import os import sys import time import modal import torch import pathlib import funcy import json import pandas as pd import numpy as np import tempfile import torchaudio import pyloudnorm as pyln from suno_utils.audio import Audio from suno_utils.gpt.generation import GenerationConfig from suno_utils.gpt.engine import Engine from suno_utils.gpt.generation_engine import make_request from suno_utils.worker.settings import s3_client # from suno_utils.worker.modal_base import MODAL_MOUNTS from suno_utils.gpt import chirp_v2_5 as chirp_v2 from suno_utils.utils.text import read_jsonl from suno_utils.utils.s3 import list_s3_dir, read_from_s3 from suno_utils.tasks.hoot import ( encode_filepaths, encode, clean_text, encode_and_align, _get_alignable_tokens, _assign_word_timings, get_aligned_lyrics, ctc_align, load_model_list, get_word_timing_from_audio_and_lyrics, _merge_into_lines, decode_logits, load_model, ) from suno_utils.tasks.shimmerscore import shimmerscore from suno_utils.utils.metrics import get_cer from suno_utils.tasks.ear import load_model from suno_utils.diffusion.generation import ( preload_dit_model, preload_tokenizer, TOKENIZER_FILEPATH, SEMANTIC_MODEL_FILEPATH, SEMANTIC_CLUSTERS_FILEPATH, _retrieve_models, # preload_ear_model, ) from suno_utils.tasks.mert_25 import ( preload_models as preload_semantic_models, encode as encode_semantic, ) from suno_utils.diffusion import generation as diffusion_gen from suno_utils.tasks.upsample_engine import UpsampleEngine, Request, Job # from suno_utils.tasks.upsample_engine_old import UpsampleEngine, Request, Job MOUNT_PATH = "/suno/models" aws_secret = modal.Secret.from_name("studio-aws") SECRETS = [ aws_secret, modal.Secret.from_dict( { "SUNO_ASSETS_PATH": "/suno/models/assets", "XDG_CACHE_HOME": "/suno/models/", } ), modal.Secret.from_name("api-callback-token"), modal.Secret.from_name("datadog-metrics"), ] base_image = ( modal.Image.from_registry("nvidia/cuda:12.4.0-devel-ubuntu22.04", add_python="3.10") .apt_install( "curl", "ffmpeg", "sox", "unzip", "libsox-fmt-mp3", "zlib1g-dev", "git", "clang" ) .run_commands( [ 'curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"', "unzip -q awscliv2.zip", "./aws/install", ] ) .dockerfile_commands( [ "COPY --from=datadog/serverless-init:1.2.1 /datadog-init /app/datadog-init", 'ENTRYPOINT ["/app/datadog-init"]', ] ) .pip_install("torch==2.5.1", "torchaudio==2.5.1") .pip_install( "flashinfer-python", index_url="https://flashinfer.ai/whl/cu124/torch2.5/" ) .pip_install_private_repos( "github.com/suno-ai/glockenspiel.git@a5dba4e50#subdirectory=descript-audio-codec&egg=descript-audio-codec", git_user="mcamac", secrets=[modal.Secret.from_name("victor-modal-github-token")], ) .pip_install_private_repos( "github.com/suno-ai/neon.git@1c83548#subdirectory=hoot", git_user="mcamac", secrets=[modal.Secret.from_name("victor-modal-github-token")], ) .pip_install( "boto3", "transformers", "tokenizers", "encodec", "ctc_segmentation", "psutil", "redis", "pydantic", "nnAudio", "rpyc", "biopython>=1.81", # TODO: don't love this depdendency, for hoot "pynvml", # for torch cuda utilization "torchsde", "ninja", "wheel", ) .pip_install_from_pyproject( "/home/christian/code/glockenspiel/suno_utils/pyproject.toml", ) .run_commands( # This is really slow "git clone https://github.com/Dao-AILab/flash-attention.git", "cd flash-attention/hopper && python setup.py install", gpu="h100", ) .apt_install( "libogg0", "libopus0", "opus-tools", ) .pip_install("transformers==4.44.0", "wandb") ) def calculate_stereo_width(waveform): # Split into left and right channels left = waveform[0] right = waveform[1] # Compute mid/side representation mid = (left + right) / 2 side = (left - right) / 2 # Compute RMS energy of mid and side channels mid_energy = torch.sqrt(torch.mean(mid**2)) side_energy = torch.sqrt(torch.mean(side**2)) # Compute stereo width based on mid/side ratio # Normalize to range 0-1 using sigmoid-like function width_ratio = (side_energy / (mid_energy + 1e-8)).item() stereo_width = 2 * (1 / (1 + np.exp(-width_ratio)) - 0.5) return stereo_width def calculate_loudness(waveform, sr): meter = pyln.Meter(sr) lufs_db = meter.integrated_loudness(waveform) return lufs_db def calculate_loudness_factor(waveform, sr): meter = pyln.Meter(sr) normalized_waveform = waveform / np.clip(np.max(np.abs(waveform)), 1e-10, None) lufs_db = meter.integrated_loudness(normalized_waveform) return lufs_db def calculate_possible_clipped_samples(waveform): # Convert to numpy if it's a torch tensor if isinstance(waveform, torch.Tensor): waveform_np = waveform.numpy() else: waveform_np = waveform # Count samples that are at or above the clipping threshold return ( np.sum(np.abs(waveform_np) >= 1.0).item() if isinstance(np.sum(np.abs(waveform_np) >= 1.0), torch.Tensor) else np.sum(np.abs(waveform_np) >= 1.0) ) def third_octave_bands(sr, fmin=20.0, fmax=None): """ Compute 1/3-octave band center frequencies and edges. """ if fmax is None: fmax = sr / 2.0 k = np.arange(-30, 30) # wide enough range f_center = 1000.0 * (2.0 ** (k / 3.0)) # ISO 1/3 octave centers f_center = f_center[(f_center >= fmin) & (f_center <= fmax)] f_lower = f_center / (2 ** (1 / 6)) f_upper = f_center * (2 ** (1 / 6)) return f_center, f_lower, f_upper def third_octave_response_db(waveform: torch.Tensor, sr: int): """ Compute 1/3 octave magnitude response in dB from waveform. Args: waveform (torch.Tensor): shape (n_samples,) or (1, n_samples) sr (int): sample rate Returns: freqs (np.ndarray): band center frequencies mags_db (torch.Tensor): band magnitudes in dB """ if waveform.ndim > 1: waveform = waveform.squeeze(0) n = waveform.numel() spec = torch.fft.rfft(waveform) mag = torch.abs(spec) / n freqs = torch.fft.rfftfreq(n, d=1.0 / sr) # Get bands f_center, f_lower, f_upper = third_octave_bands(sr) band_mags = [] for fl, fu in zip(f_lower, f_upper): idx = (freqs >= fl) & (freqs < fu) if idx.any(): band_mags.append(mag[idx].mean()) else: band_mags.append(torch.tensor(0.0)) band_mags = torch.stack(band_mags) # Convert to dB (avoid log(0)) mags_db = 20 * torch.log10(band_mags + 1e-12) return f_center, mags_db def stereo_width(waveform: torch.Tensor): """ Compute the stereo width of a waveform. """ # can you implement this? # Assume waveform shape is (2, seq_len) if waveform.ndim != 2 or waveform.shape[0] != 2: raise ValueError( "waveform must have shape (2, seq_len) for stereo width calculation" ) left = waveform[0] right = waveform[1] # Compute correlation coefficient between L and R left = left - left.mean() right = right - right.mean() numerator = (left * right).mean() denominator = torch.sqrt((left**2).mean() * (right**2).mean()) + 1e-12 corr = numerator / denominator # Stereo width: 0 = mono, 1 = fully wide (L and R uncorrelated), -1 = fully out of phase width = torch.sqrt(1 - corr**2) return width.item() def loudness(waveform: torch.Tensor, sample_rate: float): """ Compute the loudness of a waveform. """ meter = pyln.Meter(sample_rate) loudness = meter.integrated_loudness(waveform.permute(1, 0).numpy()) return loudness def analyze_audio_first_last( audio: torch.Tensor, sample_rate: int, segment_duration: int = 60, ): """ Compute 1/3-octave response and stereo width for the first and last segments of a single audio tensor, and return their deltas. Args: audio: Tensor of shape (channels, samples) or (samples,) sample_rate: sample rate in Hz segment_duration: segment length in seconds for the first/last comparison Returns: { 'center_freqs': np.ndarray, 'third_octave_first': Tensor[dB], 'third_octave_last': Tensor[dB], 'delta_response': Tensor[dB], # first - last 'total_delta': Tensor[scalar], # L1 magnitude of delta_response 'stereo_width_first': float or None, 'stereo_width_last': float or None, 'stereo_width_delta': float or None, # first - last } """ if not isinstance(audio, torch.Tensor): raise TypeError("audio must be a torch.Tensor") if audio.numel() == 0: raise ValueError("audio is empty") # Determine segment length in samples (clip to available length) seg_len = min(audio.shape[-1], segment_duration * sample_rate) # Helper: mono mixdown for 1/3-octave analysis mono = audio.mean(dim=0) if audio.ndim > 1 else audio first_mono = mono[:seg_len] last_mono = mono[-seg_len:] # 1/3-octave responses (dB) center_freqs_first, oct_first = third_octave_response_db(first_mono, sample_rate) center_freqs_last, oct_last = third_octave_response_db(last_mono, sample_rate) # Centers should match; keep the first as canonical if ( len(center_freqs_first) != len(center_freqs_last) or (center_freqs_first != center_freqs_last).any() ): raise RuntimeError( "Mismatched third-octave centers between first and last segments." ) delta = oct_first - oct_last total_delta = delta.abs().sum() # Stereo width (if stereo input) def maybe_width(x: torch.Tensor): if x.ndim == 2 and x.shape[0] == 2 and x.shape[1] > 0: return stereo_width(x) return None first_full = audio[:, :seg_len] if audio.ndim == 2 else audio last_full = audio[:, -seg_len:] if audio.ndim == 2 else audio width_first = maybe_width(first_full) width_last = maybe_width(last_full) width_delta = None if (width_first is not None) and (width_last is not None): width_delta = float(width_first - width_last) # loudness loudness_first = loudness(first_full, sample_rate) loudness_last = loudness(last_full, sample_rate) loudness_delta = None if (loudness_first is not None) and (loudness_last is not None): loudness_delta = float(loudness_first - loudness_last) return { "center_freqs": center_freqs_first, # np.ndarray "third_octave_first": oct_first, # Tensor[dB] "third_octave_last": oct_last, # Tensor[dB] "delta_response": delta, # Tensor[dB] "total_delta": total_delta, # Tensor[scalar] "stereo_width_first": width_first, # float or None "stereo_width_last": width_last, # float or None "stereo_width_delta": width_delta, # float or None "loudness_first": loudness_first, # float or None "loudness_last": loudness_last, # float or None "loudness_delta": loudness_delta, # float or None } def analyze_audio(filepath): audio, sr = torchaudio.load(filepath) if sr != 48000: audio = torchaudio.functional.resample(audio, sr, 48000) # calculate loudness lufs_db = calculate_loudness(audio.permute(1, 0).numpy(), sr) lufs_db_factor = calculate_loudness_factor(audio.permute(1, 0).numpy(), sr) # calculate stereo width stereo_width = calculate_stereo_width(audio) # clipped samples clipped_samples = calculate_possible_clipped_samples(audio) return { "lufs_db": lufs_db, "lufs_db_factor": lufs_db_factor, "stereo_width": stereo_width, "clipped_samples": clipped_samples, } def _reload_models_if_needed(hoot_filepath, hoot_tokenizer_filepath): model_list = load_model_list( checkpoint_filepath=hoot_filepath, tokenizer_filepath=hoot_tokenizer_filepath, n_gpus=None, ) model = model_list[0]["model"] tokenizer = model_list[0]["tokenizer"] return model_list, model, tokenizer class GenerateWorker: def __init__( self, dit_model_filepath: str, output_path: str, test_type: str, ): self.output_path = output_path self.test_type = test_type start_time = time.time() print("Start loading models") # load diffusion model num_gpus = torch.cuda.device_count() cuda_device = torch.cuda.current_device() print(f"Found {num_gpus} GPUs. Using GPU {cuda_device}.") tokenizer_filepath = "s3://suno-data/georg/models/tokenizers/tokenizer_60k.json" semantic_model_filepath = "s3://suno-data/georg/models/semantic/mert_25.pt" semantic_clusters_filepath = ( "s3://suno-data/georg/models/semantic/mert_25_2x4k.npy" ) codec_filepath = CODEC_FILEPATH ear_model_filepath = EAR_MODEL_FILEPATH _ = diffusion_gen.preload_dit_model( dit_model_filepath=dit_model_filepath, use_ema_if_exists=True, compile=False, weights_precision=torch.bfloat16, ) _ = preload_tokenizer(tokenizer_filepath) _ = preload_semantic_models(semantic_model_filepath, semantic_clusters_filepath) _ = preload_codec_models(codec_filepath) # _ = preload_ear_model(ear_model_filepath) # print(f"CHUNK_SIZE_SCHEDULE: {CHUNK_SIZE_SCHEDULE}") # self.diffusion_engine = UpsampleEngine(chunk_size_schedule=CHUNK_SIZE_SCHEDULE) self.diffusion_engine = UpsampleEngine(min_chunk_size=BLOCK_SIZE) # self.diffusion_engine = UpsampleEngine( # min_chunk_size=750, vae_version="v_vae_25_tuned_2" # ) # load hoot model print("Loading hoot model") hoot_filepath = "s3://suno-data/christian/checkpoints/hoot_v3.pt" hoot_tokenizer_filepath = ( "s3://suno-data/christian/checkpoints/hoot_v3_tokenizer_v3.pt" ) self.hoot_model_list, self.hoot_model, self.hoot_tokenizer = ( _reload_models_if_needed(hoot_filepath, hoot_tokenizer_filepath) ) # load ear model ear_model_filepath = "s3://suno-data/christian/checkpoints/ear/ear_v2_s3080.pt" self.ear_model = load_model(ear_model_filepath, compile=True) print( f"Finish loading models. Took {round(time.time() - start_time, 2)} seconds" ) @staticmethod def download_models(dit_model_filepath, dir_path=MOUNT_PATH): """Download