{ "cells": [ { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/home/christian/miniconda3/envs/suno_env/lib/python3.10/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", " from .autonotebook import tqdm as notebook_tqdm\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "WILL USE FLASH ATTN: True\n", "Failed to import Bio.\n" ] } ], "source": [ "import os\n", "import json\n", "import random\n", "import uuid\n", "import torchaudio\n", "import numpy as np\n", "import torch.nn.functional as F\n", "from suno_utils.audio import Audio\n", "from suno_utils.gpt.chirp_v2_5 import (\n", " GenerationConfig,\n", " decode_stream,\n", " preload_models,\n", " prep_gconf,\n", " codec_decode_stream_to_full_audio,\n", " generate,\n", ")\n", "from suno_utils.gpt.generation_engine import (\n", " align_codes,\n", " make_request,\n", ")\n", "from suno_utils.gpt.engine import Engine\n", "import torch\n", "from tqdm import tqdm\n", "\n", "from stable_audio_tools.inference.generation import (\n", " upsample_diffusion,\n", " upsample_diffusion_from_codes,\n", ")\n", "from stable_audio_tools.interface.gradio import load_model\n", "from stable_audio_tools.models.utils import apply_normalization\n", "\n", "\n", "from suno_utils.tasks.dac_2c_12cb import DAC\n", "\n", "# MERT\n", "from suno_utils.tasks.mert_25 import (\n", " preload_models as preload_semantic_models,\n", " encode as semantic_encode,\n", ")\n", "\n", "# VAE\n", "from suno_utils.tasks.dac_vae_peaq import (\n", " preload_models as preload_vae_models,\n", " encode as vae_encode,\n", " decode as vae_decode,\n", ")\n", "\n", "# DAC\n", "from suno_utils.tasks.dac_2c_12cb import (\n", " preload_models as preload_codec_models,\n", " encode as codec_encode,\n", " decode as codec_decode,\n", ")\n", "\n", "N_BATCH = 2\n", "MAX_STREAMS = N_BATCH * 4\n", "\n", "\n", "def remove_pad_tokens(arr, value: int = 2048):\n", " # Find the last index where the value is not equal to the specified value\n", " non_value_index = np.where(arr != value)[0]\n", " print(non_value_index)\n", "\n", " if non_value_index.size == 0:\n", " # If there are no values that are not equal to the specified value, return an empty array\n", " return np.array([], dtype=arr.dtype)\n", "\n", " # Get the last index where the value is not equal to the specified value\n", " last_non_value_index = non_value_index[-1]\n", " print(last_non_value_index)\n", "\n", " # Slice the array to remove trailing values\n", " return arr[: last_non_value_index]\n", "\n", "\n", "def tensorize(x):\n", " if isinstance(x, np.ndarray):\n", " x = torch.from_numpy(x.astype(np.int64))\n", " assert isinstance(x, torch.Tensor)\n", " if x.ndim == 1:\n", " x = x.unsqueeze(0)\n", "\n", " return x" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "model loaded: /app/suno/checkpoints/2024-08-11_18-28-59/last_ckpt_infer.pt\n", "model loaded: 32020.4M params\n" ] }, { "ename": "OutOfMemoryError", "evalue": "CUDA out of memory. Tried to allocate 264.00 MiB. GPU ", "output_type": "error", "traceback": [ "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m", "\u001b[0;31mOutOfMemoryError\u001b[0m Traceback (most recent call last)", "Cell \u001b[0;32mIn[3], line 28\u001b[0m\n\u001b[1;32m 25\u001b[0m gpt_30b_ft_ckpt_path \u001b[38;5;241m=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m/app/suno/checkpoints/2024-08-11_18-28-59/last_ckpt_infer.pt\u001b[39m\u001b[38;5;124m\"\u001b[39m \u001b[38;5;66;03m# genius filtered\u001b[39;00m\n\u001b[1;32m 26\u001b[0m \u001b[38;5;66;03m#tokenizer_path = \"/app/suno/data/chirp_v4_ft/base_v2/tokenizer_60k.json\" # genius_hq tokenizer (don't use)\u001b[39;00m\n\u001b[0;32m---> 28\u001b[0m \u001b[43mpreload_models\u001b[49m\u001b[43m(\u001b[49m\n\u001b[1;32m 29\u001b[0m \u001b[43m \u001b[49m\u001b[43mgpt_ckpt_path\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mgpt_30b_ft_ckpt_path\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 30\u001b[0m \u001b[43m \u001b[49m\u001b[43mtokenizer_path\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mtokenizer_path\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 31\u001b[0m \u001b[43m \u001b[49m\u001b[43mload_gpt\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[1;32m 32\u001b[0m \u001b[43m \u001b[49m\u001b[43mload_semantic\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;28;43;01mFalse\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[1;32m 33\u001b[0m \u001b[43m \u001b[49m\u001b[43mload_codec_device\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[38;5;124;43mcuda\u001b[39;49m\u001b[38;5;124;43m\"\u001b[39;49m\u001b[43m,\u001b[49m\n\u001b[1;32m 34\u001b[0m \u001b[43m)\u001b[49m\n", "File \u001b[0;32m~/code/glockenspiel/suno_utils/suno_utils/gpt/chirp_v2_5.py:114\u001b[0m, in \u001b[0;36mpreload_models\u001b[0;34m(gpt_ckpt_path, tokenizer_path, semantic_ckpt_path, semantic_centroids_path, codec_ckpt_path, fasttext_ckpt_path, hoot_ckpt_path, hoot_tokenizer_path, load_gpt, load_semantic, load_codec, load_whisper, local_whisper, fetch_only, cache_dir, use_gpu, load_codec_device, load_hoot, load_hoot_device)\u001b[0m\n\u001b[1;32m 112\u001b[0m gpt_ckpt_path \u001b[38;5;241m=\u001b[39m _get_model_if_needed(gpt_ckpt_path, cache_dir\u001b[38;5;241m=\u001b[39mcache_dir)\n\u001b[1;32m 113\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m fetch_only:\n\u001b[0;32m--> 114\u001b[0m \u001b[43mpreload_gpt_models\u001b[49m\u001b[43m(\u001b[49m\u001b[43mgpt_ckpt_path\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mtokenizer_path\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43muse_gpu\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43muse_gpu\u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 115\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m load_semantic:\n\u001b[1;32m 116\u001b[0m semantic_ckpt_path \u001b[38;5;241m=\u001b[39m _get_model_if_needed(semantic_ckpt_path, cache_dir\u001b[38;5;241m=\u001b[39mcache_dir)\n", "File \u001b[0;32m~/code/glockenspiel/suno_utils/suno_utils/gpt/generation.py:230\u001b[0m, in \u001b[0;36mpreload_models\u001b[0;34m(ckpt_path, tokenizer_path, use_gpu)\u001b[0m\n\u001b[1;32m 229\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mpreload_models\u001b[39m(ckpt_path, tokenizer_path, use_gpu\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mTrue\u001b[39;00m):\n\u001b[0;32m--> 230\u001b[0m _ \u001b[38;5;241m=\u001b[39m \u001b[43mload_model\u001b[49m\u001b[43m(\u001b[49m\n\u001b[1;32m 231\u001b[0m \u001b[43m \u001b[49m\u001b[43mckpt_path\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mckpt_path\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 232\u001b[0m \u001b[43m \u001b[49m\u001b[43mtokenizer_path\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mtokenizer_path\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 233\u001b[0m \u001b[43m \u001b[49m\u001b[43muse_gpu\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43muse_gpu\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 234\u001b[0m \u001b[43m \u001b[49m\u001b[43mforce_reload\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[38;5;28;43;01mTrue\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[1;32m 235\u001b[0m \u001b[43m \u001b[49m\u001b[43m)\u001b[49m\n", "File \u001b[0;32m~/code/glockenspiel/suno_utils/suno_utils/gpt/generation.py:192\u001b[0m, in \u001b[0;36mload_model\u001b[0;34m(ckpt_path, tokenizer_path, use_gpu, force_reload, use_tp)\u001b[0m\n\u001b[1;32m 190\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mmodel not initialized, need checkpoint path. maybe run `preload_models`?