "").strip()
+            if 'mmtag' in self.version:
+                messages[0] = (init_role, init_msg)
+                messages.insert(0, (self.roles[0], "<Image><image></Image>"))
+                messages.insert(1, (self.roles[1], "Received."))
+            else:
+                messages[0] = (init_role, "<image>\n" + init_msg)
+
+        if self.sep_style == SeparatorStyle.SINGLE:
+            ret = self.system + self.sep
+            for role, message in messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + self.sep
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.TWO:
+            seps = [self.sep, self.sep2]
+            ret = self.system + seps[0]
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = self.system
+            for i, (role, message) in enumerate(messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += message + seps[i % 2]
+                else:
+                    ret += ""
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+        return ret
+
+    def append_message(self, role, message):
+        self.messages.append([role, message])
+
+    def get_images(self, return_pil=False):
+        images = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    import base64
+                    from io import BytesIO
+                    from PIL import Image
+                    msg, image, image_process_mode = msg
+                    if image_process_mode == "Pad":
+                        def expand2square(pil_img, background_color=(122, 116, 104)):
+                            width, height = pil_img.size
+                            if width == height:
+                                return pil_img
+                            elif width > height:
+                                result = Image.new(pil_img.mode, (width, width), background_color)
+                                result.paste(pil_img, (0, (width - height) // 2))
+                                return result
+                            else:
+                                result = Image.new(pil_img.mode, (height, height), background_color)
+                                result.paste(pil_img, ((height - width) // 2, 0))
+                                return result
+                        image = expand2square(image)
+                    elif image_process_mode in ["Default", "Crop"]:
+                        pass
+                    elif image_process_mode == "Resize":
+                        image = image.resize((336, 336))
+                    else:
+                        raise ValueError(f"Invalid image_process_mode: {image_process_mode}")
+                    max_hw, min_hw = max(image.size), min(image.size)
+                    aspect_ratio = max_hw / min_hw
+                    max_len, min_len = 800, 400
+                    shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+                    longest_edge = int(shortest_edge * aspect_ratio)
+                    W, H = image.size
+                    if longest_edge != max(image.size):
+                        if H > W:
+                            H, W = longest_edge, shortest_edge
+                        else:
+                            H, W = shortest_edge, longest_edge
+                        image = image.resize((W, H))
+                    if return_pil:
+                        images.append(image)
+                    else:
+                        buffered = BytesIO()
+                        image.save(buffered, format="PNG")
+                        img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+                        images.append(img_b64_str)
+        return images
+
+    def to_gradio_chatbot(self):
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset:]):
+            if i % 2 == 0:
+                if type(msg) is tuple:
+                    import base64
+                    from io import BytesIO
+                    msg, image, image_process_mode = msg
+                    max_hw, min_hw = max(image.size), min(image.size)
+                    aspect_ratio = max_hw / min_hw
+                    max_len, min_len = 800, 400
+                    shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+                    longest_edge = int(shortest_edge * aspect_ratio)
+                    W, H = image.size
+                    if H > W:
+                        H, W = longest_edge, shortest_edge
+                    else:
+                        H, W = shortest_edge, longest_edge
+                    image = image.resize((W, H))
+                    buffered = BytesIO()
+                    image.save(buffered, format="JPEG")
+                    img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+                    img_str = f'<img src="data:image/png;base64,{img_b64_str}" alt="user upload image" />'
+                    msg = img_str + msg.replace('<image>', '').strip()
+                    ret.append([msg, None])
+                else:
+                    ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def copy(self):
+        return Conversation(
+            system=self.system,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            version=self.version)
+
+    def dict(self):
+        if len(self.get_images()) > 0:
+            return {
+                "system": self.system,
+                "roles": self.roles,
+                "messages": [[x, y[0] if type(y) is tuple else y] for x, y in self.messages],
+                "offset": self.offset,
+                "sep": self.sep,
+                "sep2": self.sep2,
+            }
+        return {
+            "system": self.system,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+            "sep": self.sep,
+            "sep2": self.sep2,
+        }
+
+
+conv_phi_v0 = Conversation(
+    system="A chat between a curious user and an artificial intelligence assistant. "
+           "The assistant gives helpful, detailed, and polite answers to the user's questions.",
+    roles=("USER", "ASSISTANT"),
+    version="v0",
+    messages=(),
+    offset=0,
+    sep_style=SeparatorStyle.TWO,
+    sep=" ",
+    sep2="<|endoftext|>",
+)
+
+conv_llava_plain = Conversation(
+    system="",
+    roles=("", ""),
+    messages=(
+    ),
+    offset=0,
+    sep_style=SeparatorStyle.PLAIN,
+    sep="\n",
+)
+
+default_conversation = conv_phi_v0
+conv_templates = {
+    "default": conv_phi_v0,
+    "v0": conv_phi_v0,
+    "phi-2_v0": conv_phi_v0,
+
+    "plain": conv_llava_plain,
+}
+
+
+if __name__ == "__main__":
+    print(default_conversation.get_prompt())

llava-phi/llava_phi/eval/eval_gpt_review.py

@@ -0,0 +1,113 @@
+import argparse
+import json
+import os
+
+import openai
+import tqdm
+import ray
+import time
+
+NUM_SECONDS_TO_SLEEP = 3
+
+@ray.remote(num_cpus=4)
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    print('success!')
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    # parser.add_argument('-a', '--answer')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    ray.init()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    review_file = open(f'{args.output}', 'w')
+
+    js_list = []
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        # if idx == 1:
+        #     break
+
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        category = json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            rule = rule_dict['default']
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        js_list.append({
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1['answer_id'],
+            'answer2_id': ans2['answer_id'],
+            'category': category})
+        idx += 1
+        handles.append(get_eval.remote(content, args.max_tokens))
+        # To avoid the rate limit set by OpenAI
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    reviews = ray.get(handles)
+    for idx, review in enumerate(reviews):
+        scores = parse_score(review)
+        js_list[idx]['content'] = review
+        js_list[idx]['tuple'] = scores
+        review_file.write(json.dumps(js_list[idx]) + '\n')
+    review_file.close()

llava-phi/llava_phi/eval/eval_gpt_review_bench.py

@@ -0,0 +1,121 @@
+import argparse
+import json
+import os
+
+import openai
+import time
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4-0314',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        inst = image_to_context[ques['image']]
+
+        if isinstance(inst['caption'], list):
+            cap_str = '\n'.join(inst['caption'])
+        else:
+            cap_str = inst['caption']
+
+        category = 'llava_bench_' + json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f"Visual QA category not found in rule file: {category}."
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans2['answer_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

llava-phi/llava_phi/eval/eval_gpt_review_visual.py

@@ -0,0 +1,118 @@
+import argparse
+import json
+import os
+
+import openai
+import time
+
+NUM_SECONDS_TO_SLEEP = 0.5
+
+
+def get_eval(content: str, max_tokens: int):
+    while True:
+        try:
+            response = openai.ChatCompletion.create(
+                model='gpt-4-0314',
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful and precise assistant for checking the quality of the answer.'
+                }, {
+                    'role': 'user',
+                    'content': content,
+                }],
+                temperature=0.2,  # TODO: figure out which temperature is best for evaluation
+                max_tokens=max_tokens,
+            )
+            break
+        except openai.error.RateLimitError:
+            pass
+        except Exception as e:
+            print(e)
+        time.sleep(NUM_SECONDS_TO_SLEEP)
+
+    return response['choices'][0]['message']['content']
+
+
+def parse_score(review):
+    try:
+        score_pair = review.split('\n')[0]
+        score_pair = score_pair.replace(',', ' ')
+        sp = score_pair.split(' ')
+        if len(sp) == 2:
+            return [float(sp[0]), float(sp[1])]
+        else:
+            print('error', review)
+            return [-1, -1]
+    except Exception as e:
+        print(e)
+        print('error', review)
+        return [-1, -1]
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-c', '--context')
+    parser.add_argument('-a', '--answer-list', nargs='+', default=[])
+    parser.add_argument('-r', '--rule')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    f_q = open(os.path.expanduser(args.question))
+    f_ans1 = open(os.path.expanduser(args.answer_list[0]))
+    f_ans2 = open(os.path.expanduser(args.answer_list[1]))
+    rule_dict = json.load(open(os.path.expanduser(args.rule), 'r'))
+
+    if os.path.isfile(os.path.expanduser(args.output)):
+        cur_reviews = [json.loads(line) for line in open(os.path.expanduser(args.output))]
+    else:
+        cur_reviews = []
+
+    review_file = open(f'{args.output}', 'a')
+
+    context_list = [json.loads(line) for line in open(os.path.expanduser(args.context))]
+    image_to_context = {context['image']: context for context in context_list}
+
+    handles = []
+    idx = 0
+    for ques_js, ans1_js, ans2_js in zip(f_q, f_ans1, f_ans2):
+        ques = json.loads(ques_js)
+        ans1 = json.loads(ans1_js)
+        ans2 = json.loads(ans2_js)
+
+        inst = image_to_context[ques['image']]
+        cap_str = '\n'.join(inst['captions'])
+        box_str = '\n'.join([f'{instance["category"]}: {instance["bbox"]}' for instance in inst['instances']])
+
+        category = json.loads(ques_js)['category']
+        if category in rule_dict:
+            rule = rule_dict[category]
+        else:
+            assert False, f"Visual QA category not found in rule file: {category}."
+        prompt = rule['prompt']
+        role = rule['role']
+        content = (f'[Context]\n{cap_str}\n\n{box_str}\n\n'
+                   f'[Question]\n{ques["text"]}\n\n'
+                   f'[{role} 1]\n{ans1["text"]}\n\n[End of {role} 1]\n\n'
+                   f'[{role} 2]\n{ans2["text"]}\n\n[End of {role} 2]\n\n'
+                   f'[System]\n{prompt}\n\n')
+        cur_js = {
+            'id': idx+1,
+            'question_id': ques['question_id'],
+            'answer1_id': ans1.get('answer_id', ans1['question_id']),
+            'answer2_id': ans2.get('answer_id', ans2['answer_id']),
+            'category': category
+        }
+        if idx >= len(cur_reviews):
+            review = get_eval(content, args.max_tokens)
+            scores = parse_score(review)
+            cur_js['content'] = review
+            cur_js['tuple'] = scores
+            review_file.write(json.dumps(cur_js) + '\n')
+            review_file.flush()
+        else:
+            print(f'Skipping {idx} as we already have it.')
+        idx += 1
+        print(idx)
+    review_file.close()

llava-phi/llava_phi/eval/eval_pope.py

@@ -0,0 +1,81 @@
+import os
+import json
+import argparse
+
+def eval_pope(answers, label_file):
+    label_list = [json.loads(q)['label'] for q in open(label_file, 'r')]
+
+    for answer in answers:
+        text = answer['text']
+
+        # Only keep the first sentence
+        if text.find('.') != -1:
+            text = text.split('.')[0]
+
+        text = text.replace(',', '')
+        words = text.split(' ')
+        if 'No' in words or 'not' in words or 'no' in words:
+            answer['text'] = 'no'
+        else:
+            answer['text'] = 'yes'
+
+    for i in range(len(label_list)):
+        if label_list[i] == 'no':
+            label_list[i] = 0
+        else:
+            label_list[i] = 1
+
+    pred_list = []
+    for answer in answers:
+        if answer['text'] == 'no':
+            pred_list.append(0)
+        else:
+            pred_list.append(1)
+
+    pos = 1
+    neg = 0
+    yes_ratio = pred_list.count(1) / len(pred_list)
+
+    TP, TN, FP, FN = 0, 0, 0, 0
+    for pred, label in zip(pred_list, label_list):
+        if pred == pos and label == pos:
+            TP += 1
+        elif pred == pos and label == neg:
+            FP += 1
+        elif pred == neg and label == neg:
+            TN += 1
+        elif pred == neg and label == pos:
+            FN += 1
+
+    print('TP\tFP\tTN\tFN\t')
+    print('{}\t{}\t{}\t{}'.format(TP, FP, TN, FN))
+
+    precision = float(TP) / float(TP + FP)
+    recall = float(TP) / float(TP + FN)
+    f1 = 2*precision*recall / (precision + recall)
+    acc = (TP + TN) / (TP + TN + FP + FN)
+    print('Accuracy: {}'.format(acc))
+    print('Precision: {}'.format(precision))
+    print('Recall: {}'.format(recall))
+    print('F1 score: {}'.format(f1))
+    print('Yes ratio: {}'.format(yes_ratio))
+    print('%.3f, %.3f, %.3f, %.3f, %.3f' % (f1, acc, precision, recall, yes_ratio) )
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--annotation-dir", type=str)
+    parser.add_argument("--question-file", type=str)
+    parser.add_argument("--result-file", type=str)
+    args = parser.parse_args()
+
+    questions = [json.loads(line) for line in open(args.question_file)]
+    questions = {question['question_id']: question for question in questions}
+    answers = [json.loads(q) for q in open(args.result_file)]
+    for file in os.listdir(args.annotation_dir):
+        assert file.startswith('coco_pope_')
+        assert file.endswith('.json')
+        category = file[10:-5]
+        cur_answers = [x for x in answers if questions[x['question_id']]['category'] == category]
+        print('Category: {}, # samples: {}'.format(category, len(cur_answers)))
+        eval_pope(cur_answers, os.path.join(args.annotation_dir, file))
+        print("====================================")

llava-phi/llava_phi/eval/eval_science_qa.py

@@ -0,0 +1,114 @@
+import argparse
+import json
+import os
+import re
+import random
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--result-file', type=str)
+    parser.add_argument('--output-file', type=str)
+    parser.add_argument('--output-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return -1
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    predictions = [json.loads(line) for line in open(args.result_file)]
+    predictions = {pred['question_id']: pred for pred in predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    results = {'correct': [], 'incorrect': []}
+    sqa_results = {}
+    sqa_results['acc'] = None
+    sqa_results['correct'] = None
+    sqa_results['count'] = None
+    sqa_results['results'] = {}
+    sqa_results['outputs'] = {}
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in predictions:
+            pred = {'text': 'FAILED', 'prompt': 'Unknown'}
+            pred_text = 'FAILED'
+        else:
+            pred = predictions[prob_id]
+            pred_text = pred['text']
+
+        if pred_text in args.options:
+            answer = pred_text
+        elif len(pred_text) >= 3 and pred_text[0] in args.options and pred_text[1:3] == ". ":
+            answer = pred_text[0]
+        else:
+            pattern = re.compile(r'The answer is ([A-Z]).')
+            res = pattern.findall(pred_text)
+            if len(res) == 1:
+                answer = res[0]  # 'A', 'B', ...
+            else:
+                answer = "FAILED"
+
+        pred_idx = get_pred_idx(answer, prob['choices'], args.options)
+
+        analysis = {
+            'question_id': prob_id,
+            'parsed_ans': answer,
+            'ground_truth': args.options[prob['answer']],
+            'question': pred['prompt'],
+            'pred': pred_text,
+            'is_multimodal': '<image>' in pred['prompt'],
+        }
+
+        sqa_results['results'][prob_id] = get_pred_idx(answer, prob['choices'], args.options)
+        sqa_results['outputs'][prob_id] = pred_text
+
+        if pred_idx == prob['answer']:
+            results['correct'].append(analysis)
+        else:
+            results['incorrect'].append(analysis)
+
+    correct = len(results['correct'])
+    total = len(results['correct']) + len(results['incorrect'])
+
+    ###### IMG ######
+    multimodal_correct = len([x for x in results['correct'] if x['is_multimodal']])
+    multimodal_incorrect = len([x for x in results['incorrect'] if x['is_multimodal']])
+    multimodal_total = multimodal_correct + multimodal_incorrect
+    ###### IMG ######
+
+    print(f'Total: {total}, Correct: {correct}, Accuracy: {correct / total * 100:.2f}%, IMG-Accuracy: {multimodal_correct / multimodal_total * 100:.2f}%')
+
+    sqa_results['acc'] = correct / total * 100
+    sqa_results['correct'] = correct
+    sqa_results['count'] = total
+
+    with open(args.output_file, 'w') as f:
+        json.dump(results, f, indent=2)
+    with open(args.output_result, 'w') as f:
+        json.dump(sqa_results, f, indent=2)

