NEW-0216-050145

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text

AR/data/bucket_sampler.py

@@ -0,0 +1,162 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/bucketsampler.py
+import itertools
+import math
+import random
+from random import shuffle
+from typing import Iterator
+from typing import Optional
+from typing import TypeVar
+
+import torch
+import torch.distributed as dist
+from torch.utils.data import Dataset
+from torch.utils.data import Sampler
+
+__all__ = [
+    "DistributedBucketSampler",
+]
+
+T_co = TypeVar("T_co", covariant=True)
+
+
+class DistributedBucketSampler(Sampler[T_co]):
+    r"""
+    sort the dataset wrt. input length
+    divide samples into buckets
+    sort within buckets
+    divide buckets into batches
+    sort batches
+    """
+
+    def __init__(
+        self,
+        dataset: Dataset,
+        num_replicas: Optional[int] = None,
+        rank: Optional[int] = None,
+        shuffle: bool = True,
+        seed: int = 0,
+        drop_last: bool = False,
+        batch_size: int = 32,
+    ) -> None:
+        if num_replicas is None:
+            if not dist.is_available():
+                raise RuntimeError("Requires distributed package to be available")
+            num_replicas = dist.get_world_size() if torch.cuda.is_available() else 1
+        if rank is None:
+            if not dist.is_available():
+                raise RuntimeError("Requires distributed package to be available")
+            rank = dist.get_rank() if torch.cuda.is_available() else 0
+            if torch.cuda.is_available():
+                torch.cuda.set_device(rank)
+        if rank >= num_replicas or rank < 0:
+            raise ValueError(
+                "Invalid rank {}, rank should be in the interval"
+                " [0, {}]".format(rank, num_replicas - 1)
+            )
+        self.dataset = dataset
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.epoch = 0
+        self.drop_last = drop_last
+        # If the dataset length is evenly divisible by # of replicas, then there
+        # is no need to drop any data, since the dataset will be split equally.
+        if (
+            self.drop_last and len(self.dataset) % self.num_replicas != 0
+        ):  # type: ignore[arg-type]
+            # Split to nearest available length that is evenly divisible.
+            # This is to ensure each rank receives the same amount of data when
+            # using this Sampler.
+            self.num_samples = math.ceil(
+                (len(self.dataset) - self.num_replicas)
+                / self.num_replicas  # type: ignore[arg-type]
+            )
+        else:
+            self.num_samples = math.ceil(
+                len(self.dataset) / self.num_replicas
+            )  # type: ignore[arg-type]
+        self.total_size = self.num_samples * self.num_replicas
+        self.shuffle = shuffle
+        self.seed = seed
+        self.batch_size = batch_size
+        self.id_with_length = self._get_sample_lengths()
+        self.id_buckets = self.make_buckets(bucket_width=2.0)
+
+    def _get_sample_lengths(self):
+        id_with_lengths = []
+        for i in range(len(self.dataset)):
+            id_with_lengths.append((i, self.dataset.get_sample_length(i)))
+        id_with_lengths.sort(key=lambda x: x[1])
+        return id_with_lengths
+
+    def make_buckets(self, bucket_width: float = 2.0):
+        buckets = []
+        cur = []
+        max_sec = bucket_width
+        for id, sec in self.id_with_length:
+            if sec < max_sec:
+                cur.append(id)
+            else:
+                buckets.append(cur)
+                cur = [id]
+                max_sec += bucket_width
+        if len(cur) > 0:
+            buckets.append(cur)
+        return buckets
+
+    def __iter__(self) -> Iterator[T_co]:
+        if self.shuffle:
+            # deterministically shuffle based on epoch and seed
+            g = torch.Generator()
+            g.manual_seed(self.seed + self.epoch)
+            random.seed(self.epoch + self.seed)
+            shuffled_bucket = []
+            for buc in self.id_buckets:
+                buc_copy = buc.copy()
+                shuffle(buc_copy)
+                shuffled_bucket.append(buc_copy)
+            grouped_batch_size = self.batch_size * self.num_replicas
+            shuffled_bucket = list(itertools.chain(*shuffled_bucket))
+            n_batch = int(math.ceil(len(shuffled_bucket) / grouped_batch_size))
+            batches = [
+                shuffled_bucket[b * grouped_batch_size : (b + 1) * grouped_batch_size]
+                for b in range(n_batch)
+            ]
+            shuffle(batches)
+            indices = list(itertools.chain(*batches))
+        else:
+            # type: ignore[arg-type]
+            indices = list(range(len(self.dataset)))
+
+        if not self.drop_last:
+            # add extra samples to make it evenly divisible
+            padding_size = self.total_size - len(indices)
+            if padding_size <= len(indices):
+                indices += indices[:padding_size]
+            else:
+                indices += (indices * math.ceil(padding_size / len(indices)))[
+                    :padding_size
+                ]
+        else:
+            # remove tail of data to make it evenly divisible.
+            indices = indices[: self.total_size]
+        assert len(indices) == self.total_size
+
+        # subsample
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        return iter(indices)
+
+    def __len__(self) -> int:
+        return self.num_samples
+
+    def set_epoch(self, epoch: int) -> None:
+        r"""
+        Sets the epoch for this sampler. When :attr:`shuffle=True`, this ensures all replicas
+        use a different random ordering for each epoch. Otherwise, the next iteration of this
+        sampler will yield the same ordering.
+
+        Args:
+            epoch (int): Epoch number.
+        """
+        self.epoch = epoch

AR/data/data_module.py

@@ -0,0 +1,74 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/data_module.py
+from pytorch_lightning import LightningDataModule
+from AR.data.bucket_sampler import DistributedBucketSampler
+from AR.data.dataset import Text2SemanticDataset
+from torch.utils.data import DataLoader
+
+
+class Text2SemanticDataModule(LightningDataModule):
+    def __init__(
+        self,
+        config,
+        train_semantic_path,
+        train_phoneme_path,
+        dev_semantic_path=None,
+        dev_phoneme_path=None,
+    ):
+        super().__init__()
+        self.config = config
+        self.train_semantic_path = train_semantic_path
+        self.train_phoneme_path = train_phoneme_path
+        self.dev_semantic_path = dev_semantic_path
+        self.dev_phoneme_path = dev_phoneme_path
+        self.num_workers = self.config["data"]["num_workers"]
+
+    def prepare_data(self):
+        pass
+
+    def setup(self, stage=None, output_logs=False):
+        self._train_dataset = Text2SemanticDataset(
+            phoneme_path=self.train_phoneme_path,
+            semantic_path=self.train_semantic_path,
+            max_sec=self.config["data"]["max_sec"],
+            pad_val=self.config["data"]["pad_val"],
+        )
+        self._dev_dataset = self._train_dataset
+        # self._dev_dataset = Text2SemanticDataset(
+        #     phoneme_path=self.dev_phoneme_path,
+        #     semantic_path=self.dev_semantic_path,
+        #     max_sample=self.config['data']['max_eval_sample'],
+        #     max_sec=self.config['data']['max_sec'],
+        #     pad_val=self.config['data']['pad_val'])
+
+    def train_dataloader(self):
+        batch_size = max(min(self.config["train"]["batch_size"],len(self._train_dataset)//4),1)#防止不保存
+        sampler = DistributedBucketSampler(self._train_dataset, batch_size=batch_size)
+        return DataLoader(
+            self._train_dataset,
+            batch_size=batch_size,
+            sampler=sampler,
+            collate_fn=self._train_dataset.collate,
+            num_workers=self.num_workers,
+            persistent_workers=True,
+            prefetch_factor=16,
+        )
+
+    def val_dataloader(self):
+        return DataLoader(
+            self._dev_dataset,
+            batch_size=1,
+            shuffle=False,
+            collate_fn=self._train_dataset.collate,
+            num_workers=max(self.num_workers, 12),
+            persistent_workers=True,
+            prefetch_factor=16,
+        )
+
+    # 这个会使用到嘛？
+    def test_dataloader(self):
+        return DataLoader(
+            self._dev_dataset,
+            batch_size=1,
+            shuffle=False,
+            collate_fn=self._train_dataset.collate,
+        )

AR/data/dataset.py

@@ -0,0 +1,320 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/t2s_dataset.py
+import pdb
+import sys
+
+# sys.path.append("/data/docker/liujing04/gpt-vits/mq-vits-s1bert_no_bert")
+import traceback, os
+from typing import Dict
+from typing import List
+
+import numpy as np
+import pandas as pd
+import torch, json
+from torch.utils.data import DataLoader
+from torch.utils.data import Dataset
+from transformers import AutoTokenizer
+
+from text import cleaned_text_to_sequence
+
+# from config import exp_dir
+
+
+def batch_sequences(sequences: List[np.array], axis: int = 0, pad_value: int = 0):
+    seq = sequences[0]
+    ndim = seq.ndim
+    if axis < 0:
+        axis += ndim
+    dtype = seq.dtype
+    pad_value = dtype.type(pad_value)
+    seq_lengths = [seq.shape[axis] for seq in sequences]
+    max_length = np.max(seq_lengths)
+
+    padded_sequences = []
+    for seq, length in zip(sequences, seq_lengths):
+        padding = (
+            [(0, 0)] * axis + [(0, max_length - length)] + [(0, 0)] * (ndim - axis - 1)
+        )
+        padded_seq = np.pad(seq, padding, mode="constant", constant_values=pad_value)
+        padded_sequences.append(padded_seq)
+    batch = np.stack(padded_sequences)
+    return batch
+
+
+class Text2SemanticDataset(Dataset):
+    """dataset class for text tokens to semantic model training."""
+
+    def __init__(
+        self,
+        phoneme_path: str,
+        semantic_path: str,
+        max_sample: int = None,
+        max_sec: int = 100,
+        pad_val: int = 1024,
+        # min value of phoneme/sec
+        min_ps_ratio: int = 3,
+        # max value of phoneme/sec
+        max_ps_ratio: int = 25,
+    ) -> None:
+        super().__init__()
+
+        self.semantic_data = pd.read_csv(
+            semantic_path, delimiter="\t", encoding="utf-8"
+        )
+        # get dict
+        self.path2 = phoneme_path  # "%s/2-name2text.txt"%exp_dir#phoneme_path
+        self.path3 = "%s/3-bert" % (
+            os.path.basename(phoneme_path)
+        )  # "%s/3-bert"%exp_dir#bert_dir
+        self.path6 = semantic_path  # "%s/6-name2semantic.tsv"%exp_dir#semantic_path
+        assert os.path.exists(self.path2)
+        assert os.path.exists(self.path6)
+        self.phoneme_data = {}
+        with open(self.path2, "r", encoding="utf8") as f:
+            lines = f.read().strip("\n").split("\n")
+
+        for line in lines:
+            tmp = line.split("\t")
+            if len(tmp) != 4:
+                continue
+            self.phoneme_data[tmp[0]] = [tmp[1], tmp[2], tmp[3]]
+
+        # self.phoneme_data = np.load(phoneme_path, allow_pickle=True).item()
+        # pad for semantic tokens
+        self.PAD: int = pad_val
+        # self.hz = 25
+        # with open("/data/docker/liujing04/gpt-vits/mq-vits-s1bert_no_bert/configs/s2.json", "r") as f:data = f.read()
+        # data=json.loads(data)["model"]["semantic_frame_rate"]#50hz
+        # self.hz=int(data[:-2])#
+        self.hz = int(os.environ.get("hz", "25hz")[:-2])
+
+        # max seconds of semantic token
+        self.max_sec = max_sec
+        self.min_ps_ratio = min_ps_ratio
+        self.max_ps_ratio = max_ps_ratio
+
+        if max_sample is not None:
+            self.semantic_data = self.semantic_data[:max_sample]
+
+        # {idx: (semantic, phoneme)}
+        # semantic list, phoneme list
+        self.semantic_phoneme = []
+        self.item_names = []
+
+        self.inited = False
+
+        if not self.inited:
+            # 调用初始化函数
+            self.init_batch()
+            self.inited = True
+            del self.semantic_data
+            del self.phoneme_data
+        # self.tokenizer = AutoTokenizer.from_pretrained("hfl/chinese-roberta-wwm-ext-large")
+        # self.tokenizer = AutoTokenizer.from_pretrained("/data/docker/liujing04/bert-vits2/Bert-VITS2-master20231106/bert/chinese-roberta-wwm-ext-large")
+
+    def init_batch(self):
+        semantic_data_len = len(self.semantic_data)
+        phoneme_data_len = len(self.phoneme_data.keys())
+        print("semantic_data_len:", semantic_data_len)
+        print("phoneme_data_len:", phoneme_data_len)
+        print(self.semantic_data)
+        idx = 0
+        num_not_in = 0
+        num_deleted_bigger = 0
+        num_deleted_ps = 0
+        for i in range(semantic_data_len):
+            # 先依次遍历
+            # get str
+            item_name = self.semantic_data.iloc[i,0]
+            # print(self.phoneme_data)
+            try:
+                phoneme, word2ph, text = self.phoneme_data[item_name]
+            except Exception:
+                traceback.print_exc()
+                # print(f"{item_name} not in self.phoneme_data !")
+                num_not_in += 1
+                continue
+
+            semantic_str = self.semantic_data.iloc[i,1]
+            # get token list
+            semantic_ids = [int(idx) for idx in semantic_str.split(" ")]
+            # (T), 是否需要变成 (1, T) -> 不需要，因为需要求 len
+            # 过滤掉太长的样本
+            if (
+                len(semantic_ids) > self.max_sec * self.hz
+            ):  #########1###根据token个数推测总时长过滤时长60s（config里）#40*25=1k
+                num_deleted_bigger += 1
+                continue
+            # (T, ), 这个速度不会很慢，所以可以在一开始就处理，无需在 __getitem__ 里面单个处理####
+            phoneme = phoneme.split(" ")
+
+            try:
+                phoneme_ids = cleaned_text_to_sequence(phoneme)
+            except:
+                traceback.print_exc()
+                # print(f"{item_name} not in self.phoneme_data !")
+                num_not_in += 1
+                continue
+            # if len(phoneme_ids) >400:###########2：改为恒定限制为semantic/2.5就行
+            if (
+                len(phoneme_ids) > self.max_sec * self.hz / 2.5
+            ):  ###########2：改为恒定限制为semantic/2.5就行
+                num_deleted_ps += 1
+                continue
+            # if len(semantic_ids) > 1000:###########3
+            #     num_deleted_bigger += 1
+            #     continue
+
+            ps_ratio = len(phoneme_ids) / (len(semantic_ids) / self.hz)
+
+            if (
+                ps_ratio > self.max_ps_ratio or ps_ratio < self.min_ps_ratio
+            ):  ##########4#3~25#每秒多少个phone
+                num_deleted_ps += 1
+                # print(item_name)
+                continue
+
+            self.semantic_phoneme.append((semantic_ids, phoneme_ids))
+            idx += 1
+            self.item_names.append(item_name)
+
+        min_num = 100  # 20直接不补#30补了也不存ckpt
+        leng = len(self.semantic_phoneme)
+        if leng < min_num:
+            tmp1 = self.semantic_phoneme
+            tmp2 = self.item_names
+            self.semantic_phoneme = []
+            self.item_names = []
+            for _ in range(max(2, int(min_num / leng))):
+                self.semantic_phoneme += tmp1
+                self.item_names += tmp2
+        if num_not_in > 0:
+            print(f"there are {num_not_in} semantic datas not in phoneme datas")
+        if num_deleted_bigger > 0:
+            print(
+                f"deleted {num_deleted_bigger} audios who's duration are bigger than {self.max_sec} seconds"
+            )
+        if num_deleted_ps > 0:
+            # 4702 for LibriTTS, LirbriTTS 是标注数据, 是否需要筛？=> 需要，有值为 100 的极端值
+            print(
+                f"deleted {num_deleted_ps} audios who's phoneme/sec are bigger than {self.max_ps_ratio} or smaller than {self.min_ps_ratio}"
+            )
+        """
+        there are 31 semantic datas not in phoneme datas
+        deleted 34 audios who's duration are bigger than 54 seconds
+        deleted 3190 audios who's phoneme/sec are bigger than 25 or smaller than 3
+        dataset.__len__(): 366463
+
+        """
+        # 345410 for LibriTTS
+        print("dataset.__len__():", self.__len__())
+
+    def __get_item_names__(self) -> List[str]:
+        return self.item_names
+
+    def __len__(self) -> int:
+        return len(self.semantic_phoneme)
+
+    def __getitem__(self, idx: int) -> Dict:
+        semantic_ids, phoneme_ids = self.semantic_phoneme[idx]
+        item_name = self.item_names[idx]
+        phoneme_ids_len = len(phoneme_ids)
+        # semantic tokens target
+        semantic_ids_len = len(semantic_ids)
+
+        flag = 0
+        path_bert = "%s/%s.pt" % (self.path3, item_name)
+        if os.path.exists(path_bert) == True:
+            bert_feature = torch.load(path_bert, map_location="cpu")
+        else:
+            flag = 1
+        if flag == 1:
+            # bert_feature=torch.zeros_like(phoneme_ids,dtype=torch.float32)
+            bert_feature = None
+        else:
+            assert bert_feature.shape[-1] == len(phoneme_ids)
+        return {
+            "idx": idx,
+            "phoneme_ids": phoneme_ids,
+            "phoneme_ids_len": phoneme_ids_len,
+            "semantic_ids": semantic_ids,
+            "semantic_ids_len": semantic_ids_len,
+            "bert_feature": bert_feature,
+        }
+
+    def get_sample_length(self, idx: int):
+        semantic_ids = self.semantic_phoneme[idx][0]
+        sec = 1.0 * len(semantic_ids) / self.hz
+        return sec
+
+    def collate(self, examples: List[Dict]) -> Dict:
+        sample_index: List[int] = []
+        phoneme_ids: List[torch.Tensor] = []
+        phoneme_ids_lens: List[int] = []
+        semantic_ids: List[torch.Tensor] = []
+        semantic_ids_lens: List[int] = []
+        # return
+
+        for item in examples:
+            sample_index.append(item["idx"])
+            phoneme_ids.append(np.array(item["phoneme_ids"], dtype=np.int64))
+            semantic_ids.append(np.array(item["semantic_ids"], dtype=np.int64))
+            phoneme_ids_lens.append(item["phoneme_ids_len"])
+            semantic_ids_lens.append(item["semantic_ids_len"])
+
+        # pad 0
+        phoneme_ids = batch_sequences(phoneme_ids)
+        semantic_ids = batch_sequences(semantic_ids, pad_value=self.PAD)
+
+        # # convert each batch to torch.tensor
+        phoneme_ids = torch.tensor(phoneme_ids)
+        semantic_ids = torch.tensor(semantic_ids)
+        phoneme_ids_lens = torch.tensor(phoneme_ids_lens)
+        semantic_ids_lens = torch.tensor(semantic_ids_lens)
+        bert_padded = torch.FloatTensor(len(examples), 1024, max(phoneme_ids_lens))
+        bert_padded.zero_()
+
+        for idx, item in enumerate(examples):
+            bert = item["bert_feature"]
+            if bert != None:
+                bert_padded[idx, :, : bert.shape[-1]] = bert
+
+        return {
+            # List[int]
+            "ids": sample_index,
+            # torch.Tensor (B, max_phoneme_length)
+            "phoneme_ids": phoneme_ids,
+            # torch.Tensor (B)
+            "phoneme_ids_len": phoneme_ids_lens,
+            # torch.Tensor (B, max_semantic_ids_length)
+            "semantic_ids": semantic_ids,
+            # torch.Tensor (B)
+            "semantic_ids_len": semantic_ids_lens,
+            # torch.Tensor (B, 1024, max_phoneme_length)
+            "bert_feature": bert_padded,
+        }
+
+
+if __name__ == "__main__":
+    root_dir = "/data/docker/liujing04/gpt-vits/prepare/dump_mix/"
+    dataset = Text2SemanticDataset(
+        phoneme_path=root_dir + "phoneme_train.npy",
+        semantic_path=root_dir + "semantic_train.tsv",
+    )
+
+    batch_size = 12
+    dataloader = DataLoader(
+        dataset, batch_size=batch_size, collate_fn=dataset.collate, shuffle=False
+    )
+    for i, batch in enumerate(dataloader):
+        if i % 1000 == 0:
+            print(i)
+        # if i == 0:
+        #     print('batch["ids"]:', batch["ids"])
+        # print('batch["phoneme_ids"]:', batch["phoneme_ids"],
+        #       batch["phoneme_ids"].shape)
+        # print('batch["phoneme_ids_len"]:', batch["phoneme_ids_len"],
+        #       batch["phoneme_ids_len"].shape)
+        # print('batch["semantic_ids"]:', batch["semantic_ids"],
+        #       batch["semantic_ids"].shape)
+        # print('batch["semantic_ids_len"]:', batch["semantic_ids_len"],
+        #       batch["semantic_ids_len"].shape)