\u001b[39m\u001b[38;5;124m\"\u001b[39m)\n\u001b[1;32m 191\u001b[0m clean_models(model_key\u001b[38;5;241m=\u001b[39mmodel_key)\n\u001b[0;32m--> 192\u001b[0m model, tokenizer \u001b[38;5;241m=\u001b[39m \u001b[43m_load_model\u001b[49m\u001b[43m(\u001b[49m\u001b[43mckpt_path\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mtokenizer_path\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mdevice\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43muse_tp\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43muse_tp\u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 193\u001b[0m models[model_key] \u001b[38;5;241m=\u001b[39m {\n\u001b[1;32m 194\u001b[0m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mmodel\u001b[39m\u001b[38;5;124m\"\u001b[39m: model,\n\u001b[1;32m 195\u001b[0m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mtokenizer\u001b[39m\u001b[38;5;124m\"\u001b[39m: tokenizer,\n\u001b[1;32m 196\u001b[0m }\n\u001b[1;32m 197\u001b[0m models[model_key][\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mmodel\u001b[39m\u001b[38;5;124m\"\u001b[39m]\u001b[38;5;241m.\u001b[39mto(device)\n", "File \u001b[0;32m~/code/glockenspiel/suno_utils/suno_utils/gpt/generation.py:167\u001b[0m, in \u001b[0;36m_load_model\u001b[0;34m(ckpt_path, tokenizer_path, device, use_tp)\u001b[0m\n\u001b[1;32m 164\u001b[0m model\u001b[38;5;241m.\u001b[39mmodel_args \u001b[38;5;241m=\u001b[39m model_args\n\u001b[1;32m 166\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m device \u001b[38;5;241m==\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mcuda\u001b[39m\u001b[38;5;124m\"\u001b[39m \u001b[38;5;129;01mand\u001b[39;00m torch\u001b[38;5;241m.\u001b[39mcuda\u001b[38;5;241m.\u001b[39mis_bf16_supported():\n\u001b[0;32m--> 167\u001b[0m \u001b[43mmodel\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mto\u001b[49m\u001b[43m(\u001b[49m\u001b[43mdevice\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43mdtype\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mtorch\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mbfloat16\u001b[49m\u001b[43m)\u001b[49m \u001b[38;5;66;03m# this should be fine since it was trained with AMP\u001b[39;00m\n\u001b[1;32m 168\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m 169\u001b[0m model\u001b[38;5;241m.\u001b[39mto(device)\n", "File \u001b[0;32m~/miniconda3/envs/suno_env/lib/python3.10/site-packages/torch/nn/modules/module.py:1173\u001b[0m, in \u001b[0;36mModule.to\u001b[0;34m(self, *args, **kwargs)\u001b[0m\n\u001b[1;32m 1170\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m 1171\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m\n\u001b[0;32m-> 1173\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_apply\u001b[49m\u001b[43m(\u001b[49m\u001b[43mconvert\u001b[49m\u001b[43m)\u001b[49m\n", "File \u001b[0;32m~/miniconda3/envs/suno_env/lib/python3.10/site-packages/torch/nn/modules/module.py:779\u001b[0m, in \u001b[0;36mModule._apply\u001b[0;34m(self, fn, recurse)\u001b[0m\n\u001b[1;32m 777\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m recurse:\n\u001b[1;32m 