llava-phi/llava_phi/eval/eval_science_qa_gpt4.py

@@ -0,0 +1,104 @@
+import argparse
+import json
+import os
+import re
+import random
+from collections import defaultdict
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--gpt4-result', type=str)
+    parser.add_argument('--our-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    our_predictions = [json.loads(line) for line in open(args.our_result)]
+    our_predictions = {pred['question_id']: pred for pred in our_predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    gpt4_predictions = json.load(open(args.gpt4_result))['outputs']
+
+    results = defaultdict(lambda: 0)
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in our_predictions:
+            continue
+        if prob_id not in gpt4_predictions:
+            continue
+        our_pred = our_predictions[prob_id]['text']
+        gpt4_pred = gpt4_predictions[prob_id]
+
+        pattern = re.compile(r'The answer is ([A-Z]).')
+        our_res = pattern.findall(our_pred)
+        if len(our_res) == 1:
+            our_answer = our_res[0]  # 'A', 'B', ...
+        else:
+            our_answer = "FAILED"
+        gpt4_res = pattern.findall(gpt4_pred)
+        if len(gpt4_res) == 1:
+            gpt4_answer = gpt4_res[0]  # 'A', 'B', ...
+        else:
+            gpt4_answer = "FAILED"
+
+        our_pred_idx = get_pred_idx(our_answer, prob['choices'], args.options)
+        gpt4_pred_idx = get_pred_idx(gpt4_answer, prob['choices'], args.options)
+
+        if gpt4_answer == 'FAILED':
+            results['gpt4_failed'] += 1
+            # continue
+            gpt4_pred_idx = our_pred_idx
+            # if our_pred_idx != prob['answer']:
+            #     print(our_predictions[prob_id]['prompt'])
+            #     print('-----------------')
+            #     print(f'LECTURE: {prob["lecture"]}')
+            #     print(f'SOLUTION: {prob["solution"]}')
+            #     print('=====================')
+        else:
+            # continue
+            pass
+        # gpt4_pred_idx = our_pred_idx
+
+        if gpt4_pred_idx == prob['answer']:
+            results['correct'] += 1
+        else:
+            results['incorrect'] += 1
+
+
+        if gpt4_pred_idx == prob['answer'] or our_pred_idx == prob['answer']:
+            results['correct_upperbound'] += 1
+
+    correct = results['correct']
+    total = results['correct'] + results['incorrect']
+    print(f'Total: {total}, Correct: {correct}, Accuracy: {correct / total * 100:.2f}%')
+    print(f'Total: {total}, Correct (upper): {results["correct_upperbound"]}, Accuracy: {results["correct_upperbound"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4 NO-ANS (RANDOM): {results["gpt4_failed"]}, Percentage: {results["gpt4_failed"] / total * 100:.2f}%')
+

llava-phi/llava_phi/eval/eval_science_qa_gpt4_requery.py

@@ -0,0 +1,149 @@
+import argparse
+import json
+import os
+import re
+import random
+from collections import defaultdict
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--base-dir', type=str)
+    parser.add_argument('--gpt4-result', type=str)
+    parser.add_argument('--requery-result', type=str)
+    parser.add_argument('--our-result', type=str)
+    parser.add_argument('--output-result', type=str)
+    parser.add_argument('--split', type=str, default='test')
+    parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
+    return parser.parse_args()
+
+
+def convert_caps(results):
+    fakecaps = []
+    for result in results:
+        image_id = result['question_id']
+        caption = result['text']
+        fakecaps.append({"image_id": int(image_id), "caption": caption})
+    return fakecaps
+
+
+def get_pred_idx(prediction, choices, options):
+    """
+    Get the index (e.g. 2) from the prediction (e.g. 'C')
+    """
+    if prediction in options[:len(choices)]:
+        return options.index(prediction)
+    else:
+        return random.choice(range(len(choices)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    base_dir = args.base_dir
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+    our_predictions = [json.loads(line) for line in open(args.our_result)]
+    our_predictions = {pred['question_id']: pred for pred in our_predictions}
+    split_problems = {idx: problems[idx] for idx in split_indices}
+
+    requery_predictions = [json.loads(line) for line in open(args.requery_result)]
+    requery_predictions = {pred['question_id']: pred for pred in requery_predictions}
+
+    gpt4_predictions = json.load(open(args.gpt4_result))['outputs']
+
+    results = defaultdict(lambda: 0)
+
+    sqa_results = {}
+    sqa_results['acc'] = None
+    sqa_results['correct'] = None
+    sqa_results['count'] = None
+    sqa_results['results'] = {}
+    sqa_results['outputs'] = {}
+
+    for prob_id, prob in split_problems.items():
+        if prob_id not in our_predictions:
+            assert False
+        if prob_id not in gpt4_predictions:
+            assert False
+        our_pred = our_predictions[prob_id]['text']
+        gpt4_pred = gpt4_predictions[prob_id]
+        if prob_id not in requery_predictions:
+            results['missing_requery'] += 1
+            requery_pred = "MISSING"
+        else:
+            requery_pred = requery_predictions[prob_id]['text']
+
+        pattern = re.compile(r'The answer is ([A-Z]).')
+        our_res = pattern.findall(our_pred)
+        if len(our_res) == 1:
+            our_answer = our_res[0]  # 'A', 'B', ...
+        else:
+            our_answer = "FAILED"
+
+        requery_res = pattern.findall(requery_pred)
+        if len(requery_res) == 1:
+            requery_answer = requery_res[0]  # 'A', 'B', ...
+        else:
+            requery_answer = "FAILED"
+
+        gpt4_res = pattern.findall(gpt4_pred)
+        if len(gpt4_res) == 1:
+            gpt4_answer = gpt4_res[0]  # 'A', 'B', ...
+        else:
+            gpt4_answer = "FAILED"
+
+        our_pred_idx = get_pred_idx(our_answer, prob['choices'], args.options)
+        gpt4_pred_idx = get_pred_idx(gpt4_answer, prob['choices'], args.options)
+        requery_pred_idx = get_pred_idx(requery_answer, prob['choices'], args.options)
+
+        results['total'] += 1
+
+        if gpt4_answer == 'FAILED':
+            results['gpt4_failed'] += 1
+            if gpt4_pred_idx == prob['answer']:
+                results['gpt4_correct'] += 1
+            if our_pred_idx == prob['answer']:
+                results['gpt4_ourvisual_correct'] += 1
+        elif gpt4_pred_idx == prob['answer']:
+            results['gpt4_correct'] += 1
+            results['gpt4_ourvisual_correct'] += 1
+
+        if our_pred_idx == prob['answer']:
+            results['our_correct'] += 1
+
+        if requery_answer == 'FAILED':
+            sqa_results['results'][prob_id] = our_pred_idx
+            if our_pred_idx == prob['answer']:
+                results['requery_correct'] += 1
+        else:
+            sqa_results['results'][prob_id] = requery_pred_idx
+            if requery_pred_idx == prob['answer']:
+                results['requery_correct'] += 1
+            else:
+                print(f"""
+Question ({args.options[prob['answer']]}): {our_predictions[prob_id]['prompt']}
+Our ({our_answer}): {our_pred}
+GPT-4 ({gpt4_answer}): {gpt4_pred}
+Requery ({requery_answer}): {requery_pred}
+print("=====================================")
+""")
+
+        if gpt4_pred_idx == prob['answer'] or our_pred_idx == prob['answer']:
+            results['correct_upperbound'] += 1
+
+    total = results['total']
+    print(f'Total: {total}, Our-Correct: {results["our_correct"]}, Accuracy: {results["our_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4-Correct: {results["gpt4_correct"]}, Accuracy: {results["gpt4_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4 NO-ANS (RANDOM): {results["gpt4_failed"]}, Percentage: {results["gpt4_failed"] / total * 100:.2f}%')
+    print(f'Total: {total}, GPT-4-OursVisual-Correct: {results["gpt4_ourvisual_correct"]}, Accuracy: {results["gpt4_ourvisual_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, Requery-Correct: {results["requery_correct"]}, Accuracy: {results["requery_correct"] / total * 100:.2f}%')
+    print(f'Total: {total}, Correct upper: {results["correct_upperbound"]}, Accuracy: {results["correct_upperbound"] / total * 100:.2f}%')
+
+    sqa_results['acc'] = results["requery_correct"] / total * 100
+    sqa_results['correct'] = results["requery_correct"]
+    sqa_results['count'] = total
+
+    with open(args.output_result, 'w') as f:
+        json.dump(sqa_results, f, indent=2)
+

llava-phi/llava_phi/eval/eval_textvqa.py

@@ -0,0 +1,65 @@
+import os
+import argparse
+import json
+import re
+
+from llava_phi.eval.m4c_evaluator import TextVQAAccuracyEvaluator
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--annotation-file', type=str)
+    parser.add_argument('--result-file', type=str)
+    parser.add_argument('--result-dir', type=str)
+    return parser.parse_args()
+
+
+def prompt_processor(prompt):
+    if prompt.startswith('OCR tokens: '):
+        pattern = r"Question: (.*?) Short answer:"
+        match = re.search(pattern, prompt, re.DOTALL)
+        question = match.group(1)
+    elif 'Reference OCR token: ' in prompt and len(prompt.split('\n')) == 3:
+        if prompt.startswith('Reference OCR token:'):
+            question = prompt.split('\n')[1]
+        else:
+            question = prompt.split('\n')[0]
+    elif len(prompt.split('\n')) == 2:
+        question = prompt.split('\n')[0]
+    else:
+        assert False
+
+    return question.lower()
+
+
+def eval_single(annotation_file, result_file):
+    experiment_name = os.path.splitext(os.path.basename(result_file))[0]
+    print(experiment_name)
+    annotations = json.load(open(annotation_file))['data']
+    annotations = {(annotation['image_id'], annotation['question'].lower()): annotation for annotation in annotations}
+    results = [json.loads(line) for line in open(result_file)]
+
+    pred_list = []
+    for result in results:
+        annotation = annotations[(result['question_id'], prompt_processor(result['prompt']))]
+        pred_list.append({
+            "pred_answer": result['text'],
+            "gt_answers": annotation['answers'],
+        })
+
+    evaluator = TextVQAAccuracyEvaluator()
+    print('Samples: {}\nAccuracy: {:.2f}%\n'.format(len(pred_list), 100. * evaluator.eval_pred_list(pred_list)))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    if args.result_file is not None:
+        eval_single(args.annotation_file, args.result_file)
+
+    if args.result_dir is not None:
+        for result_file in sorted(os.listdir(args.result_dir)):
+            if not result_file.endswith('.jsonl'):
+                print(f'Skipping {result_file}')
+                continue
+            eval_single(args.annotation_file, os.path.join(args.result_dir, result_file))