AR/models/t2s_lightning_module.py

@@ -0,0 +1,140 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/t2s_lightning_module.py
+import os, sys
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from typing import Dict
+
+import torch
+from pytorch_lightning import LightningModule
+from AR.models.t2s_model import Text2SemanticDecoder
+from AR.modules.lr_schedulers import WarmupCosineLRSchedule
+from AR.modules.optim import ScaledAdam
+
+
+class Text2SemanticLightningModule(LightningModule):
+    def __init__(self, config, output_dir, is_train=True):
+        super().__init__()
+        self.config = config
+        self.top_k = 3
+        self.model = Text2SemanticDecoder(config=config, top_k=self.top_k)
+        pretrained_s1 = config.get("pretrained_s1")
+        if pretrained_s1 and is_train:
+            # print(self.load_state_dict(torch.load(pretrained_s1,map_location="cpu")["state_dict"]))
+            print(
+                self.load_state_dict(
+                    torch.load(pretrained_s1, map_location="cpu")["weight"]
+                )
+            )
+        if is_train:
+            self.automatic_optimization = False
+            self.save_hyperparameters()
+            self.eval_dir = output_dir / "eval"
+            self.eval_dir.mkdir(parents=True, exist_ok=True)
+
+    def training_step(self, batch: Dict, batch_idx: int):
+        opt = self.optimizers()
+        scheduler = self.lr_schedulers()
+        loss, acc = self.model.forward(
+            batch["phoneme_ids"],
+            batch["phoneme_ids_len"],
+            batch["semantic_ids"],
+            batch["semantic_ids_len"],
+            batch["bert_feature"],
+        )
+        self.manual_backward(loss)
+        if batch_idx > 0 and batch_idx % 4 == 0:
+            opt.step()
+            opt.zero_grad()
+            scheduler.step()
+
+        self.log(
+            "total_loss",
+            loss,
+            on_step=True,
+            on_epoch=True,
+            prog_bar=True,
+            sync_dist=True,
+        )
+        self.log(
+            "lr",
+            scheduler.get_last_lr()[0],
+            on_epoch=True,
+            prog_bar=True,
+            sync_dist=True,
+        )
+        self.log(
+            f"top_{self.top_k}_acc",
+            acc,
+            on_step=True,
+            on_epoch=True,
+            prog_bar=True,
+            sync_dist=True,
+        )
+
+    def validation_step(self, batch: Dict, batch_idx: int):
+        return
+
+    # # get loss
+    # loss, acc = self.model.forward(
+    #     batch['phoneme_ids'], batch['phoneme_ids_len'],
+    #     batch['semantic_ids'], batch['semantic_ids_len'],
+    #     batch['bert_feature']
+    # )
+    #
+    # self.log(
+    #     "val_total_loss",
+    #     loss,
+    #     on_step=True,
+    #     on_epoch=True,
+    #     prog_bar=True,
+    #     sync_dist=True)
+    # self.log(
+    #     f"val_top_{self.top_k}_acc",
+    #     acc,
+    #     on_step=True,
+    #     on_epoch=True,
+    #     prog_bar=True,
+    #     sync_dist=True)
+    #
+    # # get infer output
+    # semantic_len = batch['semantic_ids'].size(1)
+    # prompt_len = min(int(semantic_len * 0.5), 150)
+    # prompt = batch['semantic_ids'][:, :prompt_len]
+    # pred_semantic = self.model.infer(batch['phoneme_ids'],
+    #                                  batch['phoneme_ids_len'], prompt,
+    #                                  batch['bert_feature']
+    #                                  )
+    # save_name = f'semantic_toks_{batch_idx}.pt'
+    # save_path = os.path.join(self.eval_dir, save_name)
+    # torch.save(pred_semantic.detach().cpu(), save_path)
+
+    def configure_optimizers(self):
+        model_parameters = self.model.parameters()
+        parameters_names = []
+        parameters_names.append(
+            [name_param_pair[0] for name_param_pair in self.model.named_parameters()]
+        )
+        lm_opt = ScaledAdam(
+            model_parameters,
+            lr=0.01,
+            betas=(0.9, 0.95),
+            clipping_scale=2.0,
+            parameters_names=parameters_names,
+            show_dominant_parameters=False,
+            clipping_update_period=1000,
+        )
+
+        return {
+            "optimizer": lm_opt,
+            "lr_scheduler": {
+                "scheduler": WarmupCosineLRSchedule(
+                    lm_opt,
+                    init_lr=self.config["optimizer"]["lr_init"],
+                    peak_lr=self.config["optimizer"]["lr"],
+                    end_lr=self.config["optimizer"]["lr_end"],
+                    warmup_steps=self.config["optimizer"]["warmup_steps"],
+                    total_steps=self.config["optimizer"]["decay_steps"],
+                )
+            },
+        }

AR/models/t2s_lightning_module_onnx.py

@@ -0,0 +1,106 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/t2s_lightning_module.py
+import os, sys
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from typing import Dict
+
+import torch
+from pytorch_lightning import LightningModule
+from AR.models.t2s_model_onnx import Text2SemanticDecoder
+from AR.modules.lr_schedulers import WarmupCosineLRSchedule
+from AR.modules.optim import ScaledAdam
+
+
+class Text2SemanticLightningModule(LightningModule):
+    def __init__(self, config, output_dir, is_train=True):
+        super().__init__()
+        self.config = config
+        self.top_k = 3
+        self.model = Text2SemanticDecoder(config=config, top_k=self.top_k)
+        pretrained_s1 = config.get("pretrained_s1")
+        if pretrained_s1 and is_train:
+            # print(self.load_state_dict(torch.load(pretrained_s1,map_location="cpu")["state_dict"]))
+            print(
+                self.load_state_dict(
+                    torch.load(pretrained_s1, map_location="cpu")["weight"]
+                )
+            )
+        if is_train:
+            self.automatic_optimization = False
+            self.save_hyperparameters()
+            self.eval_dir = output_dir / "eval"
+            self.eval_dir.mkdir(parents=True, exist_ok=True)
+
+    def training_step(self, batch: Dict, batch_idx: int):
+        opt = self.optimizers()
+        scheduler = self.lr_schedulers()
+        loss, acc = self.model.forward(
+            batch["phoneme_ids"],
+            batch["phoneme_ids_len"],
+            batch["semantic_ids"],
+            batch["semantic_ids_len"],
+            batch["bert_feature"],
+        )
+        self.manual_backward(loss)
+        if batch_idx > 0 and batch_idx % 4 == 0:
+            opt.step()
+            opt.zero_grad()
+            scheduler.step()
+
+        self.log(
+            "total_loss",
+            loss,
+            on_step=True,
+            on_epoch=True,
+            prog_bar=True,
+            sync_dist=True,
+        )
+        self.log(
+            "lr",
+            scheduler.get_last_lr()[0],
+            on_epoch=True,
+            prog_bar=True,
+            sync_dist=True,
+        )
+        self.log(
+            f"top_{self.top_k}_acc",
+            acc,
+            on_step=True,
+            on_epoch=True,
+            prog_bar=True,
+            sync_dist=True,
+        )
+
+    def validation_step(self, batch: Dict, batch_idx: int):
+        return
+
+    def configure_optimizers(self):
+        model_parameters = self.model.parameters()
+        parameters_names = []
+        parameters_names.append(
+            [name_param_pair[0] for name_param_pair in self.model.named_parameters()]
+        )
+        lm_opt = ScaledAdam(
+            model_parameters,
+            lr=0.01,
+            betas=(0.9, 0.95),
+            clipping_scale=2.0,
+            parameters_names=parameters_names,
+            show_dominant_parameters=False,
+            clipping_update_period=1000,
+        )
+
+        return {
+            "optimizer": lm_opt,
+            "lr_scheduler": {
+                "scheduler": WarmupCosineLRSchedule(
+                    lm_opt,
+                    init_lr=self.config["optimizer"]["lr_init"],
+                    peak_lr=self.config["optimizer"]["lr"],
+                    end_lr=self.config["optimizer"]["lr_end"],
+                    warmup_steps=self.config["optimizer"]["warmup_steps"],
+                    total_steps=self.config["optimizer"]["decay_steps"],
+                )
+            },
+        }

AR/models/t2s_model.py

@@ -0,0 +1,327 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/t2s_model.py
+import torch
+from tqdm import tqdm
+
+from AR.models.utils import make_pad_mask
+from AR.models.utils import (
+    topk_sampling,
+    sample,
+    logits_to_probs,
+    multinomial_sample_one_no_sync,
+)
+from AR.modules.embedding import SinePositionalEmbedding
+from AR.modules.embedding import TokenEmbedding
+from AR.modules.transformer import LayerNorm
+from AR.modules.transformer import TransformerEncoder
+from AR.modules.transformer import TransformerEncoderLayer
+from torch import nn
+from torch.nn import functional as F
+from torchmetrics.classification import MulticlassAccuracy
+
+default_config = {
+    "embedding_dim": 512,
+    "hidden_dim": 512,
+    "num_head": 8,
+    "num_layers": 12,
+    "num_codebook": 8,
+    "p_dropout": 0.0,
+    "vocab_size": 1024 + 1,
+    "phoneme_vocab_size": 512,
+    "EOS": 1024,
+}
+
+
+class Text2SemanticDecoder(nn.Module):
+    def __init__(self, config, norm_first=False, top_k=3):
+        super(Text2SemanticDecoder, self).__init__()
+        self.model_dim = config["model"]["hidden_dim"]
+        self.embedding_dim = config["model"]["embedding_dim"]
+        self.num_head = config["model"]["head"]
+        self.num_layers = config["model"]["n_layer"]
+        self.norm_first = norm_first
+        self.vocab_size = config["model"]["vocab_size"]
+        self.phoneme_vocab_size = config["model"]["phoneme_vocab_size"]
+        self.p_dropout = config["model"]["dropout"]
+        self.EOS = config["model"]["EOS"]
+        self.norm_first = norm_first
+        assert self.EOS == self.vocab_size - 1
+        # should be same as num of kmeans bin
+        # assert self.EOS == 1024
+        self.bert_proj = nn.Linear(1024, self.embedding_dim)
+        self.ar_text_embedding = TokenEmbedding(
+            self.embedding_dim, self.phoneme_vocab_size, self.p_dropout
+        )
+        self.ar_text_position = SinePositionalEmbedding(
+            self.embedding_dim, dropout=0.1, scale=False, alpha=True
+        )
+        self.ar_audio_embedding = TokenEmbedding(
+            self.embedding_dim, self.vocab_size, self.p_dropout
+        )
+        self.ar_audio_position = SinePositionalEmbedding(
+            self.embedding_dim, dropout=0.1, scale=False, alpha=True
+        )
+
+        self.h = TransformerEncoder(
+            TransformerEncoderLayer(
+                d_model=self.model_dim,
+                nhead=self.num_head,
+                dim_feedforward=self.model_dim * 4,
+                dropout=0.1,
+                batch_first=True,
+                norm_first=norm_first,
+            ),
+            num_layers=self.num_layers,
+            norm=LayerNorm(self.model_dim) if norm_first else None,
+        )
+
+        self.ar_predict_layer = nn.Linear(self.model_dim, self.vocab_size, bias=False)
+        self.loss_fct = nn.CrossEntropyLoss(reduction="sum")
+
+        self.ar_accuracy_metric = MulticlassAccuracy(
+            self.vocab_size,
+            top_k=top_k,
+            average="micro",
+            multidim_average="global",
+            ignore_index=self.EOS,
+        )
+
+    def forward(self, x, x_lens, y, y_lens, bert_feature):
+        """
+        x: phoneme_ids
+        y: semantic_ids
+        """
+        x = self.ar_text_embedding(x)
+        x = x + self.bert_proj(bert_feature.transpose(1, 2))
+        x = self.ar_text_position(x)
+        x_mask = make_pad_mask(x_lens)
+
+        y_mask = make_pad_mask(y_lens)
+        y_mask_int = y_mask.type(torch.int64)
+        codes = y.type(torch.int64) * (1 - y_mask_int)
+
+        # Training
+        # AR Decoder
+        y, targets = self.pad_y_eos(codes, y_mask_int, eos_id=self.EOS)
+        x_len = x_lens.max()
+        y_len = y_lens.max()
+        y_emb = self.ar_audio_embedding(y)
+        y_pos = self.ar_audio_position(y_emb)
+
+        xy_padding_mask = torch.concat([x_mask, y_mask], dim=1)
+        ar_xy_padding_mask = xy_padding_mask
+
+        x_attn_mask = F.pad(
+            torch.zeros((x_len, x_len), dtype=torch.bool, device=x.device),
+            (0, y_len),
+            value=True,
+        )
+        y_attn_mask = F.pad(
+            torch.triu(
+                torch.ones(y_len, y_len, dtype=torch.bool, device=x.device),
+                diagonal=1,
+            ),
+            (x_len, 0),
+            value=False,
+        )
+        xy_attn_mask = torch.concat([x_attn_mask, y_attn_mask], dim=0)
+        bsz, src_len = x.shape[0], x_len + y_len
+        _xy_padding_mask = (
+            ar_xy_padding_mask.view(bsz, 1, 1, src_len)
+            .expand(-1, self.num_head, -1, -1)
+            .reshape(bsz * self.num_head, 1, src_len)
+        )
+        xy_attn_mask = xy_attn_mask.logical_or(_xy_padding_mask)
+        new_attn_mask = torch.zeros_like(xy_attn_mask, dtype=x.dtype)
+        new_attn_mask.masked_fill_(xy_attn_mask, float("-inf"))
+        xy_attn_mask = new_attn_mask
+        # x 和完整的 y 一次性输入模型
+        xy_pos = torch.concat([x, y_pos], dim=1)
+        xy_dec, _ = self.h(
+            (xy_pos, None),
+            mask=xy_attn_mask,
+        )
+        logits = self.ar_predict_layer(xy_dec[:, x_len:]).permute(0, 2, 1)
+        # loss
+        # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
+        loss = F.cross_entropy(logits, targets, reduction="sum")
+        acc = self.ar_accuracy_metric(logits.detach(), targets).item()
+        return loss, acc
+
+    # 需要看下这个函数和 forward 的区别以及没有 semantic 的时候 prompts 输入什么
+    def infer(
+        self,
+        x,
+        x_lens,
+        prompts,
+        bert_feature,
+        top_k: int = -100,
+        early_stop_num: int = -1,
+        temperature: float = 1.0,
+    ):
+        x = self.ar_text_embedding(x)
+        x = x + self.bert_proj(bert_feature.transpose(1, 2))
+        x = self.ar_text_position(x)
+
+        # AR Decoder
+        y = prompts
+        prefix_len = y.shape[1]
+        x_len = x.shape[1]
+        x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
+        stop = False
+        for _ in tqdm(range(1500)):
+            y_emb = self.ar_audio_embedding(y)
+            y_pos = self.ar_audio_position(y_emb)
+            # x 和逐渐增长的 y 一起输入给模型
+            xy_pos = torch.concat([x, y_pos], dim=1)
+            y_len = y.shape[1]
+            x_attn_mask_pad = F.pad(
+                x_attn_mask,
+                (0, y_len),
+                value=True,
+            )
+            y_attn_mask = F.pad(
+                torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
+                (x_len, 0),
+                value=False,
+            )
+            xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
+                y.device
+            )
+
+            xy_dec, _ = self.h(
+                (xy_pos, None),
+                mask=xy_attn_mask,
+            )
+            logits = self.ar_predict_layer(xy_dec[:, -1])
+            samples = topk_sampling(
+                logits, top_k=top_k, top_p=1.0, temperature=temperature
+            )
+
+            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
+                print("use early stop num:", early_stop_num)
+                stop = True
+
+            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
+                # print(torch.argmax(logits, dim=-1)[0] == self.EOS, samples[0, 0] == self.EOS)
+                stop = True
+            if stop:
+                if prompts.shape[1] == y.shape[1]:
+                    y = torch.concat([y, torch.zeros_like(samples)], dim=1)
+                    print("bad zero prediction")
+                print(f"T2S Decoding EOS [{prefix_len} -> {y.shape[1]}]")
+                break
+            # 本次生成的 semantic_ids 和之前的 y 构成新的 y
+            # print(samples.shape)#[1,1]#第一个1是bs
+            # import os
+            # os._exit(2333)
+            y = torch.concat([y, samples], dim=1)
+        return y
+
+    def pad_y_eos(self, y, y_mask_int, eos_id):
+        targets = F.pad(y, (0, 1), value=0) + eos_id * F.pad(
+            y_mask_int, (0, 1), value=1
+        )
+        # 错位
+        return targets[:, :-1], targets[:, 1:]
+
+    def infer_panel(
+        self,
+        x,  #####全部文本token
+        x_lens,
+        prompts,  ####参考音频token
+        bert_feature,
+        top_k: int = -100,
+        early_stop_num: int = -1,
+        temperature: float = 1.0,
+    ):
+        x = self.ar_text_embedding(x)
+        x = x + self.bert_proj(bert_feature.transpose(1, 2))
+        x = self.ar_text_position(x)
+
+        # AR Decoder
+        y = prompts
+        prefix_len = y.shape[1]
+        x_len = x.shape[1]
+        x_attn_mask = torch.zeros((x_len, x_len), dtype=torch.bool)
+        stop = False
+        # print(1111111,self.num_layers)
+        cache = {
+            "all_stage": self.num_layers,
+            "k": [None] * self.num_layers,  ###根据配置自己手写
+            "v": [None] * self.num_layers,
+            # "xy_pos":None,##y_pos位置编码每次都不一样的没法缓存，每次都要重新拼xy_pos.主要还是写法原因，其实是可以历史统一一样的，但也没啥计算量就不管了
+            "y_emb": None,  ##只需要对最新的samples求emb，再拼历史的就行
+            # "logits":None,###原版就已经只对结尾求再拼接了，不用管
+            # "xy_dec":None,###不需要，本来只需要最后一个做logits
+            "first_infer": 1,
+            "stage": 0,
+        }
+        for idx in tqdm(range(1500)):
+            if cache["first_infer"] == 1:
+                y_emb = self.ar_audio_embedding(y)
+            else:
+                y_emb = torch.cat(
+                    [cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], 1
+                )
+            cache["y_emb"] = y_emb
+            y_pos = self.ar_audio_position(y_emb)
+            # x 和逐渐增长的 y 一起输入给模型
+            if cache["first_infer"] == 1:
+                xy_pos = torch.concat([x, y_pos], dim=1)
+            else:
+                xy_pos = y_pos[:, -1:]
+            y_len = y_pos.shape[1]
+            ###以下3个不做缓存
+            if cache["first_infer"] == 1:
+                x_attn_mask_pad = F.pad(
+                    x_attn_mask,
+                    (0, y_len),  ###xx的纯0扩展到xx纯0+xy纯1，(x,x+y)
+                    value=True,
+                )
+                y_attn_mask = F.pad(  ###yy的右上1扩展到左边xy的0,(y,x+y)
+                    torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
+                    (x_len, 0),
+                    value=False,
+                )
+                xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
+                    y.device
+                )
+            else:
+                ###最右边一列（是错的）
+                # xy_attn_mask=torch.ones((1, x_len+y_len), dtype=torch.bool,device=xy_pos.device)
+                # xy_attn_mask[:,-1]=False
+                ###最下面一行（是对的）
+                xy_attn_mask = torch.zeros(
+                    (1, x_len + y_len), dtype=torch.bool, device=xy_pos.device
+                )
+            # pdb.set_trace()
+            ###缓存重头戏
+            # print(1111,xy_pos.shape,xy_attn_mask.shape,x_len,y_len)
+            xy_dec, _ = self.h((xy_pos, None), mask=xy_attn_mask, cache=cache)
+            logits = self.ar_predict_layer(
+                xy_dec[:, -1]
+            )  ##不用改，如果用了cache的默认就是只有一帧，取最后一帧一样的
+            # samples = topk_sampling(logits, top_k=top_k, top_p=1.0, temperature=temperature)
+            if(idx==0):###第一次跑不能EOS否则没有了
+                logits = logits[:, :-1]  ###刨除1024终止符号的概率
+            samples = sample(
+                logits[0], y, top_k=top_k, top_p=1.0, repetition_penalty=1.35
+            )[0].unsqueeze(0)
+            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
+                print("use early stop num:", early_stop_num)
+                stop = True
+
+            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
+                # print(torch.argmax(logits, dim=-1)[0] == self.EOS, samples[0, 0] == self.EOS)
+                stop = True
+            if stop:
+                if prompts.shape[1] == y.shape[1]:
+                    y = torch.concat([y, torch.zeros_like(samples)], dim=1)
+                    print("bad zero prediction")
+                print(f"T2S Decoding EOS [{prefix_len} -> {y.shape[1]}]")
+                break
+            # 本次生成的 semantic_ids 和之前的 y 构成新的 y
+            # print(samples.shape)#[1,1]#第一个1是bs
+            y = torch.concat([y, samples], dim=1)
+            cache["first_infer"] = 0
+        return y, idx

