Upload 7 files

__init__.py

@@ -0,0 +1,2 @@
+from .api import generate_audio, text_to_semantic, semantic_to_waveform, save_as_prompt
+from .generation import SAMPLE_RATE, preload_models

__main__.py

@@ -0,0 +1,3 @@
+from .cli import cli
+
+cli()

api.py

@@ -0,0 +1,125 @@
+from typing import Dict, Optional, Union
+
+import numpy as np
+
+from .generation import codec_decode, generate_coarse, generate_fine, generate_text_semantic
+
+
+def text_to_semantic(
+    text: str,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    temp: float = 0.7,
+    silent: bool = False,
+):
+    """Generate semantic array from text.
+
+    Args:
+        text: text to be turned into audio
+        history_prompt: history choice for audio cloning
+        temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+
+    Returns:
+        numpy semantic array to be fed into `semantic_to_waveform`
+    """
+    x_semantic = generate_text_semantic(
+        text,
+        history_prompt=history_prompt,
+        temp=temp,
+        silent=silent,
+        use_kv_caching=True
+    )
+    return x_semantic
+
+
+def semantic_to_waveform(
+    semantic_tokens: np.ndarray,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    temp: float = 0.7,
+    silent: bool = False,
+    output_full: bool = False,
+):
+    """Generate audio array from semantic input.
+
+    Args:
+        semantic_tokens: semantic token output from `text_to_semantic`
+        history_prompt: history choice for audio cloning
+        temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+        output_full: return full generation to be used as a history prompt
+
+    Returns:
+        numpy audio array at sample frequency 24khz
+    """
+    coarse_tokens = generate_coarse(
+        semantic_tokens,
+        history_prompt=history_prompt,
+        temp=temp,
+        silent=silent,
+        use_kv_caching=True
+    )
+    fine_tokens = generate_fine(
+        coarse_tokens,
+        history_prompt=history_prompt,
+        temp=0.5,
+    )
+    audio_arr = codec_decode(fine_tokens)
+    if output_full:
+        full_generation = {
+            "semantic_prompt": semantic_tokens,
+            "coarse_prompt": coarse_tokens,
+            "fine_prompt": fine_tokens,
+        }
+        return full_generation, audio_arr
+    return audio_arr
+
+
+def save_as_prompt(filepath, full_generation):
+    assert(filepath.endswith(".npz"))
+    assert(isinstance(full_generation, dict))
+    assert("semantic_prompt" in full_generation)
+    assert("coarse_prompt" in full_generation)
+    assert("fine_prompt" in full_generation)
+    np.savez(filepath, **full_generation)
+
+
+def generate_audio(
+    text: str,
+    history_prompt: Optional[Union[Dict, str]] = None,
+    text_temp: float = 0.7,
+    waveform_temp: float = 0.7,
+    silent: bool = False,
+    output_full: bool = False,
+):
+    """Generate audio array from input text.
+
+    Args:
+        text: text to be turned into audio
+        history_prompt: history choice for audio cloning
+        text_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        waveform_temp: generation temperature (1.0 more diverse, 0.0 more conservative)
+        silent: disable progress bar
+        output_full: return full generation to be used as a history prompt
+
+    Returns:
+        numpy audio array at sample frequency 24khz
+    """
+    semantic_tokens = text_to_semantic(
+        text,
+        history_prompt=history_prompt,
+        temp=text_temp,
+        silent=silent,
+    )
+    out = semantic_to_waveform(
+        semantic_tokens,
+        history_prompt=history_prompt,
+        temp=waveform_temp,
+        silent=silent,
+        output_full=output_full,
+    )
+    if output_full:
+        full_generation, audio_arr = out
+        return full_generation, audio_arr
+    else:
+        audio_arr = out
+    return audio_arr

cli.py

@@ -0,0 +1,71 @@
+import argparse
+from typing import Dict, Optional, Union
+import os
+
+from scipy.io.wavfile import write as write_wav
+from .api import generate_audio
+from .generation import SAMPLE_RATE
+
+
+def cli():
+    """Commandline interface."""