llava-phi/llava_phi/eval/m4c_evaluator.py

@@ -0,0 +1,334 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import re
+
+from tqdm import tqdm
+
+
+class EvalAIAnswerProcessor:
+    """
+    Processes an answer similar to Eval AI
+        copied from
+        https://github.com/facebookresearch/mmf/blob/c46b3b3391275b4181567db80943473a89ab98ab/pythia/tasks/processors.py#L897
+    """
+
+    CONTRACTIONS = {
+        "aint": "ain't",
+        "arent": "aren't",
+        "cant": "can't",
+        "couldve": "could've",
+        "couldnt": "couldn't",
+        "couldn'tve": "couldn't've",
+        "couldnt've": "couldn't've",
+        "didnt": "didn't",
+        "doesnt": "doesn't",
+        "dont": "don't",
+        "hadnt": "hadn't",
+        "hadnt've": "hadn't've",
+        "hadn'tve": "hadn't've",
+        "hasnt": "hasn't",
+        "havent": "haven't",
+        "hed": "he'd",
+        "hed've": "he'd've",
+        "he'dve": "he'd've",
+        "hes": "he's",
+        "howd": "how'd",
+        "howll": "how'll",
+        "hows": "how's",
+        "Id've": "I'd've",
+        "I'dve": "I'd've",
+        "Im": "I'm",
+        "Ive": "I've",
+        "isnt": "isn't",
+        "itd": "it'd",
+        "itd've": "it'd've",
+        "it'dve": "it'd've",
+        "itll": "it'll",
+        "let's": "let's",
+        "maam": "ma'am",
+        "mightnt": "mightn't",
+        "mightnt've": "mightn't've",
+        "mightn'tve": "mightn't've",
+        "mightve": "might've",
+        "mustnt": "mustn't",
+        "mustve": "must've",
+        "neednt": "needn't",
+        "notve": "not've",
+        "oclock": "o'clock",
+        "oughtnt": "oughtn't",
+        "ow's'at": "'ow's'at",
+        "'ows'at": "'ow's'at",
+        "'ow'sat": "'ow's'at",
+        "shant": "shan't",
+        "shed've": "she'd've",
+        "she'dve": "she'd've",
+        "she's": "she's",
+        "shouldve": "should've",
+        "shouldnt": "shouldn't",
+        "shouldnt've": "shouldn't've",
+        "shouldn'tve": "shouldn't've",
+        "somebody'd": "somebodyd",
+        "somebodyd've": "somebody'd've",
+        "somebody'dve": "somebody'd've",
+        "somebodyll": "somebody'll",
+        "somebodys": "somebody's",
+        "someoned": "someone'd",
+        "someoned've": "someone'd've",
+        "someone'dve": "someone'd've",
+        "someonell": "someone'll",
+        "someones": "someone's",
+        "somethingd": "something'd",
+        "somethingd've": "something'd've",
+        "something'dve": "something'd've",
+        "somethingll": "something'll",
+        "thats": "that's",
+        "thered": "there'd",
+        "thered've": "there'd've",
+        "there'dve": "there'd've",
+        "therere": "there're",
+        "theres": "there's",
+        "theyd": "they'd",
+        "theyd've": "they'd've",
+        "they'dve": "they'd've",
+        "theyll": "they'll",
+        "theyre": "they're",
+        "theyve": "they've",
+        "twas": "'twas",
+        "wasnt": "wasn't",
+        "wed've": "we'd've",
+        "we'dve": "we'd've",
+        "weve": "we've",
+        "werent": "weren't",
+        "whatll": "what'll",
+        "whatre": "what're",
+        "whats": "what's",
+        "whatve": "what've",
+        "whens": "when's",
+        "whered": "where'd",
+        "wheres": "where's",
+        "whereve": "where've",
+        "whod": "who'd",
+        "whod've": "who'd've",
+        "who'dve": "who'd've",
+        "wholl": "who'll",
+        "whos": "who's",
+        "whove": "who've",
+        "whyll": "why'll",
+        "whyre": "why're",
+        "whys": "why's",
+        "wont": "won't",
+        "wouldve": "would've",
+        "wouldnt": "wouldn't",
+        "wouldnt've": "wouldn't've",
+        "wouldn'tve": "wouldn't've",
+        "yall": "y'all",
+        "yall'll": "y'all'll",
+        "y'allll": "y'all'll",
+        "yall'd've": "y'all'd've",
+        "y'alld've": "y'all'd've",
+        "y'all'dve": "y'all'd've",
+        "youd": "you'd",
+        "youd've": "you'd've",
+        "you'dve": "you'd've",
+        "youll": "you'll",
+        "youre": "you're",
+        "youve": "you've",
+    }
+
+    NUMBER_MAP = {
+        "none": "0",
+        "zero": "0",
+        "one": "1",
+        "two": "2",
+        "three": "3",
+        "four": "4",
+        "five": "5",
+        "six": "6",
+        "seven": "7",
+        "eight": "8",
+        "nine": "9",
+        "ten": "10",
+    }
+    ARTICLES = ["a", "an", "the"]
+    PERIOD_STRIP = re.compile(r"(?!<=\d)(\.)(?!\d)")
+    COMMA_STRIP = re.compile(r"(?<=\d)(\,)+(?=\d)")
+    PUNCTUATIONS = [
+        ";",
+        r"/",
+        "[",
+        "]",
+        '"',
+        "{",
+        "}",
+        "(",
+        ")",
+        "=",
+        "+",
+        "\\",
+        "_",
+        "-",
+        ">",
+        "<",
+        "@",
+        "`",
+        ",",
+        "?",
+        "!",
+    ]
+
+    def __init__(self, *args, **kwargs):
+        pass
+
+    def word_tokenize(self, word):
+        word = word.lower()
+        word = word.replace(",", "").replace("?", "").replace("'s", " 's")
+        return word.strip()
+
+    def process_punctuation(self, in_text):
+        out_text = in_text
+        for p in self.PUNCTUATIONS:
+            if (p + " " in in_text or " " + p in in_text) or (
+                re.search(self.COMMA_STRIP, in_text) is not None
+            ):
+                out_text = out_text.replace(p, "")
+            else:
+                out_text = out_text.replace(p, " ")
+        out_text = self.PERIOD_STRIP.sub("", out_text, re.UNICODE)
+        return out_text
+
+    def process_digit_article(self, in_text):
+        out_text = []
+        temp_text = in_text.lower().split()
+        for word in temp_text:
+            word = self.NUMBER_MAP.setdefault(word, word)
+            if word not in self.ARTICLES:
+                out_text.append(word)
+            else:
+                pass
+        for word_id, word in enumerate(out_text):
+            if word in self.CONTRACTIONS:
+                out_text[word_id] = self.CONTRACTIONS[word]
+        out_text = " ".join(out_text)
+        return out_text
+
+    def __call__(self, item):
+        item = self.word_tokenize(item)
+        item = item.replace("\n", " ").replace("\t", " ").strip()
+        item = self.process_punctuation(item)
+        item = self.process_digit_article(item)
+        return item
+
+
+class TextVQAAccuracyEvaluator:
+    def __init__(self):
+        self.answer_processor = EvalAIAnswerProcessor()
+
+    def _compute_answer_scores(self, raw_answers):
+        """
+        compute the accuracy (soft score) of human answers
+        """
+        answers = [self.answer_processor(a) for a in raw_answers]
+        assert len(answers) == 10
+        gt_answers = list(enumerate(answers))
+        unique_answers = set(answers)
+        unique_answer_scores = {}
+
+        for unique_answer in unique_answers:
+            accs = []
+            for gt_answer in gt_answers:
+                other_answers = [item for item in gt_answers if item != gt_answer]
+                matching_answers = [
+                    item for item in other_answers if item[1] == unique_answer
+                ]
+                acc = min(1, float(len(matching_answers)) / 3)
+                accs.append(acc)
+            unique_answer_scores[unique_answer] = sum(accs) / len(accs)
+
+        return unique_answer_scores
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in tqdm(pred_list):
+            pred_answer = self.answer_processor(entry["pred_answer"])
+            unique_answer_scores = self._compute_answer_scores(entry["gt_answers"])
+            score = unique_answer_scores.get(pred_answer, 0.0)
+            pred_scores.append(score)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class STVQAAccuracyEvaluator:
+    def __init__(self):
+        self.answer_processor = EvalAIAnswerProcessor()
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in pred_list:
+            pred_answer = self.answer_processor(entry["pred_answer"])
+            gts = [self.answer_processor(a) for a in entry["gt_answers"]]
+            score = 1.0 if pred_answer in gts else 0.0
+            pred_scores.append(score)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class STVQAANLSEvaluator:
+    def __init__(self):
+        import editdistance  # install with `pip install editdistance`
+
+        self.get_edit_distance = editdistance.eval
+
+    def get_anls(self, s1, s2):
+        s1 = s1.lower().strip()
+        s2 = s2.lower().strip()
+        iou = 1 - self.get_edit_distance(s1, s2) / max(len(s1), len(s2))
+        anls = iou if iou >= 0.5 else 0.0
+        return anls
+
+    def eval_pred_list(self, pred_list):
+        pred_scores = []
+        for entry in pred_list:
+            anls = max(
+                self.get_anls(entry["pred_answer"], gt) for gt in entry["gt_answers"]
+            )
+            pred_scores.append(anls)
+
+        accuracy = sum(pred_scores) / len(pred_scores)
+        return accuracy
+
+
+class TextCapsBleu4Evaluator:
+    def __init__(self):
+        # The following script requires Java 1.8.0 and pycocotools installed.
+        # The pycocoevalcap can be installed with pip as
+        # pip install git+https://github.com/ronghanghu/coco-caption.git@python23
+        # Original pycocoevalcap code is at https://github.com/tylin/coco-caption
+        # but has no python3 support yet.
+        try:
+            from pycocoevalcap.bleu.bleu import Bleu
+            from pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
+        except ModuleNotFoundError:
+            print(
+                "Please install pycocoevalcap module using "
+                "pip install git+https://github.com/ronghanghu/coco-caption.git@python23"  # noqa
+            )
+            raise
+
+        self.tokenizer = PTBTokenizer()
+        self.scorer = Bleu(4)
+
+    def eval_pred_list(self, pred_list):
+        # Create reference and hypotheses captions.
+        gts = {}
+        res = {}
+        for idx, entry in enumerate(pred_list):
+            gts[idx] = [{"caption": a} for a in entry["gt_answers"]]
+            res[idx] = [{"caption": entry["pred_answer"]}]
+
+        gts = self.tokenizer.tokenize(gts)
+        res = self.tokenizer.tokenize(res)
+        score, _ = self.scorer.compute_score(gts, res)
+
+        bleu4 = score[3]  # score is (Bleu-1, Bleu-2, Bleu-3, Bleu-4)
+        return bleu4

llava-phi/llava_phi/eval/model_qa.py

@@ -0,0 +1,88 @@
+import argparse
+from transformers import AutoTokenizer, AutoModelForCausalLM, StoppingCriteria
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava_phi.conversation import default_conversation
+from llava_phi.utils import disable_torch_init
+
+
+# new stopping implementation
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.tokenizer = tokenizer
+        self.start_len = None
+        self.input_ids = input_ids
+
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        if self.start_len is None:
+            self.start_len = self.input_ids.shape[1]
+        else:
+            outputs = self.tokenizer.batch_decode(output_ids[:, self.start_len:], skip_special_tokens=True)[0]
+            for keyword in self.keywords:
+                if keyword in outputs:
+                    return True
+        return False
+
+
+@torch.inference_mode()
+def eval_model(model_name, questions_file, answers_file):
+    # Model
+    disable_torch_init()
+    model_name = os.path.expanduser(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
+    model = AutoModelForCausalLM.from_pretrained(model_name,
+        torch_dtype=torch.float16).cuda()
+
+
+    ques_file = open(os.path.expanduser(questions_file), "r")
+    ans_file = open(os.path.expanduser(answers_file), "w")
+    for i, line in enumerate(tqdm(ques_file)):
+        idx = json.loads(line)["question_id"]
+        qs = json.loads(line)["text"]
+        cat = json.loads(line)["category"]
+        conv = default_conversation.copy()
+        conv.append_message(conv.roles[0], qs)
+        prompt = conv.get_prompt()
+        inputs = tokenizer([prompt])
+        input_ids = torch.as_tensor(inputs.input_ids).cuda()
+        # stopping_criteria = KeywordsStoppingCriteria([conv.sep], tokenizer, input_ids)
+        output_ids = model.generate(
+            input_ids,
+            do_sample=True,
+            use_cache=True,
+            temperature=0.7,
+            max_new_tokens=1024,
+            eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+            pad_token_id=tokenizer.eos_token_id,  # Pad token
+            # stopping_criteria=[stopping_criteria]
+        )
+        outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
+        try:
+            index = outputs.index(conv.sep, len(prompt))
+        except ValueError:
+            outputs += conv.sep
+            index = outputs.index(conv.sep, len(prompt))
+
+        outputs = outputs[len(prompt) + len(conv.roles[1]) + 2:index].strip()
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-name", type=str, default="facebook/opt-350m")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    args = parser.parse_args()
+
+    eval_model(args.model_name, args.question_file, args.answers_file)

llava-phi/llava_phi/eval/model_vqa.py

@@ -0,0 +1,115 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+    for line in tqdm(questions):
+        idx = line["question_id"]
+        image_file = line["image"]
+        qs = line["text"]
+        cur_prompt = qs
+        if model.config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+        image = Image.open(os.path.join(args.image_folder, image_file))
+        image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor.unsqueeze(0).cuda(),
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                num_beams=args.num_beams,
+                # no_repeat_ngram_size=3,
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.eos_token_id,  # Pad token
+                max_new_tokens=1024,
+                use_cache=True
+            )
+
+        input_token_len = input_ids.shape[1]
+        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+        if n_diff_input_output > 0:
+            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[:-len(stop_str)]
+        outputs = outputs.strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "image_id": image_file,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    args = parser.parse_args()
+
+    eval_model(args)

llava-phi/llava_phi/eval/model_vqa_loader.py

@@ -0,0 +1,144 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
+from torch.utils.data import Dataset, DataLoader
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+# Custom dataset class
+class CustomDataset(Dataset):
+    def __init__(self, questions, image_folder, tokenizer, image_processor, model_config):
+        self.questions = questions
+        self.image_folder = image_folder
+        self.tokenizer = tokenizer
+        self.image_processor = image_processor
+        self.model_config = model_config
+
+    def __getitem__(self, index):
+        line = self.questions[index]
+        image_file = line["image"]
+        qs = line["text"]
+        if self.model_config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+        image = Image.open(os.path.join(self.image_folder, image_file)).convert('RGB')
+        image_tensor = process_images([image], self.image_processor, self.model_config)[0]
+
+        input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt')
+
+        return input_ids, image_tensor
+
+    def __len__(self):
+        return len(self.questions)
+
+
+# DataLoader
+def create_data_loader(questions, image_folder, tokenizer, image_processor, model_config, batch_size=1, num_workers=4):
+    assert batch_size == 1, "batch_size must be 1"
+    dataset = CustomDataset(questions, image_folder, tokenizer, image_processor, model_config)
+    data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False)
+    return data_loader
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+    
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    data_loader = create_data_loader(questions, args.image_folder, tokenizer, image_processor, model.config)
+
+    for (input_ids, image_tensor), line in tqdm(zip(data_loader, questions), total=len(questions)):
+        idx = line["question_id"]
+        cur_prompt = line["text"]
+        
+        stop_str = conv_templates[args.conv_mode].sep if conv_templates[args.conv_mode].sep_style != SeparatorStyle.TWO else conv_templates[args.conv_mode].sep2
+        input_ids = input_ids.to(device='cuda', non_blocking=True)
+        
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor.to(device='cuda', non_blocking=True),
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                # no_repeat_ngram_size=3,
+                num_beams=args.num_beams,
+                max_new_tokens=128,
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.eos_token_id,  # Pad token
+                use_cache=True
+            )
+
+        input_token_len = input_ids.shape[1]
+        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+        
+        if n_diff_input_output > 0:
+            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[:-len(stop_str)]
+        outputs = outputs.strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        # ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="v0")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    args = parser.parse_args()
+
+    eval_model(args)

llava-phi/llava_phi/eval/model_vqa_mmbench.py

@@ -0,0 +1,173 @@
+import argparse
+import torch
+import os
+import json
+import pandas as pd
+from tqdm import tqdm
+import shortuuid
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, process_images, load_image_from_base64, get_model_name_from_path
+
+from PIL import Image
+import math
+
+
+all_options = ['A', 'B', 'C', 'D']
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def is_none(value):
+    if value is None:
+        return True
+    if type(value) is float and math.isnan(value):
+        return True
+    if type(value) is str and value.lower() == 'nan':
+        return True
+    if type(value) is str and value.lower() == 'none':
+        return True
+    return False
+
+def get_options(row, options):
+    parsed_options = []
+    for option in options:
+        option_value = row[option]
+        if is_none(option_value):
+            break
+        parsed_options.append(option_value)
+    return parsed_options
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = pd.read_table(os.path.expanduser(args.question_file))
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+
+    if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
+        args.conv_mode = args.conv_mode + '_mmtag'
+        print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
+
+    for index, row in tqdm(questions.iterrows(), total=len(questions)):
+        options = get_options(row, all_options)
+        cur_option_char = all_options[:len(options)]
+
+        if args.all_rounds:
+            num_rounds = len(options)
+        else:
+            num_rounds = 1
+
+        for round_idx in range(num_rounds):
+            idx = row['index']
+            question = row['question']
+            hint = row['hint']
+            image = load_image_from_base64(row['image'])
+            if not is_none(hint):
+                question = hint + '\n' + question
+            for option_char, option in zip(all_options[:len(options)], options):
+                question = question + '\n' + option_char + '. ' + option
+            qs = cur_prompt = question
+            if model.config.mm_use_im_start_end:
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+            else:
+                qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+            if args.single_pred_prompt:
+                if args.lang == 'cn':
+                    qs = qs + '\n' + "请直接回答选项字母。"
+                else:
+                    qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
+
+            conv = conv_templates[args.conv_mode].copy()
+            conv.append_message(conv.roles[0], qs)
+            conv.append_message(conv.roles[1], None)
+            prompt = conv.get_prompt()
+
+            input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+            image_tensor = process_images([image], image_processor, model.config)[0]
+            # image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+
+            stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+
+            with torch.inference_mode():
+                output_ids = model.generate(
+                    input_ids,
+                    images=image_tensor.unsqueeze(0).cuda(),
+                    do_sample=True if args.temperature > 0 else False,
+                    temperature=args.temperature,
+                    top_p=args.top_p,
+                    num_beams=args.num_beams,
+                    # no_repeat_ngram_size=3,
+                    eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                    pad_token_id=tokenizer.eos_token_id,  # Pad token
+                    max_new_tokens=1024,
+                    use_cache=True
+                )
+
+            input_token_len = input_ids.shape[1]
+            n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+            if n_diff_input_output > 0:
+                print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+            outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+            outputs = outputs.strip()
+            if outputs.endswith(stop_str):
+                outputs = outputs[:-len(stop_str)]
+            outputs = outputs.strip()
+
+            ans_id = shortuuid.uuid()
+            ans_file.write(json.dumps({"question_id": idx,
+                                    "round_id": round_idx,
+                                    "prompt": cur_prompt,
+                                    "text": outputs,
+                                    "options": options,
+                                    "option_char": cur_option_char,
+                                    "answer_id": ans_id,
+                                    "model_id": model_name,
+                                    "metadata": {}}) + "\n")
+            ans_file.flush()
+
+            # rotate options
+            options = options[1:] + options[:1]
+            cur_option_char = cur_option_char[1:] + cur_option_char[:1]
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    parser.add_argument("--all-rounds", action="store_true")
+    parser.add_argument("--single-pred-prompt", action="store_true")
+    parser.add_argument("--lang", type=str, default="en")
+    args = parser.parse_args()
+
+    eval_model(args)

llava-phi/llava_phi/eval/model_vqa_phi.py

@@ -0,0 +1,117 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    print(model)
+    questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)[:10]
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+    for line in tqdm(questions):
+        idx = line["question_id"]
+        image_file = line["image"]
+        qs = line["text"]
+        cur_prompt = qs
+        if model.config.mm_use_im_start_end:
+            qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+        else:
+            qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+        image = Image.open(os.path.join(args.image_folder, image_file))
+        image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor.unsqueeze(0).cuda(),
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                top_p=args.top_p,
+                num_beams=args.num_beams,
+                # no_repeat_ngram_size=3,
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.eos_token_id,  # Pad token
+                max_new_tokens=1024,
+                use_cache=True)
+
+        input_token_len = input_ids.shape[1]
+        print(output_ids[:, input_token_len:])
+        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+        if n_diff_input_output > 0:
+            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[:-len(stop_str)]
+        outputs = outputs.strip()
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "image_id": image_file,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v1")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=None)
+    parser.add_argument("--num_beams", type=int, default=1)
+    args = parser.parse_args()
+    print(args)
+
+    eval_model(args)