diffusion models.""" print("Start downloading models") _ = diffusion_gen.get_model_if_needed( CODEC_FILEPATH, cache_dir=dir_path, ) _ = diffusion_gen.get_model_if_needed( diffusion_gen.SEMANTIC_MODEL_FILEPATH, cache_dir=dir_path ) _ = diffusion_gen.get_model_if_needed( diffusion_gen.SEMANTIC_CLUSTERS_FILEPATH, cache_dir=dir_path ) _ = diffusion_gen.get_model_if_needed( diffusion_gen.CODEC_FILEPATH, cache_dir=dir_path ) _ = diffusion_gen.get_model_if_needed(dit_model_filepath, cache_dir=dir_path) # _ = chirp_v2._get_model_if_needed(gpt_ckpt, cache_dir=dir_path) # _ = chirp_v2._get_model_if_needed(chirp_v2.TOKENIZER_PATH, cache_dir=dir_path) print("Finish downloading models") def generate(self, work_item): global models """Generate audio from a work item.""" # when saving to s3 we will use a structure like this: # self.output_path/ # item_id/ # original_semantic.npz # 0_generated_audio.mp3 # 0_generated_semantic.npz # 0_metadata.json # 1_generated_audio.mp3 # 1_generated_semantic.npz # 1_metadata.json # ... # metadata.json for item in work_item: item_id = item["id"] if "s3_filepath" in item: s3_filepath = item["s3_filepath"] # read audio from s3 and then semantic encode audio = Audio.from_s3(s3_filepath, n_channels=2) if SAVE_CODES: # encode the vae latents vae_latents_cycled = codec_encode( audio.convert( sample_rate=48_000, byte_width=2, n_channels=2 ).normalize_volume(target_db=-16) ) with tempfile.TemporaryDirectory() as td: vae_latents_path = os.path.join(td, f"{item_id}_cycled_vae.npz") np.savez(vae_latents_path, vae_latents=vae_latents_cycled) s3_filepath = os.path.join( self.output_path, f"{item_id}", f"{item_id}_vae.npz", ) s3_client.upload_file( vae_latents_path, "suno-data", s3_filepath, ExtraArgs={"ContentType": "application/octet-stream"}, ) if CYCLE_ONLY: continue # also encode semantic codes codes = encode_semantic( audio.convert(sample_rate=24_000, byte_width=2, n_channels=1) ).astype(np.int64) if SAVE_CODES: with tempfile.TemporaryDirectory() as td: codes_path = os.path.join(td, f"{item_id}_semantic.npz") np.savez(codes_path, codes=codes) s3_filepath = os.path.join( self.output_path, f"{item_id}", f"{item_id}_semantic.npz", ) s3_client.upload_file( codes_path, "suno-data", s3_filepath, ExtraArgs={"ContentType": "application/octet-stream"}, ) else: # load the semantic codes from s3 s3_filepath = f"s3://suno-data-uploads/studio/uploads/{item_id}.npz" try: data = read_from_s3(s3_filepath, read_f=np.load) except Exception as e: print(f"Error loading {s3_filepath}: {e}") continue if "v3.0_raw" in data: codes = data["v3.0_raw"] elif "v3.5_raw" in data: codes = data["v3.5_raw"] elif "v4.0_raw" in data: codes = data["v4.0_raw"] elif "v4.5_raw" in data: codes = data["v4.5_raw"] elif "v5.0_raw" in data: codes = data["v5.0_raw"] else: raise ValueError("No codes found") semantic_codes = torch.from_numpy(codes[:, 0]).long() # .cuda() # semantic_codes = semantic_codes[:3000] print(semantic_codes.shape) # if tags is a list, join it into a string if isinstance(item["tags"], list): tags_str = ", ".join(item["tags"][:5]) else: tags_str = item["tags"] if tags_str is None: tags_str = "" lyrics = item.get("lyrics", None) if lyrics is None: lyrics = item.get("text", None) if lyrics is None: lyrics = "" print(tags_str) print(lyrics) # create a list of configs to run gen_cfgs = [] if self.test_type == "steps": diffusion_steps = [8, 10, 12, 14, 16, 18, 20] diffusion_seed = 42 diffusion_text_cfg_coef = 2.0 noise_ctx_level = NOISE_CTX_LEVEL noise_ctx_pad_len = NOISE_CTX_PAD_LEN rho = RHO sigma_min = SIGMA_MIN sigma_max = SIGMA_MAX for diffusion_steps in diffusion_steps: gen_cfgs.append( ( f"steps_{diffusion_steps}", diffusion_gen.DiffusionGenerationConfig( steps=diffusion_steps, lyrics=lyrics, tags=tags_str, text_cfg_coef=diffusion_text_cfg_coef, codec_scale_factor=CODEC_SCALE_FACTOR, scale_ctx_vector=SCALE_CTX_VECTOR, noise_ctx_level=noise_ctx_level, noise_ctx_pad_len=noise_ctx_pad_len, # sigma_min=sigma_min, sigma_max=sigma_max, seed=diffusion_seed, # noise_schedule=NOISE_SCHEDULE, # rank_candidates=RANK_CANDIDATES, ), ) ) elif self.test_type == "standard": # diffusion parameters if RANDOMIZE_DIFFUSION_PARAMS: diffusion_seed = np.random.randint(0, 1000000) diffusion_steps = np.random.choice([8, 10, 12, 14, 16, 18, 20]) diffusion_text_cfg_coef = np.random.choice( [1.0, 1.5, 1.75, 2.0, 2.5, 3.0] ) noise_ctx_level = np.random.choice([0.0, 0.25, 0.5, 0.75, 1.0]) rho = np.random.choice([0.9, 1.0, 1.1]) sigma_min = np.random.choice([0.05, 0.1, 0.25, 0.5]) sigma_max = np.random.choice([40.0, 50.0, 60.0, 70.0, 80.0, 100.0]) noise_ctx_pad_len = np.random.choice([0, 10, 20, 30]) else: if DIFFUSION_SEED is None: diffusion_seed = 0 for char in item_id: diffusion_seed = (diffusion_seed * 31 + ord(char)) % 1000000 print(f"Generated seed from '{item_id}': {diffusion_seed}") else: diffusion_seed = DIFFUSION_SEED print(f"Using seed from config: {diffusion_seed}") diffusion_steps = DIFFUSION_STEPS noise_ctx_level = NOISE_CTX_LEVEL noise_ctx_pad_len = NOISE_CTX_PAD_LEN rho = RHO sigma_min = SIGMA_MIN sigma_max = SIGMA_MAX sampler_type = SAMPLER_TYPE for n in range(1): # diffusion_seed = np.random.randint(0, 1000000) gen_cfgs.append( ( "", diffusion_gen.DiffusionGenerationConfig( steps=diffusion_steps, lyrics=lyrics, tags=tags_str, text_cfg_coef=TEXT_CFG_SCALE, ctx_cfg_coef=CTX_CFG_SCALE, codec_scale_factor=CODEC_SCALE_FACTOR, scale_ctx_vector=SCALE_CTX_VECTOR, noise_ctx_level=noise_ctx_level, noise_ctx_pad_len=noise_ctx_pad_len, semantic_skip_factor=SEMANTIC_SKIP_FACTOR, rho=rho, sigma_min=sigma_min, sigma_max=sigma_max, seed=diffusion_seed, objective=OBJECTIVE, sampler_type=sampler_type, # rank_candidates=RANK_CANDIDATES, ), ) ) else: raise ValueError(f"Invalid test type: {self.test_type}") # now run the requests for name, gen_cfg in gen_cfgs: request = Request( id="dummy", generation_config=gen_cfg, tokens=semantic_codes, input_tokens_finished=True, # ignore_history_every=IGNORE_HISTORY_EVERY, ) result = self.diffusion_engine.run_request(request) vae_latents = torch.concat(result.vae_latents) upsampled_audio = decode_stream_to_full_audio(vae_latents) metadata = { # "original_audio": item["s3_filepath"], "id": item_id, "text": lyrics, "tags": tags_str, "diffusion": { "steps": int(gen_cfg.steps), "seed": int(gen_cfg.seed), "text_cfg_coef": float(gen_cfg.text_cfg_coef), "noise_ctx_level": float(gen_cfg.noise_ctx_level), "codec_scale_factor": float(gen_cfg.codec_scale_factor), "scale_ctx_vector": gen_cfg.scale_ctx_vector, # "rho": float(gen_cfg.rho), }, } # we