778\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m module \u001b[38;5;129;01min\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mchildren():\n\u001b[0;32m--> 779\u001b[0m \u001b[43mmodule\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_apply\u001b[49m\u001b[43m(\u001b[49m\u001b[43mfn\u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 781\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mcompute_should_use_set_data\u001b[39m(tensor, tensor_applied):\n\u001b[1;32m 782\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m torch\u001b[38;5;241m.\u001b[39m_has_compatible_shallow_copy_type(tensor, tensor_applied):\n\u001b[1;32m 783\u001b[0m \u001b[38;5;66;03m# If the new tensor has compatible tensor type as the existing tensor,\u001b[39;00m\n\u001b[1;32m 784\u001b[0m \u001b[38;5;66;03m# the current behavior is to change the tensor in-place using `.data =`,\u001b[39;00m\n\u001b[0;32m (...)\u001b[0m\n\u001b[1;32m 789\u001b[0m \u001b[38;5;66;03m# global flag to let the user control whether they want the future\u001b[39;00m\n\u001b[1;32m 790\u001b[0m \u001b[38;5;66;03m# behavior of overwriting the existing tensor or not.\u001b[39;00m\n", "File \u001b[0;32m~/miniconda3/envs/suno_env/lib/python3.10/site-packages/torch/nn/modules/module.py:779\u001b[0m, in \u001b[0;36mModule._apply\u001b[0;34m(self, fn, recurse)\u001b[0m\n\u001b[1;32m 777\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m recurse:\n\u001b[1;32m 778\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m module \u001b[38;5;129;01min\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mchildren():\n\u001b[0;32m--> 779\u001b[0m \u001b[43mmodule\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_apply\u001b[49m\u001b[43m(\u001b[49m\u001b[43mfn\u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 781\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mcompute_should_use_set_data\u001b[39m(tensor, tensor_applied):\n\u001b[1;32m 782\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m torch\u001b[38;5;241m.\u001b[39m_has_compatible_shallow_copy_type(tensor, tensor_applied):\n\u001b[1;32m 783\u001b[0m \u001b[38;5;66;03m# If the new tensor has compatible tensor type as the existing tensor,\u001b[39;00m\n\u001b[1;32m 784\u001b[0m \u001b[38;5;66;03m# the current behavior is to change the tensor in-place using `.data =`,\u001b[39;00m\n\u001b[0;32m (...)\u001b[0m\n\u001b[1;32m 789\u001b[0m \u001b[38;5;66;03m# global flag to let the user control whether they want the future\u001b[39;00m\n\u001b[1;32m 790\u001b[0m \u001b[38;5;66;03m# behavior of overwriting the existing tensor or not.\u001b[39;00m\n", " \u001b[0;31m[... skipping similar frames: Module._apply at line 779 (2 times)]\u001b[0m\n", "File \u001b[0;32m~/miniconda3/envs/suno_env/lib/python3.10/site-packages/torch/nn/modules/module.py:779\u001b[0m, in \u001b[0;36mModule._apply\u001b[0;34m(self, fn, recurse)\u001b[0m\n\u001b[1;32m 777\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m recurse:\n\u001b[1;32m 778\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m module \u001b[38;5;129;01min\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mchildren():\n\u001b[0;32m--> 779\u001b[0m \u001b[43mmodule\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43m_apply\u001b[49m\u001b[43m(\u001b[49m\u001b[43mfn\u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 781\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mcompute_should_use_set_data\u001b[39m(tensor, tensor_applied):\n\u001b[1;32m 782\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m torch\u001b[38;5;241m.