llava-phi/llava_phi/eval/model_vqa_science.py

@@ -0,0 +1,152 @@
+import argparse
+import torch
+import os
+import json
+from tqdm import tqdm
+import shortuuid
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria
+
+from PIL import Image
+import math
+
+
+def split_list(lst, n):
+    """Split a list into n (roughly) equal-sized chunks"""
+    chunk_size = math.ceil(len(lst) / n)  # integer division
+    return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
+
+
+def get_chunk(lst, n, k):
+    chunks = split_list(lst, n)
+    return chunks[k]
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+    model_path = os.path.expanduser(args.model_path)
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name)
+
+    questions = json.load(open(os.path.expanduser(args.question_file), "r"))
+    questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
+    answers_file = os.path.expanduser(args.answers_file)
+    os.makedirs(os.path.dirname(answers_file), exist_ok=True)
+    ans_file = open(answers_file, "w")
+    for i, line in enumerate(tqdm(questions)):
+        idx = line["id"]
+        question = line['conversations'][0]
+        qs = question['value'].replace('<image>', '').strip()
+        cur_prompt = qs
+
+        if 'image' in line:
+            image_file = line["image"]
+            image = Image.open(os.path.join(args.image_folder, image_file))
+            image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            images = image_tensor.unsqueeze(0).cuda()
+            if getattr(model.config, 'mm_use_im_start_end', False):
+                qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+            else:
+                qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+            cur_prompt = '<image>' + '\n' + cur_prompt
+        else:
+            images = None
+
+        if args.single_pred_prompt:
+            qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
+            cur_prompt = cur_prompt + '\n' + "Answer with the option's letter from the given choices directly."
+
+        conv = conv_templates[args.conv_mode].copy()
+        conv.append_message(conv.roles[0], qs)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = [KeywordsStoppingCriteria(keywords, tokenizer, input_ids)] if conv.version == "v0" else None
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=images,
+                do_sample=True if args.temperature > 0 else False,
+                temperature=args.temperature,
+                max_new_tokens=1024,
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.eos_token_id,  # Pad token
+                use_cache=True
+                # stopping_criteria=stopping_criteria,
+            )
+
+        input_token_len = input_ids.shape[1]
+        n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+        if n_diff_input_output > 0:
+            print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+        outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+        outputs = outputs.strip()
+        if outputs.endswith(stop_str):
+            outputs = outputs[:-len(stop_str)]
+        outputs = outputs.strip()
+
+        # prompt for answer
+        if args.answer_prompter:
+            outputs_reasoning = outputs
+            input_ids = tokenizer_image_token(prompt + outputs_reasoning + ' ###\nANSWER:', tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+            with torch.inference_mode():
+                output_ids = model.generate(
+                    input_ids,
+                    images=images,
+                    do_sample=True if args.temperature > 0 else False,
+                    temperature=args.temperature,
+                    max_new_tokens=64,
+                    eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                    pad_token_id=tokenizer.eos_token_id,  # Pad token
+                    use_cache=True
+                    # stopping_criteria=[stopping_criteria]
+                )
+
+            input_token_len = input_ids.shape[1]
+            n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+            if n_diff_input_output > 0:
+                print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+            outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+            outputs = outputs.strip()
+            if outputs.endswith(stop_str):
+                outputs = outputs[:-len(stop_str)]
+            outputs = outputs.strip()
+            outputs = outputs_reasoning + '\n The answer is ' + outputs
+
+        ans_id = shortuuid.uuid()
+        ans_file.write(json.dumps({"question_id": idx,
+                                   "prompt": cur_prompt,
+                                   "text": outputs,
+                                   "answer_id": ans_id,
+                                   "model_id": model_name,
+                                   "metadata": {}}) + "\n")
+        ans_file.flush()
+    ans_file.close()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-folder", type=str, default="")
+    parser.add_argument("--question-file", type=str, default="tables/question.json")
+    parser.add_argument("--answers-file", type=str, default="answer.jsonl")
+    parser.add_argument("--conv-mode", type=str, default="llava_v0")
+    parser.add_argument("--num-chunks", type=int, default=1)
+    parser.add_argument("--chunk-idx", type=int, default=0)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--answer-prompter", action="store_true")
+    parser.add_argument("--single-pred-prompt", action="store_true")
+    args = parser.parse_args()
+
+    eval_model(args)

llava-phi/llava_phi/eval/qa_baseline_gpt35.py

@@ -0,0 +1,74 @@
+"""Generate answers with GPT-3.5"""
+# Note: you need to be using OpenAI Python v0.27.0 for the code below to work
+import argparse
+import json
+import os
+import time
+import concurrent.futures
+
+import openai
+import tqdm
+import shortuuid
+
+MODEL = 'gpt-3.5-turbo'
+MODEL_ID = 'gpt-3.5-turbo:20230327'
+
+def get_answer(question_id: int, question: str, max_tokens: int):
+    ans = {
+        'answer_id': shortuuid.uuid(),
+        'question_id': question_id,
+        'model_id': MODEL_ID,
+    }
+    for _ in range(3):
+        try:
+            response = openai.ChatCompletion.create(
+                model=MODEL,
+                messages=[{
+                    'role': 'system',
+                    'content': 'You are a helpful assistant.'
+                }, {
+                    'role': 'user',
+                    'content': question,
+                }],
+                max_tokens=max_tokens,
+            )
+            ans['text'] = response['choices'][0]['message']['content']
+            return ans
+        except Exception as e:
+            print('[ERROR]', e)
+            ans['text'] = '#ERROR#'
+            time.sleep(1)
+    return ans
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='ChatGPT answer generation.')
+    parser.add_argument('-q', '--question')
+    parser.add_argument('-o', '--output')
+    parser.add_argument('--max-tokens', type=int, default=1024, help='maximum number of tokens produced in the output')
+    args = parser.parse_args()
+
+    questions_dict = {}
+    with open(os.path.expanduser(args.question)) as f:
+        for line in f:
+            if not line:
+                continue
+            q = json.loads(line)
+            questions_dict[q['question_id']] = q['text']
+
+    answers = []
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=32) as executor:
+        futures = []
+        for qid, question in questions_dict.items():
+            future = executor.submit(get_answer, qid, question, args.max_tokens)
+            futures.append(future)
+
+        for future in tqdm.tqdm(concurrent.futures.as_completed(futures), total=len(futures)):
+            answers.append(future.result())
+
+    answers.sort(key=lambda x: x['question_id'])
+
+    with open(os.path.expanduser(args.output), 'w') as f:
+        table = [json.dumps(ans) for ans in answers]
+        f.write('\n'.join(table))

llava-phi/llava_phi/eval/run_llava_phi.py

@@ -0,0 +1,93 @@
+import argparse
+import torch
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, get_model_name_from_path
+
+from PIL import Image
+
+import requests
+from PIL import Image
+from io import BytesIO
+
+
+def load_image(image_file):
+    if image_file.startswith('http') or image_file.startswith('https'):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert('RGB')
+    else:
+        image = Image.open(image_file).convert('RGB')
+    return image
+
+
+def eval_model(args):
+    # Model
+    disable_torch_init()
+
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name)
+
+    qs = args.query
+    if model.config.mm_use_im_start_end:
+        qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
+    else:
+        qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
+
+    if 'phi' in model_name.lower():
+        conv_mode = "phi-2_v0"
+    else:
+        conv_mode = "default"
+
+    if args.conv_mode is not None and conv_mode != args.conv_mode:
+        print('[WARNING] the auto inferred conversation mode is {}, while `--conv-mode` is {}, using {}'.format(conv_mode, args.conv_mode, args.conv_mode))
+    else:
+        args.conv_mode = conv_mode
+
+    conv = conv_templates[args.conv_mode].copy()
+    conv.append_message(conv.roles[0], qs)
+    conv.append_message(conv.roles[1], None)
+    prompt = conv.get_prompt()
+
+    image = load_image(args.image_file)
+    image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'].cuda()
+
+    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+
+    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+
+    with torch.inference_mode():
+        output_ids = model.generate(
+            input_ids,
+            images=image_tensor,
+            do_sample=True,
+            temperature=0.2,
+            max_new_tokens=1024,
+            eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+            pad_token_id=tokenizer.eos_token_id,  # Pad token
+            use_cache=True,
+        )
+
+    input_token_len = input_ids.shape[1]
+    n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+    if n_diff_input_output > 0:
+        print(f'[Warning] {n_diff_input_output} output_ids are not the same as the input_ids')
+    outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
+    outputs = outputs.strip()
+    if outputs.endswith(stop_str):
+        outputs = outputs[:-len(stop_str)]
+    outputs = outputs.strip()
+    print(outputs)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-file", type=str, required=True)
+    parser.add_argument("--query", type=str, required=True)
+    parser.add_argument("--conv-mode", type=str, default=None)
+    args = parser.parse_args()
+
+    eval_model(args)

llava-phi/llava_phi/eval/summarize_gpt_review.py

@@ -0,0 +1,60 @@
+import json
+import os
+from collections import defaultdict
+
+import numpy as np
+
+import argparse
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+    parser.add_argument('-d', '--dir', default=None)
+    parser.add_argument('-v', '--version', default=None)
+    parser.add_argument('-s', '--select', nargs='*', default=None)
+    parser.add_argument('-f', '--files', nargs='*', default=[])
+    parser.add_argument('-i', '--ignore', nargs='*', default=[])
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = parse_args()
+
+    if args.ignore is not None:
+        args.ignore = [int(x) for x in args.ignore]
+
+    if len(args.files) > 0:
+        review_files = args.files
+    else:
+        review_files = [x for x in os.listdir(args.dir) if x.endswith('.jsonl') and (x.startswith('gpt4_text') or x.startswith('reviews_') or x.startswith('review_') or 'review' in args.dir)]
+
+    for review_file in sorted(review_files):
+        config = os.path.basename(review_file).replace('gpt4_text_', '').replace('.jsonl', '')
+        if args.select is not None and any(x not in config for x in args.select):
+            continue
+        if '0613' in config:
+            version = '0613'
+        else:
+            version = '0314'
+        if args.version is not None and args.version != version:
+            continue
+        scores = defaultdict(list)
+        print(config)
+        with open(os.path.join(args.dir, review_file) if args.dir is not None else review_file) as f:
+            for review_str in f:
+                review = json.loads(review_str)
+                if review['question_id'] in args.ignore:
+                    continue
+                if 'category' in review:
+                    scores[review['category']].append(review['tuple'])
+                    scores['all'].append(review['tuple'])
+                else:
+                    if 'tuple' in review:
+                        scores['all'].append(review['tuple'])
+                    else:
+                        scores['all'].append(review['score'])
+        for k, v in sorted(scores.items()):
+            stats = np.asarray(v).mean(0).tolist()
+            stats = [round(x, 3) for x in stats]
+            # print(k, stats, round(stats[1]/stats[0]*100, 1))
+            print(k, round(stats[1]/stats[0]*100, 1), round(stats[0] * 10, 1), round(stats[1] * 10, 1))
+        print('=================================')

llava-phi/llava_phi/eval/table/rule.json

@@ -0,0 +1,11 @@
+{
+    "coding": {"role": "Assistant", "prompt": "Your task is to evaluate the coding abilities of the above two assistants. They have been asked to implement a program to solve a given problem. Please review their code submissions, paying close attention to their problem-solving approach, code structure, readability, and the inclusion of helpful comments.\n\nPlease ensure that the assistants' submissions:\n\n1. Correctly implement the given problem statement.\n2. Contain accurate and efficient code.\n3. Include clear and concise comments that explain the code's logic and functionality.\n4. Adhere to proper coding standards and best practices.\n\nOnce you have carefully reviewed both submissions, provide detailed feedback on their strengths and weaknesses, along with any suggestions for improvement. You should first output a single line containing two scores on the scale of 1-10 (1: no code/no sense; 10: perfect) for Assistant 1 and 2, respectively. Then give extra comments starting from the next line."},
+    "math":  {"role": "Assistant", "prompt": "We would like to request your feedback on the mathematical proficiency of two AI assistants regarding the given user question.\nFirstly, please solve the problem independently, without referring to the answers provided by Assistant 1 and Assistant 2.\nAfterward, please examine the problem-solving process of Assistant 1 and Assistant 2 step-by-step to ensure their correctness, identifying any incorrect steps if present. Your evaluation should take into account not only the answer but also the problem-solving steps.\nFinally, please output a Python tuple containing two numerical scores for Assistant 1 and Assistant 2, ranging from 1 to 10, respectively. If applicable, explain the reasons for any variations in their scores and determine which assistant performed better."},
+    "default":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above.\nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."},
+    "conv":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. The user asks the question on observing an image. For your reference, the visual content in the image is represented with five descriptive sentences describing the same image and the bounding box coordinates of each object in the scene. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2) with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y, bottom right x, and bottom right y. \nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."},
+    "detail":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. The user asks the question on observing an image. For your reference, the visual content in the image is represented with five descriptive sentences describing the same image and the bounding box coordinates of each object in the scene. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2) with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y, bottom right x, and bottom right y. \nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."},
+    "complex":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. The user asks the question on observing an image. For your reference, the visual content in the image is represented with five descriptive sentences describing the same image and the bounding box coordinates of each object in the scene. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2) with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y, bottom right x, and bottom right y. \nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."},
+    "llava_bench_conv":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. The user asks the question on observing an image. For your reference, the visual content in the image is represented with a few sentences describing the image. \nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."},
+    "llava_bench_detail":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. The user asks the question on observing an image. For your reference, the visual content in the image is represented with a few sentences describing the image. \nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."},
+    "llava_bench_complex":  {"role": "Assistant", "prompt": "We would like to request your feedback on the performance of two AI assistants in response to the user question displayed above. The user asks the question on observing an image. For your reference, the visual content in the image is represented with a few sentences describing the image. \nPlease rate the helpfulness, relevance, accuracy, level of details of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.\nPlease first output a single line containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space.\nIn the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment."}
+}

llava-phi/llava_phi/mm_utils.py

@@ -0,0 +1,96 @@
+from PIL import Image
+from io import BytesIO
+import base64
+
+import torch
+from transformers import StoppingCriteria
+from llava_phi.constants import IMAGE_TOKEN_INDEX
+
+
+def load_image_from_base64(image):
+    return Image.open(BytesIO(base64.b64decode(image)))
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
+def process_images(images, image_processor, model_cfg):
+    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
+    new_images = []
+    if image_aspect_ratio == 'pad':
+        for image in images:
+            image = expand2square(image, tuple(int(x*255) for x in image_processor.image_mean))
+            image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            new_images.append(image)
+    else:
+        return image_processor(images, return_tensors='pt')['pixel_values']
+    if all(x.shape == new_images[0].shape for x in new_images):
+        new_images = torch.stack(new_images, dim=0)
+    return new_images
+
+
+def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
+    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
+
+    def insert_separator(X, sep):
+        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
+
+    input_ids = []
+    offset = 0
+    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
+        offset = 1
+        input_ids.append(prompt_chunks[0][0])
+    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
+        input_ids.extend(x[offset:])
+
+    if return_tensors is not None:
+        if return_tensors == 'pt':
+            return torch.tensor(input_ids, dtype=torch.long)
+        raise ValueError(f'Unsupported tensor type: {return_tensors}')
+    return input_ids
+
+
+def get_model_name_from_path(model_path):
+    model_path = model_path.strip("/")
+    model_paths = model_path.split("/")
+    if model_paths[-1].startswith('checkpoint-'):
+        return model_paths[-2] + "_" + model_paths[-1]
+    else:
+        return model_paths[-1]
+
+
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        self.start_len = input_ids.shape[1]
+
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        assert output_ids.shape[0] == 1, "Only support batch size 1 (yet)"  # TODO
+        offset = min(output_ids.shape[1] - self.start_len, 3)
+        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
+        for keyword_id in self.keyword_ids:
+            if output_ids[0, -keyword_id.shape[0]:] == keyword_id:
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False

