{ "nbformat": 4, "nbformat_minor": 0, "metadata": { "colab": { "provenance": [], "gpuType": "A100" }, "kernelspec": { "name": "python3", "display_name": "Python 3" }, "accelerator": "GPU" }, "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# AI 新視界 Colab 實測 1：用 Google AI Pro 送的運算單元跑本站的 AI 主播（MiniMax H3 ref2va）\n", "做的事：裝 ComfyUI → 下載 H3 ref2va 的四個權重檔（約 43 GB，Comfy-Org 公開倉庫，不用登入）→ 用本站固定的主播圖＋一段台灣口音旁白，生成一段對嘴影片 → 記錄每一步時間、GPU 尖峰記憶體、運算單元。\n", "\n", "怎麼用：右上角先選 GPU（A100 優先，沒有就 L4），然後「執行階段 → 全部執行」。中途要做的只有：拿到 GPU 後，把「資源」面板上的「每小時使用 N 個運算單元」填到第 1 格的 UNITS_PER_HOUR，再重跑第 1 格（或一開始就填）。最後把第 8 格印出的整段文字貼回來。\n", "\n", "本站 A6000 的對照值：8 秒 480×864、20 步約 10 分鐘；GB10 約 15 分鐘。" ] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 1. 看看拿到什麼機器（先在右上角「執行階段 → 變更執行階段類型」選 GPU；建議 A100，拿不到就 L4）\n", "import subprocess, time, json, os, shutil, re\n", "T0 = time.time()\n", "UNITS_PER_HOUR = 0 #@param {type:\"number\"}\n", "VARIANTS = \"pruned_int8\" #@param [\"pruned_int8\", \"int8\", \"bf16\", \"pruned_int8,int8,bf16\", \"int8,bf16\"]\n", "#@markdown ↑ 要跑哪些版本（逗號分隔會依序跑）：pruned_int8＝本站地端用的修剪版；int8＝未修剪；bf16＝完整版（66 GB 權重＋51 GB 文字編碼器，要 A100 80 GB）。\n", "#@markdown ↑ 把右上角「資源」面板顯示的「每小時使用 N 個運算單元」填進來（拿到 GPU 後才會顯示），用來換算每支影片的成本；不填就只記時間。\n", "gpu = subprocess.run([\"nvidia-smi\",\"--query-gpu=name,memory.total,driver_version\",\"--format=csv,noheader\"],capture_output=True,text=True).stdout.strip()\n", "ram = int(next(l for l in open(\"/proc/meminfo\") if l.startswith(\"MemTotal\")).split()[1])//2**20\n", "disk = shutil.disk_usage(\"/content\").free//2**30\n", "import torch; print(\"GPU:\", gpu); print(f\"RAM {ram} GB, 可用磁碟 {disk} GB, torch {torch.__version__}, cuda {torch.version.cuda}\")\n", "RESULT = {\"gpu\": gpu, \"ram_gb\": ram, \"disk_free_gb\": disk, \"torch\": torch.__version__, \"units_per_hour\": UNITS_PER_HOUR, \"t\": {}}\n", "def mark(k): RESULT[\"t\"][k] = round(time.time()-T0); print(f\"[{RESULT['t'][k]:>5}s] {k}\")\n", "mark(\"probe\")" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 2. 掛 Google Drive（結果影片與紀錄存到 MyDrive/aiwv-colab/，跑完才不會隨機器消失）\n", "from google.colab import drive\n", "drive.mount(\"/content/drive\")\n", "OUT = \"/content/drive/MyDrive/aiwv-colab\"; os.makedirs(OUT, exist_ok=True)\n", "mark(\"drive\")" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 3. 安裝 ComfyUI（和本站 A6000 上的同一套工具鏈）\n", "%cd /content\n", "if not os.path.exists(\"/content/ComfyUI\"):\n", " !git clone -q --depth 1 https://github.com/comfyanonymous/ComfyUI.git\n", "%cd /content/ComfyUI\n", "!pip install -q -r requirements.txt 2>&1 | tail -2\n", "!pip install -q hf_transfer huggingface_hub 2>&1 | tail -1\n", "mark(\"install_comfyui\")" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 4. 下載 H3 權重（Comfy-Org/MiniMax-H3；修剪版 43 GB、未修剪 67 GB、完整版 123 GB；記下載速度）\n", "os.environ[\"HF_HUB_ENABLE_HF_TRANSFER\"] = \"1\"\n", "from huggingface_hub import hf_hub_download\n", "# 每個版本配對的 ref2va 權重與文字編碼器（修剪版配 nvfp4，和本站地端完全相同；未修剪配 int8；完整版配 bf16）\n", "VAR = {\"pruned_int8\": (\"minimax_h3_ref2va_pruned_int8_convrot.safetensors\", \"qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors\"),\n", " \"int8\": (\"minimax_h3_ref2va_int8_convrot.safetensors\", \"qwen3vl_32b_minimax_h3_int8_convrot.safetensors\"),\n", " \"bf16\": (\"minimax_h3_ref2va_bf16.safetensors\", \"qwen3vl_32b_minimax_h3_bf16.safetensors\")}\n", "RUN = [v.strip() for v in VARIANTS.split(\",\") if v.strip() in VAR]\n", "FILES = [(\"vae/minimax_h3_video_vae_fp16.safetensors\", \"vae\"), (\"vae/minimax_h3_audio_vae_fp32.safetensors\", \"vae\")]\n", "for