AR/models/t2s_model_onnx.py

@@ -0,0 +1,337 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/t2s_model.py
+import torch
+from tqdm import tqdm
+
+from AR.modules.embedding_onnx import SinePositionalEmbedding
+from AR.modules.embedding_onnx import TokenEmbedding
+from AR.modules.transformer_onnx import LayerNorm
+from AR.modules.transformer_onnx import TransformerEncoder
+from AR.modules.transformer_onnx import TransformerEncoderLayer
+from torch import nn
+from torch.nn import functional as F
+from torchmetrics.classification import MulticlassAccuracy
+
+default_config = {
+    "embedding_dim": 512,
+    "hidden_dim": 512,
+    "num_head": 8,
+    "num_layers": 12,
+    "num_codebook": 8,
+    "p_dropout": 0.0,
+    "vocab_size": 1024 + 1,
+    "phoneme_vocab_size": 512,
+    "EOS": 1024,
+}
+
+inf_tensor_value = torch.FloatTensor([-float("Inf")]).float()
+
+def logits_to_probs(
+    logits,
+    previous_tokens = None,
+    temperature: float = 1.0,
+    top_k = None,
+    top_p = None,
+    repetition_penalty: float = 1.0,
+):
+    previous_tokens = previous_tokens.squeeze()
+    if previous_tokens is not None and repetition_penalty != 1.0:
+        previous_tokens = previous_tokens.long()
+        score = torch.gather(logits, dim=0, index=previous_tokens)
+        score = torch.where(
+            score < 0, score * repetition_penalty, score / repetition_penalty
+        )
+        logits.scatter_(dim=0, index=previous_tokens, src=score)
+
+    if top_p is not None and top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cum_probs = torch.cumsum(
+            torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1
+        )
+        sorted_indices_to_remove = cum_probs > top_p
+        sorted_indices_to_remove[0] = False  # keep at least one option
+        indices_to_remove = sorted_indices_to_remove.scatter(
+            dim=0, index=sorted_indices, src=sorted_indices_to_remove
+        )
+        logits = logits.masked_fill(indices_to_remove, -float("Inf"))
+
+    logits = logits / max(temperature, 1e-5)
+
+    if top_k is not None:
+        v, _ = torch.topk(logits, top_k)
+        pivot = v.select(-1, -1).unsqueeze(-1)
+        logits = torch.where(logits < pivot, inf_tensor_value, logits)
+
+    probs = torch.nn.functional.softmax(logits, dim=-1)
+    return probs
+
+
+def multinomial_sample_one_no_sync(
+    probs_sort
+):  # Does multinomial sampling without a cuda synchronization
+    q = torch.randn_like(probs_sort)
+    return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)
+
+
+def sample(
+    logits,
+    previous_tokens,
+    **sampling_kwargs,
+):
+    probs = logits_to_probs(
+        logits=logits, previous_tokens=previous_tokens, **sampling_kwargs
+    )
+    idx_next = multinomial_sample_one_no_sync(probs)
+    return idx_next, probs
+
+
+class OnnxEncoder(nn.Module):
+    def __init__(self, ar_text_embedding, bert_proj, ar_text_position):
+        super().__init__()
+        self.ar_text_embedding = ar_text_embedding
+        self.bert_proj = bert_proj
+        self.ar_text_position = ar_text_position
+    
+    def forward(self, x, bert_feature):
+        x = self.ar_text_embedding(x)
+        x = x + self.bert_proj(bert_feature.transpose(1, 2))
+        return self.ar_text_position(x)
+
+
+class T2SFirstStageDecoder(nn.Module):
+    def __init__(self, ar_audio_embedding, ar_audio_position, h, ar_predict_layer, loss_fct, ar_accuracy_metric,
+    top_k, early_stop_num, num_layers):
+        super().__init__()
+        self.ar_audio_embedding = ar_audio_embedding
+        self.ar_audio_position = ar_audio_position
+        self.h = h
+        self.ar_predict_layer = ar_predict_layer
+        self.loss_fct = loss_fct
+        self.ar_accuracy_metric = ar_accuracy_metric
+        self.top_k = top_k
+        self.early_stop_num = early_stop_num
+        self.num_layers = num_layers
+    
+    def forward(self, x, prompt):
+        y = prompt
+        x_example = x[:,:,0] * 0.0
+        #N, 1, 512
+        cache = {
+            "all_stage": self.num_layers,
+            "k": None,
+            "v": None,
+            "y_emb": None,
+            "first_infer": 1,
+            "stage": 0,
+        }
+
+        y_emb = self.ar_audio_embedding(y)
+
+        cache["y_emb"] = y_emb
+        y_pos = self.ar_audio_position(y_emb)
+
+        xy_pos = torch.concat([x, y_pos], dim=1)
+
+        y_example = y_pos[:,:,0] * 0.0
+        x_attn_mask = torch.matmul(x_example.transpose(0, 1) , x_example).bool()
+        y_attn_mask = torch.ones_like(torch.matmul(y_example.transpose(0, 1), y_example), dtype=torch.int64)
+        y_attn_mask = torch.cumsum(y_attn_mask, dim=1) - torch.cumsum(
+            torch.ones_like(y_example.transpose(0, 1), dtype=torch.int64), dim=0
+        )
+        y_attn_mask = y_attn_mask > 0
+
+        x_y_pad = torch.matmul(x_example.transpose(0, 1), y_example).bool()
+        y_x_pad = torch.matmul(y_example.transpose(0, 1), x_example).bool()
+        x_attn_mask_pad = torch.cat([x_attn_mask, torch.ones_like(x_y_pad)], dim=1)
+        y_attn_mask = torch.cat([y_x_pad, y_attn_mask], dim=1)
+        xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0)
+        cache["k"] = torch.matmul(x_attn_mask_pad[0].float().unsqueeze(-1), torch.zeros((1, 512)))\
+        .unsqueeze(1).repeat(self.num_layers, 1, 1, 1)
+        cache["v"] = torch.matmul(x_attn_mask_pad[0].float().unsqueeze(-1), torch.zeros((1, 512)))\
+        .unsqueeze(1).repeat(self.num_layers, 1, 1, 1)
+
+        xy_dec = self.h(xy_pos, mask=xy_attn_mask, cache=cache)
+        logits = self.ar_predict_layer(xy_dec[:, -1])
+        samples = sample(logits[0], y, top_k=self.top_k, top_p=1.0, repetition_penalty=1.35)[0].unsqueeze(0)
+
+        y = torch.concat([y, samples], dim=1)
+
+        return y, cache["k"], cache["v"], cache["y_emb"], x_example
+
+
+class T2SStageDecoder(nn.Module):
+    def __init__(self, ar_audio_embedding, ar_audio_position, h, ar_predict_layer, loss_fct, ar_accuracy_metric,
+    top_k, early_stop_num, num_layers):
+        super().__init__()
+        self.ar_audio_embedding = ar_audio_embedding
+        self.ar_audio_position = ar_audio_position
+        self.h = h
+        self.ar_predict_layer = ar_predict_layer
+        self.loss_fct = loss_fct
+        self.ar_accuracy_metric = ar_accuracy_metric
+        self.top_k = top_k
+        self.early_stop_num = early_stop_num
+        self.num_layers = num_layers
+
+    def forward(self, y, k, v, y_emb, x_example):
+        cache = {
+            "all_stage": self.num_layers,
+            "k": torch.nn.functional.pad(k, (0, 0, 0, 0, 0, 1)),
+            "v": torch.nn.functional.pad(v, (0, 0, 0, 0, 0, 1)),
+            "y_emb": y_emb,
+            "first_infer": 0,
+            "stage": 0,
+        }
+
+        y_emb = torch.cat(
+            [cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], 1
+        )
+        cache["y_emb"] = y_emb
+        y_pos = self.ar_audio_position(y_emb)
+
+        xy_pos = y_pos[:, -1:]
+        
+        y_example = y_pos[:,:,0] * 0.0
+
+        xy_attn_mask = torch.cat([x_example, y_example], dim=1)
+        xy_attn_mask = torch.zeros_like(xy_attn_mask, dtype=torch.bool)
+
+        xy_dec = self.h(xy_pos, mask=xy_attn_mask, cache=cache)
+        logits = self.ar_predict_layer(xy_dec[:, -1])
+        samples = sample(logits[0], y, top_k=self.top_k, top_p=1.0, repetition_penalty=1.35)[0].unsqueeze(0)
+
+        y = torch.concat([y, samples], dim=1)
+
+        return y, cache["k"], cache["v"], cache["y_emb"], logits, samples
+
+
+class Text2SemanticDecoder(nn.Module):
+    def __init__(self, config, norm_first=False, top_k=3):
+        super(Text2SemanticDecoder, self).__init__()
+        self.model_dim = config["model"]["hidden_dim"]
+        self.embedding_dim = config["model"]["embedding_dim"]
+        self.num_head = config["model"]["head"]
+        self.num_layers = config["model"]["n_layer"]
+        self.norm_first = norm_first
+        self.vocab_size = config["model"]["vocab_size"]
+        self.phoneme_vocab_size = config["model"]["phoneme_vocab_size"]
+        self.p_dropout = float(config["model"]["dropout"])
+        self.EOS = config["model"]["EOS"]
+        self.norm_first = norm_first
+        assert self.EOS == self.vocab_size - 1
+        self.bert_proj = nn.Linear(1024, self.embedding_dim)
+        self.ar_text_embedding = TokenEmbedding(self.embedding_dim, self.phoneme_vocab_size, self.p_dropout)
+        self.ar_text_position = SinePositionalEmbedding(self.embedding_dim, dropout=0.1, scale=False, alpha=True)
+        self.ar_audio_embedding = TokenEmbedding(self.embedding_dim, self.vocab_size, self.p_dropout)
+        self.ar_audio_position = SinePositionalEmbedding(self.embedding_dim, dropout=0.1, scale=False, alpha=True)
+        self.h = TransformerEncoder(
+            TransformerEncoderLayer(
+                d_model=self.model_dim,
+                nhead=self.num_head,
+                dim_feedforward=self.model_dim * 4,
+                dropout=0.1,
+                batch_first=True,
+                norm_first=norm_first,
+            ),
+            num_layers=self.num_layers,
+            norm=LayerNorm(self.model_dim) if norm_first else None,
+        )
+        self.ar_predict_layer = nn.Linear(self.model_dim, self.vocab_size, bias=False)
+        self.loss_fct = nn.CrossEntropyLoss(reduction="sum")
+        self.ar_accuracy_metric = MulticlassAccuracy(
+            self.vocab_size,
+            top_k=top_k,
+            average="micro",
+            multidim_average="global",
+            ignore_index=self.EOS,
+        )
+        self.top_k = torch.LongTensor([1])
+        self.early_stop_num = torch.LongTensor([-1])
+
+    def init_onnx(self):
+        self.onnx_encoder = OnnxEncoder(self.ar_text_embedding, self.bert_proj, self.ar_text_position)
+        self.first_stage_decoder = T2SFirstStageDecoder(self.ar_audio_embedding, self.ar_audio_position, self.h, 
+            self.ar_predict_layer, self.loss_fct, self.ar_accuracy_metric, self.top_k, self.early_stop_num,
+            self.num_layers)
+        self.stage_decoder = T2SStageDecoder(self.ar_audio_embedding, self.ar_audio_position, self.h, 
+            self.ar_predict_layer, self.loss_fct, self.ar_accuracy_metric, self.top_k, self.early_stop_num,
+            self.num_layers)
+
+    def forward(self, x, prompts, bert_feature):
+        early_stop_num = self.early_stop_num
+        prefix_len = prompts.shape[1]
+
+        x = self.onnx_encoder(x, bert_feature)
+        y, k, v, y_emb, stage, x_example = self.first_stage_decoder(x, prompts)
+
+        stop = False
+        for idx in range(1, 1500):
+            enco = self.stage_decoder(y, k, v, y_emb, stage, x_example)
+            y, k, v, y_emb, stage, logits, samples = enco
+            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
+                stop = True
+            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
+                stop = True
+            if stop:
+                break
+        y[0, -1] = 0
+        return y, idx
+
+    def infer(self, x, prompts, bert_feature):
+        top_k = self.top_k
+        early_stop_num = self.early_stop_num
+
+        x = self.onnx_encoder(x, bert_feature)
+
+        y = prompts
+        prefix_len = y.shape[1]
+        x_len = x.shape[1]
+        x_example = x[:,:,0] * 0.0
+        x_attn_mask = torch.matmul(x_example.transpose(0, 1), x_example)
+        x_attn_mask = torch.zeros_like(x_attn_mask, dtype=torch.bool)
+
+        stop = False
+        cache = {
+            "all_stage": self.num_layers,
+            "k": [None] * self.num_layers,
+            "v": [None] * self.num_layers,
+            "y_emb": None,
+            "first_infer": 1,
+            "stage": 0,
+        }
+        for idx in range(1500):
+            if cache["first_infer"] == 1:
+                y_emb = self.ar_audio_embedding(y)
+            else:
+                y_emb = torch.cat(
+                    [cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], 1
+                )
+            cache["y_emb"] = y_emb
+            y_pos = self.ar_audio_position(y_emb)
+            if cache["first_infer"] == 1:
+                xy_pos = torch.concat([x, y_pos], dim=1)
+            else:
+                xy_pos = y_pos[:, -1:]
+            y_len = y_pos.shape[1]
+            if cache["first_infer"] == 1:
+                x_attn_mask_pad = F.pad(x_attn_mask, (0, y_len), value=True)
+                y_attn_mask = F.pad(
+                    torch.triu(torch.ones(y_len, y_len, dtype=torch.bool), diagonal=1),
+                    (x_len, 0), value=False
+                )
+                xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0)
+            else:
+                xy_attn_mask = torch.zeros((1, x_len + y_len), dtype=torch.bool)
+            xy_dec = self.h(xy_pos, mask=xy_attn_mask, cache=cache)
+            logits = self.ar_predict_layer(xy_dec[:, -1])
+            samples = sample(logits[0], y, top_k=top_k, top_p=1.0, repetition_penalty=1.35)[0].unsqueeze(0)
+            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
+                stop = True
+            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
+                stop = True
+            if stop:
+                if prompts.shape[1] == y.shape[1]:
+                    y = torch.concat([y, torch.zeros_like(samples)], dim=1)
+                break
+            y = torch.concat([y, samples], dim=1)
+            cache["first_infer"] = 0
+        return y, idx