llava-phi/llava_phi/model/__init__.py

@@ -0,0 +1,2 @@
+from .language_model.llava_phi import LlavaPhiForCausalLM
+from .language_model.configuration_llava_phi import LlavaPhiConfig, LlavaPhiVisionConfig, ProjectorConfig

llava-phi/llava_phi/model/builder.py

@@ -0,0 +1,121 @@
+import os
+import warnings
+import shutil
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig, CLIPImageProcessor
+import torch
+from llava_phi.model import *
+from llava_phi.constants import DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+
+def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="cuda", device="cuda"):
+    kwargs = {"device_map": device_map}
+    if load_8bit:
+        kwargs['load_in_8bit'] = True
+    elif load_4bit:
+        kwargs['load_in_4bit'] = True
+        kwargs['quantization_config'] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4'
+        )
+    # else:  # TODO: after fine-tuning LLava-Phi, load the model weights with fp16 will pose nan
+    #     kwargs['torch_dtype'] = torch.float16
+
+    if 'phi' in model_name.lower():
+        # Load LLaVA-Phi model
+        if 'lora' in model_name.lower() and model_base is None:
+            warnings.warn('There is `lora` in model name but no `model_base` is provided. If you are loading a LoRA model, please provide the `model_base` argument.')
+        if 'lora' in model_name.lower() and model_base is not None:
+            lora_cfg_pretrained = AutoConfig.from_pretrained(model_path)
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            print('Loading LLaVA-Phi from base model...')
+            model = LlavaPhiForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=lora_cfg_pretrained, **kwargs)
+            token_num, tokem_dim = model.lm_head.out_features, model.lm_head.in_features
+            if model.lm_head.weight.shape[0] != token_num:
+                model.lm_head.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
+                model.model.embed_tokens.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
+
+            print('Loading additional LLaVA-Phi weights...')
+            if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')):
+                non_lora_trainables = torch.load(os.path.join(model_path, 'non_lora_trainables.bin'), map_location='cpu')
+            else:
+                # this is probably from HF Hub
+                from huggingface_hub import hf_hub_download
+                def load_from_hf(repo_id, filename, subfolder=None):
+                    cache_file = hf_hub_download(
+                        repo_id=repo_id,
+                        filename=filename,
+                        subfolder=subfolder)
+                    return torch.load(cache_file, map_location='cpu')
+                non_lora_trainables = load_from_hf(model_path, 'non_lora_trainables.bin')
+            non_lora_trainables = {(k[11:] if k.startswith('base_model.') else k): v for k, v in non_lora_trainables.items()}
+            if any(k.startswith('model.model.') for k in non_lora_trainables):
+                non_lora_trainables = {(k[6:] if k.startswith('model.') else k): v for k, v in non_lora_trainables.items()}
+            model.load_state_dict(non_lora_trainables, strict=False)
+
+            from peft import PeftModel
+            print('Loading LoRA weights...')
+            model = PeftModel.from_pretrained(model, model_path)
+            print('Merging LoRA weights...')
+            model = model.merge_and_unload()
+            print('Model is loaded...')
+        elif model_base is not None:
+            # this may be mm projector only
+            print('Loading LLaVA-Phi from base model...')
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            cfg_pretrained = AutoConfig.from_pretrained(model_path)
+            model = LlavaPhiForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=cfg_pretrained, **kwargs)
+
+            mm_projector_weights = torch.load(os.path.join(model_path, 'mm_projector.bin'), map_location='cpu')
+            mm_projector_weights = {k: v.to(torch.float16) for k, v in mm_projector_weights.items()}
+            model.load_state_dict(mm_projector_weights, strict=False)
+        else:
+            print("load llaVA-Phi MLLM!!!")
+            config = LlavaPhiConfig.from_pretrained(model_path, trust_remote_code=True)
+            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True)
+            model = LlavaPhiForCausalLM.from_pretrained(
+                model_path, 
+                config=config, 
+                use_safetensors=True, 
+                **kwargs).to("cuda")
+    else:
+        # Load language model
+        if model_base is not None:
+            # PEFT model
+            from peft import PeftModel
+            tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(model_base, torch_dtype=torch.float16, low_cpu_mem_usage=True, device_map="auto")
+            print(f"Loading LoRA weights from {model_path}")
+            model = PeftModel.from_pretrained(model, model_path)
+            print(f"Merging weights")
+            model = model.merge_and_unload()
+            print('Convert to FP16...')
+            model.to(torch.float16)
+        else:
+            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+            model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, **kwargs)
+
+    image_processor = CLIPImageProcessor.from_pretrained(model_path)
+
+    if 'phi' in model_name.lower():
+        mm_use_im_start_end = getattr(model.config, "mm_use_im_start_end", False)
+        mm_use_im_patch_token = getattr(model.config, "mm_use_im_patch_token", True)
+
+        # TODO: the tokenizer length of phi-2 is 50295, but the output class of lm_head is 51200
+        if mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+        if mm_use_im_start_end:
+            tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            # model.resize_token_embeddings(len(tokenizer))
+    else:
+        raise ValueError(f"Unsupported model name: {model_name}")
+
+    if hasattr(model.config, "max_sequence_length"):
+        context_len = model.config.max_sequence_length
+    else:
+        context_len = 2048
+    model.to(device="cuda")
+    print(kwargs)
+    return tokenizer, model, image_processor, context_len