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument("--text", type=str, help="text to be turned into audio")
+    parser.add_argument(
+        "--output_filename",
+        type=str,
+        default="bark_generation.wav",
+        help="output audio file name",
+    )
+    parser.add_argument("--output_dir", type=str, default=".", help="directory to save the outputs")
+    parser.add_argument(
+        "--history_prompt",
+        type=str,
+        default=None,
+        help="history choice for audio cloning, be path to the .npz file.",
+    )
+    parser.add_argument(
+        "--text_temp",
+        default=0.7,
+        type=float,
+        help="generation temperature (1.0 more diverse, 0.0 more conservative)",
+    )
+    parser.add_argument(
+        "--waveform_temp",
+        default=0.7,
+        type=float,
+        help="generation temperature (1.0 more diverse, 0.0 more conservative)",
+    )
+    parser.add_argument("--silent", default=False, type=bool, help="disable progress bar")
+    parser.add_argument(
+        "--output_full",
+        default=False,
+        type=bool,
+        help="return full generation to be used as a history prompt",
+    )
+
+    args = vars(parser.parse_args())
+    input_text: str = args.get("text")
+    output_filename: str = args.get("output_filename")
+    output_dir: str = args.get("output_dir")
+    history_prompt: str = args.get("history_prompt")
+    text_temp: float = args.get("text_temp")
+    waveform_temp: float = args.get("waveform_temp")
+    silent: bool = args.get("silent")
+    output_full: bool = args.get("output_full")
+
+    try:
+        os.makedirs(output_dir, exist_ok=True)
+        generated_audio = generate_audio(
+            input_text,
+            history_prompt=history_prompt,
+            text_temp=text_temp,
+            waveform_temp=waveform_temp,
+            silent=silent,
+            output_full=output_full,
+        )
+        output_file_path = os.path.join(output_dir, output_filename)
+        write_wav(output_file_path, SAMPLE_RATE, generated_audio)
+        print(f"Done! Output audio file is saved at: '{output_file_path}'")
+    except Exception as e:
+        print(f"Oops, an error occurred: {e}")

generation.py

@@ -0,0 +1,820 @@
+import contextlib
+import gc
+import os
+import re
+
+from encodec import EncodecModel
+import funcy
+import logging
+import numpy as np
+from scipy.special import softmax
+import torch
+import torch.nn.functional as F
+import tqdm
+from transformers import BertTokenizer
+from huggingface_hub import hf_hub_download
+
+from .model import GPTConfig, GPT
+from .model_fine import FineGPT, FineGPTConfig
+
+if (
+    torch.cuda.is_available() and
+    hasattr(torch.cuda, "amp") and
+    hasattr(torch.cuda.amp, "autocast") and
+    hasattr(torch.cuda, "is_bf16_supported") and
+    torch.cuda.is_bf16_supported()
+):
+    autocast = funcy.partial(torch.cuda.amp.autocast, dtype=torch.bfloat16)
+else:
+    @contextlib.contextmanager
+    def autocast():
+        yield
+
+
+# hold models in global scope to lazy load
+global models
+models = {}
+
+global models_devices
+models_devices = {}
+
+
+CONTEXT_WINDOW_SIZE = 1024
+
+SEMANTIC_RATE_HZ = 49.9
+SEMANTIC_VOCAB_SIZE = 10_000
+
+CODEBOOK_SIZE = 1024
+N_COARSE_CODEBOOKS = 2
+N_FINE_CODEBOOKS = 8
+COARSE_RATE_HZ = 75
+
+SAMPLE_RATE = 24_000
+
+
+SUPPORTED_LANGS = [
+    ("English", "en"),
+    ("German", "de"),
+    ("Spanish", "es"),
+    ("French", "fr"),
+    ("Hindi", "hi"),
+    ("Italian", "it"),
+    ("Japanese", "ja"),
+    ("Korean", "ko"),
+    ("Polish", "pl"),
+    ("Portuguese", "pt"),
+    ("Russian", "ru"),
+    ("Turkish", "tr"),
+    ("Chinese", "zh"),
+]
+
+ALLOWED_PROMPTS = {"announcer"}
+for _, lang in SUPPORTED_LANGS:
+    for prefix in ("", f"v2{os.path.sep}"):
+        for n in range(10):
+            ALLOWED_PROMPTS.add(f"{prefix}{lang}_speaker_{n}")
+
+
+logger = logging.getLogger(__name__)
+
+
+CUR_PATH = os.path.dirname(os.path.abspath(__file__))
+
+
+default_cache_dir = os.path.join(os.path.expanduser("~"), ".cache")
+CACHE_DIR = os.path.join(os.getenv("XDG_CACHE_HOME", default_cache_dir), "suno", "bark_v0")
+
+
+def _cast_bool_env_var(s):
+    return s.lower() in ('true', '1', 't')
+
+
+USE_SMALL_MODELS = _cast_bool_env_var(os.environ.get("SUNO_USE_SMALL_MODELS", "False"))
+GLOBAL_ENABLE_MPS = _cast_bool_env_var(os.environ.get("SUNO_ENABLE_MPS", "False"))
+OFFLOAD_CPU = _cast_bool_env_var(os.environ.get("SUNO_OFFLOAD_CPU", "False"))
+
+
+REMOTE_MODEL_PATHS = {
+    "text_small": {
+        "repo_id": "suno/bark",
+        "file_name": "text.pt",
+    },
+    "coarse_small": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse.pt",
+    },
+    "fine_small": {
+        "repo_id": "suno/bark",
+        "file_name": "fine.pt",
+    },
+    "text": {
+        "repo_id": "suno/bark",
+        "file_name": "text_2.pt",
+    },
+    "coarse": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse_2.pt",
+    },
+    "fine": {
+        "repo_id": "suno/bark",
+        "file_name": "fine_2.pt",
+    },
+}
+
+
+if not hasattr(torch.nn.functional, 'scaled_dot_product_attention') and torch.cuda.is_available():
+    logger.warning(
+        "torch version does not support flash attention. You will get faster" +
+        " inference speed by upgrade torch to newest nightly version."
+    )
+
+
+def _grab_best_device(use_gpu=True):
+    if torch.cuda.device_count() > 0 and use_gpu:
+        device = "cuda"
+    elif torch.backends.mps.is_available() and use_gpu and GLOBAL_ENABLE_MPS:
+        device = "mps"
+    else:
+        device = "cpu"
+    return device
+
+
+def _get_ckpt_path(model_type, use_small=False):
+    key = model_type
+    if use_small or USE_SMALL_MODELS:
+        key += "_small"
+    return os.path.join(CACHE_DIR, REMOTE_MODEL_PATHS[key]["file_name"])
+
+
+def _download(from_hf_path, file_name):
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR)
+
+
+class InferenceContext:
+    def __init__(self, benchmark=False):
+        # we can't expect inputs to be the same length, so disable benchmarking by default
+        self._chosen_cudnn_benchmark = benchmark
+        self._cudnn_benchmark = None
+
+    def __enter__(self):
+        self._cudnn_benchmark = torch.backends.cudnn.benchmark
+        torch.backends.cudnn.benchmark = self._chosen_cudnn_benchmark
+
+    def __exit__(self, exc_type, exc_value, exc_traceback):
+        torch.backends.cudnn.benchmark = self._cudnn_benchmark
+
+
+if torch.cuda.is_available():
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+
+@contextlib.contextmanager
+def _inference_mode():
+    with InferenceContext(), torch.inference_mode(), torch.no_grad(), autocast():
+        yield
+
+
+def _clear_cuda_cache():
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+
+def clean_models(model_key=None):
+    global models
+    model_keys = [model_key] if model_key is not None else list(models.keys())
+    for k in model_keys:
+        if k in models:
+            del models[k]
+    _clear_cuda_cache()
+    gc.collect()
+
+
+def _load_model(ckpt_path, device, use_small=False, model_type="text"):
+    if model_type == "text":
+        ConfigClass = GPTConfig
+        ModelClass = GPT
+    elif model_type == "coarse":
+        ConfigClass = GPTConfig
+        ModelClass = GPT
+    elif model_type == "fine":
+        ConfigClass = FineGPTConfig
+        ModelClass = FineGPT
+    else:
+        raise NotImplementedError()
+    model_key = f"{model_type}_small" if use_small or USE_SMALL_MODELS else model_type
+    model_info = REMOTE_MODEL_PATHS[model_key]
+    if not os.path.exists(ckpt_path):
+        logger.info(f"{model_type} model not found, downloading into `{CACHE_DIR}`.")