want to save out # audio file of the final audio # npz of the estimated semantics # metadata json with the original prompt and tags # copy some stuff # we want to copy the original audio and the npz of the original semantics with tempfile.TemporaryDirectory() as td: # save audio to s3 upsampled_audio_path = os.path.join(td, f"{item_id}_{name}.mp3") upsampled_audio.write_hq_mp3(upsampled_audio_path) # run hoot evaluation (CER) # check if the lyrics are not empty if lyrics != "": upsampled_audio_obj = Audio.from_file( upsampled_audio_path, n_channels=2 ) if False: aligned_lyrics = encode_and_align( upsampled_audio_obj, prior_texts=lyrics.lower(), ) print(aligned_lyrics) out = encode(upsampled_audio_obj, return_logits=True) basic_cleaned_lyrics = lyrics.lower() decoded_preds = self.hoot_tokenizer.decode_logits( out, prior_text=basic_cleaned_lyrics ) true_text_norm = clean_text(basic_cleaned_lyrics) cer_val = round(get_cer(true_text_norm, decoded_preds), 3) metadata["hoot_cer"] = cer_val # add cer to metadata print("hoot cer: ", metadata["hoot_cer"]) else: metadata["hoot_cer"] = None # run ear evaluation (quality score) # check if the length of the audio is greater than 5 seconds if upsampled_audio.duration_s > 5.0: quality_score = self.ear_model.get_score(upsampled_audio_path) metadata["ear_score"] = ( quality_score # add quality score to metadata ) print("ear score: ", metadata["ear_score"]) else: metadata["ear_score"] = None # run shimmer score evaluation shimmer_score = shimmerscore(upsampled_audio_path) metadata["shimmer_score"] = shimmer_score # run the other evals audio_analysis = analyze_audio(upsampled_audio_path) # add all the keys in audio_analysis to metadata for key, value in audio_analysis.items(): metadata[key] = value s3_filepath = os.path.join( self.output_path, f"{item_id}", ( f"{item_id}_{CHECKPOINT_NAME}_{name}.mp3" if name != "" else f"{item_id}_{CHECKPOINT_NAME}.mp3" ), ) print(f"Uploading to {s3_filepath}") s3_client.upload_file( upsampled_audio_path, "suno-data", s3_filepath, ExtraArgs={ "ContentType": "audio/mpeg", }, ) if SAVE_CODES: print("Saving vae latents...") # save the vae latents vae_latents_path = os.path.join( td, f"{item_id}_upsampled_vae.npz" ) np.savez( vae_latents_path, vae_latents=vae_latents.cpu().numpy() ) s3_filepath = os.path.join( self.output_path, f"{item_id}", f"{item_id}_upsampled_vae.npz", ) s3_client.upload_file( vae_latents_path, "suno-data", s3_filepath, ExtraArgs={ "ContentType": "application/octet-stream", }, ) # save the metadata metadata_path = os.path.join(td, f"{item_id}_{name}_metadata.npz") np.savez(metadata_path, **metadata) # move the metadata to s3 s3_filepath = os.path.join( self.output_path, f"{item_id}", ( f"{item_id}_{CHECKPOINT_NAME}_{name}__metadata.npz" if name != "" else f"{item_id}_{CHECKPOINT_NAME}__metadata.npz" ), ) s3_client.upload_file( metadata_path, "suno-data", s3_filepath, ExtraArgs={ "ContentType": "application/octet-stream", }, ) # prod diffusion model # DIT_MODEL_FILEPATH = "s3://suno-data/tony/tmp/diff/dit_v3_dpo_t10_3k_5e6_b100_t25.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/georg/tmp/2b_prefix_ft.pt" # 2b v45 # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/dpo/v1_t6_1E6_beta100_n8_bt4.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t6_2E6_beta100_n8_bt4.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t6_5E6_beta100_n16_bt4.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t6_5E6_beta100_n8_bt4_30k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t10_5E6_beta100_n8_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v45_2b_shared_ctx_s3784.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_infill_t6_5E6_beta100_n16_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_infill_t11_5E6_beta100_n16_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_infill_t6_5E6_beta100_n16_bt4_6k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t12_5E6_beta100_n16_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t13_5E6_beta100_n16_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t13_5E6_beta100_n16_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t15_5E6_beta100_n16_bt4_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_10k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_20k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_infill_shared_base_200k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_infill_v1_t6_5E6_beta100_n16_bt4_30k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_100k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_5e5_150k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_infill_shared_base_600k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_t3_5e5_200k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_base_ear_sft_50k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_infill_v1_t17_5E6_beta100_n16_bt4_9k_repro_main_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t17_5E6_beta100_n16_bt4_9k_repro_mai_ema.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_infill_v1_t17_5E6_beta100_n16_bt4_9k_repro_main_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_base_ear_sft_100k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t17_5E6_beta1000_n16_bt4_9k_repro_main_6k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t17_5E6_beta1000_n16_bt4_9k_repro_main_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t17_5E6_beta1000_n16_bt4_2k_repro_main.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t6_5E6_beta1000_n16_bt4_2k_repro_main.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_base_ear_sft_t3+t5_40k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_t4_5e5_180k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_ft_ear_t4_5e5_270k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t6_2E6_beta100_n16_bt4_1k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_base_ear_sft_t3_190k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_v45_infill_ear_sft_3e5_t6_100k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_base_ear_sft_t3_210k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t17_2E6_beta100_n16_bt4_1p5k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t17_2E6_beta5000_n8_bt4_1p5k_fix.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_base_ear_sft_t3_230k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t6_2E6_beta5000_n8_bt4_3k_fix.