llava-phi/llava_phi/model/language_model/configuration_llava_phi.py

@@ -0,0 +1,179 @@
+import os
+from typing import Union
+from transformers import PretrainedConfig, PhiConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class LlavaPhiVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`CLIPVisionModel`]. It is used to instantiate a
+    CLIP vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the CLIP
+    [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 32):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1.0):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+        mm_vision_select_feature (`str`, *optional*, defaults to `"patch"`):
+            The feature to select from the vision encoder output. Can be one of `"patch"` or `"cls_patch"`.
+        mm_vision_select_layer (`int`, *optional*, defaults to `-2`):
+            The layer to select from the vision encoder output.
+
+    Example:
+
+    ```python
+    >>> from transformers import CLIPVisionConfig, CLIPVisionModel
+
+    >>> # Initializing a CLIPVisionConfig with openai/clip-vit-base-patch32 style configuration
+    >>> configuration = CLIPVisionConfig()
+
+    >>> # Initializing a CLIPVisionModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
+    >>> model = CLIPVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "llava_phi_clip_vision_model"
+
+    def __init__(
+            self,
+            hidden_size=768,
+            intermediate_size=3072,
+            projection_dim=512,
+            num_hidden_layers=12,
+            num_attention_heads=12,
+            num_channels=3,
+            image_size=224,
+            patch_size=32,
+            hidden_act="quick_gelu",
+            layer_norm_eps=1e-5,
+            attention_dropout=0.0,
+            initializer_range=0.02,
+            initializer_factor=1.0,
+            mm_vision_select_feature="patch",
+            mm_vision_select_layer=-2,
+            **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.mm_vision_select_feature = mm_vision_select_feature
+        self.mm_vision_select_layer = mm_vision_select_layer
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "llava_phi-phi":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class ProjectorConfig(PretrainedConfig):
+    model_type = "llava_phi_projector"
+
+    def __init__(
+            self,
+            mm_projector_type="linear",
+            mm_hidden_size=768,
+            hidden_size=2560,
+            **kwargs
+    ):
+        self.mm_projector_type = mm_projector_type
+        self.mm_hidden_size = mm_hidden_size
+        self.hidden_size = hidden_size
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from CLIPConfig
+        if config_dict.get("model_type") == "llava_phi-phi":
+            config_dict = config_dict["projector_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+DEFAULT_VISUAL_CONFIG = {
+    "vision_tower": LlavaPhiVisionConfig().to_dict(),
+    "mm_projector": ProjectorConfig().to_dict()
+}
+
+
+class LlavaPhiConfig(PhiConfig):
+    model_type = "llava_phi"
+
+    def __init__(self, vision_config=None, **kwargs):
+        if vision_config is None:
+            self.vision_config = DEFAULT_VISUAL_CONFIG
+        else:
+            self.vision_config = vision_config
+
+        super().__init__(**kwargs)
+
+
+if __name__ == "__main__":
+    print(LlavaPhiVisionConfig())

llava-phi/llava_phi/model/language_model/llava_phi.py

@@ -0,0 +1,126 @@
+import os
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss
+
+from transformers import AutoConfig, AutoModelForCausalLM, \
+    PhiModel, PhiPreTrainedModel
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from ..llava_arch import LlavaMetaModel, LlavaMetaForCausalLM
+from transformers.utils import logging
+from .configuration_llava_phi import LlavaPhiConfig
+
+logger = logging.get_logger(__name__)
+
+
+class LLavaPhiModel(LlavaMetaModel, PhiModel):
+    config_class = LlavaPhiConfig
+
+    def __init__(self, config):
+        super(LLavaPhiModel, self).__init__(config)
+
+
+class LlavaPhiForCausalLM(PhiPreTrainedModel, LlavaMetaForCausalLM):
+    config_class = LlavaPhiConfig
+
+    def __init__(self, config):
+        super(PhiPreTrainedModel, self).__init__(config)
+        self.model = LLavaPhiModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=True)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_model(self):
+        return self.model
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            labels: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            images: Optional[torch.FloatTensor] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        input_ids, attention_mask, past_key_values, inputs_embeds, labels = self.prepare_inputs_labels_for_multimodal(
+            input_ids, attention_mask, past_key_values, labels, images)
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model/pipeline parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+        
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+            self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+                "images": kwargs.get("images", None),
+            }
+        )
+        return model_inputs
+
+
+AutoConfig.register("llava_phi", LlavaPhiConfig)
+AutoModelForCausalLM.register(LlavaPhiConfig, LlavaPhiForCausalLM)

llava-phi/llava_phi/model/llava_arch.py

@@ -0,0 +1,208 @@
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from abc import ABC, abstractmethod
+
+import torch
+
+from .multimodal_encoder.clip_encoder import CLIPVisionTower
+from .multimodal_projector.builder import build_vision_projector
+from .language_model.configuration_llava_phi import LlavaPhiConfig, LlavaPhiVisionConfig, ProjectorConfig
+from llava_phi.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+
+
+class LlavaMetaModel:
+    def __init__(self, config):
+        super(LlavaMetaModel, self).__init__(config)
+        self.vision_tower = CLIPVisionTower(
+            LlavaPhiVisionConfig(**config.vision_config["vision_tower"])
+        )
+        self.mm_projector = build_vision_projector(
+            ProjectorConfig(**config.vision_config["mm_projector"])
+        )
+
+    def get_vision_tower(self):
+        vision_tower = getattr(self, 'vision_tower', None)
+        if type(vision_tower) is list:
+            vision_tower = vision_tower[0]
+        return vision_tower
+
+
+class LlavaMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def get_vision_tower(self):
+        return self.get_model().get_vision_tower()
+
+    def encode_images(self, images):
+        image_features = self.get_model().get_vision_tower()(images)
+        image_features = self.get_model().mm_projector(image_features)
+        return image_features
+
+    def prepare_inputs_labels_for_multimodal(
+        self, input_ids, attention_mask, past_key_values, labels, images
+    ):
+        vision_tower = self.get_vision_tower()
+        if vision_tower is None or images is None or input_ids.shape[1] == 1:
+            if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
+                attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device)
+            return input_ids, attention_mask, past_key_values, None, labels
+
+        if type(images) is list or images.ndim == 5:
+            concat_images = torch.cat([image for image in images], dim=0)
+            image_features = self.encode_images(concat_images)
+            split_sizes = [image.shape[0] for image in images]
+            image_features = torch.split(image_features, split_sizes, dim=0)
+            image_features = [x.flatten(0, 1) for x in image_features]
+        else:
+            image_features = self.encode_images(images)
+
+        new_input_embeds = []
+        new_labels = [] if labels is not None else None
+        cur_image_idx = 0
+        for batch_idx, cur_input_ids in enumerate(input_ids):
+            if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
+                # multimodal LLM, but the current sample is not multimodal
+                # FIXME: this is a hacky fix, for deepspeed zero3 to work
+                half_len = cur_input_ids.shape[0] // 2
+                cur_image_features = image_features[cur_image_idx]
+                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids[:half_len])
+                cur_input_embeds_2 = self.get_model().embed_tokens(cur_input_ids[half_len:])
+                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0], cur_input_embeds_2], dim=0)
+                new_input_embeds.append(cur_input_embeds)
+                if labels is not None:
+                    new_labels.append(labels[batch_idx])
+                cur_image_idx += 1
+                continue
+            image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
+            cur_new_input_embeds = []
+            if labels is not None:
+                cur_labels = labels[batch_idx]
+                cur_new_labels = []
+                assert cur_labels.shape == cur_input_ids.shape
+            while image_token_indices.numel() > 0:
+                cur_image_features = image_features[cur_image_idx]
+                image_token_start = image_token_indices[0]
+                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start-1]).detach())
+                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start-1:image_token_start]))
+                    cur_new_input_embeds.append(cur_image_features)
+                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start+1:image_token_start+2]))
+                    if labels is not None:
+                        cur_new_labels.append(cur_labels[:image_token_start])
+                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
+                        cur_new_labels.append(cur_labels[image_token_start:image_token_start+1])
+                        cur_labels = cur_labels[image_token_start+2:]
+                else:
+                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start]))
+                    cur_new_input_embeds.append(cur_image_features)
+                    if labels is not None:
+                        cur_new_labels.append(cur_labels[:image_token_start])
+                        cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
+                        cur_labels = cur_labels[image_token_start+1:]
+                cur_image_idx += 1
+                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+                    cur_input_ids = cur_input_ids[image_token_start+2:]
+                else:
+                    cur_input_ids = cur_input_ids[image_token_start+1:]
+                image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
+            if cur_input_ids.numel() > 0:
+                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
+                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids).detach())
+                else:
+                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids))
+                if labels is not None:
+                    cur_new_labels.append(cur_labels)
+            cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds]
+            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
+            new_input_embeds.append(cur_new_input_embeds)
+            if labels is not None:
+                cur_new_labels = torch.cat(cur_new_labels, dim=0)
+                new_labels.append(cur_new_labels)
+
+        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
+            max_len = max(x.shape[0] for x in new_input_embeds)
+
+            new_input_embeds_align = []
+            for cur_new_embed in new_input_embeds:
+                cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)
+                new_input_embeds_align.append(cur_new_embed)
+            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)
+
+            if labels is not None:
+                new_labels_align = []
+                _new_labels = new_labels
+                for cur_new_label in new_labels:
+                    cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0)
+                    new_labels_align.append(cur_new_label)
+                new_labels = torch.stack(new_labels_align, dim=0)
+
+            if attention_mask is not None:
+                new_attention_mask = []
+                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels):
+                    new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device)
+                    new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device)
+                    cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0)
+                    new_attention_mask.append(cur_new_attention_mask)
+                attention_mask = torch.stack(new_attention_mask, dim=0)
+                assert attention_mask.shape == new_labels.shape
+        else:
+            new_input_embeds = torch.stack(new_input_embeds, dim=0)
+            if labels is not None:
+                new_labels  = torch.stack(new_labels, dim=0)
+
+            if attention_mask is not None:
+                new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device)
+                attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1)
+                assert attention_mask.shape == new_input_embeds.shape[:2]
+
+        return None, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    def initialize_vision_tokenizer(self, model_args, tokenizer):
+        if model_args.mm_use_im_patch_token:
+            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+        if model_args.mm_use_im_start_end:
+            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
+            self.resize_token_embeddings(len(tokenizer))
+
+            if num_new_tokens > 0:
+                input_embeddings = self.get_input_embeddings().weight.data
+                output_embeddings = self.get_output_embeddings().weight.data
+
+                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+                    dim=0, keepdim=True)
+
+                input_embeddings[-num_new_tokens:] = input_embeddings_avg
+                output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = True
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False
+                    
+        elif model_args.mm_use_im_patch_token:
+            if model_args.tune_mm_mlp_adapter:
+                for p in self.get_input_embeddings().parameters():
+                    p.requires_grad = False
+                for p in self.get_output_embeddings().parameters():
+                    p.requires_grad = False

llava-phi/llava_phi/model/multimodal_encoder/clip_encoder.py

@@ -0,0 +1,89 @@
+from abc import ABC
+
+import torch
+import torch.nn as nn
+
+from transformers import CLIPPreTrainedModel, CLIPVisionConfig
+from transformers.models.clip.modeling_clip import CLIPVisionTransformer
+from llava_phi.model.language_model.configuration_llava_phi import LlavaPhiVisionConfig
+
+
+class CLIPVisionTower(CLIPPreTrainedModel):
+    config_class = LlavaPhiVisionConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.vision_model = CLIPVisionTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[
+            self.config.mm_vision_select_layer
+        ]
+        if self.config.mm_vision_select_feature == "patch":
+            image_features = image_features[:, 1:]
+        elif self.config.mm_vision_select_feature == "cls_patch":
+            image_features = image_features
+        else:
+            raise ValueError(
+                f"Unexpected select feature: {self.config.mm_vision_select_feature}"
+            )
+        return image_features
+
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_model(
+                    image.to(device=self.device, dtype=self.dtype).unsqueeze(0),
+                    output_hidden_states=True,
+                )
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_model(
+                images.to(device=self.device, dtype=self.dtype),
+                output_hidden_states=True,
+            )
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return list(self.vision_model.parameters())[0].dtype
+
+    @property
+    def device(self):
+        return list(self.vision_model.parameters())[0].device
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+
+if __name__ == "__main__":
+    clip_config = CLIPVisionConfig.from_pretrained(
+        "/data/private/zhumj/GPTcode/mm-phi/openai/clip-vit-large-patch14-336"
+    )
+    print("################ clip_config ##############")
+    print(clip_config)
+    phi_vis_config = LlavaPhiVisionConfig(**clip_config.to_dict())
+    print("################ phi_vis_config ##############")
+    print(phi_vis_config)
+
+    model = CLIPVisionTower(clip_config)
+    # print(list(model.vision_model.parameters())[0].dtype)

llava-phi/llava_phi/model/multimodal_projector/builder.py

@@ -0,0 +1,50 @@
+import torch
+import torch.nn as nn
+import re
+
+
+class IdentityMap(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+    @property
+    def config(self):
+        return {"mm_projector_type": "identity"}
+
+
+class SimpleResBlock(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.pre_norm = nn.LayerNorm(channels)
+
+        self.proj = nn.Sequential(
+            nn.Linear(channels, channels), nn.GELU(), nn.Linear(channels, channels)
+        )
+
+    def forward(self, x):
+        x = self.pre_norm(x)
+        return x + self.proj(x)
+
+
+def build_vision_projector(config):
+    projector_type = getattr(config, "mm_projector_type", "linear")
+
+    if projector_type == "linear":
+        return nn.Linear(config.mm_hidden_size, config.hidden_size)
+
+    mlp_gelu_match = re.match(r"^mlp(\d+)x_gelu$", projector_type)
+    if mlp_gelu_match:
+        mlp_depth = int(mlp_gelu_match.group(1))
+        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
+        for _ in range(1, mlp_depth):
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
+        return nn.Sequential(*modules)
+
+    if projector_type == "identity":
+        return IdentityMap()
+
+    raise ValueError(f"Unknown projector type: {projector_type}")

llava-phi/llava_phi/serve/app.py

@@ -0,0 +1,354 @@
+import argparse
+import hashlib
+import json
+import os
+import time
+from threading import Thread
+
+import gradio as gr
+import torch
+from llava_phi.constants import (DEFAULT_IM_END_TOKEN, DEFAULT_IM_START_TOKEN,
+                                 DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_INDEX)
+from llava_phi.conversation import (SeparatorStyle, conv_templates,
+                                    default_conversation)
+from llava_phi.mm_utils import (KeywordsStoppingCriteria, load_image_from_base64,
+                                process_images, tokenizer_image_token)
+from llava_phi.model.builder import load_pretrained_model
+from transformers import TextIteratorStreamer
+
+print(gr.__version__)
+
+block_css = """
+
+#buttons button {
+    min-width: min(120px,100%);
+}
+"""
+title_markdown = ("""
+# LLaVA-Phi: Efficient Multi-Modal Assistant with Small Language Model
+[[Code](https://github.com/zhuyiche/llava-phi)] | 📚 [[Paper](https://arxiv.org/pdf/2401.02330)]
+""")
+tos_markdown = ("""
+### Terms of use
+By using this service, users are required to agree to the following terms:
+The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes.
+For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality.
+""")
+learn_more_markdown = ("""
+### License
+The service is a research preview intended for non-commercial use only, subject to the model [License](https://huggingface.co/microsoft/phi-2) of Phi-2. Please contact us if you find any potential violation.
+""")
+ack_markdown = ("""
+### Acknowledgement
+The template for this web demo is from [LLaVA](https://github.com/haotian-liu/LLaVA), and we are very grateful to LLaVA for their open source contributions to the community!
+""")
+
+
+def regenerate(state, image_process_mode):
+    state.messages[-1][-1] = None
+    prev_human_msg = state.messages[-2]
+    if type(prev_human_msg[1]) in (tuple, list):
+        prev_human_msg[1] = (*prev_human_msg[1][:2], image_process_mode)
+    state.skip_next = False
+    return (state, state.to_gradio_chatbot(), "", None)
+
+
+def clear_history():
+    state = default_conversation.copy()
+    return (state, state.to_gradio_chatbot(), "", None)
+
+
+def add_text(state, text, image, image_process_mode):
+    if len(text) <= 0 and image is None:
+        state.skip_next = True
+        return (state, state.to_gradio_chatbot(), "", None)
+
+    text = text[:1536]  # Hard cut-off
+    if image is not None:
+        text = text[:1200]  # Hard cut-off for images
+        if '<image>' not in text:
+            # text = '<Image><image></Image>' + text
+            text = text + '\n<image>'
+        text = (text, image, image_process_mode)
+        if len(state.get_images(return_pil=True)) > 0:
+            state = default_conversation.copy()
+    state.append_message(state.roles[0], text)
+    state.append_message(state.roles[1], None)
+    state.skip_next = False
+    return (state, state.to_gradio_chatbot(), "", None)
+
+
+def load_demo():
+    state = default_conversation.copy()
+    return state
+
+
+@torch.inference_mode()
+def get_response(params):
+    prompt = params["prompt"]
+    ori_prompt = prompt
+    images = params.get("images", None)
+    num_image_tokens = 0
+    if images is not None and len(images) > 0:
+        if len(images) > 0:
+            if len(images) != prompt.count(DEFAULT_IMAGE_TOKEN):
+                raise ValueError(
+                    "Number of images does not match number of <image> tokens in prompt")
+
+            images = [load_image_from_base64(image) for image in images]
+            images = process_images(images, image_processor, model.config)
+
+            if type(images) is list:
+                images = [image.to(model.device, dtype=torch.float16)
+                          for image in images]
+            else:
+                images = images.to(model.device, dtype=torch.float16)
+
+            replace_token = DEFAULT_IMAGE_TOKEN
+            if getattr(model.config, 'mm_use_im_start_end', False):
+                replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
+            prompt = prompt.replace(DEFAULT_IMAGE_TOKEN, replace_token)
+
+            num_image_tokens = prompt.count(
+                replace_token) * model.get_vision_tower().num_patches
+        else:
+            images = None
+        image_args = {"images": images}
+    else:
+        images = None
+        image_args = {}
+
+    temperature = float(params.get("temperature", 1.0))
+    top_p = float(params.get("top_p", 1.0))
+    max_context_length = getattr(
+        model.config, 'max_position_embeddings', 2048)
+    max_new_tokens = min(int(params.get("max_new_tokens", 256)), 1024)
