{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "6511a91c-ed20-41ff-befb-699bda1912a3",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-03-25T05:42:29.023013Z",
     "iopub.status.busy": "2026-03-25T05:42:29.022863Z",
     "iopub.status.idle": "2026-03-25T05:42:40.880280Z",
     "shell.execute_reply": "2026-03-25T05:42:40.879248Z",
     "shell.execute_reply.started": "2026-03-25T05:42:29.022998Z"
    },
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "8310548c3b0d460899adcb96ee4af2e1",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Downloading (incomplete total...): 0.00B [00:00, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "663ea1161c934235a53948b93d224495",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Fetching 2 files:   0%|          | 0/2 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "667df34dda224931ac9ccd442a5d42f0",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading weights:   0%|          | 0/824 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[text] batch=0, tokens=4, pos=0..3 (t=h=w): [0, 1, 2, 3]\n",
      "get_vision_position_ids: grid_thw=tensor([ 1, 18, 18], device='cuda:0'), llm_grid_thw=(1, 9, 9), start_position=4\n",
      "  temp_merge_size=1, spatial_merge_size=2\n",
      "  image_seq_length=81\n",
      "  position_width (repeat)=[4, 5, 6, 7, 8, 9, 10, 11, 12, 4]...[12, 4, 5, 6, 7, 8, 9, 10, 11, 12]\n",
      "  position_height (repeat_interleave)=[4, 4, 4, 4, 4, 4, 4, 4, 4, 5]...[11, 12, 12, 12, 12, 12, 12, 12, 12, 12]\n",
      "  position_temporal (torch.full) (before spacing)=[4, 4, 4, 4, 4, 4, 4, 4, 4, 4]...[4, 4, 4, 4, 4, 4, 4, 4, 4, 4]\n",
      "  time_interval=2\n",
      "  position_temporal (after spacing)=[8, 8, 8, 8, 8, 8, 8, 8, 8, 8]...[8, 8, 8, 8, 8, 8, 8, 8, 8, 8]\n",
      "[vision pos] grid_thw=tensor([ 1, 18, 18], device='cuda:0'), start=4\n",
      "  t: [8, 8, 8, 8, 8, 8, 8, 8, 8, 8]...[8, 8, 8, 8, 8, 8, 8, 8, 8, 8]\n",
      "  h: [4, 4, 4, 4, 4, 4, 4, 4, 4, 5]...[11, 12, 12, 12, 12, 12, 12, 12, 12, 12]\n",
      "  w: [4, 5, 6, 7, 8, 9, 10, 11, 12, 4]...[12, 4, 5, 6, 7, 8, 9, 10, 11, 12]\n",
      "[text] batch=0, tokens=9, pos=13..21 (t=h=w): [13, 14, 15, 16, 17, 18, 19, 20, 21]\n",
      "[LLM prefill] position_ids shape: torch.Size([3, 1, 94])  (3=t/h/w, bs, seq_len)\n",
      "  batch 0 (shape: 94):\n",
      "    t: [0, 1, 2, 3, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 13, 14, 15, 16, 17, 18, 19, 20, 21] \n",
      "    h: [0, 1, 2, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10, 10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21] \n",
      "    w: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21] \n"
     ]
    },
    {
     "ename": "SystemExit",
     "evalue": "Debugging: Terminate after 1st decoder saved cos and sin tensors.",
     "output_type": "error",
     "traceback": [
      "An exception has occurred, use %tb to see the full traceback.\n",
      "\u001b[31mSystemExit\u001b[39m\u001b[31m:\u001b[39m Debugging: Terminate after 1st decoder saved cos and sin tensors.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/ubuntu/miniconda3/envs/dc_airnd/lib/python3.12/site-packages/IPython/core/interactiveshell.py:3755: UserWarning: To exit: use 'exit', 'quit', or Ctrl-D.\n",
      "  warn(\"To exit: use 'exit', 'quit', or Ctrl-D.\", stacklevel=1)\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor\n",
    "from qwen_vl_utils import process_vision_info\n",
    "\n",
    "# 1. Load Model and Processor\n",
    "model_name = \"Qwen/Qwen2.5-VL-3B-Instruct\"\n",
    "model = Qwen2_5_VLForConditionalGeneration.from_pretrained(\n",
    "    model_name, torch_dtype=torch.float16, device_map=\"auto\"\n",