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t6_5E6_beta100_n8_bt4_3k_acc16_fix_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t6_5E6_beta100_n8_bt4_3k_acc16_fix_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_v45_infill_ear_sft_3e5_t6_250k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t18_1E6_beta100_n32_bt4_3k_acc4_fix_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_sft_t6_dpo_t18_1E6_beta100_n32_bt4_1k_acc8_fix_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_sft_t6_dpo_t18_1E6_beta100_n32_bt4_1k_acc8_fix_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t18_1E6_beta100_n32_bt4_1k_acc8_fix_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_genius_t6_sampled_10k_5E6_beta100_n12_bt4_3k_acc4_fix_6k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_sft_t18_dpo_t18_5E6_beta100_n16_bt4_1k_acc8_fix_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/dit_v6_dpo_t11_9k_5e6_b100.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v1_t6_5E6_beta100_n16_bt4.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t18_1E6_beta100_n32_bt4_1k_acc8_fix_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_sft_t6_dpo_t6_5E6_beta100_n32_bt4_1k_acc4_fix_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v2_t3_1E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_v45_infill_ear_sft_5e5_t6_c0_2_30k.pt" # sft model # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/diff_v2_2b_2mil_ft_infill_20250421_v2.pt" # current prod model # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v45_2b_step_2mil_ft_8k_infill_apr21_t1_18_cs0.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t18+syn_5E6_beta100_n16_bt2_3k_acc_4_6k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t18+syn_5E6_beta100_n16_bt2_1k_acc_8_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_v45_infill_ear_sft_1e5_t7_70k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t18+syn_5E6_beta100_n16_bt2_1k_acc4_sft.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_syn_corr_5E6_beta100_n16_bt2_3k_acc4_sft_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_v45_infill_ear_sft_1e5_t7_380k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_1E6_beta100_n32_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_1E6_beta100_n32_bt4_acc4_3k_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_5E6_beta100_n16_bt2_acc8_1k_sft_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_5E6_beta100_n16_bt2_acc8_1k_sft_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_syn_5E6_beta100_n16_bt2_acc8_1k_sft_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_syn_5E6_beta100_n16_bt2_acc8_1k_sft_0_1_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_5E6_beta100_n32_bt4_acc4_3k_1k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_1E6_beta100_n16_bt2_acc8_1k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_1E6_beta100_n32_bt4_acc4_3k_1k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t2_1E6_beta100_n16_bt4_acc4_1k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t2_1E6_beta100_n16_bt4_acc4_1k_1k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_ear_1E6_beta100_n4_bt2_acc8_1k_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_ear_neg_0_5_1E6_beta100_n4_bt2_acc8_1k_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_ear_1E6_beta100_n8_bt2_acc8_2k_16k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_apr21_t1_18_cs0_t1_ear_1E6_beta100_n4_bt2_acc8_4k_32k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v3_t1_ear_1E6_beta100_n4_bt2_acc8_4k_32k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t6_1E6_beta100_n16_bt4_acc4_1k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t7_1E6_beta100_n16_bt4_acc4_1k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_og_d3_t7_1E6_beta100_n16_bt4_acc4_1k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_og_d3_t9_1E6_beta100_n16_bt4_acc4_1k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_ear_t1_1E6_beta100_n4_bt2_acc8_4k_32k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t9_2E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t19_1E6_beta100_n16_bt4_acc4_3k_12k.pt" # # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t10_1E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_resume_750k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t12_1E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t13_1E6_beta100_n16_bt4_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t15_1E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t16_1E6_beta100_n16_bt4_acc4_3k_10k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t17_1E6_beta100_n16_bt4_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t17_1E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t19_1E6_beta100_n8_bt4_acc8_3k_24k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t18_1E6_beta100_n8_bt4_acc8_3k_24k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t20_1E6_beta100_n16_bt16_acc4_6k_24k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t21_1E6_beta100_n16_bt16_acc4_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_resume_1_25m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d3_t10_1E6_beta100_n16_bt16_acc8_3k_24k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_v1_t19_5E6_beta100_n16_bt4_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d4_2_5E6_beta100_n16_bt2_acc4_4k_16k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_4b_1m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_resume_1_5m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_4b_1_25m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d4_3_5E6_beta100_n16_bt2_acc4_4k_16k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d4_t6_freeze_5E6_beta100_n16_bt2_acc4_4k_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_v45_infill_5e5_sft_log_snr_20k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_resume_1_75m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_4b_resume_2m.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_v45_distill_adv_100k.pt" # ) # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v45_2b_step_2mil_ft_8k_infill_apr21_d3_v10.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_v45_distill_adv2.pt" # ) # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_infill_shared_flow_5e5_sft_50k.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/8n_2b_adv_bs2_N5_5e6_001_50k.pt" # ) # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/4n_2b_adv_bs2_N5_5e6_acc4_t1.pt" # ) # /app/suno/modal/models/tony/tmp/diff/v45_2b_step_2mil_ft_8k_infill_apr21_t1_18_cs0.