+    stop_str = params.get("stop", None)
+    do_sample = True if temperature > 0.001 else False
+
+    input_ids = tokenizer_image_token(
+        prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
+    keywords = [stop_str]
+    stopping_criteria = KeywordsStoppingCriteria(
+        keywords, tokenizer, input_ids)
+    streamer = TextIteratorStreamer(
+        tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=15)
+
+    max_new_tokens = min(max_new_tokens, max_context_length -
+                         input_ids.shape[-1] - num_image_tokens)
+
+    if max_new_tokens < 1:
+        yield json.dumps({"text": ori_prompt + "Exceeds max token length. Please start a new conversation, thanks.",
+                          "error_code": 0}).encode() + b"\0"
+        return
+
+    # local inference
+    thread = Thread(target=model.generate, kwargs=dict(
+        inputs=input_ids,
+        do_sample=do_sample,
+        temperature=temperature,
+        top_p=top_p,
+        max_new_tokens=max_new_tokens,
+        streamer=streamer,
+        stopping_criteria=[stopping_criteria],
+        use_cache=True,
+        **image_args
+    ))
+    thread.start()
+
+    generated_text = ori_prompt
+    for new_text in streamer:
+        generated_text += new_text
+        if generated_text.endswith(stop_str):
+            generated_text = generated_text[:-len(stop_str)]
+        yield json.dumps({"text": generated_text, "error_code": 0}).encode()
+
+
+def http_bot(state, temperature, top_p, max_new_tokens):
+    if state.skip_next:
+        # This generate call is skipped due to invalid inputs
+        yield (state, state.to_gradio_chatbot())
+        return
+
+    if len(state.messages) == state.offset + 2:
+        # First round of conversation
+        if "phi" in model_name.lower():
+            template_name = "phi-2_v0"
+        else:
+            template_name = "phi-2_v0"
+        new_state = conv_templates[template_name].copy()
+        new_state.append_message(new_state.roles[0], state.messages[-2][1])
+        new_state.append_message(new_state.roles[1], None)
+        state = new_state
+
+    # Construct prompt
+    prompt = state.get_prompt()
+
+    all_images = state.get_images(return_pil=True)
+    all_image_hash = [hashlib.md5(image.tobytes()).hexdigest()
+                      for image in all_images]
+
+    # Make requests
+    pload = {
+        "model": model_name,
+        "prompt": prompt,
+        "temperature": float(temperature),
+        "top_p": float(top_p),
+        "max_new_tokens": min(int(max_new_tokens), 1536),
+        "stop": state.sep if state.sep_style in [SeparatorStyle.SINGLE, SeparatorStyle.MPT] else state.sep2,
+        "images": f'List of {len(state.get_images())} images: {all_image_hash}',
+    }
+
+    pload['images'] = state.get_images()
+
+    state.messages[-1][-1] = "▌"
+    yield (state, state.to_gradio_chatbot())
+
+    # for stream
+    output = get_response(pload)
+    for chunk in output:
+        if chunk:
+            data = json.loads(chunk.decode())
+            if data["error_code"] == 0:
+                output = data["text"][len(prompt):].strip()
+                state.messages[-1][-1] = output + "▌"
+                yield (state, state.to_gradio_chatbot())
+            else:
+                output = data["text"] + \
+                         f" (error_code: {data['error_code']})"
+                state.messages[-1][-1] = output
+                yield (state, state.to_gradio_chatbot())
+                return
+            time.sleep(0.03)
+
+    state.messages[-1][-1] = state.messages[-1][-1][:-1]
+    yield (state, state.to_gradio_chatbot())
+
+
+def build_demo():
+    textbox = gr.Textbox(
+        show_label=False, placeholder="Enter text and press ENTER", container=False)
+    with gr.Blocks(title="LLaVA-Phi", theme=gr.themes.Default(), css=block_css) as demo:
+        state = gr.State()
+        gr.Markdown(title_markdown)
+
+        with gr.Row():
+            with gr.Column(scale=5):
+                with gr.Row(elem_id="Model ID"):
+                    gr.Dropdown(
+                        choices=['LLaVA-Phi-3B'],
+                        value='LLaVA-Phi-3B',
+                        interactive=True,
+                        label='Model ID',
+                        container=False)
+                imagebox = gr.Image(type="pil")
+                image_process_mode = gr.Radio(
+                    ["Crop", "Resize", "Pad", "Default"],
+                    value="Default",
+                    label="Preprocess for non-square image", visible=False)
+
+                cur_dir = os.path.dirname(os.path.abspath(__file__))
+                gr.Examples(examples=[
+                    [f"{cur_dir}/examples/extreme_ironing.jpg",
+                     "What is unusual about this image?"],
+                    [f"{cur_dir}/examples/waterview.jpg",
+                     "What are the things I should be cautious about when I visit here?"],
+                ], inputs=[imagebox, textbox])
+
+                with gr.Accordion("Parameters", open=False) as _:
+                    temperature = gr.Slider(
+                        minimum=0.0, maximum=1.0, value=0.2, step=0.1, interactive=True, label="Temperature", )
+                    top_p = gr.Slider(
+                        minimum=0.0, maximum=1.0, value=0.7, step=0.1, interactive=True, label="Top P", )
+                    max_output_tokens = gr.Slider(
+                        minimum=0, maximum=1024, value=512, step=64, interactive=True, label="Max output tokens", )
+
+            with gr.Column(scale=8):
+                chatbot = gr.Chatbot(
+                    elem_id="chatbot", label="LLaVA-Phi Chatbot", height=550)
+                with gr.Row():
+                    with gr.Column(scale=8):
+                        textbox.render()
+                    with gr.Column(scale=1, min_width=50):
+                        submit_btn = gr.Button(value="Send", variant="primary")
+                with gr.Row(elem_id="buttons") as _:
+                    regenerate_btn = gr.Button(
+                        value="🔄  Regenerate", interactive=True)
+                    clear_btn = gr.Button(value="🗑️  Clear", interactive=True)
+
+        gr.Markdown(tos_markdown)
+        gr.Markdown(learn_more_markdown)
+        gr.Markdown(ack_markdown)
+
+        regenerate_btn.click(
+            regenerate,
+            [state, image_process_mode],
+            [state, chatbot, textbox, imagebox],
+            queue=False
+        ).then(
+            http_bot,
+            [state, temperature, top_p, max_output_tokens],
+            [state, chatbot]
+        )
+
+        clear_btn.click(
+            clear_history,
+            None,
+            [state, chatbot, textbox, imagebox],
+            queue=False
+        )
+
+        textbox.submit(
+            add_text,
+            [state, textbox, imagebox, image_process_mode],
+            [state, chatbot, textbox, imagebox],
+            queue=False
+        ).then(
+            http_bot,
+            [state, temperature, top_p, max_output_tokens],
+            [state, chatbot]
+        )
+
+        submit_btn.click(
+            add_text,
+            [state, textbox, imagebox, image_process_mode],
+            [state, chatbot, textbox, imagebox],
+            queue=False
+        ).then(
+            http_bot,
+            [state, temperature, top_p, max_output_tokens],
+            [state, chatbot]
+        )
+
+        demo.load(
+            load_demo,
+            None,
+            [state],
+            queue=False
+        )
+    return demo
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="0.0.0.0")
+    parser.add_argument("--port", type=int, default=7860)
+    parser.add_argument("--share", default=True)
+    parser.add_argument("--model-path", type=str,
+                        default="checkpoints/llavaPhi-v0-3b-finetune")
+    parser.add_argument("--model-name", type=str,
+                        default="llavaPhi-v0-3b")
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    model_name = args.model_name
+    tokenizer, model, image_processor, context_len = load_pretrained_model(
+        args.model_path, None, args.model_name, False, False)
+    demo = build_demo()
+    demo.queue()
+    demo.launch(server_name=args.host,
+                server_port=args.port,
+                share=args.share)

llava-phi/llava_phi/serve/cli.py

@@ -0,0 +1,121 @@
+import argparse
+import torch
+
+from llava_phi.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
+from llava_phi.conversation import conv_templates, SeparatorStyle
+from llava_phi.model.builder import load_pretrained_model
+from llava_phi.utils import disable_torch_init
+from llava_phi.mm_utils import tokenizer_image_token, get_model_name_from_path, KeywordsStoppingCriteria
+
+from PIL import Image
+
+import requests
+from PIL import Image
+from io import BytesIO
+from transformers import TextStreamer
+
+
+def load_image(image_file):
+    if image_file.startswith('http') or image_file.startswith('https'):
+        response = requests.get(image_file)
+        image = Image.open(BytesIO(response.content)).convert('RGB')
+    else:
+        image = Image.open(image_file).convert('RGB')
+    return image
+
+
+def main(args):
+    # Model
+    disable_torch_init()
+
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name, args.load_8bit, args.load_4bit)
+
+    if 'llama-2' in model_name.lower():
+        conv_mode = "llava_llama_2"
+    elif "v1" in model_name.lower():
+        conv_mode = "llava_v1"
+    elif "mpt" in model_name.lower():
+        conv_mode = "mpt"
+    else:
+        conv_mode = "llava_v0"
+    conv_mode="vicuna_v1"
+    if args.conv_mode is not None and conv_mode != args.conv_mode:
+        print('[WARNING] the auto inferred conversation mode is {}, while `--conv-mode` is {}, using {}'.format(conv_mode, args.conv_mode, args.conv_mode))
+    else:
+        args.conv_mode = conv_mode
+
+    conv = conv_templates[args.conv_mode].copy()
+    if "mpt" in model_name.lower():
+        roles = ('user', 'assistant')
+    else:
+        roles = conv.roles
+
+    image = load_image(args.image_file)
+    image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'].cuda()
+
+    while True:
+        try:
+            inp = input(f"{roles[0]}: ")
+        except EOFError:
+            inp = ""
+        if not inp:
+            print("exit...")
+            break
+
+        print(f"{roles[1]}: ", end="")
+
+        if image is not None:
+            # first message
+            if model.config.mm_use_im_start_end:
+                inp = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + inp
+            else:
+                inp = DEFAULT_IMAGE_TOKEN + '\n' + inp
+            conv.append_message(conv.roles[0], inp)
+            image = None
+        else:
+            # later messages
+            conv.append_message(conv.roles[0], inp)
+        conv.append_message(conv.roles[1], None)
+        prompt = conv.get_prompt()
+
+        input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
+        stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
+        keywords = [stop_str]
+        stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids)
+        streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+
+        with torch.inference_mode():
+            output_ids = model.generate(
+                input_ids,
+                images=image_tensor,
+                do_sample=True,
+                temperature=0.2,
+                max_new_tokens=1024,
+                streamer=streamer,
+                use_cache=True,
+                eos_token_id=tokenizer.eos_token_id,  # End of sequence token
+                pad_token_id=tokenizer.eos_token_id,  # Pad token
+                stopping_criteria=[stopping_criteria])
+
+        outputs = tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
+        conv.messages[-1][-1] = outputs
+
+        if args.debug:
+            print("\n", {"prompt": prompt, "outputs": outputs}, "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
+    parser.add_argument("--model-base", type=str, default=None)
+    parser.add_argument("--image-file", type=str, required=True)
+    parser.add_argument("--num-gpus", type=int, default=1)
+    parser.add_argument("--conv-mode", type=str, default=None)
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--max-new-tokens", type=int, default=512)
+    parser.add_argument("--load-8bit", action="store_true")
+    parser.add_argument("--load-4bit", action="store_true")
+    parser.add_argument("--debug", action="store_true")
+    args = parser.parse_args()
+    main(args)

llava-phi/llava_phi/train/convert_model2base_llava_phi.py

@@ -0,0 +1,767 @@
+# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
+# Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
+#    Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+import os
+import copy
+from dataclasses import dataclass, field
+import json
+import logging
+import pathlib
+from typing import Dict, Optional, Sequence, List
+
+import torch
+
+import transformers
+
+from llava_phi.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, \
+    DEFAULT_IM_END_TOKEN
+from torch.utils.data import Dataset
+from llava_phi.train.llava_phi_trainer import LLaVAPhiTrainer
+
+from llava_phi import conversation as conversation_lib
+from llava_phi.model import *
+from llava_phi.mm_utils import tokenizer_image_token
+from transformers import CLIPVisionConfig, CLIPImageProcessor
+
+from PIL import Image
+
+local_rank = None
+
+
+def rank0_print(*args):
+    if local_rank == 0:
+        print(*args)
+
+
+@dataclass
+class ModelArguments:
+    model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
+    version: Optional[str] = field(default="v0")
+    freeze_backbone: bool = field(default=False)
+    tune_mm_mlp_adapter: bool = field(default=False)
+    freeze_vision_tower: bool = field(default=False)
+    vision_tower: Optional[str] = field(default=None)
+    mm_vision_select_layer: Optional[int] = field(default=-1)  # default to the last layer
+    mm_vision_select_feature: Optional[str] = field(default="patch")
+    pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
+    mm_use_im_start_end: bool = field(default=False)
+    mm_use_im_patch_token: bool = field(default=True)
+
+
+
+@dataclass
+class ProjectorArguments:
+    mm_projector_type: Optional[str] = field(default='linear')
+
+
+@dataclass
+class DataArguments:
+    data_path: str = field(default=None,
+                           metadata={"help": "Path to the training data."})
+    lazy_preprocess: bool = False
+    is_multimodal: bool = False
+    image_folder: Optional[str] = field(default=None)
+    image_aspect_ratio: str = 'square'
+
+
+@dataclass
+class TrainingArguments(transformers.TrainingArguments):
+    cache_dir: Optional[str] = field(default=None)
+    optim: str = field(default="adamw_torch")
+    adam_beta1: float = field(default=0.9)
+    adam_beta2: float = field(default=0.98)
+    adam_epsilon: float = field(default=1e-7)
+    remove_unused_columns: bool = field(default=False)
+
+    # freeze_mm_mlp_adapter: bool = field(default=False)
+    model_max_length: int = field(
+        default=512,
+        metadata={
+            "help":
+                "Maximum sequence length. Sequences will be right padded (and possibly truncated)."
+        },
+    )
+    double_quant: bool = field(
+        default=True,
+        metadata={"help": "Compress the quantization statistics through double quantization."}
+    )
+    quant_type: str = field(
+        default="nf4",
+        metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
+    )
+    bits: int = field(
+        default=16,
+        metadata={"help": "How many bits to use."}
+    )
+    lora_enable: bool = False
+    lora_r: int = 64
+    lora_alpha: int = 16
+    lora_dropout: float = 0.05
+    lora_weight_path: str = ""
+    lora_bias: str = "none"
+    mm_projector_lr: Optional[float] = None
+    group_by_modality_length: bool = field(default=False)
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+# Borrowed from peft.utils.get_peft_model_state_dict
+def get_peft_state_maybe_zero_3(named_params, bias):
+    if bias == "none":
+        to_return = {k: t for k, t in named_params if "lora_" in k}
+    elif bias == "all":
+        to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
+    elif bias == "lora_only":
+        to_return = {}
+        maybe_lora_bias = {}
+        lora_bias_names = set()
+        for k, t in named_params:
+            if "lora_" in k:
+                to_return[k] = t
+                bias_name = k.split("lora_")[0] + "bias"
+                lora_bias_names.add(bias_name)
+            elif "bias" in k:
+                maybe_lora_bias[k] = t
+        for k, t in maybe_lora_bias:
+            if bias_name in lora_bias_names:
+                to_return[bias_name] = t
+    else:
+        raise NotImplementedError
+    to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
+    return to_return
+
+
+def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
+    to_return = {k: t for k, t in named_params if "lora_" not in k}
+    if require_grad_only:
+        to_return = {k: t for k, t in to_return.items() if t.requires_grad}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def find_all_linear_names(model):
+    cls = torch.nn.Linear
+    lora_module_names = set()
+    multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler']
+    for name, module in model.named_modules():
+        if any(mm_keyword in name for mm_keyword in multimodal_keywords):
+            continue
+        if isinstance(module, cls):
+            names = name.split('.')
+            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+
+    if 'lm_head' in lora_module_names:  # needed for 16-bit
+        lora_module_names.remove('lm_head')
+    return list(lora_module_names)
+
+
+def safe_save_model_for_hf_trainer(trainer: transformers.Trainer,
+                                   output_dir: str):
+    """Collects the state dict and dump to disk."""
+
+    # if getattr(trainer.args, "tune_mm_mlp_adapter", False):
+    #     # Only save Adapter
+    #     keys_to_match = ['mm_projector']
+    #     if getattr(trainer.args, "use_im_start_end", False):
+    #         keys_to_match.extend(['embed_tokens', 'embed_in'])
+    #
+    #     weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
+    #     trainer.model.config.save_pretrained(output_dir)
+    #
+    #     current_folder = output_dir.split('/')[-1]
+    #     parent_folder = os.path.dirname(output_dir)
+    #     if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
+    #         if current_folder.startswith('checkpoint-'):
+    #             mm_projector_folder = os.path.join(parent_folder, "mm_projector")
+    #             os.makedirs(mm_projector_folder, exist_ok=True)
+    #             torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin'))
+    #         else:
+    #             torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
+    #
+    #     if getattr(trainer.args, "freeze_vision_model", False):
+    #         pass
+    #     return
+
+    if trainer.deepspeed:
+        torch.cuda.synchronize()
+        trainer.save_model(output_dir)
+        return
+
+    state_dict = trainer.model.state_dict()
+    if trainer.args.should_save:
+        cpu_state_dict = {
+            key: value.cpu()
+            for key, value in state_dict.items()
+        }
+        del state_dict
+        trainer._save(output_dir, state_dict=cpu_state_dict)  # noqa
+
+
+def smart_tokenizer_and_embedding_resize(
+        special_tokens_dict: Dict,
+        tokenizer: transformers.PreTrainedTokenizer,
+        model: transformers.PreTrainedModel,
+):
+    """Resize tokenizer and embedding.
+
+    Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
+    """
+    num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
+    model.resize_token_embeddings(len(tokenizer))
+
+    if num_new_tokens > 0:
+        input_embeddings = model.get_input_embeddings().weight.data
+        output_embeddings = model.get_output_embeddings().weight.data
+
+        input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
+            dim=0, keepdim=True)
+        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
+            dim=0, keepdim=True)
+
+        input_embeddings[-num_new_tokens:] = input_embeddings_avg
+        output_embeddings[-num_new_tokens:] = output_embeddings_avg
+
+
+def _tokenize_fn(strings: Sequence[str],
+                 tokenizer: transformers.PreTrainedTokenizer) -> Dict:
+    """Tokenize a list of strings."""
+    tokenized_list = [
+        tokenizer(
+            text,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ) for text in strings
+    ]
+    input_ids = labels = [
+        tokenized.input_ids[0] for tokenized in tokenized_list
+    ]
+    input_ids_lens = labels_lens = [
+        tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
+        for tokenized in tokenized_list
+    ]
+    return dict(
+        input_ids=input_ids,
+        labels=labels,
+        input_ids_lens=input_ids_lens,
+        labels_lens=labels_lens,
+    )
+
+
+def _mask_targets(target, tokenized_lens, speakers):
+    # cur_idx = 0
+    cur_idx = tokenized_lens[0]
+    tokenized_lens = tokenized_lens[1:]
+    target[:cur_idx] = IGNORE_INDEX
+    for tokenized_len, speaker in zip(tokenized_lens, speakers):
+        if speaker == "human":
+            target[cur_idx + 2:cur_idx + tokenized_len] = IGNORE_INDEX
+        cur_idx += tokenized_len
+
+
+def _add_speaker_and_signal(header, source, get_conversation=True):
+    """Add speaker and start/end signal on each round."""
+    BEGIN_SIGNAL = "### "
+    END_SIGNAL = "\n"
+    conversation = header
+    for sentence in source:
+        from_str = sentence["from"]
+        if from_str.lower() == "human":
+            from_str = conversation_lib.default_conversation.roles[0]
+        elif from_str.lower() == "gpt":
+            from_str = conversation_lib.default_conversation.roles[1]
+        else:
+            from_str = 'unknown'
+        sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
+                             sentence["value"] + END_SIGNAL)
+        if get_conversation:
+            conversation += sentence["value"]
+    conversation += BEGIN_SIGNAL
+    return conversation
+
+
+def preprocess_multimodal(
+        sources: Sequence[str],
+        data_args: DataArguments
+) -> Dict:
+    is_multimodal = data_args.is_multimodal
+    if not is_multimodal:
+        return sources
+
+    for source in sources:
+        for sentence in source:
+            if DEFAULT_IMAGE_TOKEN in sentence['value']:
+                sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '').strip()