AR/models/utils.py

@@ -0,0 +1,160 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/utils.py\
+import torch
+import torch.nn.functional as F
+
+
+def sequence_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def make_pad_mask(lengths: torch.Tensor, max_len: int = 0) -> torch.Tensor:
+    """
+    Args:
+      lengths:
+        A 1-D tensor containing sentence lengths.
+      max_len:
+        The length of masks.
+    Returns:
+      Return a 2-D bool tensor, where masked positions
+      are filled with `True` and non-masked positions are
+      filled with `False`.
+
+    #>>> lengths = torch.tensor([1, 3, 2, 5])
+    #>>> make_pad_mask(lengths)
+    tensor([[False,  True,  True,  True,  True],
+            [False, False, False,  True,  True],
+            [False, False,  True,  True,  True],
+            [False, False, False, False, False]])
+    """
+    assert lengths.ndim == 1, lengths.ndim
+    max_len = max(max_len, lengths.max())
+    n = lengths.size(0)
+    seq_range = torch.arange(0, max_len, device=lengths.device)
+    expaned_lengths = seq_range.unsqueeze(0).expand(n, max_len)
+
+    return expaned_lengths >= lengths.unsqueeze(-1)
+
+
+# https://github.com/microsoft/unilm/blob/master/xtune/src/transformers/modeling_utils.py
+def top_k_top_p_filtering(
+    logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1
+):
+    """Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+    Args:
+        logits: logits distribution shape (batch size, vocabulary size)
+        if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+        if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+            Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+        Make sure we keep at least min_tokens_to_keep per batch example in the output
+    From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    if top_k > 0:
+        top_k = min(max(top_k, min_tokens_to_keep), logits.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > top_p
+        if min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., :min_tokens_to_keep] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(
+            1, sorted_indices, sorted_indices_to_remove
+        )
+        logits[indices_to_remove] = filter_value
+    return logits
+
+
+def topk_sampling(logits, top_k=10, top_p=1.0, temperature=1.0):
+    # temperature: (`optional`) float
+    #     The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.
+    # top_k: (`optional`) int
+    #     The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
+    # top_p: (`optional`) float
+    #     The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.
+
+    # Temperature (higher temperature => more likely to sample low probability tokens)
+    if temperature != 1.0:
+        logits = logits / temperature
+    # Top-p/top-k filtering
+    logits = top_k_top_p_filtering(logits, top_k=top_k, top_p=top_p)
+    # Sample
+    token = torch.multinomial(F.softmax(logits, dim=-1), num_samples=1)
+    return token
+
+
+from typing import Optional, Tuple
+
+
+def multinomial_sample_one_no_sync(
+    probs_sort,
+):  # Does multinomial sampling without a cuda synchronization
+    q = torch.empty_like(probs_sort).exponential_(1)
+    return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)
+
+
+def logits_to_probs(
+    logits,
+    previous_tokens: Optional[torch.Tensor] = None,
+    temperature: float = 1.0,
+    top_k: Optional[int] = None,
+    top_p: Optional[int] = None,
+    repetition_penalty: float = 1.0,
+):
+    previous_tokens = previous_tokens.squeeze()
+    # print(logits.shape,previous_tokens.shape)
+    # pdb.set_trace()
+    if previous_tokens is not None and repetition_penalty != 1.0:
+        previous_tokens = previous_tokens.long()
+        score = torch.gather(logits, dim=0, index=previous_tokens)
+        score = torch.where(
+            score < 0, score * repetition_penalty, score / repetition_penalty
+        )
+        logits.scatter_(dim=0, index=previous_tokens, src=score)
+
+    if top_p is not None and top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cum_probs = torch.cumsum(
+            torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1
+        )
+        sorted_indices_to_remove = cum_probs > top_p
+        sorted_indices_to_remove[0] = False  # keep at least one option
+        indices_to_remove = sorted_indices_to_remove.scatter(
+            dim=0, index=sorted_indices, src=sorted_indices_to_remove
+        )
+        logits = logits.masked_fill(indices_to_remove, -float("Inf"))
+
+    logits = logits / max(temperature, 1e-5)
+
+    if top_k is not None:
+        v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+        pivot = v.select(-1, -1).unsqueeze(-1)
+        logits = torch.where(logits < pivot, -float("Inf"), logits)
+
+    probs = torch.nn.functional.softmax(logits, dim=-1)
+    return probs
+
+
+def sample(
+    logits,
+    previous_tokens: Optional[torch.Tensor] = None,
+    **sampling_kwargs,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    probs = logits_to_probs(
+        logits=logits, previous_tokens=previous_tokens, **sampling_kwargs
+    )
+    idx_next = multinomial_sample_one_no_sync(probs)
+    return idx_next, probs

AR/modules/activation.py

@@ -0,0 +1,428 @@
+# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/activation.py
+from typing import Optional
+from typing import Tuple
+import torch
+from torch import Tensor
+from torch.nn import Linear
+from torch.nn import Module
+from torch.nn.init import constant_
+from torch.nn.init import xavier_normal_
+from torch.nn.init import xavier_uniform_
+from torch.nn.modules.linear import NonDynamicallyQuantizableLinear
+from torch.nn.parameter import Parameter
+
+from torch.nn import functional as F
+from AR.modules.patched_mha_with_cache import multi_head_attention_forward_patched
+
+F.multi_head_attention_forward = multi_head_attention_forward_patched
+
+
+class MultiheadAttention(Module):
+    r"""Allows the model to jointly attend to information
+    from different representation subspaces as described in the paper:
+    `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_.
+
+    Multi-Head Attention is defined as:
+
+    .. math::
+        \text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
+
+    where :math:`head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)`.
+
+    ``forward()`` will use a special optimized implementation if all of the following
+    conditions are met:
+
+    - self attention is being computed (i.e., ``query``, ``key``, and ``value`` are the same tensor. This
+      restriction will be loosened in the future.)
+    - Either autograd is disabled (using ``torch.inference_mode`` or ``torch.no_grad``) or no tensor argument ``requires_grad``
+    - training is disabled (using ``.eval()``)
+    - dropout is 0
+    - ``add_bias_kv`` is ``False``
+    - ``add_zero_attn`` is ``False``
+    - ``batch_first`` is ``True`` and the input is batched
+    - ``kdim`` and ``vdim`` are equal to ``embed_dim``
+    - at most one of ``key_padding_mask`` or ``attn_mask`` is passed
+    - if a `NestedTensor <https://pytorch.org/docs/stable/nested.html>`_ is passed, neither ``key_padding_mask``
+      nor ``attn_mask`` is passed
+
+    If the optimized implementation is in use, a
+    `NestedTensor <https://pytorch.org/docs/stable/nested.html>`_ can be passed for
+    ``query``/``key``/``value`` to represent padding more efficiently than using a
+    padding mask. In this case, a `NestedTensor <https://pytorch.org/docs/stable/nested.html>`_
+    will be returned, and an additional speedup proportional to the fraction of the input
+    that is padding can be expected.
+
+    Args:
+        embed_dim: Total dimension of the model.
+        num_heads: Number of parallel attention heads. Note that ``embed_dim`` will be split
+            across ``num_heads`` (i.e. each head will have dimension ``embed_dim // num_heads``).
+        dropout: Dropout probability on ``attn_output_weights``. Default: ``0.0`` (no dropout).
+        bias: If specified, adds bias to input / output projection layers. Default: ``True``.
+        add_bias_kv: If specified, adds bias to the key and value sequences at dim=0. Default: ``False``.
+        add_zero_attn: If specified, adds a new batch of zeros to the key and value sequences at dim=1.
+            Default: ``False``.
+        kdim: Total number of features for keys. Default: ``None`` (uses ``kdim=embed_dim``).
+        vdim: Total number of features for values. Default: ``None`` (uses ``vdim=embed_dim``).
+        batch_first: If ``True``, then the input and output tensors are provided
+            as (batch, seq, feature). Default: ``False`` (seq, batch, feature).
+
+    Examples::
+
+        >>> # xdoctest: +SKIP
+        >>> multihead_attn = nn.MultiheadAttention(embed_dim, num_heads)
+        >>> attn_output, attn_output_weights = multihead_attn(query, key, value)
+
+    """
+    __constants__ = ["batch_first"]
+    bias_k: Optional[torch.Tensor]
+    bias_v: Optional[torch.Tensor]
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        dropout=0.0,
+        bias=True,
+        add_bias_kv=False,
+        add_zero_attn=False,
+        kdim=None,
+        vdim=None,
+        batch_first=False,
+        linear1_cls=Linear,
+        linear2_cls=Linear,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(MultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.batch_first = batch_first
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        if add_bias_kv:
+            self.bias_k = Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
+            self.bias_v = Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
+        else:
+            self.bias_k = self.bias_v = None
+
+        if linear1_cls == Linear:
+            if not self._qkv_same_embed_dim:
+                self.q_proj_weight = Parameter(
+                    torch.empty((embed_dim, embed_dim), **factory_kwargs)
+                )
+                self.k_proj_weight = Parameter(
+                    torch.empty((embed_dim, self.kdim), **factory_kwargs)
+                )
+                self.v_proj_weight = Parameter(
+                    torch.empty((embed_dim, self.vdim), **factory_kwargs)
+                )
+                self.register_parameter("in_proj_weight", None)
+            else:
+                self.in_proj_weight = Parameter(
+                    torch.empty((3 * embed_dim, embed_dim), **factory_kwargs)
+                )
+                self.register_parameter("q_proj_weight", None)
+                self.register_parameter("k_proj_weight", None)
+                self.register_parameter("v_proj_weight", None)
+
+            if bias:
+                self.in_proj_bias = Parameter(
+                    torch.empty(3 * embed_dim, **factory_kwargs)
+                )
+            else:
+                self.register_parameter("in_proj_bias", None)
+            self.out_proj = NonDynamicallyQuantizableLinear(
+                embed_dim, embed_dim, bias=bias, **factory_kwargs
+            )
+
+            self._reset_parameters()
+        else:
+            if not self._qkv_same_embed_dim:
+                raise NotImplementedError
+            else:
+                self.in_proj_linear = linear1_cls(
+                    embed_dim, 3 * embed_dim, bias=bias, **factory_kwargs
+                )
+                self.in_proj_weight = self.in_proj_linear.weight
+
+                self.register_parameter("q_proj_weight", None)
+                self.register_parameter("k_proj_weight", None)
+                self.register_parameter("v_proj_weight", None)
+
+                if bias:
+                    self.in_proj_bias = self.in_proj_linear.bias
+                else:
+                    self.register_parameter("in_proj_bias", None)
+
+            self.out_proj = linear2_cls(
+                embed_dim, embed_dim, bias=bias, **factory_kwargs
+            )
+
+            if self.bias_k is not None:
+                xavier_normal_(self.bias_k)
+            if self.bias_v is not None:
+                xavier_normal_(self.bias_v)
+
+        self.add_zero_attn = add_zero_attn
+
+    def _reset_parameters(self):
+        if self._qkv_same_embed_dim:
+            xavier_uniform_(self.in_proj_weight)
+        else:
+            xavier_uniform_(self.q_proj_weight)
+            xavier_uniform_(self.k_proj_weight)
+            xavier_uniform_(self.v_proj_weight)
+
+        if self.in_proj_bias is not None:
+            constant_(self.in_proj_bias, 0.0)
+            constant_(self.out_proj.bias, 0.0)
+
+        if self.bias_k is not None:
+            xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            xavier_normal_(self.bias_v)
+
+    def __setstate__(self, state):
+        # Support loading old MultiheadAttention checkpoints generated by v1.1.0
+        if "_qkv_same_embed_dim" not in state:
+            state["_qkv_same_embed_dim"] = True
+
+        super(MultiheadAttention, self).__setstate__(state)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        average_attn_weights: bool = True,
+        cache=None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        r"""
+        Args:
+            query: Query embeddings of shape :math:`(L, E_q)` for unbatched input, :math:`(L, N, E_q)` when ``batch_first=False``
+                or :math:`(N, L, E_q)` when ``batch_first=True``, where :math:`L` is the target sequence length,
+                :math:`N` is the batch size, and :math:`E_q` is the query embedding dimension ``embed_dim``.
+                Queries are compared against key-value pairs to produce the output.
+                See "Attention Is All You Need" for more details.
+            key: Key embeddings of shape :math:`(S, E_k)` for unbatched input, :math:`(S, N, E_k)` when ``batch_first=False``
+                or :math:`(N, S, E_k)` when ``batch_first=True``, where :math:`S` is the source sequence length,
+                :math:`N` is the batch size, and :math:`E_k` is the key embedding dimension ``kdim``.
+                See "Attention Is All You Need" for more details.
+            value: Value embeddings of shape :math:`(S, E_v)` for unbatched input, :math:`(S, N, E_v)` when
+                ``batch_first=False`` or :math:`(N, S, E_v)` when ``batch_first=True``, where :math:`S` is the source
+                sequence length, :math:`N` is the batch size, and :math:`E_v` is the value embedding dimension ``vdim``.
+                See "Attention Is All You Need" for more details.
+            key_padding_mask: If specified, a mask of shape :math:`(N, S)` indicating which elements within ``key``
+                to ignore for the purpose of attention (i.e. treat as "padding"). For unbatched `query`, shape should be :math:`(S)`.
+                Binary and byte masks are supported.
+                For a binary mask, a ``True`` value indicates that the corresponding ``key`` value will be ignored for
+                the purpose of attention. For a float mask, it will be directly added to the corresponding ``key`` value.
+            need_weights: If specified, returns ``attn_output_weights`` in addition to ``attn_outputs``.
+                Default: ``True``.
+            attn_mask: If specified, a 2D or 3D mask preventing attention to certain positions. Must be of shape
+                :math:`(L, S)` or :math:`(N\cdot\text{num\_heads}, L, S)`, where :math:`N` is the batch size,
+                :math:`L` is the target sequence length, and :math:`S` is the source sequence length. A 2D mask will be
+                broadcasted across the batch while a 3D mask allows for a different mask for each entry in the batch.
+                Binary, byte, and float masks are supported. For a binary mask, a ``True`` value indicates that the
+                corresponding position is not allowed to attend. For a byte mask, a non-zero value indicates that the
+                corresponding position is not allowed to attend. For a float mask, the mask values will be added to
+                the attention weight.
+            average_attn_weights: If true, indicates that the returned ``attn_weights`` should be averaged across
+                heads. Otherwise, ``attn_weights`` are provided separately per head. Note that this flag only has an
+                effect when ``need_weights=True``. Default: ``True`` (i.e. average weights across heads)
+
+        Outputs:
+            - **attn_output** - Attention outputs of shape :math:`(L, E)` when input is unbatched,
+              :math:`(L, N, E)` when ``batch_first=False`` or :math:`(N, L, E)` when ``batch_first=True``,
+              where :math:`L` is the target sequence length, :math:`N` is the batch size, and :math:`E` is the
+              embedding dimension ``embed_dim``.
+            - **attn_output_weights** - Only returned when ``need_weights=True``. If ``average_attn_weights=True``,
+              returns attention weights averaged across heads of shape :math:`(L, S)` when input is unbatched or
+              :math:`(N, L, S)`, where :math:`N` is the batch size, :math:`L` is the target sequence length, and
+              :math:`S` is the source sequence length. If ``average_attn_weights=False``, returns attention weights per
+              head of shape :math:`(\text{num\_heads}, L, S)` when input is unbatched or :math:`(N, \text{num\_heads}, L, S)`.
+
+            .. note::
+                `batch_first` argument is ignored for unbatched inputs.
+        """
+        is_batched = query.dim() == 3
+        if key_padding_mask is not None:
+            _kpm_dtype = key_padding_mask.dtype
+            if _kpm_dtype != torch.bool and not torch.is_floating_point(
+                key_padding_mask
+            ):
+                raise AssertionError(
+                    "only bool and floating types of key_padding_mask are supported"
+                )
+        why_not_fast_path = ""
+        if not is_batched:
+            why_not_fast_path = (
+                f"input not batched; expected query.dim() of 3 but got {query.dim()}"
+            )
+        elif query is not key or key is not value:
+            # When lifting this restriction, don't forget to either
+            # enforce that the dtypes all match or test cases where
+            # they don't!
+            why_not_fast_path = "non-self attention was used (query, key, and value are not the same Tensor)"
+        elif self.in_proj_bias is not None and query.dtype != self.in_proj_bias.dtype:
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
+        elif (
+            self.in_proj_weight is not None and query.dtype != self.in_proj_weight.dtype
+        ):
+            # this case will fail anyway, but at least they'll get a useful error message.
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
+        elif self.training:
+            why_not_fast_path = "training is enabled"
+        elif not self.batch_first:
+            why_not_fast_path = "batch_first was not True"
+        elif self.bias_k is not None:
+            why_not_fast_path = "self.bias_k was not None"
+        elif self.bias_v is not None:
+            why_not_fast_path = "self.bias_v was not None"
+        elif self.dropout:
+            why_not_fast_path = f"dropout was {self.dropout}, required zero"
+        elif self.add_zero_attn:
+            why_not_fast_path = "add_zero_attn was enabled"
+        elif not self._qkv_same_embed_dim:
+            why_not_fast_path = "_qkv_same_embed_dim was not True"
+        elif attn_mask is not None:
+            why_not_fast_path = "attn_mask was not None"
+        elif query.is_nested and key_padding_mask is not None:
+            why_not_fast_path = (
+                "key_padding_mask is not supported with NestedTensor input"
+            )
+        elif self.num_heads % 2 == 1:
+            why_not_fast_path = "num_heads is odd"
+        elif torch.is_autocast_enabled():
+            why_not_fast_path = "autocast is enabled"
+
+        if not why_not_fast_path:
+            tensor_args = (
+                query,
+                key,
+                value,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.out_proj.weight,
+                self.out_proj.bias,
+            )
+            # We have to use list comprehensions below because TorchScript does not support
+            # generator expressions.
+            if torch.overrides.has_torch_function(tensor_args):
+                why_not_fast_path = "some Tensor argument has_torch_function"
+            elif not all(
+                [
+                    (x is None or x.is_cuda or "cpu" in str(x.device))
+                    for x in tensor_args
+                ]
+            ):
+                why_not_fast_path = "some Tensor argument is neither CUDA nor CPU"
+            elif torch.is_grad_enabled() and any(
+                [x is not None and x.requires_grad for x in tensor_args]
+            ):
+                why_not_fast_path = (
+                    "grad is enabled and at least one of query or the "
+                    "input/output projection weights or biases requires_grad"
+                )
+            if not why_not_fast_path:
+                return torch._native_multi_head_attention(
+                    query,
+                    key,
+                    value,
+                    self.embed_dim,
+                    self.num_heads,
+                    self.in_proj_weight,
+                    self.in_proj_bias,
+                    self.out_proj.weight,
+                    self.out_proj.bias,
+                    key_padding_mask if key_padding_mask is not None else attn_mask,
+                    need_weights,
+                    average_attn_weights,
+                    1
+                    if key_padding_mask is not None
+                    else 0
+                    if attn_mask is not None
+                    else None,
+                )
+
+        any_nested = query.is_nested or key.is_nested or value.is_nested
+        assert not any_nested, (
+            "MultiheadAttention does not support NestedTensor outside of its fast path. "
+            + f"The fast path was not hit because {why_not_fast_path}"
+        )
+
+        if self.batch_first and is_batched:
+            # make sure that the transpose op does not affect the "is" property
+            if key is value:
+                if query is key:
+                    query = key = value = query.transpose(1, 0)
+                else:
+                    query, key = [x.transpose(1, 0) for x in (query, key)]
+                    value = key
+            else:
+                query, key, value = [x.transpose(1, 0) for x in (query, key, value)]
+
+        if not self._qkv_same_embed_dim:
+            attn_output, attn_output_weights = F.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout,
+                self.out_proj.weight,
+                self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                use_separate_proj_weight=True,
+                q_proj_weight=self.q_proj_weight,
+                k_proj_weight=self.k_proj_weight,
+                v_proj_weight=self.v_proj_weight,
+                average_attn_weights=average_attn_weights,
+                cache=cache,
+            )
+        else:
+            attn_output, attn_output_weights = F.multi_head_attention_forward(
+                query,
+                key,
+                value,
+                self.embed_dim,
+                self.num_heads,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.bias_k,
+                self.bias_v,
+                self.add_zero_attn,
+                self.dropout,
+                self.out_proj.weight,
+                self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                average_attn_weights=average_attn_weights,
+                cache=cache,
+            )
+        if self.batch_first and is_batched:
+            return attn_output.transpose(1, 0), attn_output_weights
+        else:
+            return attn_output, attn_output_weights