\u001b[39m_has_compatible_shallow_copy_type(tensor, tensor_applied):\n\u001b[1;32m 783\u001b[0m \u001b[38;5;66;03m# If the new tensor has compatible tensor type as the existing tensor,\u001b[39;00m\n\u001b[1;32m 784\u001b[0m \u001b[38;5;66;03m# the current behavior is to change the tensor in-place using `.data =`,\u001b[39;00m\n\u001b[0;32m (...)\u001b[0m\n\u001b[1;32m 789\u001b[0m \u001b[38;5;66;03m# global flag to let the user control whether they want the future\u001b[39;00m\n\u001b[1;32m 790\u001b[0m \u001b[38;5;66;03m# behavior of overwriting the existing tensor or not.\u001b[39;00m\n", "File \u001b[0;32m~/miniconda3/envs/suno_env/lib/python3.10/site-packages/torch/nn/modules/module.py:804\u001b[0m, in \u001b[0;36mModule._apply\u001b[0;34m(self, fn, recurse)\u001b[0m\n\u001b[1;32m 800\u001b[0m \u001b[38;5;66;03m# Tensors stored in modules are graph leaves, and we don't want to\u001b[39;00m\n\u001b[1;32m 801\u001b[0m \u001b[38;5;66;03m# track autograd history of `param_applied`, so we have to use\u001b[39;00m\n\u001b[1;32m 802\u001b[0m \u001b[38;5;66;03m# `with torch.no_grad():`\u001b[39;00m\n\u001b[1;32m 803\u001b[0m \u001b[38;5;28;01mwith\u001b[39;00m torch\u001b[38;5;241m.\u001b[39mno_grad():\n\u001b[0;32m--> 804\u001b[0m param_applied \u001b[38;5;241m=\u001b[39m \u001b[43mfn\u001b[49m\u001b[43m(\u001b[49m\u001b[43mparam\u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 805\u001b[0m p_should_use_set_data \u001b[38;5;241m=\u001b[39m compute_should_use_set_data(param, param_applied)\n\u001b[1;32m 807\u001b[0m \u001b[38;5;66;03m# subclasses may have multiple child tensors so we need to use swap_tensors\u001b[39;00m\n", "File \u001b[0;32m~/miniconda3/envs/suno_env/lib/python3.10/site-packages/torch/nn/modules/module.py:1159\u001b[0m, in \u001b[0;36mModule.to..convert\u001b[0;34m(t)\u001b[0m\n\u001b[1;32m 1152\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m convert_to_format \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m \u001b[38;5;129;01mand\u001b[39;00m t\u001b[38;5;241m.\u001b[39mdim() \u001b[38;5;129;01min\u001b[39;00m (\u001b[38;5;241m4\u001b[39m, \u001b[38;5;241m5\u001b[39m):\n\u001b[1;32m 1153\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m t\u001b[38;5;241m.\u001b[39mto(\n\u001b[1;32m 1154\u001b[0m device,\n\u001b[1;32m 1155\u001b[0m dtype \u001b[38;5;28;01mif\u001b[39;00m t\u001b[38;5;241m.\u001b[39mis_floating_point() \u001b[38;5;129;01mor\u001b[39;00m t\u001b[38;5;241m.\u001b[39mis_complex() \u001b[38;5;28;01melse\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m,\n\u001b[1;32m 1156\u001b[0m non_blocking,\n\u001b[1;32m 1157\u001b[0m memory_format\u001b[38;5;241m=\u001b[39mconvert_to_format,\n\u001b[1;32m 1158\u001b[0m )\n\u001b[0;32m-> 1159\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[43mt\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mto\u001b[49m\u001b[43m(\u001b[49m\n\u001b[1;32m 1160\u001b[0m \u001b[43m \u001b[49m\u001b[43mdevice\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 1161\u001b[0m \u001b[43m \u001b[49m\u001b[43mdtype\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;28;43;01mif\u001b[39;49;00m\u001b[43m \u001b[49m\u001b[43mt\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mis_floating_point\u001b[49m\u001b[43m(\u001b[49m\u001b[43m)\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;129;43;01mor\u001b[39;49;00m\u001b[43m \u001b[49m\u001b[43mt\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mis_complex\u001b[49m\u001b[43m(\u001b[49m\u001b[43m)\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;28;43;01melse\u001b[39;49;00m\u001b[43m \u001b[49m\u001b[38;5;28;43;01mNone\u001b[39;49;00m\u001b[43m,\u001b[49m\n\u001b[1;32m 1162\u001b[0m \u001b[43m \u001b[49m\u001b[43mnon_blocking\u001b[49m\u001b[43m,\u001b[49m\n\u001b[1;32m 1163\u001b[0m \u001b[43m \u001b[49m\u001b[43m)\u001b[49m\n\u001b[1;32m 1164\u001b[0m \u001b[38;5;28;01mexcept\u001b[39;00m \u001b[38;5;167;01mNotImplementedError\u001b[39;00m \u001b[38;5;28;01mas\u001b[39;00m e:\n\u001b[1;32m 1165\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28mstr\u001b[39m(e) \u001b[38;5;241m==\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mCannot copy out of meta tensor; no data!\u001b[39m\u001b[38;5;124m\"\u001b[39m:\n", "\u001b[0;31mOutOfMemoryError\u001b[0m: CUDA out of memory. Tried to allocate 264.00 MiB. GPU " ] } ], "source": [ "import os\n", "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"2\"\n", "os.environ[\"PYTORCH_CUDA_ALLOC_CONF\"] = \"expandable_segments:True\"\n", "\n", "timestamp = \"20240701\"\n", "exp_name = f\"model_13b_full_ft\"\n", "# exp_name = \"model_13b_full\"\n", "os.makedirs(f\"outputs/{timestamp}/{exp_name}\", exist_ok=True)\n", "\n", "# load VAE\n", "_ = preload_vae_models(\"/app/suno/checkpoints/dac_mw/mw_vae_peaq_128_fix.pth\")\n", "\n", "# load DAC\n", "_ = preload_codec_models(\n", " checkpoint_filepath=\"s3://suno-data/georg/models/codec/dac_2c_25x12.pt\",\n", ")\n", "\n", "#gpt_ckpt_path = \"/app/suno/data/dpo/models/model_13b_full.pt\" # model before fine-tuning\n", "gpt_ckpt_path = \"/app/suno/data/dpo/models/model_13b_full.pt\" # start of finetuning\n", "\n", "#gpt_13b_ft_ckpt_path = \"/app/suno/checkpoints/2024-08-08_19-18-28/last_ckpt_infer.pt\" # fine tune for audio quality control tags\n", "#gpt_13b_ft_ckpt_path = \"/app/suno/checkpoints/2024-08-09_20-41-33/last_ckpt_infer.pt\" # fine tune for audio quality control tags\n", "tokenizer_path = \"/app/suno/data/dpo/models/tokenizer_60k.json\"\n", "gpt_13b_ft_ckpt_path = \"/app/suno/checkpoints/2024-08-10_11-19-21/best_ckpt.pt\" # genius filtered\n", "gpt_30b_ft_ckpt_path = \"/app/suno/checkpoints/2024-08-11_18-28-59/last_ckpt_infer.pt\" # genius filtered\n", "#tokenizer_path = \"/app/suno/data/chirp_v4_ft/base_v2/tokenizer_60k.json\" # genius_hq tokenizer (don't use)\n", "\n", "preload_models(\n", " gpt_ckpt_path=gpt_30b_ft_ckpt_path,\n", " tokenizer_path=tokenizer_path,\n", " load_gpt=True,\n", " load_semantic=False,\n", " load_codec_device=\"cuda\",\n", ")\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "text_tags_list = [\n", " \"indie-pop, energetic, rock, psychedelic\",\n", " \"jazz, female vocal, smooth\",\n", " \"r&b, soul, funk\",\n", " \"pop, high energy, female vocal\",\n", " \"rock, high energy, male vocal\",\n", " \"indie rock, female vocal, psychedelic\",\n", " \"bluegrass, female vocal, guitar\",\n", " \"moody, blues, soulful\",\n", " \"classical, cinematic, female vocal\",\n", " \"metal, hardcore, dark\",\n", " \"male vocal, bluegrass, guitar\",\n", " \"metal, hardcore, dark\",\n", " \"pop, rock\",\n", " \"heavy rock\"\n", " \"female vocal, voice, singing, guitar, folk, jazz, r&b\",\n", " \"Cyberpunk ballad, kaleidoscopic funk, dreamwave fusion, existential groove\",\n", " \"#Ambient glitch-pop, neo-baroque jazz, surreal groove, vaporwave funk\",\n", " \"Indie-electro, playful pulse, starry-eyed synths, upbeat\",\n", " #\"Jazz-hop, smooth satire, cocktail lounge cool, lyrical swing\",\n", " #\"Lo-fi chillwave, campfire pop, mellow