+        _download(model_info["repo_id"], model_info["file_name"])
+    checkpoint = torch.load(ckpt_path, map_location=device)
+    # this is a hack
+    model_args = checkpoint["model_args"]
+    if "input_vocab_size" not in model_args:
+        model_args["input_vocab_size"] = model_args["vocab_size"]
+        model_args["output_vocab_size"] = model_args["vocab_size"]
+        del model_args["vocab_size"]
+    gptconf = ConfigClass(**checkpoint["model_args"])
+    model = ModelClass(gptconf)
+    state_dict = checkpoint["model"]
+    # fixup checkpoint
+    unwanted_prefix = "_orig_mod."
+    for k, v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            state_dict[k[len(unwanted_prefix) :]] = state_dict.pop(k)
+    extra_keys = set(state_dict.keys()) - set(model.state_dict().keys())
+    extra_keys = set([k for k in extra_keys if not k.endswith(".attn.bias")])
+    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
+    missing_keys = set([k for k in missing_keys if not k.endswith(".attn.bias")])
+    if len(extra_keys) != 0:
+        raise ValueError(f"extra keys found: {extra_keys}")
+    if len(missing_keys) != 0:
+        raise ValueError(f"missing keys: {missing_keys}")
+    model.load_state_dict(state_dict, strict=False)
+    n_params = model.get_num_params()
+    val_loss = checkpoint["best_val_loss"].item()
+    logger.info(f"model loaded: {round(n_params/1e6,1)}M params, {round(val_loss,3)} loss")
+    model.eval()
+    model.to(device)
+    del checkpoint, state_dict
+    _clear_cuda_cache()
+    if model_type == "text":
+        tokenizer = BertTokenizer.from_pretrained("bert-base-multilingual-cased")
+        return {
+            "model": model,
+            "tokenizer": tokenizer,
+        }
+    return model
+
+
+def _load_codec_model(device):
+    model = EncodecModel.encodec_model_24khz()
+    model.set_target_bandwidth(6.0)
+    model.eval()
+    model.to(device)
+    _clear_cuda_cache()
+    return model
+
+
+def load_model(use_gpu=True, use_small=False, force_reload=False, model_type="text"):
+    _load_model_f = funcy.partial(_load_model, model_type=model_type, use_small=use_small)
+    if model_type not in ("text", "coarse", "fine"):
+        raise NotImplementedError()
+    global models
+    global models_devices
+    device = _grab_best_device(use_gpu=use_gpu)
+    model_key = f"{model_type}"
+    if OFFLOAD_CPU:
+        models_devices[model_key] = device
+        device = "cpu"
+    if model_key not in models or force_reload:
+        ckpt_path = _get_ckpt_path(model_type, use_small=use_small)
+        clean_models(model_key=model_key)
+        model = _load_model_f(ckpt_path, device)
+        models[model_key] = model
+    if model_type == "text":
+        models[model_key]["model"].to(device)
+    else:
+        models[model_key].to(device)
+    return models[model_key]
+
+
+def load_codec_model(use_gpu=True, force_reload=False):
+    global models
+    global models_devices
+    device = _grab_best_device(use_gpu=use_gpu)
+    if device == "mps":
+        # encodec doesn't support mps
+        device = "cpu"
+    model_key = "codec"
+    if OFFLOAD_CPU:
+        models_devices[model_key] = device
+        device = "cpu"
+    if model_key not in models or force_reload:
+        clean_models(model_key=model_key)
+        model = _load_codec_model(device)
+        models[model_key] = model
+    models[model_key].to(device)
+    return models[model_key]
+
+
+def preload_models(
+    text_use_gpu=True,
+    text_use_small=False,
+    coarse_use_gpu=True,
+    coarse_use_small=False,
+    fine_use_gpu=True,
+    fine_use_small=False,
+    codec_use_gpu=True,
+    force_reload=False,
+):
+    """Load all the necessary models for the pipeline."""
+    if _grab_best_device() == "cpu" and (
+        text_use_gpu or coarse_use_gpu or fine_use_gpu or codec_use_gpu
+    ):
+        logger.warning("No GPU being used. Careful, inference might be very slow!")
