{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "0", "metadata": {}, "outputs": [], "source": [ "import os\n", "\n", "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"6\"\n", "from suno_utils.gpt.generation_prompt import ALL_AUDIO_PROMPTS" ] }, { "cell_type": "code", "execution_count": null, "id": "1", "metadata": {}, "outputs": [], "source": [ "import numpy as np\n", "import torch\n", "from suno_utils.audio import Audio\n", "from suno_utils.gpt.generation import GenerationConfig, CfgGenerationConfig\n", "from suno_utils.gpt.engine import Engine\n", "from suno_utils.gpt.generation_engine import make_request\n", "\n", "from suno_utils.diffusion.generation import (\n", " preload_models as preload_diff_models,\n", " generate,\n", " DiffusionGenerationConfig,\n", ")\n", "from suno_utils.diffusion.generation import encode_semantic\n", "from suno_utils.tasks.upsample_engine import UpsampleEngine, Request, Job\n", "\n", "\n", "# gpt_model_fp = \"/app/suno/checkpoints/2025-02-04_21-04-31/last_ckpt_infer.pt\" # base\n", "gpt_model_fp = \"/app/suno/checkpoints/2025-04-12_03-10-00/last_ckpt_infer.pt\" # sft\n", "# gpt_model_fp = \"/app/suno/checkpoints/2025-04-08_19-36-46/last_ckpt_infer.pt\" # dpo\n", "\n", "\n", "diff_model_fp = \"/app/suno/data/dpo/models/diff_vae_25_peaq_v7_mar24.pt\"\n", "\n", "\n", "# GPT\n", "N_BATCH = 2\n", "n_skip_semantic = 1\n", "engine = Engine(\n", " gpt_model_fp,\n", " \"/app/suno/models/chirp_v2/tokenizer_60k.json\",\n", " max_sequences=8 * N_BATCH,\n", " compile=False,\n", ")\n", "cfg = engine.model.config\n", "\n", "# Diffusion\n", "_ = preload_diff_models(\n", " tokenizer_filepath=\"/home/georg/notebooks/gpu_nb/tmp/tokenizer_60k.json\",\n", " semantic_model_filepath=\"/home/georg/notebooks/gpu_nb/tmp/mert_25.pt\",\n", " semantic_clusters_filepath=\"/home/georg/notebooks/gpu_nb/tmp/mert_25_2x4k.npy\",\n", " codec_filepath=\"/app/suno/data/dpo/models/25hz_vae_peaq_kl_0.005.pth\",\n", " dit_model_filepath=diff_model_fp,\n", " weights_precision=torch.bfloat16,\n", " compile=False,\n", ")\n", "\n", "up_engine = UpsampleEngine(min_chunk_size=25 * 30)" ] }, { "cell_type": "code", "execution_count": null, "id": "2", "metadata": {}, "outputs": [], "source": [ "from suno_utils.utils.clip import SunoClip\n", "\n", "sc = SunoClip(\"37f34bc3-a6e9-40b9-8a51-635c268eddb4\")\n", "\n", "sc.audio().play()\n", "\n", "semantic_arr = sc.full_arr()[:, :1]\n", "\n", "print(semantic_arr.shape)" ] }, { "cell_type": "code", "execution_count": null, "id": "3", "metadata": {}, "outputs": [], "source": [ "CONTEXT_TOKENS = 350\n", "b_left_idx = 25 * 60\n", "a_left_idx = b_left_idx - CONTEXT_TOKENS\n", "b_right_idx = int(25 * 101)\n", "c_right_idx = b_right_idx + CONTEXT_TOKENS\n", "history_arr = semantic_arr[a_left_idx:b_left_idx].copy()\n", "future_arr = semantic_arr[b_right_idx:c_right_idx].copy()\n", "tags = \"slow, incredibly sad, face melting, epic, beautifully sung intimate anthem for the heat death of the universe\"\n", "infill_text = \"\"\"\n", "Returnal\n", "You've never left, give that weasel some wine\n", "Returnal\n", "You've never left, you've been here the whole time\n", "\n", "[Instrumental interlude]\n", "\n", "[Verse 2]\n", "Another silo full\n", "\"\"\"\n", "sample_duration_s = int(round((c_right_idx - a_left_idx) / 25))" ] }, { "cell_type": "code", "execution_count": null, "id": "4", "metadata": {}, "outputs": [], "source": [ "gconf = GenerationConfig(\n", " text=infill_text,\n", " text_tags=tags,\n", " history_arr=history_arr,\n", " future_arr=future_arr,\n", " cfg_coef=1.2,\n", " text_start_control_tags=\"{duration:\" + str(sample_duration_s) + \"}\",\n", " # min_eos_p=0.1,\n", " cfg_coef_max_steps=None,\n", " cfg_coef_tags_max_steps=None,\n", " n_repeat_tags=1,\n", " n_repeat_neg_tags=1,\n", " n_skip_semantic=n_skip_semantic,\n", " cfg_coef_neg_tags=0.0,\n", " text_neg_tags=None,\n", " temp_semantic=0.92,\n", " top_k_semantic=None,\n", " top_p_semantic=None,\n", " min_p_semantic=0.005,\n", " n_batch=1,\n", " min_text_offset=0,\n", " eos_pad_duration_s=0,\n", " max_gen_duration_s=int(2 * 60 / n_skip_semantic), # just incase\n", " random_seed=100,\n", " cfg_streams=[\n", " CfgGenerationConfig(\n", " stream_type=\"custom\",\n", " prompts=[\"control_start\", \"tag\", \"lyrics\", \"control_end\"]\n", " + ALL_AUDIO_PROMPTS,\n", " null_prompts=[\"tag\", \"lyrics_no_tags\"] + ALL_AUDIO_PROMPTS,\n", " weight=2.0,\n", " max_steps=25 * 60,\n", " ),\n", " ],\n", " custom_null_fields=ALL_AUDIO_PROMPTS,\n", ")" ] }, { "cell_type": "code", "execution_count": null, "id": "5", "metadata": {}, "outputs": [], "source": [ "requests = [\n", " make_request(f\"{i}\", gconf, engine.model.config, engine.tokenizer)\n", " for i in range(N_BATCH)\n", "]\n", "\n", "jobs = engine.run_request(requests, tqdm_enabled=True)\n", "out_gpt = []\n", "for n, job in enumerate(jobs):\n", " stream = engine.token_generator(job)\n", " arr = torch.stack(list(stream))[:, 1]\n", " if arr[-1] == 4000:\n", " arr = arr[:-1]\n", " print(f\"{round(arr.shape[-1]/25*n_skip_semantic)}s for track {n}\")\n", " # do stuff incase skip\n", " arr2 = (\n", " torch.zeros(arr.shape[0] * n_skip_semantic, dtype=arr.dtype)\n", " + cfg.semantic_pad_token\n", " )\n", " arr2[::n_skip_semantic] = arr\n", " # add\n", " out_gpt.append(arr2)" ] }, { "cell_type": "code", "execution_count": null, "id": "6", "metadata": {}, "outputs": [], "source": [ "from suno_utils.tasks.dac_vae_100hz_peaq import decode as decode_vae\n", "\n", "infill_text_with_pad = \"\"\"\n", "she's the cure\n", "Underground, search for more\n", "\n", "[Chorus]\n", "Returnal\n", "You've never left, give that weasel some wine\n", "Returnal\n", "You've never left, you've been here the whole time\n", "\n", "[Verse 2]\n", "Another silo full another dark dawn\n", "Bending the air, love is so full\n", "\"\"\".strip()\n", "\n", "# upsample with 10s pad on each side just incase\n", "# note: for simplicity let's upsample the entire clip rather than just the infill with context\n", "out_audio = []\n", "for in_sem_arr in out_gpt:\n", " gen_cfg = DiffusionGenerationConfig(\n", " lyrics=infill_text_with_pad,\n", " tags=tags,\n", " text_cfg_coef=4.0,\n", " steps=32,\n", " # codec_scale_factor=0.4,\n", " scale_ctx_vector=False,\n", " noise_ctx_level=0.0,\n", " drop_semantic_tokens=False,\n", " seed=0,\n", " )\n", "\n", " request = Request(\n", " id=\"dummy\",\n", " generation_config=gen_cfg,\n", " tokens=torch.cat(\n", " [\n", " torch.from_numpy(history_arr[-25 * 10 :, 0].astype(np.int32)),\n", " in_sem_arr,\n", " torch.from_numpy(future_arr[: 25 * 10, 0].astype(np.int32)),\n", " ],\n", " dim=0,\n", " ),\n", " input_tokens_finished=True,\n", " )\n", "\n", " result = up_engine.run_request(request).vae_latents\n", " decode_vae(torch.cat(result, dim=0)).play()" ] }, { "cell_type": "code", "execution_count": null, "id": "7", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "8", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "suno_clean", "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": "3.10.15" } }, "nbformat": 4, "nbformat_minor": 5 }