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_bootstrap_t2_1E6_beta100_n16_bt2_acc2_4k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_bootstrap_t2_1E6_beta100_n16_bt2_acc2_4k_8k.pt" DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_infill_d4_t39_1E6_beta100_n16_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_r1_1E6_beta100_n4_bt2_acc2_4k_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_infill_shared_flow_revert_resume_1_4m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e7_acc2_ema_alt_dmd_cfg2_400k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diff_v2_2b_2mil_ft_v0.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_resume_1_75m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_2b_flow_sft_bs4_t8_5e6_300k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_25hz_v45_infill_shared_flow_60s_180k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_bootstrap_t4_1E6_beta100_n8_bt2_acc2_4k_test_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_bootstrap_t4_1E6_beta100_n8_bt2_acc2_4k_test_9k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg2_cosine_210k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_2_cosine_nu1_0p1_160k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg2_cosine_400k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_2b_flow_1e6_syn_sft_t1_30k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_2b_flow_5e5_sft_t8_250k.pt" # # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_2b_flow_1e6_syn_sft_t2_7k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg2_cosine_700k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_2_cosine_nu1_0p1_500k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1_cosine_nu1_0p1_residual_5k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_2_cosine_nu1_0p1_residual_6k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_2_cosine_nu1_0p1_residual_13k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p5_cosine_nu1_0p1_residual_10k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p5_cosine_nu1_0p1_residual_50k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p5_cosine_nu1_0p1_residual_70k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p25_cosine_nu1_0p1_residual_20k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_1E6_beta100_n8_bt2_acc2_4k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p25_residual_aug_80k.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_2b_flow_5e5_sft_t8_500k.pt" # ) # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_1E6_beta100_n8_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p5_cosine_nu1_0p1_residual_170k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p25_cosine_nu1_0p1_residual_120k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e7_alt_dmd_cfg_1p25_residual_sft_50k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e7_alt_dmd_cfg_1p25_residual_sft_100k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e7_alt_dmd_cfg_2_residual_sft_rm_50k.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_2b_flow_5e5_sft_t8_725k.pt" # ) # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e7_alt_dmd_cfg_1p25_residual_sft_120k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_2b_flow_distill_bs1_N5_c5e5_g1e6_alt_dmd_cfg_2_residual_sft_65k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_2b_flow_distill_bs1_N5_c5e5_g1e6_alt_dmd_cfg_2_residual_sft_10k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_ctx_t0_1E6_beta100_n4_bt2_acc2_4k_4k.pt" # # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_2b_flow_distill_bs1_N5_c5e5_g1e6_alt_dmd_cfg_2_residual_sft_40k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_ctx_t0_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_2b_flow_distill_bs1_N5_c5e5_g1e6_alt_dmd_cfg_2_residual_sft_125k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_t1_1E6_beta100_n4_bt2_acc2_8k_6k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_t0_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_2b_flow_distill_bs1_N5_c5e5_g1e6_alt_dmd_cfg_2_residual_sft_220k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_v1_t18_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_v2_d4_v39_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_v1_t18_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_v1_t18_1E6_beta100_n15_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_lm_t0_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/4n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_residual_sft_45k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_lm_t1_0_6_1E6_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_lm_t1_0_6_5E7_beta100_n4_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_lm_t1_0_6_0822_5E7_beta100_n8_bt2_acc2_4k_8k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_lm_t1_0_6_0822_5E7_beta100_n8_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_6_0822_5E7_beta100_n8_bt2_acc2_4k_last.pt" # # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_6_0822_1_sft_0p1_5E7_beta100_n8_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_6_0823_0_0p1_5E7_beta100_n8_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_6_0824_0_1E6_beta100_n8_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_7_cut_0824_0_1E6_beta100_n8_bt2_acc4_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_7_cut_0824_0_1E6_beta100_n8_bt2_acc4_2k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_7_cut_history_4x_1E6_beta100_n8_bt2_acc4_2k_last.pt" # # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0_7_cut_history_4x_1E6_beta100_n8_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_carp_t1_v1_1E6_beta100_n4_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_lm_t1_0p6_cut_history_4x_1E6_beta100_n4_bt2_acc4_2k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3395_rd2_t1_merge_1E6_beta100_n8_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_lm_t2_0_7_cut_history_4x_1E6_beta100_n8_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_merge_t1_1E6_beta100_n8_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t0_1E6_beta100_n8_bt2_acc4_3k_12k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_merge_t1_1E6_beta100_n8_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t0_1E6_beta100_n8_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t1_1E6_beta100_n16_bt2_acc1_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t1_1E6_beta50_n16_bt2_acc1_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t1_1E6_beta25_n16_bt2_acc1_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t2_1E6_beta100_n16_bt2_acc1_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3177_rd2_hoot_t2_1E6_beta100_n16_bt2_acc1_6k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_residual_sft_t8_200k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s1039_rd2_lm_t2_1E6_beta100_n16_bt2_acc1_6k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_distill_s3242_rd1_v1_t18_1E6_beta100_n16_bt2_acc1_6k_last.