    ")\n",
    "processor = AutoProcessor.from_pretrained(model_name)\n",
    "\n",
    "# 2. Define your inputs manually\n",
    "image_url = \"./car-1_256_0.jpg\"\n",
    "user_query = \"Describe the image\"\n",
    "\n",
    "# 3. Construct the prompt string manually\n",
    "# Qwen2.5-VL expects specific tokens to wrap system, user, and assistant roles.\n",
    "# Note: The <|vision_start|> and <|vision_end|> tags tell the processor \n",
    "# where to inject the image features.\n",
    "prompt = (\n",
    "    \"<|im_start|>user\\n\"\n",
    "    \"<|vision_start|><|image_pad|><|vision_end|>\"\n",
    "    f\"{user_query}<|im_end|>\\n\"\n",
    "    \"<|im_start|>assistant\\n\"\n",
    ")\n",
    "\n",
    "# 4. Process the vision information\n",
    "# We still use this utility to fetch the image and handle resizing logic\n",
    "messages = [{\"role\": \"user\", \"content\": [{\"type\": \"image\", \"image\": image_url}]}]\n",
    "image_inputs, _ = process_vision_info(messages)\n",
    "\n",
    "# 5. Tokenize and Prepare Tensors\n",
    "inputs = processor(\n",
    "    text=[prompt],\n",
    "    images=image_inputs,\n",
    "    videos=None,\n",
    "    padding=True,\n",
    "    return_tensors=\"pt\",\n",
    ")\n",
    "inputs = inputs.to(model.device)\n",
    "\n",
    "# 6. Generate\n",
    "generated_ids = model.generate(**inputs, max_new_tokens=100)\n",
    "\n",
    "# Trim the prompt tokens from the result\n",
    "generated_ids_trimmed = [\n",
    "    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)\n",
    "]\n",
    "\n",
    "output_text = processor.batch_decode(\n",
    "    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False\n",
    ")\n",
    "\n",
    "print(f\"\\nManual Prompt Response: {output_text[0]}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "f45df021-6302-4f47-9e06-8070577885a2",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-03-25T04:36:13.766580Z",
     "iopub.status.busy": "2026-03-25T04:36:13.766400Z",
     "iopub.status.idle": "2026-03-25T04:36:13.770145Z",
     "shell.execute_reply": "2026-03-25T04:36:13.769588Z",
     "shell.execute_reply.started": "2026-03-25T04:36:13.766563Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'<|im_start|>user\\n<|vision_start|><|image_pad|><|vision_end|>Describe the image<|im_end|>\\n<|im_start|>assistant\\n'"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "prompt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "504fa71b-42b4-4f53-8988-25fcfba38d13",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-03-25T05:43:53.839325Z",
     "iopub.status.busy": "2026-03-25T05:43:53.839044Z",
     "iopub.status.idle": "2026-03-25T05:43:53.843214Z",
     "shell.execute_reply": "2026-03-25T05:43:53.842555Z",
     "shell.execute_reply.started": "2026-03-25T05:43:53.839304Z"
    }
   },
   "outputs": [],
   "source": [
    "cos = torch.load('cos.pt')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "642d9dcf-e591-4d70-96af-b69bf955d9e1",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-03-25T05:43:54.296041Z",
     "iopub.status.busy": "2026-03-25T05:43:54.295869Z",
     "iopub.status.idle": "2026-03-25T05:43:54.299276Z",
     "shell.execute_reply": "2026-03-25T05:43:54.298634Z",
     "shell.execute_reply.started": "2026-03-25T05:43:54.296029Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(torch.Size([1, 1, 94, 128]), torch.float16)"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cos.shape, cos.dtype"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f44460e3-58e9-4fd2-898a-06e8a00f9365",
   "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": "3.12.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}