+    _ = load_model(
+        model_type="text", use_gpu=text_use_gpu, use_small=text_use_small, force_reload=force_reload
+    )
+    _ = load_model(
+        model_type="coarse",
+        use_gpu=coarse_use_gpu,
+        use_small=coarse_use_small,
+        force_reload=force_reload,
+    )
+    _ = load_model(
+        model_type="fine", use_gpu=fine_use_gpu, use_small=fine_use_small, force_reload=force_reload
+    )
+    _ = load_codec_model(use_gpu=codec_use_gpu, force_reload=force_reload)
+
+
+####
+# Generation Functionality
+####
+
+
+def _tokenize(tokenizer, text):
+    return tokenizer.encode(text, add_special_tokens=False)
+
+
+def _detokenize(tokenizer, enc_text):
+    return tokenizer.decode(enc_text)
+
+
+def _normalize_whitespace(text):
+    return re.sub(r"\s+", " ", text).strip()
+
+
+TEXT_ENCODING_OFFSET = 10_048
+SEMANTIC_PAD_TOKEN = 10_000
+TEXT_PAD_TOKEN = 129_595
+SEMANTIC_INFER_TOKEN = 129_599
+
+
+def _load_history_prompt(history_prompt_input):
+    if isinstance(history_prompt_input, str) and history_prompt_input.endswith(".npz"):
+        history_prompt = np.load(history_prompt_input)
+    elif isinstance(history_prompt_input, str):
+        # make sure this works on non-ubuntu
+        history_prompt_input = os.path.join(*history_prompt_input.split("/"))
+        if history_prompt_input not in ALLOWED_PROMPTS:
+            raise ValueError("history prompt not found")
+        history_prompt = np.load(
+            os.path.join(CUR_PATH, "assets", "prompts", f"{history_prompt_input}.npz")
+        )
+    elif isinstance(history_prompt_input, dict):
+        assert("semantic_prompt" in history_prompt_input)
+        assert("coarse_prompt" in history_prompt_input)
+        assert("fine_prompt" in history_prompt_input)
+        history_prompt = history_prompt_input
+    else:
+        raise ValueError("history prompt format unrecognized")
+    return history_prompt
+
+
+def generate_text_semantic(
+    text,
+    history_prompt=None,
+    temp=0.7,
+    top_k=None,
+    top_p=None,
+    silent=False,
+    min_eos_p=0.2,
+    max_gen_duration_s=None,
+    allow_early_stop=True,
+    use_kv_caching=False,
+):
+    """Generate semantic tokens from text."""
+    assert isinstance(text, str)
+    text = _normalize_whitespace(text)
+    assert len(text.strip()) > 0
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        semantic_history = history_prompt["semantic_prompt"]
+        assert (
+            isinstance(semantic_history, np.ndarray)
+            and len(semantic_history.shape) == 1
+            and len(semantic_history) > 0
+            and semantic_history.min() >= 0
+            and semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
+        )
+    else:
+        semantic_history = None
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "text" not in models:
+        preload_models()
+    model_container = models["text"]
+    model = model_container["model"]
+    tokenizer = model_container["tokenizer"]
+    encoded_text = np.array(_tokenize(tokenizer, text)) + TEXT_ENCODING_OFFSET
+    if OFFLOAD_CPU:
+        model.to(models_devices["text"])
+    device = next(model.parameters()).device
+    if len(encoded_text) > 256:
+        p = round((len(encoded_text) - 256) / len(encoded_text) * 100, 1)
+        logger.warning(f"warning, text too long, lopping of last {p}%")
+        encoded_text = encoded_text[:256]
+    encoded_text = np.pad(
+        encoded_text,
+        (0, 256 - len(encoded_text)),
+        constant_values=TEXT_PAD_TOKEN,
+        mode="constant",
+    )
+    if semantic_history is not None:
+        semantic_history = semantic_history.astype(np.int64)
+        # lop off if history is too long, pad if needed
+        semantic_history = semantic_history[-256:]
+        semantic_history = np.pad(
+            semantic_history,
+            (0, 256 - len(semantic_history)),
+            constant_values=SEMANTIC_PAD_TOKEN,
+            mode="constant",
+        )
+    else:
+        semantic_history = np.array([SEMANTIC_PAD_TOKEN] * 256)
+    x = torch.from_numpy(
+        np.hstack([
+            encoded_text, semantic_history, np.array([SEMANTIC_INFER_TOKEN])
+        ]).astype(np.int64)
+    )[None]
+    assert x.shape[1] == 256 + 256 + 1
+    with _inference_mode():
+        x = x.to(device)
+        n_tot_steps = 768
+        # custom tqdm updates since we don't know when eos will occur
+        pbar = tqdm.tqdm(disable=silent, total=n_tot_steps)
+        pbar_state = 0
+        tot_generated_duration_s = 0
+        kv_cache = None
+        for n in range(n_tot_steps):
+            if use_kv_caching and kv_cache is not None:
+                x_input = x[:, [-1]]
+            else:
+                x_input = x
+            logits, kv_cache = model(
+                x_input, merge_context=True, use_cache=use_kv_caching, past_kv=kv_cache
+            )
+            relevant_logits = logits[0, 0, :SEMANTIC_VOCAB_SIZE]
+            if allow_early_stop:
+                relevant_logits = torch.hstack(
+                    (relevant_logits, logits[0, 0, [SEMANTIC_PAD_TOKEN]])  # eos
+                )
+            if top_p is not None:
+                # faster to convert to numpy
+                original_device = relevant_logits.device
+                relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
+                sorted_indices = np.argsort(relevant_logits)[::-1]
+                sorted_logits = relevant_logits[sorted_indices]
+                cumulative_probs = np.cumsum(softmax(sorted_logits))
+                sorted_indices_to_remove = cumulative_probs > top_p