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/4n_25hz_2b_flow_5e5_sft_t8_500k.pt" # ) # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_v1_t18_1E6_beta100_n16_bt2_acc2_4k_4k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_v1_t18_1E6_beta100_n16_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t2_merge_1E6_beta100_n16_bt2_acc2_4k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_v1_t18_1E6_beta100_n16_bt2_acc4_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_v1_t18_5E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d0_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d1_1E6_beta100_n16_bt2_9k_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d1_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d3_1E6_beta100_n8_bt2_9k_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d4_1E6_beta100_n14_bt2_9k_2k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d4_5E7_beta100_n14_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d5_5E7_beta100_n16_bt2_6k_3k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d5_5E7_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d6_5E7_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d6_5E7_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d8_5E7_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d9_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d9_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d11_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d10_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d11_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d11_1E6_beta100_n16_bt2_1p5k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d12_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d13_1E6_beta100_n14_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d14_1E6_beta100_n14_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d15_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d15_1E6_beta50_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d16_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d18_1E6_beta100_n12_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d18_1E6_beta200_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d19_1E6_noise1_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d20_1E6_noise1_beta100_n16_bt2_1p5k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d21_1E6_noise1_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d21_5E7_noise1_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d21_5E7_noise1_beta100_n8_bt2_1k_500.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d22_1E6_noise1_beta100_n8_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d22_1E6_noise1_beta100_n8_bt2_1k_500.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_d23_1E6_noise1_beta100_n8_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_base_rd2_d24_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_base_rd2_d24_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_base_rd1_d25_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d25_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_base_rd1_d26_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_base_rd1_d27_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d28_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d30_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d30_1E6_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d31_1E6_beta100_n8_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d32_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g5e6_alt_dmd_cfg_1p5_residual_sft_t8_rm_s4862_10k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d33_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d34_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d34_5E7_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d34_5E7_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_diffv2_v1_t18_1E6_beta100_n16_bt2_sft0p01_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_d35_1E6_beta200_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t0_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t1_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t2_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t2_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t2_1E6_beta100_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t3_1E6_beta100_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t4_1E6_beta100_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t5_1E6_beta100_n16_bt2_noise_1k_last.pt" # IT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t5_1E6_beta100_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t5_1E6_beta200_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t5_1E6_beta100_n16_bt2_noise_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_pair_t9_1E6_beta100_n8_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_pair_t9_1E6_beta200_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_pair_t10_1E6_beta100_n16_bt2_noise_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_t3_pair_t10_1E6_beta100_n16_bt2_noise_500_last.pt" # DIT_MODEL_FILEPATH = ( # "s3://suno-data/christian/checkpoints/diffusion/v45_2b_step_2_600_000.pt" # ) # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/16n_25hz_v45_infill_shared_flow_resume_1_75m.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t14_1E6_sft_1E4_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t14_1E6_sft_only_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t14_1E6_sft_1E4_beta100_n16_bt2_9k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_t14_1E6_sft_1E4_beta100_n16_bt2_3k_last.pt" # $DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_t14_1E6_sft_1E4_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_t15_1E6_sft_1E4_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t14_1E6_sft_1E3_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd2_t15_1E6_sft_1E2_beta100_n16_bt2_acc4_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t15_1E6_sft_1E2_beta100_n16_bt2_acc4_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/8n_2b_flow_distill_bs1_N5_c5e5_g2e6_alt_dmd_cfg_1p5_residual_s3431_100k.