AR/modules/activation_onnx.py

@@ -0,0 +1,178 @@
+# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/activation.py
+from typing import Optional
+from typing import Tuple
+import torch
+from torch import Tensor
+from torch.nn import Linear
+from torch.nn import Module
+from torch.nn.init import constant_
+from torch.nn.init import xavier_normal_
+from torch.nn.init import xavier_uniform_
+from torch.nn.modules.linear import NonDynamicallyQuantizableLinear
+from torch.nn.parameter import Parameter
+
+from torch.nn import functional as F
+from AR.modules.patched_mha_with_cache_onnx import multi_head_attention_forward_patched
+
+
+class MultiheadAttention(Module):
+    __constants__ = ["batch_first"]
+    bias_k: Optional[torch.Tensor]
+    bias_v: Optional[torch.Tensor]
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        dropout=0.0,
+        bias=True,
+        add_bias_kv=False,
+        add_zero_attn=False,
+        kdim=None,
+        vdim=None,
+        batch_first=False,
+        linear1_cls=Linear,
+        linear2_cls=Linear,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(MultiheadAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.kdim = kdim if kdim is not None else embed_dim
+        self.vdim = vdim if vdim is not None else embed_dim
+        self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
+
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.batch_first = batch_first
+        self.head_dim = embed_dim // num_heads
+        assert (
+            self.head_dim * num_heads == self.embed_dim
+        ), "embed_dim must be divisible by num_heads"
+
+        if add_bias_kv:
+            self.bias_k = Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
+            self.bias_v = Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
+        else:
+            self.bias_k = self.bias_v = None
+
+        if linear1_cls == Linear:
+            if not self._qkv_same_embed_dim:
+                self.q_proj_weight = Parameter(
+                    torch.empty((embed_dim, embed_dim), **factory_kwargs)
+                )
+                self.k_proj_weight = Parameter(
+                    torch.empty((embed_dim, self.kdim), **factory_kwargs)
+                )
+                self.v_proj_weight = Parameter(
+                    torch.empty((embed_dim, self.vdim), **factory_kwargs)
+                )
+                self.register_parameter("in_proj_weight", None)
+            else:
+                self.in_proj_weight = Parameter(
+                    torch.empty((3 * embed_dim, embed_dim), **factory_kwargs)
+                )
+                self.register_parameter("q_proj_weight", None)
+                self.register_parameter("k_proj_weight", None)
+                self.register_parameter("v_proj_weight", None)
+
+            if bias:
+                self.in_proj_bias = Parameter(
+                    torch.empty(3 * embed_dim, **factory_kwargs)
+                )
+            else:
+                self.register_parameter("in_proj_bias", None)
+            self.out_proj = NonDynamicallyQuantizableLinear(
+                embed_dim, embed_dim, bias=bias, **factory_kwargs
+            )
+
+            self._reset_parameters()
+        else:
+            if not self._qkv_same_embed_dim:
+                raise NotImplementedError
+            else:
+                self.in_proj_linear = linear1_cls(
+                    embed_dim, 3 * embed_dim, bias=bias, **factory_kwargs
+                )
+                self.in_proj_weight = self.in_proj_linear.weight
+
+                self.register_parameter("q_proj_weight", None)
+                self.register_parameter("k_proj_weight", None)
+                self.register_parameter("v_proj_weight", None)
+
+                if bias:
+                    self.in_proj_bias = self.in_proj_linear.bias
+                else:
+                    self.register_parameter("in_proj_bias", None)
+
+            self.out_proj = linear2_cls(
+                embed_dim, embed_dim, bias=bias, **factory_kwargs
+            )
+
+            if self.bias_k is not None:
+                xavier_normal_(self.bias_k)
+            if self.bias_v is not None:
+                xavier_normal_(self.bias_v)
+
+        self.add_zero_attn = add_zero_attn
+
+    def _reset_parameters(self):
+        if self._qkv_same_embed_dim:
+            xavier_uniform_(self.in_proj_weight)
+        else:
+            xavier_uniform_(self.q_proj_weight)
+            xavier_uniform_(self.k_proj_weight)
+            xavier_uniform_(self.v_proj_weight)
+
+        if self.in_proj_bias is not None:
+            constant_(self.in_proj_bias, 0.0)
+            constant_(self.out_proj.bias, 0.0)
+
+        if self.bias_k is not None:
+            xavier_normal_(self.bias_k)
+        if self.bias_v is not None:
+            xavier_normal_(self.bias_v)
+
+    def __setstate__(self, state):
+        # Support loading old MultiheadAttention checkpoints generated by v1.1.0
+        if "_qkv_same_embed_dim" not in state:
+            state["_qkv_same_embed_dim"] = True
+
+        super(MultiheadAttention, self).__setstate__(state)
+
+    def forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        average_attn_weights: bool = True,
+        cache=None,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        any_nested = query.is_nested or key.is_nested or value.is_nested
+        query = key = value = query.transpose(1, 0)
+        attn_output = multi_head_attention_forward_patched(
+            query,
+            key,
+            value,
+            self.embed_dim,
+            self.num_heads,
+            self.in_proj_weight,
+            self.in_proj_bias,
+            self.bias_k,
+            self.bias_v,
+            self.add_zero_attn,
+            self.dropout,
+            self.out_proj.weight,
+            self.out_proj.bias,
+            training=self.training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            average_attn_weights=average_attn_weights,
+            cache=cache,
+        )
+        return attn_output.transpose(1, 0)

AR/modules/embedding.py

@@ -0,0 +1,81 @@
+# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/embedding.py
+import math
+
+import torch
+from torch import nn
+
+
+class TokenEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        vocab_size: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+
+        self.vocab_size = vocab_size
+        self.embedding_dim = embedding_dim
+
+        self.dropout = torch.nn.Dropout(p=dropout)
+        self.word_embeddings = nn.Embedding(self.vocab_size, self.embedding_dim)
+
+    @property
+    def weight(self) -> torch.Tensor:
+        return self.word_embeddings.weight
+
+    def embedding(self, index: int) -> torch.Tensor:
+        return self.word_embeddings.weight[index : index + 1]
+
+    def forward(self, x: torch.Tensor):
+        x = self.word_embeddings(x)
+        x = self.dropout(x)
+        return x
+
+
+class SinePositionalEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        dropout: float = 0.0,
+        scale: bool = False,
+        alpha: bool = False,
+    ):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.x_scale = math.sqrt(embedding_dim) if scale else 1.0
+        self.alpha = nn.Parameter(torch.ones(1), requires_grad=alpha)
+        self.dropout = torch.nn.Dropout(p=dropout)
+
+        self.reverse = False
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, 4000))
+
+    def extend_pe(self, x):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.embedding_dim)
+        if self.reverse:
+            position = torch.arange(
+                x.size(1) - 1, -1, -1.0, dtype=torch.float32
+            ).unsqueeze(1)
+        else:
+            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.embedding_dim, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.embedding_dim)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        self.pe = pe.to(device=x.device, dtype=x.dtype).detach()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        self.extend_pe(x)
+        output = x.unsqueeze(-1) if x.ndim == 2 else x
+        output = output * self.x_scale + self.alpha * self.pe[:, : x.size(1)]
+        return self.dropout(output)

AR/modules/embedding_onnx.py

@@ -0,0 +1,63 @@
+# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/embedding.py
+import math
+
+import torch
+from torch import nn
+
+
+class TokenEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        vocab_size: int,
+        dropout: float = 0.0,
+    ):
+        super().__init__()
+
+        self.vocab_size = vocab_size
+        self.embedding_dim = embedding_dim
+
+        self.dropout = torch.nn.Dropout(p=dropout)
+        self.word_embeddings = nn.Embedding(self.vocab_size, self.embedding_dim)
+
+    @property
+    def weight(self) -> torch.Tensor:
+        return self.word_embeddings.weight
+
+    def embedding(self, index: int) -> torch.Tensor:
+        return self.word_embeddings.weight[index : index + 1]
+
+    def forward(self, x: torch.Tensor):
+        x = self.word_embeddings(x)
+        x = self.dropout(x)
+        return x
+
+
+class SinePositionalEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        dropout: float = 0.0,
+        scale: bool = False,
+        alpha: bool = False,
+    ):
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.x_scale = math.sqrt(embedding_dim) if scale else 1.0
+        self.alpha = nn.Parameter(torch.ones(1), requires_grad=alpha)
+        self.dropout = torch.nn.Dropout(p=dropout)
+        self.reverse = False
+        self.div_term = torch.exp(torch.arange(0, self.embedding_dim, 2) * -(math.log(10000.0) / self.embedding_dim))
+
+    def extend_pe(self, x):
+        position = torch.cumsum(torch.ones_like(x[:,:,0]), dim=1).transpose(0, 1)
+        scpe = (position * self.div_term).unsqueeze(0)
+        pe = torch.cat([torch.sin(scpe), torch.cos(scpe)]).permute(1, 2, 0)
+        pe = pe.contiguous().view(1, -1, self.embedding_dim)
+        return pe
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pe = self.extend_pe(x)
+        output = x.unsqueeze(-1) if x.ndim == 2 else x
+        output = output * self.x_scale + self.alpha * pe
+        return self.dropout(output)

AR/modules/lr_schedulers.py

@@ -0,0 +1,82 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/model/lr_schedulers.py
+import math
+
+import torch
+from matplotlib import pyplot as plt
+from torch import nn
+from torch.optim import Adam
+
+
+class WarmupCosineLRSchedule(torch.optim.lr_scheduler._LRScheduler):
+    """
+    Implements Warmup learning rate schedule until 'warmup_steps', going from 'init_lr' to 'peak_lr' for multiple optimizers.
+    """
+
+    def __init__(
+        self,
+        optimizer,
+        init_lr,
+        peak_lr,
+        end_lr,
+        warmup_steps=10000,
+        total_steps=400000,
+        current_step=0,
+    ):
+        self.init_lr = init_lr
+        self.peak_lr = peak_lr
+        self.end_lr = end_lr
+        self.optimizer = optimizer
+        self._warmup_rate = (peak_lr - init_lr) / warmup_steps
+        self._decay_rate = (end_lr - peak_lr) / (total_steps - warmup_steps)
+        self._current_step = current_step
+        self.lr = init_lr
+        self.warmup_steps = warmup_steps
+        self.total_steps = total_steps
+        self._last_lr = [self.lr]
+
+    def set_lr(self, lr):
+        self._last_lr = [g["lr"] for g in self.optimizer.param_groups]
+        for g in self.optimizer.param_groups:
+            # g['lr'] = lr
+            g["lr"] = self.end_lr  ###锁定用线性
+
+    def step(self):
+        if self._current_step < self.warmup_steps:
+            lr = self.init_lr + self._warmup_rate * self._current_step
+
+        elif self._current_step > self.total_steps:
+            lr = self.end_lr
+
+        else:
+            decay_ratio = (self._current_step - self.warmup_steps) / (
+                self.total_steps - self.warmup_steps
+            )
+            if decay_ratio < 0.0 or decay_ratio > 1.0:
+                raise RuntimeError(
+                    "Decay ratio must be in [0.0, 1.0]. Fix LR scheduler settings."
+                )
+            coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
+            lr = self.end_lr + coeff * (self.peak_lr - self.end_lr)
+
+        self.lr = lr = self.end_lr = 0.002  ###锁定用线性###不听话，直接锁定！
+        self.set_lr(lr)
+        self.lr = lr
+        self._current_step += 1
+        return self.lr
+
+
+if __name__ == "__main__":
+    m = nn.Linear(10, 10)
+    opt = Adam(m.parameters(), lr=1e-4)
+    s = WarmupCosineLRSchedule(
+        opt, 1e-6, 2e-4, 1e-6, warmup_steps=2000, total_steps=20000, current_step=0
+    )
+    lrs = []
+    for i in range(25000):
+        s.step()
+        lrs.append(s.lr)
+        print(s.lr)
+
+    plt.plot(lrs)
+    plt.plot(range(0, 25000), lrs)
+    plt.show()