+                sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                sorted_indices_to_remove[0] = False
+                relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
+                relevant_logits = torch.from_numpy(relevant_logits)
+                relevant_logits = relevant_logits.to(original_device)
+            if top_k is not None:
+                v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
+                relevant_logits[relevant_logits < v[-1]] = -float("Inf")
+            probs = F.softmax(relevant_logits / temp, dim=-1)
+            item_next = torch.multinomial(probs, num_samples=1).to(torch.int32)
+            if allow_early_stop and (
+                item_next == SEMANTIC_VOCAB_SIZE
+                or (min_eos_p is not None and probs[-1] >= min_eos_p)
+            ):
+                # eos found, so break
+                pbar.update(n - pbar_state)
+                break
+            x = torch.cat((x, item_next[None]), dim=1)
+            tot_generated_duration_s += 1 / SEMANTIC_RATE_HZ
+            if max_gen_duration_s is not None and tot_generated_duration_s > max_gen_duration_s:
+                pbar.update(n - pbar_state)
+                break
+            if n == n_tot_steps - 1:
+                pbar.update(n - pbar_state)
+                break
+            del logits, relevant_logits, probs, item_next
+
+            if n > pbar_state:
+                if n > pbar.total:
+                    pbar.total = n
+                pbar.update(n - pbar_state)
+            pbar_state = n
+        pbar.total = n
+        pbar.refresh()
+        pbar.close()
+        out = x.detach().cpu().numpy().squeeze()[256 + 256 + 1 :]
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    assert all(0 <= out) and all(out < SEMANTIC_VOCAB_SIZE)
+    _clear_cuda_cache()
+    return out
+
+
+def _flatten_codebooks(arr, offset_size=CODEBOOK_SIZE):
+    assert len(arr.shape) == 2
+    arr = arr.copy()
+    if offset_size is not None:
+        for n in range(1, arr.shape[0]):
+            arr[n, :] += offset_size * n
+    flat_arr = arr.ravel("F")
+    return flat_arr
+
+
+COARSE_SEMANTIC_PAD_TOKEN = 12_048
+COARSE_INFER_TOKEN = 12_050
+
+
+def generate_coarse(
+    x_semantic,
+    history_prompt=None,
+    temp=0.7,
+    top_k=None,
+    top_p=None,
+    silent=False,
+    max_coarse_history=630,  # min 60 (faster), max 630 (more context)
+    sliding_window_len=60,
+    use_kv_caching=False,
+):
+    """Generate coarse audio codes from semantic tokens."""
+    assert (
+        isinstance(x_semantic, np.ndarray)
+        and len(x_semantic.shape) == 1
+        and len(x_semantic) > 0
+        and x_semantic.min() >= 0
+        and x_semantic.max() <= SEMANTIC_VOCAB_SIZE - 1
+    )
+    assert 60 <= max_coarse_history <= 630
+    assert max_coarse_history + sliding_window_len <= 1024 - 256
+    semantic_to_coarse_ratio = COARSE_RATE_HZ / SEMANTIC_RATE_HZ * N_COARSE_CODEBOOKS
+    max_semantic_history = int(np.floor(max_coarse_history / semantic_to_coarse_ratio))
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        x_semantic_history = history_prompt["semantic_prompt"]
+        x_coarse_history = history_prompt["coarse_prompt"]
+        assert (
+            isinstance(x_semantic_history, np.ndarray)
+            and len(x_semantic_history.shape) == 1
+            and len(x_semantic_history) > 0
+            and x_semantic_history.min() >= 0
+            and x_semantic_history.max() <= SEMANTIC_VOCAB_SIZE - 1
+            and isinstance(x_coarse_history, np.ndarray)
+            and len(x_coarse_history.shape) == 2
+            and x_coarse_history.shape[0] == N_COARSE_CODEBOOKS
+            and x_coarse_history.shape[-1] >= 0
+            and x_coarse_history.min() >= 0
+            and x_coarse_history.max() <= CODEBOOK_SIZE - 1
+            and (
+                round(x_coarse_history.shape[-1] / len(x_semantic_history), 1)
+                == round(semantic_to_coarse_ratio / N_COARSE_CODEBOOKS, 1)
+            )
+        )
+        x_coarse_history = _flatten_codebooks(x_coarse_history) + SEMANTIC_VOCAB_SIZE
+        # trim histories correctly
+        n_semantic_hist_provided = np.min(
+            [
+                max_semantic_history,
+                len(x_semantic_history) - len(x_semantic_history) % 2,
+                int(np.floor(len(x_coarse_history) / semantic_to_coarse_ratio)),
+            ]
+        )
+        n_coarse_hist_provided = int(round(n_semantic_hist_provided * semantic_to_coarse_ratio))
+        x_semantic_history = x_semantic_history[-n_semantic_hist_provided:].astype(np.int32)
+        x_coarse_history = x_coarse_history[-n_coarse_hist_provided:].astype(np.int32)
+        # TODO: bit of a hack for time alignment (sounds better)
+        x_coarse_history = x_coarse_history[:-2]
+    else:
+        x_semantic_history = np.array([], dtype=np.int32)
+        x_coarse_history = np.array([], dtype=np.int32)
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "coarse" not in models:
+        preload_models()
+    model = models["coarse"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["coarse"])
+    device = next(model.parameters()).device
+    # start loop
+    n_steps = int(
+        round(
+            np.floor(len(x_semantic) * semantic_to_coarse_ratio / N_COARSE_CODEBOOKS)
+            * N_COARSE_CODEBOOKS
+        )
+    )
+    assert n_steps > 0 and n_steps % N_COARSE_CODEBOOKS == 0
+    x_semantic = np.hstack([x_semantic_history, x_semantic]).astype(np.int32)