bounce, conversational\",\n", "]\n", "\n", "text = \"\"\"\n", "[verse]\n", "Almost Heaven, West Virginia\n", "Blue Ridge Mountains, Shenandoah River\n", "Life is old there, older than the trees\n", "Younger than the mountains, growing like a breeze\n", "\n", "Country roads, take me home\n", "To the place I belong\n", "West Virginia, mountain mama\n", "Take me home, country roads\n", "\"\"\"\n", "\n", "text = \"\"\"\n", "[verse]\n", "oh, my love\n", "My friend you know\n", "it's been a while\n", "Without thinking of you\n", "but the thought makes me smile\n", "\n", "[chorus]\n", "I'm so tired of wanting\n", "wanting more than this\n", "i know it but what am i to do\n", "i need some space to breathe,\n", "so give me some room\n", "\n", "[verse]\n", "oh, my love\n", "you have a heart of stone\n", "cause since i've come home\n", "i've never felt so alone\n", "but the thought makes me smile\n", "\n", "[chorus]\n", "I'm so tired of wanting\n", "wanting more than this\n", "i know it but what am i to do\n", "i need some space to breathe,\n", "so give me some room\n", "\"\"\"\n", "\n", "temps_coarse = [0.8, 0.9, 0.95, 1.0, 1.1]\n", "\n", "# generate uuid\n", "uid = uuid.uuid4()\n", "\n", "temp_coarse = random.choice(temps_coarse)\n", "\n", "# generate high quality and low quality\n", "\n", "for n in range(5):\n", " random_seed = np.random.randint(0, 1000000)\n", " text_tags = random.choice(text_tags_list)\n", " print(random_seed)\n", "\n", " for quality in [\"standard\", \"high\"]:\n", " print(f\"Tags: {text_tags} Quality: {quality}\")\n", " #text_start_control_tags = \"{start;start:0;vocals:start;audio_quality:\" + f\"{quality}\" + \"}\"\n", " text_start_control_tags = \"{start;audio_quality:\" + f\"{quality}\" + \"}\"\n", " #text_start_control_tags = \"{start:\" + \"}\"\n", " print(text_start_control_tags)\n", "\n", " general_config = dict(\n", " #cfg_coef=3.0, # no text cfg for dpo stream\n", " #min_eos_p=0.1,\n", " #eos_pad_duration_s=0,\n", " #cfg_coef=2.0,\n", " #cfg_coef_tags=2,\n", " #n_repeat_tags=3,\n", " # cfg_coef_tags_max_steps=None, # collect the data for now\n", " #cfg_coef_tags=0,\n", " #cfg_coef_neg_tags=0,\n", " #text_neg_tags=\"repetitive, loop, noise, noisy\",\n", " use_whisper=False,\n", " text_start_control_tags=text_start_control_tags,\n", " #text_end_control_tags=\"{end}\",\n", " random_seed=random_seed,\n", " n_batch=1,\n", " )\n", "\n", " #temp_coarse_tensor = torch.tensor(\n", " # [0.9, 1.0, 1.25, 1.5, 0.6, 0.5, 0.4, 0.3, 0.3, 0.3, 0.3, 0.3]\n", " #).view(1, 12, 1)\n", "\n", " gconfig = GenerationConfig(\n", " text=text,\n", " text_tags=text_tags,\n", " max_gen_duration_s=240,\n", " **general_config,\n", " return_semantic=True,\n", " temp_coarse=1.0,\n", " #top_k_coarse=None,\n", " #top_p_coarse=None,\n", " )\n", "\n", " semantic_codes, codec_codes = generate(gconfig)\n", " semantic_codes = torch.from_numpy(semantic_codes[0]).squeeze()\n", " #codec_codes = torch.from_numpy(codec_codes[0])\n", " codec_codes = codec_codes[0]\n", " # convert dac tokens back to audio\n", " audio = codec_decode(codec_codes)\n", " audio.play()\n", " #audio = torch.from_numpy(audio.array_float)\n", " #audio /= audio.abs().max()\n", " #audio = apply_normalization(audio, 48000, target_loudness_lufs_db=-16.0)\n", " #gpt_out = os.path.join(\"outputs\", timestamp, exp_name, f\"{uid}-output.wav\")\n", " #torchaudio.save(gpt_out, audio, 48000)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "suno_env", "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": "3.10.9" } }, "nbformat": 4, "nbformat_minor": 2 }