AR/modules/optim.py

@@ -0,0 +1,622 @@
+# Copyright      2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import contextlib
+import logging
+from collections import defaultdict
+from typing import List
+from typing import Tuple
+
+import torch
+from torch import Tensor
+from torch.optim import Optimizer
+
+
+class BatchedOptimizer(Optimizer):
+    """
+    This class adds to class Optimizer the capability to optimize parameters in batches:
+    it will stack the parameters and their grads for you so the optimizer can work
+    on tensors with an extra leading dimension.  This is intended for speed with GPUs,
+    as it reduces the number of kernels launched in the optimizer.
+
+    Args:
+      params:
+    """
+
+    def __init__(self, params, defaults):
+        super(BatchedOptimizer, self).__init__(params, defaults)
+
+    @contextlib.contextmanager
+    def batched_params(self, param_group, group_params_names):
+        """
+        This function returns (technically, yields) a list of
+          of tuples (p, state), where
+        p is a `fake` parameter that is stacked (over axis 0) from real parameters
+        that share the same shape, and its gradient is also stacked;
+        `state` is the state corresponding to this batch of parameters
+        (it will be physically located in the "state" for one of the real
+        parameters, the last one that has any particular shape and dtype).
+
+        This function is decorated as a context manager so that it can
+        write parameters back to their "real" locations.
+
+        The idea is, instead of doing:
+        <code>
+          for p in group["params"]:
+             state = self.state[p]
+             ...
+        </code>
+        you can do:
+        <code>
+          with self.batched_params(group["params"]) as batches:
+             for p, state, p_names in batches:
+                 ...
+        </code>
+
+        Args:
+          group: a parameter group, which is a list of parameters; should be
+                one of self.param_groups.
+          group_params_names: name for each parameter in group,
+                which is List[str].
+        """
+        batches = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
+        batches_names = defaultdict(
+            list
+        )  # `batches` maps from tuple (dtype_as_str,*shape) to list of str
+
+        assert len(param_group) == len(group_params_names)
+        for p, named_p in zip(param_group, group_params_names):
+            key = (str(p.dtype), *p.shape)
+            batches[key].append(p)
+            batches_names[key].append(named_p)
+
+        batches_names_keys = list(batches_names.keys())
+        sorted_idx = sorted(
+            range(len(batches_names)), key=lambda i: batches_names_keys[i])
+        batches_names = [
+            batches_names[batches_names_keys[idx]] for idx in sorted_idx
+        ]
+        batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
+
+        stacked_params_dict = dict()
+
+        # turn batches into a list, in deterministic order.
+        # tuples will contain tuples of (stacked_param, state, stacked_params_names),
+        # one for each batch in `batches`.
+        tuples = []
+
+        for batch, batch_names in zip(batches, batches_names):
+            p = batch[0]
+            # we arbitrarily store the state in the
+            # state corresponding to the 1st parameter in the
+            # group.  class Optimizer will take care of saving/loading state.
+            state = self.state[p]
+            p_stacked = torch.stack(batch)
+            grad = torch.stack([
+                torch.zeros_like(p) if p.grad is None else p.grad for p in batch
+            ])
+            p_stacked.grad = grad
+            stacked_params_dict[key] = p_stacked
+            tuples.append((p_stacked, state, batch_names))
+
+        yield tuples  # <-- calling code will do the actual optimization here!
+
+        for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
+            for i, p in enumerate(batch):  # batch is list of Parameter
+                p.copy_(stacked_params[i])
+
+
+class ScaledAdam(BatchedOptimizer):
+    """
+     Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update
+     proportional to the norm of that parameter; and also learn the scale of the parameter,
+     in log space, subject to upper and lower limits (as if we had factored each parameter as
+     param = underlying_param * log_scale.exp())
+
+
+     Args:
+          params:  The parameters or param_groups to optimize (like other Optimizer subclasses)
+              lr:  The learning rate.  We will typically use a learning rate schedule that starts
+                   at 0.03 and decreases over time, i.e. much higher than other common
+                   optimizers.
+     clipping_scale: (e.g. 2.0)
+                   A scale for gradient-clipping: if specified, the normalized gradients
+                   over the whole model will be clipped to have 2-norm equal to
+                   `clipping_scale` times the median 2-norm over the most recent period
+                   of `clipping_update_period` minibatches.  By "normalized gradients",
+                   we mean after multiplying by the rms parameter value for this tensor
+                   [for non-scalars]; this is appropriate because our update is scaled
+                   by this quantity.
+            betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad.
+                   Must satisfy 0 < beta <= beta2 < 1.
+     scalar_lr_scale: A scaling factor on the learning rate, that we use to update the
+                   scale of each parameter tensor and scalar parameters of the mode..
+                   If each parameter were decomposed
+                   as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale
+                   would be a the scaling factor on the learning rate of p_scale.
+              eps:  A general-purpose epsilon to prevent division by zero
+    param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be >= this value)
+    param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of
+                   learning the scale on the parameters (we'll constrain the rms of each non-scalar
+                   parameter tensor to be <= this value)
+       scalar_max: Maximum absolute value for scalar parameters (applicable if your
+                   model has any parameters with numel() == 1).
+    size_update_period: The periodicity, in steps, with which we update the size (scale)
+                   of the parameter tensor.  This is provided to save a little time
+                   in the update.
+     clipping_update_period: if clipping_scale is specified, this is the period
+    """
+
+    def __init__(
+            self,
+            params,
+            lr=3e-02,
+            clipping_scale=None,
+            betas=(0.9, 0.98),
+            scalar_lr_scale=0.1,
+            eps=1.0e-08,
+            param_min_rms=1.0e-05,
+            param_max_rms=3.0,
+            scalar_max=10.0,
+            size_update_period=4,
+            clipping_update_period=100,
+            parameters_names=None,
+            show_dominant_parameters=True, ):
+
+        assert parameters_names is not None, (
+            "Please prepare parameters_names,"
+            "which is a List[List[str]]. Each List[str] is for a group"
+            "and each str is for a parameter")
+        defaults = dict(
+            lr=lr,
+            clipping_scale=clipping_scale,
+            betas=betas,
+            scalar_lr_scale=scalar_lr_scale,
+            eps=eps,
+            param_min_rms=param_min_rms,
+            param_max_rms=param_max_rms,
+            scalar_max=scalar_max,
+            size_update_period=size_update_period,
+            clipping_update_period=clipping_update_period, )
+
+        super(ScaledAdam, self).__init__(params, defaults)
+        assert len(self.param_groups) == len(parameters_names)
+        self.parameters_names = parameters_names
+        self.show_dominant_parameters = show_dominant_parameters
+
+    def __setstate__(self, state):
+        super(ScaledAdam, self).__setstate__(state)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+
+        Arguments:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        batch = True
+
+        for group, group_params_names in zip(self.param_groups,
+                                             self.parameters_names):
+
+            with self.batched_params(group["params"],
+                                     group_params_names) as batches:
+
+                # batches is list of pairs (stacked_param, state).  stacked_param is like
+                # a regular parameter, and will have a .grad, but the 1st dim corresponds to
+                # a stacking dim, it is not a real dim.
+
+                if (len(batches[0][1]) ==
+                        0):  # if len(first state) == 0: not yet initialized
+                    clipping_scale = 1
+                else:
+                    clipping_scale = self._get_clipping_scale(group, batches)
+
+                for p, state, _ in batches:
+                    # Perform optimization step.
+                    # grad is not going to be None, we handled that when creating the batches.
+                    grad = p.grad
+                    if grad.is_sparse:
+                        raise RuntimeError(
+                            "ScaledAdam optimizer does not support sparse gradients"
+                        )
+                    # State initialization
+                    if len(state) == 0:
+                        self._init_state(group, p, state)
+
+                    self._step_one_batch(group, p, state, clipping_scale)
+
+        return loss
+
+    def _init_state(self, group: dict, p: Tensor, state: dict):
+        """
+        Initializes state dict for parameter 'p'.  Assumes that dim 0 of tensor p
+        is actually the batch dimension, corresponding to batched-together
+        parameters of a given shape.
+
+
+        Args:
+           group:   Dict to look up configuration values.
+               p: The parameter that we are initializing the state for
+           state: Dict from string to whatever state we are initializing
+        """
+        size_update_period = group["size_update_period"]
+
+        state["step"] = 0
+
+        kwargs = {"device": p.device, "dtype": p.dtype}
+
+        # 'delta' implements conventional momentum.  There are
+        # several different kinds of update going on, so rather than
+        # compute "exp_avg" like in Adam, we store and decay a
+        # parameter-change "delta", which combines all forms of
+        # update.  this is equivalent to how it's done in Adam,
+        # except for the first few steps.
+        state["delta"] = torch.zeros_like(
+            p, memory_format=torch.preserve_format)
+
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        numel = p.numel()
+
+        if numel > 1:
+            # "param_rms" just periodically records the scalar root-mean-square value of
+            # the parameter tensor.
+            # it has a shape like (batch_size, 1, 1, 1, 1)
+            param_rms = (
+                (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt())
+            state["param_rms"] = param_rms
+
+            state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
+            state["scale_grads"] = torch.zeros(size_update_period,
+                                               *param_rms.shape, **kwargs)
+
+        # exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
+        state["exp_avg_sq"] = torch.zeros_like(
+            p, memory_format=torch.preserve_format)
+
+    def _get_clipping_scale(self,
+                            group: dict,
+                            tuples: List[Tuple[Tensor, dict, List[str]]]
+                            ) -> float:
+        """
+        Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
+        by this amount before applying the rest of the update.
+
+        Args:
+           group: the parameter group, an item in self.param_groups
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+        """
+        assert len(tuples) >= 1
+        clipping_scale = group["clipping_scale"]
+        (first_p, first_state, _) = tuples[0]
+        step = first_state["step"]
+        if clipping_scale is None or step == 0:
+            # no clipping.  return early on step == 0 because the other
+            # parameters' state won't have been initialized yet.
+            return 1.0
+        clipping_update_period = group["clipping_update_period"]
+
+        tot_sumsq = torch.tensor(0.0, device=first_p.device)
+        for (p, state, param_names) in tuples:
+            grad = p.grad
+            if grad.is_sparse:
+                raise RuntimeError(
+                    "ScaledAdam optimizer does not support sparse gradients")
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                tot_sumsq += (grad**2).sum()  # sum() to change shape [1] to []
+            else:
+                tot_sumsq += ((grad * state["param_rms"])**2).sum()
+
+        tot_norm = tot_sumsq.sqrt()
+        if "model_norms" not in first_state:
+            first_state["model_norms"] = torch.zeros(
+                clipping_update_period, device=p.device)
+        first_state["model_norms"][step % clipping_update_period] = tot_norm
+
+        if step % clipping_update_period == 0:
+            # Print some stats.
+            # We don't reach here if step == 0 because we would have returned
+            # above.
+            sorted_norms = first_state["model_norms"].sort()[0].to("cpu")
+            quartiles = []
+            for n in range(0, 5):
+                index = min(
+                    clipping_update_period - 1,
+                    (clipping_update_period // 4) * n, )
+                quartiles.append(sorted_norms[index].item())
+
+            median = quartiles[2]
+            threshold = clipping_scale * median
+            first_state["model_norm_threshold"] = threshold
+            percent_clipped = (first_state["num_clipped"] * 100.0 /
+                               clipping_update_period
+                               if "num_clipped" in first_state else 0.0)
+            first_state["num_clipped"] = 0
+            quartiles = " ".join(["%.3e" % x for x in quartiles])
+            logging.info(
+                f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
+                f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
+            )
+
+        if step < clipping_update_period:
+            return 1.0  # We have not yet estimated a norm to clip to.
+        else:
+            try:
+                model_norm_threshold = first_state["model_norm_threshold"]
+            except KeyError:
+                logging.info(
+                    "Warning: model_norm_threshold not in state: possibly "
+                    "you changed config when restarting, adding clipping_scale option?"
+                )
+                return 1.0
+            ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
+            if ans < 1.0:
+                first_state["num_clipped"] += 1
+            if ans < 0.1:
+                logging.warn(
+                    f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
+                )
+                if self.show_dominant_parameters:
+                    assert p.shape[0] == len(param_names)
+                    self._show_gradient_dominating_parameter(tuples, tot_sumsq)
+            return ans
+
+    def _show_gradient_dominating_parameter(
+            self, tuples: List[Tuple[Tensor, dict, List[str]]],
+            tot_sumsq: Tensor):
+        """
+        Show information of parameter wihch dominanting tot_sumsq.
+
+        Args:
+           tuples: a list of tuples of (param, state, param_names)
+                where param is a batched set of parameters,
+                with a .grad (1st dim is batch dim)
+                and state is the state-dict where optimization parameters are kept.
+                param_names is a List[str] while each str is name for a parameter
+                in batched set of parameters "param".
+            tot_sumsq: sumsq of all parameters. Though it's could be calculated
+                from tuples, we still pass it to save some time.
+        """
+        all_sumsq_orig = {}
+        for (p, state, batch_param_names) in tuples:
+            # p is a stacked batch parameters.
+            batch_grad = p.grad
+            if p.numel() == p.shape[0]:  # a batch of scalars
+                batch_sumsq_orig = batch_grad**2
+                # Dummpy values used by following `zip` statement.
+                batch_rms_orig = torch.ones(p.shape[0])
+            else:
+                batch_rms_orig = state["param_rms"]
+                batch_sumsq_orig = ((batch_grad * batch_rms_orig)**2).sum(
+                    dim=list(range(1, batch_grad.ndim)))
+
+            for name, sumsq_orig, rms, grad in zip(batch_param_names,
+                                                   batch_sumsq_orig,
+                                                   batch_rms_orig, batch_grad):
+
+                proportion_orig = sumsq_orig / tot_sumsq
+                all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
+
+        assert torch.isclose(
+            sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
+            torch.tensor(1.0), )
+        sorted_by_proportion = {
+            k: v
+            for k, v in sorted(
+                all_sumsq_orig.items(),
+                key=lambda item: item[1][0],
+                reverse=True, )
+        }
+        dominant_param_name = next(iter(sorted_by_proportion))
+        (dominant_proportion, dominant_sumsq, dominant_rms,
+         dominant_grad, ) = sorted_by_proportion[dominant_param_name]
+        logging.info(f"Parameter Dominanting tot_sumsq {dominant_param_name}"
+                     f" with proportion {dominant_proportion:.2f},"
+                     f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
+                     f"={dominant_sumsq:.3e},"
+                     f" grad_sumsq = {(dominant_grad**2).sum():.3e},"
+                     f" orig_rms_sq={(dominant_rms**2).item():.3e}")
+
+    def _step_one_batch(self,
+                        group: dict,
+                        p: Tensor,
+                        state: dict,
+                        clipping_scale: float):
+        """
+        Do the step for one parameter, which is actually going to be a batch of
+        `real` parameters, with dim 0 as the batch dim.
+        Args:
+                  group:  dict to look up configuration values
+                    p: parameter to update (actually multiple parameters stacked together
+                       as a batch)
+                  state: state-dict for p, to look up the optimizer state
+        """
+        lr = group["lr"]
+        size_update_period = group["size_update_period"]
+        beta1 = group["betas"][0]
+
+        grad = p.grad
+        if clipping_scale != 1.0:
+            grad = grad * clipping_scale
+        step = state["step"]
+        delta = state["delta"]
+
+        delta.mul_(beta1)
+        batch_size = p.shape[0]
+        numel = p.numel() // batch_size
+        if numel > 1:
+            # Update the size/scale of p, and set param_rms
+            scale_grads = state["scale_grads"]
+            scale_grads[step % size_update_period] = (p * grad).sum(
+                dim=list(range(1, p.ndim)), keepdim=True)
+            if step % size_update_period == size_update_period - 1:
+                param_rms = state["param_rms"]  # shape: (batch_size, 1, 1, ..)
+                param_rms.copy_((p**2)
+                                .mean(dim=list(range(1, p.ndim)), keepdim=True)
+                                .sqrt())
+                if step > 0:
+                    # self._size_update() learns the overall scale on the
+                    # parameter, by shrinking or expanding it.
+                    self._size_update(group, scale_grads, p, state)
+
+        if numel == 1:
+            # For parameters with 1 element we just use regular Adam.
+            # Updates delta.
+            self._step_scalar(group, p, state)
+        else:
+            self._step(group, p, state)
+
+        state["step"] = step + 1
+
+    def _size_update(self,
+                     group: dict,
+                     scale_grads: Tensor,
+                     p: Tensor,
+                     state: dict) -> None:
+        """
+               Called only where p.numel() > 1, this updates the scale of the parameter.
+               If we imagine: p =  underlying_param * scale.exp(), and we are doing
+               gradient descent on underlying param and on scale, this function does the update
+               on `scale`.
+
+               Args:
+              group: dict to look up configuration values
+        scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing
+                      grads w.r.t. the scales.
+                  p:  The parameter to update
+               state: The state-dict of p
+        """
+
+        param_rms = state["param_rms"]
+        beta1, beta2 = group["betas"]
+        size_lr = group["lr"] * group["scalar_lr_scale"]
+        param_min_rms = group["param_min_rms"]
+        param_max_rms = group["param_max_rms"]
+        eps = group["eps"]
+        step = state["step"]
+        batch_size = p.shape[0]
+
+        size_update_period = scale_grads.shape[0]
+        # correct beta2 for the size update period: we will have
+        # faster decay at this level.
+        beta2_corr = beta2**size_update_period
+
+        scale_exp_avg_sq = state[
+            "scale_exp_avg_sq"]  # shape: (batch_size, 1, 1, ..)
+        scale_exp_avg_sq.mul_(beta2_corr).add_(
+            (scale_grads**2).mean(dim=0),  # mean over dim `size_update_period`
+            alpha=1 - beta2_corr, )  # shape is (batch_size, 1, 1, ...)
+
+        # The 1st time we reach here is when size_step == 1.
+        size_step = (step + 1) // size_update_period
+        bias_correction2 = 1 - beta2_corr**size_step
+        # we don't bother with bias_correction1; this will help prevent divergence
+        # at the start of training.
+
+        denom = scale_exp_avg_sq.sqrt() + eps
+
+        scale_step = (-size_lr * (bias_correction2**0.5) *
+                      scale_grads.sum(dim=0) / denom)
+
+        is_too_small = param_rms < param_min_rms
+        is_too_large = param_rms > param_max_rms
+
+        # when the param gets too small, just don't shrink it any further.
+        scale_step.masked_fill_(is_too_small, 0.0)
+        # when it gets too large, stop it from getting any larger.
+        scale_step.masked_fill_(is_too_large, -size_lr * size_update_period)
+        delta = state["delta"]
+        # the factor of (1-beta1) relates to momentum.
+        delta.add_(p * scale_step, alpha=(1 - beta1))
+
+    def _step(self, group: dict, p: Tensor, state: dict):
+        """
+        This function does the core update of self.step(), in the case where the members of
+        the batch have more than 1 element.
+
+        Args:
+            group: A dict which will be used to look up configuration values
+                p: The parameter to be updated
+             grad: The grad of p
+            state: The state-dict corresponding to parameter p
+
+        This function modifies p.
+        """
+        grad = p.grad
+        lr = group["lr"]
+        beta1, beta2 = group["betas"]
+        eps = group["eps"]
+        param_min_rms = group["param_min_rms"]
+        step = state["step"]
+
+        exp_avg_sq = state["exp_avg_sq"]
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
+
+        this_step = state["step"] - (state["zero_step"]
+                                     if "zero_step" in state else 0)
+        bias_correction2 = 1 - beta2**(this_step + 1)
+        if bias_correction2 < 0.99:
+            # note: not in-place.
+            exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
+
+        denom = exp_avg_sq.sqrt()
+        denom += eps
+        grad = grad / denom
+
+        alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms)
+
+        delta = state["delta"]
+        delta.add_(grad * alpha)
+        p.add_(delta)
+
+    def _step_scalar(self, group: dict, p: Tensor, state: dict):
+        """
+        A simplified form of the core update for scalar tensors, where we cannot get a good
+        estimate of the parameter rms.
+        """
+        beta1, beta2 = group["betas"]
+        scalar_max = group["scalar_max"]
+        eps = group["eps"]
+        lr = group["lr"] * group["scalar_lr_scale"]
+        grad = p.grad
+
+        exp_avg_sq = state["exp_avg_sq"]  # shape: (batch_size,)
+        exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+        # bias_correction2 is like in Adam.  Don't bother with bias_correction1;
+        # slower update at the start will help stability anyway.
+        bias_correction2 = 1 - beta2**(state["step"] + 1)
+        denom = (exp_avg_sq / bias_correction2).sqrt() + eps
+
+        delta = state["delta"]
+        delta.add_(grad / denom, alpha=-lr * (1 - beta1))
+        p.clamp_(min=-scalar_max, max=scalar_max)
+        p.add_(delta)