+                sentence['value'] = DEFAULT_IMAGE_TOKEN + '\n' + sentence['value']
+                sentence['value'] = sentence['value'].strip()
+                if "mmtag" in conversation_lib.default_conversation.version:
+                    sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN,
+                                                                  '<Image>' + DEFAULT_IMAGE_TOKEN + '</Image>')
+            replace_token = DEFAULT_IMAGE_TOKEN
+            if data_args.mm_use_im_start_end:
+                replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
+            sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token)
+
+    return sources
+
+
+def preprocess_v0(
+        sources,
+        tokenizer: transformers.PreTrainedTokenizer,
+        has_image: bool = False
+) -> Dict:
+    conv = conversation_lib.default_conversation.copy()
+    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+    # Apply prompt templates
+    conversations = []
+    for i, source in enumerate(sources):
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], f"{i}"
+            conv.append_message(role, sentence["value"])
+        conversations.append(conv.get_prompt())
+
+    # Tokenize conversations
+    if has_image:
+        input_ids = torch.stack(
+            [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
+    else:
+        input_ids = tokenizer(
+            conversations,
+            return_tensors="pt",
+            padding="longest",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+        ).input_ids
+
+    targets = input_ids.clone()
+
+    assert conv.sep_style == conversation_lib.SeparatorStyle.TWO
+
+    # Mask targets
+    sep = conv.sep + conv.roles[1] + ": "
+    for conversation, target in zip(conversations, targets):
+        total_len = int(target.ne(tokenizer.pad_token_id).sum()) + conversation.count(
+            conv.sep2)  # in phi-2, pad_token_id == eos_token_id
+
+        rounds = conversation.split(conv.sep2)
+        cur_len = 0
+        if cur_len > 0:
+            target[:cur_len] = IGNORE_INDEX
+        for i, rou in enumerate(rounds):
+            if rou == "":
+                break
+
+            parts = rou.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+
+            if has_image:
+                round_len = len(tokenizer_image_token(rou, tokenizer)) + 1  # +1 for <|endoftext|>
+                instruction_len = len(tokenizer_image_token(parts[0], tokenizer))
+            else:
+                round_len = len(tokenizer(rou).input_ids) + 1  # +1 for <|endoftext|>
+                instruction_len = len(tokenizer(parts[0]).input_ids)
+
+            target[cur_len: cur_len + instruction_len] = IGNORE_INDEX
+
+            cur_len += round_len
+        target[cur_len:] = IGNORE_INDEX
+
+        if cur_len < tokenizer.model_max_length:
+            if cur_len != total_len:
+                target[:] = IGNORE_INDEX
+                print(conversation)
+                print(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+    return dict(
+        input_ids=input_ids,
+        labels=targets,
+    )
+
+
+def preprocess_plain(
+        sources: Sequence[str],
+        tokenizer: transformers.PreTrainedTokenizer,
+) -> Dict:
+    # add end signal and concatenate together
+    conversations = []
+    # print(sources)
+    # time.sleep(5)
+    for source in sources:
+        assert len(source) == 2
+        assert DEFAULT_IMAGE_TOKEN in source[0]['value']
+        source[0]['value'] = DEFAULT_IMAGE_TOKEN
+        conversation = source[0]['value'] + source[1]['value'] + conversation_lib.default_conversation.sep
+        conversations.append(conversation)
+    # tokenize conversations
+    # print(conversations)
+    input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        tokenized_len = len(tokenizer_image_token(source[0]['value'], tokenizer))
+        target[:tokenized_len] = IGNORE_INDEX
+    return dict(input_ids=input_ids, labels=targets)
+
+
+def preprocess(
+        sources: Sequence[str],
+        tokenizer: transformers.PreTrainedTokenizer,
+        has_image: bool = False
+) -> Dict:
+    """
+    Given a list of sources, each is a conversation list. This transform:
+    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
+    2. Concatenate conversations together;
+    3. Tokenize the concatenated conversation;
+    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
+    """
+    if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
+        return preprocess_plain(sources, tokenizer)
+    if conversation_lib.default_conversation.version.startswith("v0"):
+        return preprocess_v0(sources, tokenizer, has_image=has_image)
+    # add end signal and concatenate together
+    conversations = []
+    for source in sources:
+        header = f"{conversation_lib.default_conversation.system}\n\n"
+        conversation = _add_speaker_and_signal(header, source)
+        conversations.append(conversation)
+
+    # tokenize conversations
+    def get_tokenize_len(prompts):
+        return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts]
+
+    if has_image:
+        input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
+    else:
+        conversations_tokenized = _tokenize_fn(conversations, tokenizer)
+        input_ids = conversations_tokenized["input_ids"]
+
+    targets = copy.deepcopy(input_ids)
+    for target, source in zip(targets, sources):
+        if has_image:
+            tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source])
+        else:
+            tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"]
+        speakers = [sentence["from"] for sentence in source]
+        _mask_targets(target, tokenized_lens, speakers)
+
+    return dict(input_ids=input_ids, labels=targets)
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(self, data_path: str,
+                 tokenizer: transformers.PreTrainedTokenizer,
+                 data_args: DataArguments):
+        super(LazySupervisedDataset, self).__init__()
+        list_data_dict = json.load(open(data_path, "r"))
+
+        rank0_print("Formatting inputs...Skip in lazy mode")
+        self.tokenizer = tokenizer
+        self.list_data_dict = list_data_dict
+        self.data_args = data_args
+
+    def __len__(self):
+        return len(self.list_data_dict)
+
+    @property
+    def lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            img_tokens = 128 if 'image' in sample else 0
+            length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens)
+        return length_list
+
+    @property
+    def modality_lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            cur_len = sum(len(conv['value'].split()) for conv in sample['conversations'])
+            cur_len = cur_len if 'image' in sample else -cur_len
+            length_list.append(cur_len)
+        return length_list
+
+    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
+        sources = self.list_data_dict[i]
+        if isinstance(i, int):
+            sources = [sources]
+        assert len(sources) == 1, "Don't know why it is wrapped to a list"  # FIXME
+        if 'image' in sources[0]:
+            image_file = self.list_data_dict[i]['image']
+            image_folder = self.data_args.image_folder
+            processor = self.data_args.image_processor
+            image = Image.open(os.path.join(image_folder, image_file)).convert('RGB')
+            if self.data_args.image_aspect_ratio == 'pad':
+                def expand2square(pil_img, background_color):
+                    width, height = pil_img.size
+                    if width == height:
+                        return pil_img
+                    elif width > height:
+                        result = Image.new(pil_img.mode, (width, width), background_color)
+                        result.paste(pil_img, (0, (width - height) // 2))
+                        return result
+                    else:
+                        result = Image.new(pil_img.mode, (height, height), background_color)
+                        result.paste(pil_img, ((height - width) // 2, 0))
+                        return result
+
+                image = expand2square(image, tuple(int(x * 255) for x in processor.image_mean))
+                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            else:
+                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
+            sources = preprocess_multimodal(
+                copy.deepcopy([e["conversations"] for e in sources]),
+                self.data_args)
+        else:
+            sources = copy.deepcopy([e["conversations"] for e in sources])
+        data_dict = preprocess(
+            sources,
+            self.tokenizer,
+            has_image=('image' in self.list_data_dict[i]))
+        if isinstance(i, int):
+            data_dict = dict(input_ids=data_dict["input_ids"][0],
+                             labels=data_dict["labels"][0])
+
+        # image exist in the data
+        if 'image' in self.list_data_dict[i]:
+            data_dict['image'] = image
+        elif self.data_args.is_multimodal:
+            # image does not exist in the data, but the model is multimodal
+            crop_size = self.data_args.image_processor.crop_size
+            data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width'])
+        return data_dict
+
+
+@dataclass
+class DataCollatorForSupervisedDataset(object):
+    """Collate examples for supervised fine-tuning."""
+
+    tokenizer: transformers.PreTrainedTokenizer
+
+    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
+        input_ids, labels = tuple([instance[key] for instance in instances]
+                                  for key in ("input_ids", "labels"))
+        temp_pad_token_id = 51000
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=temp_pad_token_id)
+        labels = torch.nn.utils.rnn.pad_sequence(labels,
+                                                 batch_first=True,
+                                                 padding_value=IGNORE_INDEX)
+        input_ids = input_ids[:, :self.tokenizer.model_max_length]
+        labels = labels[:, :self.tokenizer.model_max_length]
+        batch = dict(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=input_ids.ne(temp_pad_token_id),
+        )
+
+        if 'image' in instances[0]:
+            images = [instance['image'] for instance in instances]
+            if all(x is not None and x.shape == images[0].shape for x in images):
+                batch['images'] = torch.stack(images)
+            else:
+                batch['images'] = images
+
+        return batch
+
+
+def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer,
+                                data_args) -> Dict:
+    """Make dataset and collator for supervised fine-tuning."""
+    train_dataset = LazySupervisedDataset(tokenizer=tokenizer,
+                                          data_path=data_args.data_path,
+                                          data_args=data_args)
+    data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
+    return dict(train_dataset=train_dataset,
+                eval_dataset=None,
+                data_collator=data_collator)
+
+
+def train():
+    global local_rank
+
+    parser = transformers.HfArgumentParser(
+        (ModelArguments, DataArguments, TrainingArguments, ProjectorArguments))
+    model_args, data_args, training_args, projector_args = parser.parse_args_into_dataclasses()
+    local_rank = training_args.local_rank
+    compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
+
+    bnb_model_from_pretrained_args = {}
+    if training_args.bits in [4, 8]:
+        from transformers import BitsAndBytesConfig
+        bnb_model_from_pretrained_args.update(dict(
+            device_map={"": training_args.device},
+            load_in_4bit=training_args.bits == 4,
+            load_in_8bit=training_args.bits == 8,
+            quantization_config=BitsAndBytesConfig(
+                load_in_4bit=training_args.bits == 4,
+                load_in_8bit=training_args.bits == 8,
+                llm_int8_skip_modules=["mm_projector"],
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=False,
+                bnb_4bit_compute_dtype=compute_dtype,
+                bnb_4bit_use_double_quant=training_args.double_quant,
+                bnb_4bit_quant_type=training_args.quant_type  # {'fp4', 'nf4'}
+            )
+        ))
+
+    if model_args.vision_tower is not None:
+        config = LlavaPhiConfig.from_pretrained(model_args.model_name_or_path, trust_remote_code=True)
+        clip_config = CLIPVisionConfig.from_pretrained(model_args.vision_tower)
+        vis_config = LlavaPhiVisionConfig(**clip_config.to_dict())
+        config.vision_config["vision_tower"] = vis_config.to_dict()
+        config.vision_config["vision_tower"]["mm_vision_select_feature"] = model_args.mm_vision_select_feature
+        config.vision_config["vision_tower"]["mm_vision_select_layer"] = model_args.mm_vision_select_layer
+
+        config.vision_config["mm_projector"]["mm_projector_type"] = projector_args.mm_projector_type
+        config.vision_config["mm_projector"]["mm_hidden_size"] = vis_config.hidden_size
+        config.vision_config["mm_projector"]["hidden_size"] = config.hidden_size
+
+        model = LlavaPhiForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            config=config,
+            cache_dir=training_args.cache_dir,
+            trust_remote_code=True,
+            **bnb_model_from_pretrained_args
+        )
+        rank0_print(model)
+        clip_model_param = torch.load(os.path.join(model_args.vision_tower, "pytorch_model.bin"), map_location='cpu')
+        model.get_model().vision_tower.load_state_dict(clip_model_param, strict=False)
+    else:
+        model = transformers.PhiForCausalLM.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            **bnb_model_from_pretrained_args
+        )
+    model.config.use_cache = False
+
+    if model_args.freeze_backbone:
+        model.model.requires_grad_(False)
+
+    if training_args.gradient_checkpointing:
+        if hasattr(model, "enable_input_require_grads"):
+            model.enable_input_require_grads()
+        else:
+            def make_inputs_require_grad(module, input, output):
+                output.requires_grad_(True)
+
+            model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+
+    if 'phi' in model_args.model_name_or_path:
+        tokenizer = transformers.AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            model_max_length=training_args.model_max_length,
+            padding_side="right"
+        )
+    else:
+        tokenizer = transformers.AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path,
+            cache_dir=training_args.cache_dir,
+            model_max_length=training_args.model_max_length,
+            padding_side="right",
+            use_fast=False,
+        )
+
+    tokenizer.pad_token = tokenizer.unk_token
+    if model_args.version in conversation_lib.conv_templates:
+        conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version]
+    else:
+        conversation_lib.default_conversation = conversation_lib.conv_templates["phi-2_v0"]
+
+    assert model_args.vision_tower is not None, "llava_phi-phi only supports multi-modal models"
+    if model_args.vision_tower is not None:
+
+        vision_tower = model.get_vision_tower()
+        vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device)
+
+        data_args.image_processor = CLIPImageProcessor.from_pretrained(model_args.vision_tower)
+        data_args.is_multimodal = True
+
+        model.config.image_aspect_ratio = data_args.image_aspect_ratio
+        model.config.tokenizer_padding_side = tokenizer.padding_side
+        model.config.tokenizer_model_max_length = tokenizer.model_max_length
+
+        model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
+        model.requires_grad_(False)
+        if model_args.tune_mm_mlp_adapter:
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = True
+
+        model.config.freeze_vision_tower = training_args.freeze_vision_tower = model_args.freeze_vision_tower
+        if not model_args.freeze_vision_tower:
+            for p in model.get_model().vision_tower.parameters():
+                p.requires_grad = True
+
+        if training_args.bits in [4, 8]:
+            model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device)
+
+        model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
+        model.config.mm_projector_lr = training_args.mm_projector_lr
+        training_args.use_im_start_end = model_args.mm_use_im_start_end
+        model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token
+        model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)
+
+    data_module = make_supervised_data_module(tokenizer=tokenizer,
+                                              data_args=data_args)
+
+    trainer = LLaVAPhiTrainer(model=model,
+                              tokenizer=tokenizer,
+                              args=training_args,
+                              **data_module)
+    # integrate the MLLM
+    trainer.save_state()
+
+    model.config.use_cache = True
+
+    safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir)
+
+
+if __name__ == "__main__":
+    train()

llava-phi/llava_phi/train/llava_phi_trainer.py

@@ -0,0 +1,156 @@
+import os
+import torch
+
+from torch.utils.data import Sampler
+
+from transformers import Trainer
+from transformers.trainer import (
+    has_length,
+)
+from typing import List, Optional
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                print(name, 'no ignore status')
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True, name=k).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def split_to_even_chunks(indices, lengths, num_chunks):
+    """
+    Split a list of indices into `chunks` chunks of roughly equal lengths.
+    """
+
+    if len(indices) % num_chunks != 0:
+        return [indices[i::num_chunks] for i in range(num_chunks)]
+
+    num_indices_per_chunk = len(indices) // num_chunks
+
+    chunks = [[] for _ in range(num_chunks)]
+    chunks_lengths = [0 for _ in range(num_chunks)]
+    for index in indices:
+        shortest_chunk = chunks_lengths.index(min(chunks_lengths))
+        chunks[shortest_chunk].append(index)
+        chunks_lengths[shortest_chunk] += lengths[index]
+        if len(chunks[shortest_chunk]) == num_indices_per_chunk:
+            chunks_lengths[shortest_chunk] = float("inf")
+
+    return chunks
+
+
+def get_modality_length_grouped_indices(lengths, batch_size, world_size, generator=None):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    assert all(l != 0 for l in lengths), "Should not have zero length."
+    # assert all(l > 0 for l in lengths) or all(l < 0 for l in lengths), "Should have only positive or negative lengths."
+
+    mm_indices, mm_lengths = zip(*[(i, l) for i, l in enumerate(lengths) if l > 0])
+    lang_indices, lang_lengths = zip(*[(i, -l) for i, l in enumerate(lengths) if l < 0])
+
+    assert len(mm_indices) > 0, "Should have at least one multimodal sample."
+    assert len(lang_indices) > 0, "Should have at least one language sample."
+
+    mm_shuffle = [mm_indices[i] for i in get_length_grouped_indices(mm_lengths, batch_size, world_size, generator=None)]
+    lang_shuffle = [lang_indices[i] for i in get_length_grouped_indices(lang_lengths, batch_size, world_size, generator=None)]
+    megabatch_size = world_size * batch_size
+    mm_megabatches = [mm_shuffle[i : i + megabatch_size] for i in range(0, len(mm_shuffle), megabatch_size)]
+    lang_megabatches = [lang_shuffle[i : i + megabatch_size] for i in range(0, len(lang_shuffle), megabatch_size)]
+
+    last_mm = mm_megabatches[-1]
+    last_lang = lang_megabatches[-1]
+    additional_batch = last_mm + last_lang
+    megabatches = mm_megabatches[:-1] + lang_megabatches[:-1]
+    megabatch_indices = torch.randperm(len(megabatches), generator=generator)
+    megabatches = [megabatches[i] for i in megabatch_indices]
+
+    if len(additional_batch) >= megabatch_size:
+        megabatches = [additional_batch[:megabatch_size]] + megabatches
+        additional_batch = additional_batch[megabatch_size:]
+
+    if len(additional_batch) > 0:
+        megabatches.append(additional_batch)
+
+    return [i for megabatch in megabatches for i in megabatch]
+
+
+def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    indices = torch.randperm(len(lengths), generator=generator)
+    megabatch_size = world_size * batch_size
+    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
+    megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches]
+    megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches]
+
+    return [i for megabatch in megabatches for batch in megabatch for i in batch]
+
+
+class LengthGroupedSampler(Sampler):
+    r"""
+    Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while
+    keeping a bit of randomness.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        world_size: int,
+        lengths: Optional[List[int]] = None,
+        generator=None,
+        group_by_modality: bool = False,
+    ):
+        if lengths is None:
+            raise ValueError("Lengths must be provided.")
+
+        self.batch_size = batch_size
+        self.world_size = world_size
+        self.lengths = lengths
+        self.generator = generator
+        self.group_by_modality = group_by_modality
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def __iter__(self):
+        if self.group_by_modality:
+            indices = get_modality_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        else:
+            indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        return iter(indices)
+
+
+class LLaVAPhiTrainer(Trainer):
+
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+        if self.train_dataset is None or not has_length(self.train_dataset):
+            return None
+
+        if self.args.group_by_modality_length:
+            lengths = self.train_dataset.modality_lengths
+            return LengthGroupedSampler(
+                # self.args.train_batch_size * self.args.gradient_accumulation_steps, # TODO: seems that we should not have gradient_accumulation_steps
+                self.args.train_batch_size,
+                world_size=self.args.world_size,
+                lengths=lengths,
+                group_by_modality=True,
+            )
+        else:
+            return super()._get_train_sampler()
+
+    def _save_checkpoint(self, model, trial, metrics=None):
+        super(LLaVAPhiTrainer, self)._save_checkpoint(model, trial, metrics)
+
+    def _save(self, output_dir: Optional[str] = None, state_dict=None):
+            super(LLaVAPhiTrainer, self)._save(output_dir, state_dict)