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t16_1E6_sft_1E2_beta100_n16_bt2_acc4_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t17_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t18_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t19_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t19_1E6_beta100_n16_bt2_3k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t20_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t21_1E6_beta100_n16_bt2_1k_last.pt" # DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v2_bass_d5_t10_1E6_beta100_n16_bt2_acc4_noise1_3k_off_d4_v40_last.pt" DIT_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/diffusion/v3_flow_sft_t8_rd1_t22_1E6_beta100_n16_bt2_3k_last.pt" CHECKPOINT_NAME = DIT_MODEL_FILEPATH.split("/")[-1].replace(".pt", "") # OUTPUT_STR = "farm-speaker-desk" # OUTPUT_STR = "pencil-helmet-window" # OUTPUT_STR = "walking-flowers" # OUTPUT_STR = "reference-audio" # OUTPUT_STR = "auk-test" # OUTPUT_STR = "auk-clips-up-u-2" OUTPUT_STR = "bluejay-lang-balanced-200" # OUTPUT_STR = "v2-infill-data-v1" # OUTPUT_STR = "v2-diff-step-test" # OUTPUT_STR = "real-audio-test" TEST_TYPE = "standard" # standard, steps SAVE_CODES = False CYCLE_ONLY = False GPU_TYPE = "H100" # H100, A10G OBJECTIVE = "rectified_flow" if "flow" in DIT_MODEL_FILEPATH else "v" SAMPLER_TYPE = "pingpong" if "distill" in DIT_MODEL_FILEPATH else "dpmpp" # extra params RHO = 1.0 SIGMA_MIN = 0.5 SIGMA_MAX = 50.0 BLOCK_SIZE = 25 * 30 # 30s chunks DIFFUSION_SEED = None # 42 is default RANDOMIZE_DIFFUSION_PARAMS = False USE_ALTERNATE_CFG = False TEXT_CFG_SCALE = 2.0 if "distill" not in DIT_MODEL_FILEPATH else 1.0 CTX_CFG_SCALE = 1.0 CTX_STEP_PERCENT = 1.0 RANK_CANDIDATES = 1 # IGNORE_HISTORY_EVERY = 2 EAR_MODEL_FILEPATH = "s3://suno-data/christian/checkpoints/ear/ear_v3_s8558.pt" # use this to add a suffix to the checkpoint name EXTRA_NAME_SUFFIX = "_step10" if "distill" not in DIT_MODEL_FILEPATH else "_step2" CHECKPOINT_NAME = CHECKPOINT_NAME + EXTRA_NAME_SUFFIX if "dit_v6_dpo_t11_9k_5e6_b100" in DIT_MODEL_FILEPATH: CODEC_SCALE_FACTOR = 2.5 SCALE_CTX_VECTOR = False NOISE_CTX_LEVEL = 0.0 NOISE_CTX_PAD_LEN = 0 DIFFUSION_STEPS = 10 SEMANTIC_SKIP_FACTOR = 1 CODEC_FILEPATH = "s3://suno-data/christian/25hz_vae_peaq_kl_0.005.pth" from suno_utils.tasks.dac_vae_100hz_peaq import ( preload_models as preload_codec_models, decode as codec_decode, encode as codec_encode, decode_stream_to_full_audio, ) else: CODEC_SCALE_FACTOR = 0.4 SCALE_CTX_VECTOR = True NOISE_CTX_LEVEL = 0.75 NOISE_CTX_PAD_LEN = 0 DIFFUSION_STEPS = 10 if "distill" not in DIT_MODEL_FILEPATH else 2 SEMANTIC_SKIP_FACTOR = 1 NOISE_SCHEDULE = "polyexponential" # NOISE_SCHEDULE = "cosine" # CHUNK_SIZE_SCHEDULE = [10 * 25, 20 * 25, 30 * 25] # 10s, 20s, 30s chunks # CHUNK_SIZE_SCHEDULE = [30 * 25, 30 * 25, 30 * 25] # 30s chunks CODEC_FILEPATH = "s3://suno-data/minz/models/dac_vae_tuned_25hz.pth" from suno_utils.tasks.dac_vae_fixed_25hz import ( preload_models as preload_codec_models, decode as codec_decode, encode as codec_encode, decode_stream_to_full_audio, ) def download_model_wrapper_d(): # this print is necessary to have modal rerun this when MODEL changes # Modal tracks referenced global variables # Change the name of the function to force a rerun print("Downloading model for history encoder") GenerateWorker.download_models(DIT_MODEL_FILEPATH) image = base_image.run_function(download_model_wrapper_d, secrets=SECRETS) APP_NAME = f"batch-eval-diffusion" app = modal.App(APP_NAME, image=image, secrets=SECRETS) N_MAX_REPLICAS = 16 @app.cls( # gpu=modal.gpu.H100(count=1), gpu=GPU_TYPE, cpu=4, secrets=SECRETS, timeout=2 * 60 * 60, container_idle_timeout=240, # mounts=MODAL_MOUNTS, memory=15000, concurrency_limit=N_MAX_REPLICAS, ) class GenerateStub: def __init__(self, dit_ckpt: str, output_path: str, test_type: str): import torch num_gpus = torch.cuda.device_count() print(f"Found {num_gpus} GPUs.") self.worker = GenerateWorker( dit_ckpt, output_path, test_type, ) @modal.method() def generate(self, work_item: list[dict]): return self.worker.generate(work_item) @app.local_entrypoint() def main(): # checkpoints in s3://suno-data/christian/checkpoints # prompts in s3://suno-data/christian/prompts # outputs in s3://suno-data/christian/outputs # load the positive prompts # work_items = read_from_s3( # "s3://suno-data/christian/sft/pos_interesting_clips_up_u_1_20241201_full.jsonl", # read_f=read_jsonl, # ) # work_items = read_jsonl( # "/home/christian/code/christian/metadata/sft/auk_clips_up_u_1_20241201_pos.jsonl", # ) work_items = read_jsonl( "/home/christian/code/christian/metadata/sft/interesting_clips_bluejay_t1_20250811_public_only_lang_balanced_200.jsonl", ) # work_items = read_jsonl( # "/home/christian/audio/reference-audio.jsonl", # ) # work_items = read_jsonl( # "/home/christian/code/christian/metadata/discogs_subset_sampled_metas.jsonl" # # work_items = read_jsonl( # "/home/christian/code/christian/metadata/ear/genius_t6_sampled_10k.jsonl" # ) # just do the first 128 work_items = work_items # [:128] print(f"Total work items: {len(work_items)}") # first check for existing ids in the output path # existing_ids = list_s3_dir( # f"s3://suno-data/christian/outputs/{OUTPUT_STR}/", # ) # existing_ids = [os.path.dirname(f[0]).split("/")[-1] for f in existing_ids] # existing_ids = list(set(existing_ids)) # print(f"Total existing ids: {len(existing_ids)}") ## now remove these from the work items # work_items = [f for f in work_items if f["id"] not in existing_ids] print(f"Total work items remaining: {len(work_items)}") chunksize = 8 # num of prompts per worker worker = GenerateStub( DIT_MODEL_FILEPATH, f"christian/outputs/{OUTPUT_STR}", test_type=TEST_TYPE ) work_items = list(funcy.chunks(chunksize, work_items)) print(f"Chunksize: {chunksize}, total chunks: {len(work_items)}") # print("Testing inference...") # t0 = time.time() # for work_item in work_items[:1]: # _ = worker.generate.remote(work_item) # print(f"{int(round(time.time()-t0))}s for test") print("Running batch inference...") t0 = time.time() _ = list(worker.generate.map(work_items)) print(round((time.time() - t0) / 60 / 60), "h total for batch generation")