v in RUN:\n", " FILES += [(f\"diffusion_models/{VAR[v][0]}\", \"diffusion_models\"), (f\"text_encoders/{VAR[v][1]}\", \"text_encoders\")]\n", "FILES = list(dict.fromkeys(FILES))\n", "print(\"要跑的版本：\", RUN)\n", "t = time.time(); total = 0\n", "for f, sub in FILES:\n", " p = hf_hub_download(\"Comfy-Org/MiniMax-H3\", f, local_dir=\"/content/ComfyUI/models\")\n", " total += os.path.getsize(p); print(f\" {f} {os.path.getsize(p)/2**30:.1f} GB\")\n", "dl = time.time()-t\n", "RESULT[\"download_gb\"] = round(total/2**30, 1); RESULT[\"download_mbps\"] = round(total/2**20/dl)\n", "print(f\"下載 {RESULT['download_gb']} GB 用 {dl:.0f} 秒 ≈ {RESULT['download_mbps']} MB/s\"); mark(\"download\")" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 5. 啟動 ComfyUI（背景）並等它就緒\n", "import urllib.request\n", "log = open(\"/content/comfyui.log\",\"w\")\n", "P = subprocess.Popen([\"python\",\"main.py\",\"--listen\",\"127.0.0.1\",\"--port\",\"8188\",\"--disable-auto-launch\"], cwd=\"/content/ComfyUI\", stdout=log, stderr=subprocess.STDOUT)\n", "for _ in range(120):\n", " time.sleep(2)\n", " try:\n", " urllib.request.urlopen(\"http://127.0.0.1:8188/system_stats\", timeout=2).read(); break\n", " except Exception: pass\n", "else:\n", " print(open(\"/content/comfyui.log\").read()[-3000:]); raise SystemExit(\"ComfyUI 沒起來\")\n", "print(\"ComfyUI 就緒\"); mark(\"comfyui_ready\")\n", "\n", "import threading\n", "PEAK = {\"mb\": 0, \"stop\": False}\n", "def _watch():\n", " while not PEAK[\"stop\"]:\n", " try:\n", " mb = int(subprocess.run([\"nvidia-smi\",\"--query-gpu=memory.used\",\"--format=csv,noheader,nounits\"],capture_output=True,text=True).stdout.split()[0]); PEAK[\"mb\"] = max(PEAK[\"mb\"], mb)\n", " except Exception: pass\n", " time.sleep(3)\n", "threading.Thread(target=_watch, daemon=True).start()\n", "\n", "def submit(graph):\n", " req = urllib.request.Request(\"http://127.0.0.1:8188/prompt\", data=json.dumps({\"prompt\": graph}).encode(), headers={\"Content-Type\": \"application/json\"})\n", " with urllib.request.urlopen(req) as r: return json.load(r)[\"prompt_id\"]\n", "def wait(pid, every=5):\n", " t0 = time.time()\n", " while True:\n", " with urllib.request.urlopen(f\"http://127.0.0.1:8188/history/{pid}\") as r: h = json.load(r)\n", " if pid in h:\n", " st = h[pid].get(\"status\", {}); outs = []\n", " for n in h[pid].get(\"outputs\", {}).values():\n", " for k in (\"images\",\"video\",\"audio\",\"gifs\"):\n", " for f in n.get(k, []): outs.append(f\"{f.get('subfolder','')}/{f['filename']}\".lstrip(\"/\"))\n", " return st.get(\"status_str\"), round(time.time()-t0), outs, st.get(\"messages\", [])\n", " time.sleep(every)" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 6. 抓本站的主播圖與旁白（公開測試檔），算影格數\n", "for name in (\"persona-f.png\", \"narration-10s.wav\"):\n", " urllib.request.urlretrieve(f\"https://aiview.tw/media/lab/colab/{name}\", f\"/content/ComfyUI/input/{name}\")\n", "import wave\n", "with wave.open(\"/content/ComfyUI/input/narration-10s.wav\") as w: dur = w.getnframes()/w.getframerate()\n", "PAD, W, H, STEPS, SEED = 0.6, 480, 864, 20, 1\n", "n = max(5, round((dur + PAD) * 24))\n", "while n % 17 != 5: n += 1\n", "print(f\"旁白 {dur:.2f} 秒 → {n} 影格（{n/24:.2f} 秒）{W}x{H} {STEPS} 步\")\n", "RESULT.update({\"audio_s\": round(dur,2), \"frames\": n, \"size\": f\"{W}x{H}\", \"steps\": STEPS})" ], "execution_count": null, "outputs": [] }, { "cell_type": "code", "metadata": {}, "source": [ "#@title 7. 生成（和本站 host.py 同一張圖、同一個提示、同樣 20 步與種子）；每個版本跑一次，跑完先卸載模型\n", "PROMPT = (\"Use as the presenter and