AR/modules/patched_mha_with_cache.py

@@ -0,0 +1,463 @@
+from torch.nn.functional import *
+from torch.nn.functional import (
+    _mha_shape_check,
+    _canonical_mask,
+    _none_or_dtype,
+    _in_projection_packed,
+)
+
+# import torch
+# Tensor = torch.Tensor
+# from typing import Callable, List, Optional, Tuple, Union
+
+
+def multi_head_attention_forward_patched(
+    query: Tensor,
+    key: Tensor,
+    value: Tensor,
+    embed_dim_to_check: int,
+    num_heads: int,
+    in_proj_weight: Optional[Tensor],
+    in_proj_bias: Optional[Tensor],
+    bias_k: Optional[Tensor],
+    bias_v: Optional[Tensor],
+    add_zero_attn: bool,
+    dropout_p: float,
+    out_proj_weight: Tensor,
+    out_proj_bias: Optional[Tensor],
+    training: bool = True,
+    key_padding_mask: Optional[Tensor] = None,
+    need_weights: bool = True,
+    attn_mask: Optional[Tensor] = None,
+    use_separate_proj_weight: bool = False,
+    q_proj_weight: Optional[Tensor] = None,
+    k_proj_weight: Optional[Tensor] = None,
+    v_proj_weight: Optional[Tensor] = None,
+    static_k: Optional[Tensor] = None,
+    static_v: Optional[Tensor] = None,
+    average_attn_weights: bool = True,
+    is_causal: bool = False,
+    cache=None,
+) -> Tuple[Tensor, Optional[Tensor]]:
+    r"""
+    Args:
+        query, key, value: map a query and a set of key-value pairs to an output.
+            See "Attention Is All You Need" for more details.
+        embed_dim_to_check: total dimension of the model.
+        num_heads: parallel attention heads.
+        in_proj_weight, in_proj_bias: input projection weight and bias.
+        bias_k, bias_v: bias of the key and value sequences to be added at dim=0.
+        add_zero_attn: add a new batch of zeros to the key and
+                       value sequences at dim=1.
+        dropout_p: probability of an element to be zeroed.
+        out_proj_weight, out_proj_bias: the output projection weight and bias.
+        training: apply dropout if is ``True``.
+        key_padding_mask: if provided, specified padding elements in the key will
+            be ignored by the attention. This is an binary mask. When the value is True,
+            the corresponding value on the attention layer will be filled with -inf.
+        need_weights: output attn_output_weights.
+            Default: `True`
+            Note: `needs_weight` defaults to `True`, but should be set to `False`
+            For best performance when attention weights are not nedeeded.
+            *Setting needs_weights to `True`
+            leads to a significant performance degradation.*
+        attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
+            the batches while a 3D mask allows to specify a different mask for the entries of each batch.
+        is_causal: If specified, applies a causal mask as attention mask, and ignores
+            attn_mask for computing scaled dot product attention.
+            Default: ``False``.
+            .. warning::
+                is_causal is provides a hint that the attn_mask is the
+                causal mask.Providing incorrect hints can result in
+                incorrect execution, including forward and backward
+                compatibility.
+        use_separate_proj_weight: the function accept the proj. weights for query, key,
+            and value in different forms. If false, in_proj_weight will be used, which is
+            a combination of q_proj_weight, k_proj_weight, v_proj_weight.
+        q_proj_weight, k_proj_weight, v_proj_weight, in_proj_bias: input projection weight and bias.
+        static_k, static_v: static key and value used for attention operators.
+        average_attn_weights: If true, indicates that the returned ``attn_weights`` should be averaged across heads.
+            Otherwise, ``attn_weights`` are provided separately per head. Note that this flag only has an effect
+            when ``need_weights=True.``. Default: True
+
+
+    Shape:
+        Inputs:
+        - query: :math:`(L, E)` or :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
+          the embedding dimension.
+        - key: :math:`(S, E)` or :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
+          the embedding dimension.
+        - value: :math:`(S, E)` or :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
+          the embedding dimension.
+        - key_padding_mask: :math:`(S)` or :math:`(N, S)` where N is the batch size, S is the source sequence length.
+          If a FloatTensor is provided, it will be directly added to the value.
+          If a BoolTensor is provided, the positions with the
+          value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
+        - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
+          3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
+          S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
+          positions. If a BoolTensor is provided, positions with ``True``
+          are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
+          is provided, it will be added to the attention weight.
+        - static_k: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
+          N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
+        - static_v: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
+          N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
+
+        Outputs:
+        - attn_output: :math:`(L, E)` or :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
+          E is the embedding dimension.
+        - attn_output_weights: Only returned when ``need_weights=True``. If ``average_attn_weights=True``, returns
+          attention weights averaged across heads of shape :math:`(L, S)` when input is unbatched or
+          :math:`(N, L, S)`, where :math:`N` is the batch size, :math:`L` is the target sequence length, and
+          :math:`S` is the source sequence length. If ``average_attn_weights=False``, returns attention weights per
+          head of shape :math:`(num_heads, L, S)` when input is unbatched or :math:`(N, num_heads, L, S)`.
+    """
+    tens_ops = (
+        query,
+        key,
+        value,
+        in_proj_weight,
+        in_proj_bias,
+        bias_k,
+        bias_v,
+        out_proj_weight,
+        out_proj_bias,
+    )
+    if has_torch_function(tens_ops):
+        return handle_torch_function(
+            multi_head_attention_forward,
+            tens_ops,
+            query,
+            key,
+            value,
+            embed_dim_to_check,
+            num_heads,
+            in_proj_weight,
+            in_proj_bias,
+            bias_k,
+            bias_v,
+            add_zero_attn,
+            dropout_p,
+            out_proj_weight,
+            out_proj_bias,
+            training=training,
+            key_padding_mask=key_padding_mask,
+            need_weights=need_weights,
+            attn_mask=attn_mask,
+            is_causal=is_causal,
+            use_separate_proj_weight=use_separate_proj_weight,
+            q_proj_weight=q_proj_weight,
+            k_proj_weight=k_proj_weight,
+            v_proj_weight=v_proj_weight,
+            static_k=static_k,
+            static_v=static_v,
+            average_attn_weights=average_attn_weights,
+            cache=cache,
+        )
+
+    is_batched = _mha_shape_check(
+        query, key, value, key_padding_mask, attn_mask, num_heads
+    )
+
+    # For unbatched input, we unsqueeze at the expected batch-dim to pretend that the input
+    # is batched, run the computation and before returning squeeze the
+    # batch dimension so that the output doesn't carry this temporary batch dimension.
+    if not is_batched:
+        # unsqueeze if the input is unbatched
+        query = query.unsqueeze(1)
+        key = key.unsqueeze(1)
+        value = value.unsqueeze(1)
+        if key_padding_mask is not None:
+            key_padding_mask = key_padding_mask.unsqueeze(0)
+
+    # set up shape vars
+    tgt_len, bsz, embed_dim = query.shape
+    src_len, _, _ = key.shape
+
+    key_padding_mask = _canonical_mask(
+        mask=key_padding_mask,
+        mask_name="key_padding_mask",
+        other_type=_none_or_dtype(attn_mask),
+        other_name="attn_mask",
+        target_type=query.dtype,
+    )
+
+    if is_causal and attn_mask is None:
+        raise RuntimeError(
+            "Need attn_mask if specifying the is_causal hint. "
+            "You may use the Transformer module method "
+            "`generate_square_subsequent_mask` to create this mask."
+        )
+
+    if is_causal and key_padding_mask is None and not need_weights:
+        # when we have a kpm or need weights, we need attn_mask
+        # Otherwise, we use the is_causal hint go as is_causal
+        # indicator to SDPA.
+        attn_mask = None
+    else:
+        attn_mask = _canonical_mask(
+            mask=attn_mask,
+            mask_name="attn_mask",
+            other_type=None,
+            other_name="",
+            target_type=query.dtype,
+            check_other=False,
+        )
+
+        if key_padding_mask is not None:
+            # We have the attn_mask, and use that to merge kpm into it.
+            # Turn off use of is_causal hint, as the merged mask is no
+            # longer causal.
+            is_causal = False
+
+    assert (
+        embed_dim == embed_dim_to_check
+    ), f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
+    if isinstance(embed_dim, torch.Tensor):
+        # embed_dim can be a tensor when JIT tracing
+        head_dim = embed_dim.div(num_heads, rounding_mode="trunc")
+    else:
+        head_dim = embed_dim // num_heads
+    assert (
+        head_dim * num_heads == embed_dim
+    ), f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
+    if use_separate_proj_weight:
+        # allow MHA to have different embedding dimensions when separate projection weights are used
+        assert (
+            key.shape[:2] == value.shape[:2]
+        ), f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
+    else:
+        assert (
+            key.shape == value.shape
+        ), f"key shape {key.shape} does not match value shape {value.shape}"
+
+    #
+    # compute in-projection
+    #
+    if not use_separate_proj_weight:
+        assert (
+            in_proj_weight is not None
+        ), "use_separate_proj_weight is False but in_proj_weight is None"
+        q, k, v = _in_projection_packed(query, key, value, in_proj_weight, in_proj_bias)
+    else:
+        assert (
+            q_proj_weight is not None
+        ), "use_separate_proj_weight is True but q_proj_weight is None"
+        assert (
+            k_proj_weight is not None
+        ), "use_separate_proj_weight is True but k_proj_weight is None"
+        assert (
+            v_proj_weight is not None
+        ), "use_separate_proj_weight is True but v_proj_weight is None"
+        if in_proj_bias is None:
+            b_q = b_k = b_v = None
+        else:
+            b_q, b_k, b_v = in_proj_bias.chunk(3)
+        q, k, v = _in_projection(
+            query,
+            key,
+            value,
+            q_proj_weight,
+            k_proj_weight,
+            v_proj_weight,
+            b_q,
+            b_k,
+            b_v,
+        )
+    if cache != None:
+        if cache["first_infer"] == 1:
+            cache["k"][cache["stage"]] = k
+            # print(0,cache["k"].shape)
+            cache["v"][cache["stage"]] = v
+        else:  ###12个layer每个都要留自己的cache_kv
+            # print(1,cache["k"].shape)
+            cache["k"][cache["stage"]] = torch.cat(
+                [cache["k"][cache["stage"]], k], 0
+            )  ##本来时序是1，但是proj的时候可能transpose了所以时序到0维了
+            cache["v"][cache["stage"]] = torch.cat([cache["v"][cache["stage"]], v], 0)
+            # print(2, cache["k"].shape)
+            src_len = cache["k"][cache["stage"]].shape[0]
+            k = cache["k"][cache["stage"]]
+            v = cache["v"][cache["stage"]]
+            # if attn_mask is not None:
+            #     attn_mask=attn_mask[-1:,]
+            # print(attn_mask.shape,attn_mask)
+        cache["stage"] = (cache["stage"] + 1) % cache["all_stage"]
+    # print(2333,cache)
+    # prep attention mask
+
+    attn_mask = _canonical_mask(
+        mask=attn_mask,
+        mask_name="attn_mask",
+        other_type=None,
+        other_name="",
+        target_type=q.dtype,
+        check_other=False,
+    )
+
+    if attn_mask is not None:
+        # ensure attn_mask's dim is 3
+        if attn_mask.dim() == 2:
+            correct_2d_size = (tgt_len, src_len)
+            if attn_mask.shape != correct_2d_size:
+                raise RuntimeError(
+                    f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}."
+                )
+            attn_mask = attn_mask.unsqueeze(0)
+        elif attn_mask.dim() == 3:
+            correct_3d_size = (bsz * num_heads, tgt_len, src_len)
+            if attn_mask.shape != correct_3d_size:
+                raise RuntimeError(
+                    f"The shape of the 3D attn_mask is {attn_mask.shape}, but should be {correct_3d_size}."
+                )
+        else:
+            raise RuntimeError(
+                f"attn_mask's dimension {attn_mask.dim()} is not supported"
+            )
+
+    # add bias along batch dimension (currently second)
+    if bias_k is not None and bias_v is not None:
+        assert static_k is None, "bias cannot be added to static key."
+        assert static_v is None, "bias cannot be added to static value."
+        k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
+        v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
+        if attn_mask is not None:
+            attn_mask = pad(attn_mask, (0, 1))
+        if key_padding_mask is not None:
+            key_padding_mask = pad(key_padding_mask, (0, 1))
+    else:
+        assert bias_k is None
+        assert bias_v is None
+
+    #
+    # reshape q, k, v for multihead attention and make em batch first
+    #
+    q = q.view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
+    if static_k is None:
+        k = k.view(k.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
+    else:
+        # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+        assert (
+            static_k.size(0) == bsz * num_heads
+        ), f"expecting static_k.size(0) of {bsz * num_heads}, but got {static_k.size(0)}"
+        assert (
+            static_k.size(2) == head_dim
+        ), f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
+        k = static_k
+    if static_v is None:
+        v = v.view(v.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
+    else:
+        # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+        assert (
+            static_v.size(0) == bsz * num_heads
+        ), f"expecting static_v.size(0) of {bsz * num_heads}, but got {static_v.size(0)}"
+        assert (
+            static_v.size(2) == head_dim
+        ), f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
+        v = static_v
+
+    # add zero attention along batch dimension (now first)
+    if add_zero_attn:
+        zero_attn_shape = (bsz * num_heads, 1, head_dim)
+        k = torch.cat(
+            [k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1
+        )
+        v = torch.cat(
+            [v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1
+        )
+        if attn_mask is not None:
+            attn_mask = pad(attn_mask, (0, 1))
+        if key_padding_mask is not None:
+            key_padding_mask = pad(key_padding_mask, (0, 1))
+
+    # update source sequence length after adjustments
+    src_len = k.size(1)
+
+    # merge key padding and attention masks
+    if key_padding_mask is not None:
+        assert key_padding_mask.shape == (
+            bsz,
+            src_len,
+        ), f"expecting key_padding_mask shape of {(bsz, src_len)}, but got {key_padding_mask.shape}"
+        key_padding_mask = (
+            key_padding_mask.view(bsz, 1, 1, src_len)
+            .expand(-1, num_heads, -1, -1)
+            .reshape(bsz * num_heads, 1, src_len)
+        )
+        if attn_mask is None:
+            attn_mask = key_padding_mask
+        else:
+            attn_mask = attn_mask + key_padding_mask
+
+    # adjust dropout probability
+    if not training:
+        dropout_p = 0.0
+
+    #
+    # (deep breath) calculate attention and out projection
+    #
+
+    if need_weights:
+        B, Nt, E = q.shape
+        q_scaled = q / math.sqrt(E)
+
+        assert not (
+            is_causal and attn_mask is None
+        ), "FIXME: is_causal not implemented for need_weights"
+
+        if attn_mask is not None:
+            attn_output_weights = torch.baddbmm(
+                attn_mask, q_scaled, k.transpose(-2, -1)
+            )
+        else:
+            attn_output_weights = torch.bmm(q_scaled, k.transpose(-2, -1))
+        attn_output_weights = softmax(attn_output_weights, dim=-1)
+        if dropout_p > 0.0:
+            attn_output_weights = dropout(attn_output_weights, p=dropout_p)
+
+        attn_output = torch.bmm(attn_output_weights, v)
+
+        attn_output = (
+            attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
+        )
+        attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
+        attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+
+        # optionally average attention weights over heads
+        attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
+        if average_attn_weights:
+            attn_output_weights = attn_output_weights.mean(dim=1)
+
+        if not is_batched:
+            # squeeze the output if input was unbatched
+            attn_output = attn_output.squeeze(1)
+            attn_output_weights = attn_output_weights.squeeze(0)
+        return attn_output, attn_output_weights
+    else:
+        # attn_mask can be either (L,S) or (N*num_heads, L, S)
+        # if attn_mask's shape is (1, L, S) we need to unsqueeze to (1, 1, L, S)
+        # in order to match the input for SDPA of (N, num_heads, L, S)
+        if attn_mask is not None:
+            if attn_mask.size(0) == 1 and attn_mask.dim() == 3:
+                attn_mask = attn_mask.unsqueeze(0)
+            else:
+                attn_mask = attn_mask.view(bsz, num_heads, -1, src_len)
+
+        q = q.view(bsz, num_heads, tgt_len, head_dim)
+        k = k.view(bsz, num_heads, src_len, head_dim)
+        v = v.view(bsz, num_heads, src_len, head_dim)
+
+        attn_output = scaled_dot_product_attention(
+            q, k, v, attn_mask, dropout_p, is_causal
+        )
+        attn_output = (
+            attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
+        )
+
+        attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
+        attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+        if not is_batched:
+            # squeeze the output if input was unbatched
+            attn_output = attn_output.squeeze(1)
+        return attn_output, None

AR/modules/patched_mha_with_cache_onnx.py

@@ -0,0 +1,92 @@
+from torch.nn.functional import *
+from torch.nn.functional import (
+    _mha_shape_check,
+    _canonical_mask,
+    _none_or_dtype,
+    _in_projection_packed,
+)
+
+def multi_head_attention_forward_patched(
+    query,
+    key,
+    value,
+    embed_dim_to_check: int,
+    num_heads: int,
+    in_proj_weight,
+    in_proj_bias: Optional[Tensor],
+    bias_k: Optional[Tensor],
+    bias_v: Optional[Tensor],
+    add_zero_attn: bool,
+    dropout_p: float,
+    out_proj_weight: Tensor,
+    out_proj_bias: Optional[Tensor],
+    training: bool = True,
+    key_padding_mask: Optional[Tensor] = None,
+    need_weights: bool = True,
+    attn_mask: Optional[Tensor] = None,
+    use_separate_proj_weight: bool = False,
+    q_proj_weight: Optional[Tensor] = None,
+    k_proj_weight: Optional[Tensor] = None,
+    v_proj_weight: Optional[Tensor] = None,
+    static_k: Optional[Tensor] = None,
+    static_v: Optional[Tensor] = None,
+    average_attn_weights: bool = True,
+    is_causal: bool = False,
+    cache=None,
+) -> Tuple[Tensor, Optional[Tensor]]:
+
+    # set up shape vars
+    _, _, embed_dim = query.shape
+    attn_mask = _canonical_mask(
+        mask=attn_mask,
+        mask_name="attn_mask",
+        other_type=None,
+        other_name="",
+        target_type=query.dtype,
+        check_other=False,
+    )
+    head_dim = embed_dim // num_heads
+
+    proj_qkv = linear(query, in_proj_weight, in_proj_bias)
+    proj_qkv = proj_qkv.unflatten(-1, (3, query.size(-1))).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
+    q, k, v = proj_qkv[0], proj_qkv[1], proj_qkv[2]
+
+    if cache["first_infer"] == 1:
+        cache["k"][cache["stage"]] = k
+        cache["v"][cache["stage"]] = v
+    else:
+        cache["k"][cache["stage"]] = torch.cat([cache["k"][cache["stage"]][:-1], k], 0)
+        cache["v"][cache["stage"]] = torch.cat([cache["v"][cache["stage"]][:-1], v], 0)
+        k = cache["k"][cache["stage"]]
+        v = cache["v"][cache["stage"]]
+    cache["stage"] = (cache["stage"] + 1) % cache["all_stage"]
+
+    attn_mask = _canonical_mask(
+        mask=attn_mask,
+        mask_name="attn_mask",
+        other_type=None,
+        other_name="",
+        target_type=q.dtype,
+        check_other=False,
+    )
+    attn_mask = attn_mask.unsqueeze(0)
+
+    q = q.view(-1, num_heads, head_dim).transpose(0, 1)
+    k = k.view(-1, num_heads, head_dim).transpose(0, 1)
+    v = v.view(-1, num_heads, head_dim).transpose(0, 1)
+
+    dropout_p = 0.0
+    attn_mask = attn_mask.unsqueeze(0)
+    q = q.view(num_heads, -1, head_dim).unsqueeze(0)
+    k = k.view(num_heads, -1, head_dim).unsqueeze(0)
+    v = v.view(num_heads, -1, head_dim).unsqueeze(0)
+    attn_output = scaled_dot_product_attention(
+        q, k, v, attn_mask, dropout_p, is_causal
+    )
+    attn_output = (
+        attn_output.permute(2, 0, 1, 3).contiguous().view(-1, embed_dim)
+    )
+    attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
+    attn_output = attn_output.view(-1, 1, attn_output.size(1))
+
+    return attn_output

AR/modules/scaling.py

@@ -0,0 +1,335 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#import logging
+import math
+import random
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+
+class DoubleSwishFunction(torch.autograd.Function):
+    """
+      double_swish(x) = x * torch.sigmoid(x-1)
+    This is a definition, originally motivated by its close numerical
+    similarity to swish(swish(x)), where swish(x) =  x * sigmoid(x).
+
+    Memory-efficient derivative computation:
+     double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1)
+     double_swish'(x) = d/dx double_swish(x) =  x * s'(x) + x' * s(x) = x * s'(x) + s(x).
+     Now, s'(x) = s(x) * (1-s(x)).
+     double_swish'(x) =  x * s'(x) + s(x).
+                      =  x * s(x) * (1-s(x)) + s(x).
+                     = double_swish(x) * (1-s(x)) + s(x)
+     ... so we just need to remember s(x) but not x itself.
+    """
+
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        requires_grad = x.requires_grad
+        x_dtype = x.dtype
+        if x.dtype == torch.float16:
+            x = x.to(torch.float32)
+
+        s = torch.sigmoid(x - 1.0)
+        y = x * s
+
+        if requires_grad:
+            deriv = y * (1 - s) + s
+            # notes on derivative of x * sigmoid(x - 1):
+            # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29
+            # min \simeq -0.043638.  Take floor as -0.043637 so it's a lower bund
+            # max \simeq 1.1990.   Take ceil to be 1.2 so it's an upper bound.
+            # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which
+            # floors), should be expectation-preserving.
+            floor = -0.043637
+            ceil = 1.2
+            d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
+                deriv
+            )
+            if __name__ == "__main__":
+                # for self-testing only.
+                assert d_scaled.min() >= 0.0
+                assert d_scaled.max() < 256.0
+            d_int = d_scaled.to(torch.uint8)
+            ctx.save_for_backward(d_int)
+        if x.dtype == torch.float16 or torch.is_autocast_enabled():
+            y = y.to(torch.float16)
+        return y
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        (d,) = ctx.saved_tensors
+        # the same constants as used in forward pass.
+        floor = -0.043637
+        ceil = 1.2
+        d = d * ((ceil - floor) / 255.0) + floor
+        return y_grad * d
+
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+        that we approximate closely with x * sigmoid(x-1).
+        """
+        if torch.jit.is_scripting() or torch.jit.is_tracing():
+            return x * torch.sigmoid(x - 1.0)
+        return DoubleSwishFunction.apply(x)
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x: Tensor,
+        scale_factor: Tensor,
+        sign_factor: Optional[Tensor],
+        channel_dim: int,
+    ) -> Tensor:
+        if channel_dim < 0:
+            channel_dim += x.ndim
+        ctx.channel_dim = channel_dim
+        xgt0 = x > 0
+        if sign_factor is None:
+            ctx.save_for_backward(xgt0, scale_factor)
+        else:
+            ctx.save_for_backward(xgt0, scale_factor, sign_factor)
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]:
+        if len(ctx.saved_tensors) == 3:
+            xgt0, scale_factor, sign_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+                sign_factor = sign_factor.unsqueeze(-1)
+            factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        else:
+            xgt0, scale_factor = ctx.saved_tensors
+            for _ in range(ctx.channel_dim, x_grad.ndim - 1):
+                scale_factor = scale_factor.unsqueeze(-1)
+            factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5)
+        neg_delta_grad = x_grad.abs() * factor
+        return (
+            x_grad - neg_delta_grad,
+            None,
+            None,
+            None,
+        )
+
+
+def _compute_scale_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_abs: float,
+    max_abs: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32)
+
+    if min_abs == 0.0:
+        below_threshold = 0.0
+    else:
+        # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if
+        # x_abs)_mean , min_abs.
+        below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp(
+            min=0, max=max_factor
+        )
+
+    above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp(
+        min=0, max=max_factor
+    )
+
+    return below_threshold - above_threshold
+
+
+def _compute_sign_factor(
+    x: Tensor,
+    channel_dim: int,
+    min_positive: float,
+    max_positive: float,
+    gain_factor: float,
+    max_factor: float,
+) -> Tensor:
+    if channel_dim < 0:
+        channel_dim += x.ndim
+    sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+    proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims)
+    if min_positive == 0.0:
+        factor1 = 0.0
+    else:
+        # 0 if proportion_positive >= min_positive, else can be
+        # as large as max_factor.
+        factor1 = (
+            (min_positive - proportion_positive) * (gain_factor / min_positive)
+        ).clamp_(min=0, max=max_factor)
+
+    if max_positive == 1.0:
+        factor2 = 0.0
+    else:
+        # 0 if self.proportion_positive <= max_positive, else can be
+        # as large as -max_factor.
+        factor2 = (
+            (proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive))
+        ).clamp_(min=0, max=max_factor)
+    sign_factor = factor1 - factor2
+    # require min_positive != 0 or max_positive != 1:
+    assert not isinstance(sign_factor, float)
+    return sign_factor
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+    Args:
+           num_channels: the number of channels
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), above which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           sign_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_positive and max_positive
+              are violated.
+           scale_gain_factor: determines the 'gain' with which we increase the
+              change in gradient once the constraints on min_abs and max_abs
+              are violated.
+           min_abs:  the minimum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+           max_abs:  the maximum average-absolute-value difference from the mean
+               value per channel, which we allow, before we start to modify
+               the derivatives to prevent this.
+          min_prob: determines the minimum probability with which we modify the
+             gradients for the {min,max}_positive and {min,max}_abs constraints,
+             on each forward().  This is done randomly to prevent all layers
+             from doing it at the same time.  Early in training we may use
+             higher probabilities than this; it will decay to this value.
+    """
+
+    def __init__(
+        self,
+        num_channels: int,
+        channel_dim: int,
+        min_positive: float = 0.05,
+        max_positive: float = 0.95,
+        max_factor: float = 0.04,
+        sign_gain_factor: float = 0.01,
+        scale_gain_factor: float = 0.02,
+        min_abs: float = 0.2,
+        max_abs: float = 100.0,
+        min_prob: float = 0.1,
+    ):
+        super(ActivationBalancer, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+        self.min_prob = min_prob
+        self.sign_gain_factor = sign_gain_factor
+        self.scale_gain_factor = scale_gain_factor
+
+        # count measures how many times the forward() function has been called.
+        # We occasionally sync this to a tensor called `count`, that exists to
+        # make sure it is synced to disk when we load and save the model.
+        self.cpu_count = 0
+        self.register_buffer("count", torch.tensor(0, dtype=torch.int64))
+
+    def forward(self, x: Tensor) -> Tensor:
+        if torch.jit.is_scripting() or not x.requires_grad or torch.jit.is_tracing():
+            return _no_op(x)
+
+        count = self.cpu_count
+        self.cpu_count += 1
+
+        if random.random() < 0.01:
+            # Occasionally sync self.cpu_count with self.count.
+            # count affects the decay of 'prob'.  don't do this on every iter,
+            # because syncing with the GPU is slow.
+            self.cpu_count = max(self.cpu_count, self.count.item())
+            self.count.fill_(self.cpu_count)
+
+        # the prob of doing some work exponentially decreases from 0.5 till it hits
+        # a floor at min_prob (==0.1, by default)
+        prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0)))
+
+        if random.random() < prob:
+            sign_gain_factor = 0.5
+            if self.min_positive != 0.0 or self.max_positive != 1.0:
+                sign_factor = _compute_sign_factor(
+                    x,
+                    self.channel_dim,
+                    self.min_positive,
+                    self.max_positive,
+                    gain_factor=self.sign_gain_factor / prob,
+                    max_factor=self.max_factor,
+                )
+            else:
+                sign_factor = None
+
+            scale_factor = _compute_scale_factor(
+                x.detach(),
+                self.channel_dim,
+                min_abs=self.min_abs,
+                max_abs=self.max_abs,
+                gain_factor=self.scale_gain_factor / prob,
+                max_factor=self.max_factor,
+            )
+            return ActivationBalancerFunction.apply(
+                x,
+                scale_factor,
+                sign_factor,
+                self.channel_dim,
+            )
+        else:
+            return _no_op(x)
+
+
+def BalancedDoubleSwish(
+    d_model, channel_dim=-1, max_abs=10.0, min_prob=0.25
+) -> nn.Sequential:
+    """
+    ActivationBalancer -> DoubleSwish
+    """
+    balancer = ActivationBalancer(
+        d_model, channel_dim=channel_dim, max_abs=max_abs, min_prob=min_prob
+    )
+    return nn.Sequential(
+        balancer,
+        DoubleSwish(),
+    )