+    x_coarse = x_coarse_history.astype(np.int32)
+    base_semantic_idx = len(x_semantic_history)
+    with _inference_mode():
+        x_semantic_in = torch.from_numpy(x_semantic)[None].to(device)
+        x_coarse_in = torch.from_numpy(x_coarse)[None].to(device)
+        n_window_steps = int(np.ceil(n_steps / sliding_window_len))
+        n_step = 0
+        for _ in tqdm.tqdm(range(n_window_steps), total=n_window_steps, disable=silent):
+            semantic_idx = base_semantic_idx + int(round(n_step / semantic_to_coarse_ratio))
+            # pad from right side
+            x_in = x_semantic_in[:, np.max([0, semantic_idx - max_semantic_history]) :]
+            x_in = x_in[:, :256]
+            x_in = F.pad(
+                x_in,
+                (0, 256 - x_in.shape[-1]),
+                "constant",
+                COARSE_SEMANTIC_PAD_TOKEN,
+            )
+            x_in = torch.hstack(
+                [
+                    x_in,
+                    torch.tensor([COARSE_INFER_TOKEN])[None].to(device),
+                    x_coarse_in[:, -max_coarse_history:],
+                ]
+            )
+            kv_cache = None
+            for _ in range(sliding_window_len):
+                if n_step >= n_steps:
+                    continue
+                is_major_step = n_step % N_COARSE_CODEBOOKS == 0
+
+                if use_kv_caching and kv_cache is not None:
+                    x_input = x_in[:, [-1]]
+                else:
+                    x_input = x_in
+
+                logits, kv_cache = model(x_input, use_cache=use_kv_caching, past_kv=kv_cache)
+                logit_start_idx = (
+                    SEMANTIC_VOCAB_SIZE + (1 - int(is_major_step)) * CODEBOOK_SIZE
+                )
+                logit_end_idx = (
+                    SEMANTIC_VOCAB_SIZE + (2 - int(is_major_step)) * CODEBOOK_SIZE
+                )
+                relevant_logits = logits[0, 0, logit_start_idx:logit_end_idx]
+                if top_p is not None:
+                    # faster to convert to numpy
+                    original_device = relevant_logits.device
+                    relevant_logits = relevant_logits.detach().cpu().type(torch.float32).numpy()
+                    sorted_indices = np.argsort(relevant_logits)[::-1]
+                    sorted_logits = relevant_logits[sorted_indices]
+                    cumulative_probs = np.cumsum(softmax(sorted_logits))
+                    sorted_indices_to_remove = cumulative_probs > top_p
+                    sorted_indices_to_remove[1:] = sorted_indices_to_remove[:-1].copy()
+                    sorted_indices_to_remove[0] = False
+                    relevant_logits[sorted_indices[sorted_indices_to_remove]] = -np.inf
+                    relevant_logits = torch.from_numpy(relevant_logits)
+                    relevant_logits = relevant_logits.to(original_device)
+                if top_k is not None:
+                    v, _ = torch.topk(relevant_logits, min(top_k, relevant_logits.size(-1)))
+                    relevant_logits[relevant_logits < v[-1]] = -float("Inf")
+                probs = F.softmax(relevant_logits / temp, dim=-1)
+                item_next = torch.multinomial(probs, num_samples=1).to(torch.int32)
+                item_next += logit_start_idx
+                x_coarse_in = torch.cat((x_coarse_in, item_next[None]), dim=1)
+                x_in = torch.cat((x_in, item_next[None]), dim=1)
+                del logits, relevant_logits, probs, item_next
+                n_step += 1
+            del x_in
+        del x_semantic_in
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    gen_coarse_arr = x_coarse_in.detach().cpu().numpy().squeeze()[len(x_coarse_history) :]
+    del x_coarse_in
+    assert len(gen_coarse_arr) == n_steps
+    gen_coarse_audio_arr = gen_coarse_arr.reshape(-1, N_COARSE_CODEBOOKS).T - SEMANTIC_VOCAB_SIZE
+    for n in range(1, N_COARSE_CODEBOOKS):
+        gen_coarse_audio_arr[n, :] -= n * CODEBOOK_SIZE
+    _clear_cuda_cache()
+    return gen_coarse_audio_arr
+
+
+def generate_fine(
+    x_coarse_gen,
+    history_prompt=None,
+    temp=0.5,
+    silent=True,
+):
+    """Generate full audio codes from coarse audio codes."""
+    assert (
+        isinstance(x_coarse_gen, np.ndarray)
+        and len(x_coarse_gen.shape) == 2
+        and 1 <= x_coarse_gen.shape[0] <= N_FINE_CODEBOOKS - 1
+        and x_coarse_gen.shape[1] > 0
+        and x_coarse_gen.min() >= 0
+        and x_coarse_gen.max() <= CODEBOOK_SIZE - 1
+    )
+    if history_prompt is not None:
+        history_prompt = _load_history_prompt(history_prompt)
+        x_fine_history = history_prompt["fine_prompt"]
+        assert (
+            isinstance(x_fine_history, np.ndarray)
+            and len(x_fine_history.shape) == 2
+            and x_fine_history.shape[0] == N_FINE_CODEBOOKS
+            and x_fine_history.shape[1] >= 0
+            and x_fine_history.min() >= 0
+            and x_fine_history.max() <= CODEBOOK_SIZE - 1
+        )
+    else:
+        x_fine_history = None
+    n_coarse = x_coarse_gen.shape[0]
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "fine" not in models:
+        preload_models()
+    model = models["fine"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["fine"])
+    device = next(model.parameters()).device
+    # make input arr
+    in_arr = np.vstack(
+        [
+            x_coarse_gen,
+            np.zeros((N_FINE_CODEBOOKS - n_coarse, x_coarse_gen.shape[1]))
+            + CODEBOOK_SIZE,  # padding
+        ]
+    ).astype(np.int32)
+    # prepend history if available (max 512)
+    if x_fine_history is not None:
+        x_fine_history = x_fine_history.astype(np.int32)
+        in_arr = np.hstack(
+            [
+                x_fine_history[:, -512:].astype(np.int32),
+                in_arr,
+            ]
+        )
+        n_history = x_fine_history[:, -512:].shape[1]
+    else:
+        n_history = 0
+    n_remove_from_end = 0
+    # need to pad if too short (since non-causal model)
+    if in_arr.shape[1] < 1024:
+        n_remove_from_end = 1024 - in_arr.shape[1]
+        in_arr = np.hstack(
+            [
+                in_arr,
+                np.zeros((N_FINE_CODEBOOKS, n_remove_from_end), dtype=np.int32) + CODEBOOK_SIZE,
+            ]
+        )
+    # we can be lazy about fractional loop and just keep overwriting codebooks
+    n_loops = np.max([0, int(np.ceil((x_coarse_gen.shape[1] - (1024 - n_history)) / 512))]) + 1
+    with _inference_mode():
+        in_arr = torch.tensor(in_arr.T).to(device)
+        for n in tqdm.tqdm(range(n_loops), disable=silent):
+            start_idx = np.min([n * 512, in_arr.shape[0] - 1024])
+            start_fill_idx = np.min([n_history + n * 512, in_arr.shape[0] - 512])
+            rel_start_fill_idx = start_fill_idx - start_idx
+            in_buffer = in_arr[start_idx : start_idx + 1024, :][None]
+            for nn in range(n_coarse, N_FINE_CODEBOOKS):
+                logits = model(nn, in_buffer)
+                if temp is None:
+                    relevant_logits = logits[0, rel_start_fill_idx:, :CODEBOOK_SIZE]
+                    codebook_preds = torch.argmax(relevant_logits, -1)
+                else:
+                    relevant_logits = logits[0, :, :CODEBOOK_SIZE] / temp
+                    probs = F.softmax(relevant_logits, dim=-1)
+                    codebook_preds = torch.multinomial(
+                        probs[rel_start_fill_idx:1024], num_samples=1
+                    ).reshape(-1)
+                codebook_preds = codebook_preds.to(torch.int32)
+                in_buffer[0, rel_start_fill_idx:, nn] = codebook_preds
+                del logits, codebook_preds
+            # transfer over info into model_in and convert to numpy
+            for nn in range(n_coarse, N_FINE_CODEBOOKS):
+                in_arr[
+                    start_fill_idx : start_fill_idx + (1024 - rel_start_fill_idx), nn
+                ] = in_buffer[0, rel_start_fill_idx:, nn]
+            del in_buffer
+        gen_fine_arr = in_arr.detach().cpu().numpy().squeeze().T
+        del in_arr
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    gen_fine_arr = gen_fine_arr[:, n_history:]
+    if n_remove_from_end > 0:
+        gen_fine_arr = gen_fine_arr[:, :-n_remove_from_end]
+    assert gen_fine_arr.shape[-1] == x_coarse_gen.shape[-1]
+    _clear_cuda_cache()
+    return gen_fine_arr
+
+
+def codec_decode(fine_tokens):
+    """Turn quantized audio codes into audio array using encodec."""
+    # load models if not yet exist
+    global models
+    global models_devices
+    if "codec" not in models:
+        preload_models()
+    model = models["codec"]
+    if OFFLOAD_CPU:
+        model.to(models_devices["codec"])
+    device = next(model.parameters()).device
+    arr = torch.from_numpy(fine_tokens)[None]
+    arr = arr.to(device)
+    arr = arr.transpose(0, 1)
+    emb = model.quantizer.decode(arr)
+    out = model.decoder(emb)
+    audio_arr = out.detach().cpu().numpy().squeeze()
+    del arr, emb, out
+    if OFFLOAD_CPU:
+        model.to("cpu")
+    return audio_arr

model.py

@@ -0,0 +1,218 @@
+"""
+Much of this code is adapted from Andrej Karpathy's NanoGPT
+(https://github.com/karpathy/nanoGPT)
+"""
+import math
+from dataclasses import dataclass
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+class LayerNorm(nn.Module):
+    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """
+
+    def __init__(self, ndim, bias):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(ndim))
+        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
+
+    def forward(self, input):
+        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch nightly and still a bit scary
+        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
+        if not self.flash:
+            # print("WARNING: using slow attention. Flash Attention atm needs PyTorch nightly and dropout=0.0")
+            # causal mask to ensure that attention is only applied to the left in the input sequence
+            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
+                                        .view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x, past_kv=None, use_cache=False):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k ,v  = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        if past_kv is not None:
+            past_key = past_kv[0]
+            past_value = past_kv[1]
+            k = torch.cat((past_key, k), dim=-2)
+            v = torch.cat((past_value, v), dim=-2)
+
+        FULL_T = k.shape[-2]
+
+        if use_cache is True:
+            present = (k, v)
+        else:
+            present = None
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            if past_kv is not None:
+                # When `past_kv` is provided, we're doing incremental decoding and `q.shape[2] == 1`: q only contains
+                # the query for the last token. scaled_dot_product_attention interprets this as the first token in the
+                # sequence, so if is_causal=True it will mask out all attention from it. This is not what we want, so 
+                # to work around this we set is_causal=False.
+                is_causal = False
+            else:
+                is_causal = True
+
+            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout, is_causal=is_causal)
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = att.masked_fill(self.bias[:,:,FULL_T-T:FULL_T,:FULL_T] == 0, float('-inf'))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return (y, present)
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
+        self.dropout = nn.Dropout(config.dropout)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        x = self.dropout(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
+        self.mlp = MLP(config)
+        self.layer_idx = layer_idx
+
+    def forward(self, x, past_kv=None, use_cache=False):
+        attn_output, prev_kvs = self.attn(self.ln_1(x), past_kv=past_kv, use_cache=use_cache)
+        x = x + attn_output
+        x = x + self.mlp(self.ln_2(x))
+        return (x, prev_kvs)
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024
+    input_vocab_size: int = 10_048
+    output_vocab_size: int = 10_048
+    n_layer: int = 12
+    n_head: int = 12
+    n_embd: int = 768
+    dropout: float = 0.0
+    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.input_vocab_size is not None
+        assert config.output_vocab_size is not None
+        assert config.block_size is not None
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.input_vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            drop = nn.Dropout(config.dropout),
+            h = nn.ModuleList([Block(config, idx) for idx in range(config.n_layer)]),
+            ln_f = LayerNorm(config.n_embd, bias=config.bias),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            n_params -= self.transformer.wte.weight.numel()
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+    def forward(self, idx, merge_context=False, past_kv=None, position_ids=None, use_cache=False):
+        device = idx.device
+        b, t = idx.size()
+        if past_kv is not None:
+            assert t == 1
+            tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+        else:
+            if merge_context:
+                assert(idx.shape[1] >= 256+256+1)
+                t = idx.shape[1] - 256
+            else:
+                assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+
+            # forward the GPT model itself
+            if merge_context:
+                tok_emb = torch.cat([
+                    self.transformer.wte(idx[:,:256]) + self.transformer.wte(idx[:,256:256+256]),
+                    self.transformer.wte(idx[:,256+256:])
+                ], dim=1)
+            else:
+                tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
+
+        if past_kv is None:
+            past_length = 0
+            past_kv = tuple([None] * len(self.transformer.h))
+        else:
+            past_length = past_kv[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(past_length, t + past_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0) # shape (1, t)
+            assert position_ids.shape == (1, t)
+
+        pos_emb = self.transformer.wpe(position_ids) # position embeddings of shape (1, t, n_embd)
+
+        x = self.transformer.drop(tok_emb + pos_emb)
+
+        new_kv = () if use_cache else None
+
+        for i, (block, past_layer_kv) in enumerate(zip(self.transformer.h, past_kv)):
+            x, kv = block(x, past_kv=past_layer_kv, use_cache=use_cache)
+
+            if use_cache:
+                new_kv = new_kv + (kv,)
+
+        x = self.transformer.ln_f(x)
+
+        # inference-time mini-optimization: only forward the lm_head on the very last position
+        logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
+
+        return (logits, new_kv)

model_fine.py

@@ -0,0 +1,149 @@
+"""
+Much of this code is adapted from Andrej Karpathy's NanoGPT
+(https://github.com/karpathy/nanoGPT)
+"""
+from dataclasses import dataclass
+import math
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from .model import GPT, GPTConfig, MLP
+
+
+class NonCausalSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0
+        self.flash = (
+            hasattr(torch.nn.functional, "scaled_dot_product_attention")
+        )
+
+    def forward(self, x):
+        B, T, C = x.size()  # batch size, sequence length, embedding dimensionality (n_embd)
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        if self.flash:
+            # efficient attention using Flash Attention CUDA kernels
+            y = torch.nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=None, dropout_p=self.dropout, is_causal=False
+            )
+        else:
+            # manual implementation of attention
+            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+            att = F.softmax(att, dim=-1)
+            att = self.attn_dropout(att)
+            y = att @ v  # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = (
+            y.transpose(1, 2).contiguous().view(B, T, C)
+        )  # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_dropout(self.c_proj(y))
+        return y
+
+
+class FineBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = NonCausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class FineGPT(GPT):
+    def __init__(self, config):
+        super().__init__(config)
+        del self.lm_head
+        self.config = config
+        self.n_codes_total = config.n_codes_total
+        self.transformer = nn.ModuleDict(
+            dict(
+                wtes=nn.ModuleList(
+                    [
+                        nn.Embedding(config.input_vocab_size, config.n_embd)
+                        for _ in range(config.n_codes_total)
+                    ]
+                ),
+                wpe=nn.Embedding(config.block_size, config.n_embd),
+                drop=nn.Dropout(config.dropout),
+                h=nn.ModuleList([FineBlock(config) for _ in range(config.n_layer)]),
+                ln_f=nn.LayerNorm(config.n_embd),
+            )
+        )
+        self.lm_heads = nn.ModuleList(
+            [
+                nn.Linear(config.n_embd, config.output_vocab_size, bias=False)
+                for _ in range(config.n_codes_given, self.n_codes_total)
+            ]
+        )
+        for i in range(self.n_codes_total - config.n_codes_given):
+            self.transformer.wtes[i + 1].weight = self.lm_heads[i].weight
+
+    def forward(self, pred_idx, idx):
+        device = idx.device
+        b, t, codes = idx.size()
+        assert (
+            t <= self.config.block_size
+        ), f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
+        assert pred_idx > 0, "cannot predict 0th codebook"
+        assert codes == self.n_codes_total, (b, t, codes)
+        pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0)  # shape (1, t)
+
+        # forward the GPT model itself
+        tok_embs = [
+            wte(idx[:, :, i]).unsqueeze(-1) for i, wte in enumerate(self.transformer.wtes)
+        ]  # token embeddings of shape (b, t, n_embd)
+        tok_emb = torch.cat(tok_embs, dim=-1)
+        pos_emb = self.transformer.wpe(pos)  # position embeddings of shape (1, t, n_embd)
+        x = tok_emb[:, :, :, : pred_idx + 1].sum(dim=-1)
+        x = self.transformer.drop(x + pos_emb)
+        for block in self.transformer.h:
+            x = block(x)
+        x = self.transformer.ln_f(x)
+        logits = self.lm_heads[pred_idx - self.config.n_codes_given](x)
+        return logits
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            for wte in self.transformer.wtes:
+                n_params -= wte.weight.numel()
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+
+@dataclass
+class FineGPTConfig(GPTConfig):
+    n_codes_total: int = 8
+    n_codes_given: int = 1