AR/modules/transformer.py

@@ -0,0 +1,378 @@
+# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/transformer.py
+import copy
+import numbers
+from functools import partial
+from typing import Any
+from typing import Callable
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+from AR.modules.activation import MultiheadAttention
+from AR.modules.scaling import BalancedDoubleSwish
+from torch import nn
+from torch import Tensor
+from torch.nn import functional as F
+
+_shape_t = Union[int, List[int], torch.Size]
+
+
+class LayerNorm(nn.Module):
+    __constants__ = ["normalized_shape", "eps", "elementwise_affine"]
+    normalized_shape: Tuple[int, ...]
+    eps: float
+    elementwise_affine: bool
+
+    def __init__(
+        self,
+        normalized_shape: _shape_t,
+        eps: float = 1e-5,
+        elementwise_affine: bool = True,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(LayerNorm, self).__init__()
+        if isinstance(normalized_shape, numbers.Integral):
+            # mypy error: incompatible types in assignment
+            normalized_shape = (normalized_shape,)  # type: ignore[assignment]
+        self.normalized_shape = tuple(normalized_shape)  # type: ignore[arg-type]
+        self.eps = eps
+        self.elementwise_affine = elementwise_affine
+        if self.elementwise_affine:
+            self.weight = nn.Parameter(
+                torch.empty(self.normalized_shape, **factory_kwargs)
+            )
+            self.bias = nn.Parameter(
+                torch.empty(self.normalized_shape, **factory_kwargs)
+            )
+        else:
+            self.register_parameter("weight", None)
+            self.register_parameter("bias", None)
+
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+            nn.init.zeros_(self.bias)
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            return (
+                F.layer_norm(
+                    input,
+                    self.normalized_shape,
+                    self.weight,
+                    self.bias,
+                    self.eps,
+                ),
+                embedding,
+            )
+
+        assert embedding is None
+        return F.layer_norm(
+            input, self.normalized_shape, self.weight, self.bias, self.eps
+        )
+
+    def extra_repr(self) -> str:
+        return (
+            "{normalized_shape}, eps={eps}, "
+            "elementwise_affine={elementwise_affine}".format(**self.__dict__)
+        )
+
+
+class IdentityNorm(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        eps: float = 1e-5,
+        device=None,
+        dtype=None,
+    ) -> None:
+        super(IdentityNorm, self).__init__()
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            return input
+
+        assert embedding is None
+        return input
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers. Users can build the
+    BERT(https://arxiv.org/abs/1810.04805) model with corresponding parameters.
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+        enable_nested_tensor: if True, input will automatically convert to nested tensor
+            (and convert back on output). This will improve the overall performance of
+            TransformerEncoder when padding rate is high. Default: ``True`` (enabled).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+    __constants__ = ["norm"]
+
+    def __init__(self, encoder_layer, num_layers, norm=None):
+        super(TransformerEncoder, self).__init__()
+        self.layers = _get_clones(encoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+
+    def forward(
+        self,
+        src: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+        cache=None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layers in turn.
+
+        Args:
+            src: the sequence to the encoder (required).
+            mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+            return_layer_states: return layers' state (optional).
+
+        Shape:
+            see the docs in Transformer class.
+        """
+        if return_layer_states:
+            layer_states = []  # layers' output
+            output = src
+            for mod in self.layers:
+                output = mod(
+                    output,
+                    src_mask=mask,
+                    src_key_padding_mask=src_key_padding_mask,
+                    cache=cache,
+                )
+                layer_states.append(output[0])
+
+            if self.norm is not None:
+                output = self.norm(output)
+
+            return layer_states, output
+
+        output = src
+        for mod in self.layers:
+            output = mod(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                cache=cache,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class TransformerEncoderLayer(nn.Module):
+    __constants__ = ["batch_first", "norm_first"]
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: Union[str, Callable[[Tensor], Tensor]] = F.relu,
+        batch_first: bool = False,
+        norm_first: bool = False,
+        device=None,
+        dtype=None,
+        linear1_self_attention_cls: nn.Module = nn.Linear,
+        linear2_self_attention_cls: nn.Module = nn.Linear,
+        linear1_feedforward_cls: nn.Module = nn.Linear,
+        linear2_feedforward_cls: nn.Module = nn.Linear,
+        layer_norm_cls: nn.Module = LayerNorm,
+        layer_norm_eps: float = 1e-5,
+        adaptive_layer_norm=False,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(TransformerEncoderLayer, self).__init__()
+        # print(233333333333,d_model,nhead)
+        # import os
+        # os._exit(2333333)
+        self.self_attn = MultiheadAttention(
+            d_model,  # 512 16
+            nhead,
+            dropout=dropout,
+            batch_first=batch_first,
+            linear1_cls=linear1_self_attention_cls,
+            linear2_cls=linear2_self_attention_cls,
+            **factory_kwargs,
+        )
+
+        # Implementation of Feedforward model
+        self.linear1 = linear1_feedforward_cls(
+            d_model, dim_feedforward, **factory_kwargs
+        )
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = linear2_feedforward_cls(
+            dim_feedforward, d_model, **factory_kwargs
+        )
+
+        self.norm_first = norm_first
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        # Legacy string support for activation function.
+        if isinstance(activation, str):
+            activation = _get_activation_fn(activation)
+        elif isinstance(activation, partial):
+            activation = activation(d_model)
+        elif activation == BalancedDoubleSwish:
+            activation = BalancedDoubleSwish(d_model)
+
+        # # We can't test self.activation in forward() in TorchScript,
+        # # so stash some information about it instead.
+        # if activation is F.relu or isinstance(activation, torch.nn.ReLU):
+        #     self.activation_relu_or_gelu = 1
+        # elif activation is F.gelu or isinstance(activation, torch.nn.GELU):
+        #     self.activation_relu_or_gelu = 2
+        # else:
+        #     self.activation_relu_or_gelu = 0
+        self.activation = activation
+
+        norm1 = layer_norm_cls(d_model, eps=layer_norm_eps, **factory_kwargs)
+        if layer_norm_cls == IdentityNorm:
+            norm2 = BalancedBasicNorm(d_model, eps=layer_norm_eps, **factory_kwargs)
+        else:
+            norm2 = layer_norm_cls(d_model, eps=layer_norm_eps, **factory_kwargs)
+
+        if adaptive_layer_norm:
+            self.norm1 = AdaptiveLayerNorm(d_model, norm1)
+            self.norm2 = AdaptiveLayerNorm(d_model, norm2)
+        else:
+            self.norm1 = norm1
+            self.norm2 = norm2
+
+    def __setstate__(self, state):
+        super(TransformerEncoderLayer, self).__setstate__(state)
+        if not hasattr(self, "activation"):
+            self.activation = F.relu
+
+    def forward(
+        self,
+        src: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        cache=None,
+    ) -> Tensor:
+        r"""Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional).
+
+        Shape:
+            see the docs in Transformer class.
+        """
+        x, stage_embedding = src, None
+        is_src_tuple = False
+        if isinstance(src, tuple):
+            x, stage_embedding = src
+            is_src_tuple = True
+
+        if src_key_padding_mask is not None:
+            _skpm_dtype = src_key_padding_mask.dtype
+            if _skpm_dtype != torch.bool and not torch.is_floating_point(
+                src_key_padding_mask
+            ):
+                raise AssertionError(
+                    "only bool and floating types of key_padding_mask are supported"
+                )
+
+        if self.norm_first:
+            x = x + self._sa_block(
+                self.norm1(x, stage_embedding),
+                src_mask,
+                src_key_padding_mask,
+                cache=cache,
+            )
+            x = x + self._ff_block(self.norm2(x, stage_embedding))
+        else:
+            x = self.norm1(
+                x + self._sa_block(x, src_mask, src_key_padding_mask, cache=cache),
+                stage_embedding,
+            )
+            x = self.norm2(x + self._ff_block(x), stage_embedding)
+
+        if is_src_tuple:
+            return (x, stage_embedding)
+        return x
+
+    # self-attention block
+    def _sa_block(
+        self,
+        x: Tensor,
+        attn_mask: Optional[Tensor],
+        key_padding_mask: Optional[Tensor],
+        cache=None,
+    ) -> Tensor:
+        # print(x.shape,attn_mask.shape,key_padding_mask)
+        # torch.Size([1, 188, 512]) torch.Size([188, 188]) None
+        # import os
+        # os._exit(23333)
+        x = self.self_attn(
+            x,
+            x,
+            x,
+            attn_mask=attn_mask,
+            key_padding_mask=key_padding_mask,
+            need_weights=False,
+            cache=cache,
+        )[0]
+        return self.dropout1(x)
+
+    # feed forward block
+    def _ff_block(self, x: Tensor) -> Tensor:
+        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
+        return self.dropout2(x)
+
+
+class AdaptiveLayerNorm(nn.Module):
+    r"""Adaptive Layer Normalization"""
+
+    def __init__(self, d_model, norm) -> None:
+        super(AdaptiveLayerNorm, self).__init__()
+        self.project_layer = nn.Linear(d_model, 2 * d_model)
+        self.norm = norm
+        self.d_model = d_model
+        self.eps = self.norm.eps
+
+    def forward(self, input: Tensor, embedding: Tensor = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            weight, bias = torch.split(
+                self.project_layer(embedding),
+                split_size_or_sections=self.d_model,
+                dim=-1,
+            )
+            return (weight * self.norm(input) + bias, embedding)
+
+        weight, bias = torch.split(
+            self.project_layer(embedding),
+            split_size_or_sections=self.d_model,
+            dim=-1,
+        )
+        return weight * self.norm(input) + bias
+
+
+def _get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])

AR/modules/transformer_onnx.py

@@ -0,0 +1,292 @@
+# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/transformer.py
+import copy
+import numbers
+from functools import partial
+from typing import Any
+from typing import Callable
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+from AR.modules.activation_onnx import MultiheadAttention
+from AR.modules.scaling import BalancedDoubleSwish
+from torch import nn
+from torch import Tensor
+from torch.nn import functional as F
+
+_shape_t = Union[int, List[int], torch.Size]
+
+
+class LayerNorm(nn.Module):
+    __constants__ = ["normalized_shape", "eps", "elementwise_affine"]
+    normalized_shape: Tuple[int, ...]
+    eps: float
+    elementwise_affine: bool
+
+    def __init__(
+        self,
+        normalized_shape: _shape_t,
+        eps: float = 1e-5,
+        elementwise_affine: bool = True,
+        device=None,
+        dtype=None,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(LayerNorm, self).__init__()
+        if isinstance(normalized_shape, numbers.Integral):
+            # mypy error: incompatible types in assignment
+            normalized_shape = (normalized_shape,)  # type: ignore[assignment]
+        self.normalized_shape = tuple(normalized_shape)  # type: ignore[arg-type]
+        self.eps = eps
+        self.elementwise_affine = elementwise_affine
+        if self.elementwise_affine:
+            self.weight = nn.Parameter(
+                torch.empty(self.normalized_shape, **factory_kwargs)
+            )
+            self.bias = nn.Parameter(
+                torch.empty(self.normalized_shape, **factory_kwargs)
+            )
+        else:
+            self.register_parameter("weight", None)
+            self.register_parameter("bias", None)
+
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        if self.elementwise_affine:
+            nn.init.ones_(self.weight)
+            nn.init.zeros_(self.bias)
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            return (
+                F.layer_norm(
+                    input,
+                    self.normalized_shape,
+                    self.weight,
+                    self.bias,
+                    self.eps,
+                ),
+                embedding,
+            )
+
+        assert embedding is None
+        return F.layer_norm(
+            input, self.normalized_shape, self.weight, self.bias, self.eps
+        )
+
+    def extra_repr(self) -> str:
+        return (
+            "{normalized_shape}, eps={eps}, "
+            "elementwise_affine={elementwise_affine}".format(**self.__dict__)
+        )
+
+
+class IdentityNorm(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        eps: float = 1e-5,
+        device=None,
+        dtype=None,
+    ) -> None:
+        super(IdentityNorm, self).__init__()
+
+    def forward(self, input: Tensor, embedding: Any = None) -> Tensor:
+        if isinstance(input, tuple):
+            return input
+
+        assert embedding is None
+        return input
+
+
+class TransformerEncoder(nn.Module):
+    r"""TransformerEncoder is a stack of N encoder layers. Users can build the
+    BERT(https://arxiv.org/abs/1810.04805) model with corresponding parameters.
+
+    Args:
+        encoder_layer: an instance of the TransformerEncoderLayer() class (required).
+        num_layers: the number of sub-encoder-layers in the encoder (required).
+        norm: the layer normalization component (optional).
+        enable_nested_tensor: if True, input will automatically convert to nested tensor
+            (and convert back on output). This will improve the overall performance of
+            TransformerEncoder when padding rate is high. Default: ``True`` (enabled).
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> transformer_encoder = TransformerEncoder(encoder_layer, num_layers=6)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = transformer_encoder(src)
+    """
+    __constants__ = ["norm"]
+
+    def __init__(self, encoder_layer, num_layers, norm=None):
+        super(TransformerEncoder, self).__init__()
+        self.layers = _get_clones(encoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = norm
+
+    def forward(
+        self,
+        src: Tensor,
+        mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        return_layer_states: bool = False,
+        cache=None,
+    ) -> Tensor:
+        output = src
+        for mod in self.layers:
+            output = mod(
+                output,
+                src_mask=mask,
+                src_key_padding_mask=src_key_padding_mask,
+                cache=cache,
+            )
+
+        if self.norm is not None:
+            output = self.norm(output)
+
+        return output
+
+
+class TransformerEncoderLayer(nn.Module):
+    __constants__ = ["batch_first", "norm_first"]
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: Union[str, Callable[[Tensor], Tensor]] = F.relu,
+        batch_first: bool = False,
+        norm_first: bool = False,
+        device=None,
+        dtype=None,
+        linear1_self_attention_cls: nn.Module = nn.Linear,
+        linear2_self_attention_cls: nn.Module = nn.Linear,
+        linear1_feedforward_cls: nn.Module = nn.Linear,
+        linear2_feedforward_cls: nn.Module = nn.Linear,
+        layer_norm_cls: nn.Module = LayerNorm,
+        layer_norm_eps: float = 1e-5,
+        adaptive_layer_norm=False,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = MultiheadAttention(
+            d_model,  # 512 16
+            nhead,
+            dropout=dropout,
+            batch_first=batch_first,
+            linear1_cls=linear1_self_attention_cls,
+            linear2_cls=linear2_self_attention_cls,
+            **factory_kwargs,
+        )
+        self.linear1 = linear1_feedforward_cls(
+            d_model, dim_feedforward, **factory_kwargs
+        )
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = linear2_feedforward_cls(
+            dim_feedforward, d_model, **factory_kwargs
+        )
+        self.norm_first = norm_first
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        if isinstance(activation, str):
+            activation = _get_activation_fn(activation)
+        elif isinstance(activation, partial):
+            activation = activation(d_model)
+        elif activation == BalancedDoubleSwish:
+            activation = BalancedDoubleSwish(d_model)
+        self.activation = activation
+
+        norm1 = layer_norm_cls(d_model, eps=layer_norm_eps, **factory_kwargs)
+        if layer_norm_cls == IdentityNorm:
+            norm2 = BalancedBasicNorm(d_model, eps=layer_norm_eps, **factory_kwargs)
+        else:
+            norm2 = layer_norm_cls(d_model, eps=layer_norm_eps, **factory_kwargs)
+
+        if adaptive_layer_norm:
+            self.norm1 = AdaptiveLayerNorm(d_model, norm1)
+            self.norm2 = AdaptiveLayerNorm(d_model, norm2)
+        else:
+            self.norm1 = norm1
+            self.norm2 = norm2
+
+    def __setstate__(self, state):
+        super(TransformerEncoderLayer, self).__setstate__(state)
+        if not hasattr(self, "activation"):
+            self.activation = F.relu
+
+    def forward(
+        self,
+        src: Tensor,
+        src_mask: Optional[Tensor] = None,
+        src_key_padding_mask: Optional[Tensor] = None,
+        cache=None,
+    ) -> Tensor:
+        x = src
+        stage_embedding = None
+        x = self.norm1(
+            x + self._sa_block(x, src_mask, src_key_padding_mask, cache=cache),
+            stage_embedding,
+        )
+        x = self.norm2(x + self._ff_block(x), stage_embedding)
+
+        return x
+
+    def _sa_block(
+        self,
+        x: Tensor,
+        attn_mask: Optional[Tensor],
+        key_padding_mask: Optional[Tensor],
+        cache=None,
+    ) -> Tensor:
+        x = self.self_attn(
+            x,
+            x,
+            x,
+            attn_mask=attn_mask,
+            key_padding_mask=key_padding_mask,
+            need_weights=False,
+            cache=cache,
+        )
+        return self.dropout1(x)
+
+    def _ff_block(self, x: Tensor) -> Tensor:
+        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
+        return self.dropout2(x)
+
+
+class AdaptiveLayerNorm(nn.Module):
+    r"""Adaptive Layer Normalization"""
+
+    def __init__(self, d_model, norm) -> None:
+        super(AdaptiveLayerNorm, self).__init__()
+        self.project_layer = nn.Linear(d_model, 2 * d_model)
+        self.norm = norm
+        self.d_model = d_model
+        self.eps = self.norm.eps
+
+    def forward(self, input: Tensor, embedding: Tensor = None) -> Tensor:
+        if isinstance(input, tuple):
+            input, embedding = input
+            weight, bias = torch.split(
+                self.project_layer(embedding),
+                split_size_or_sections=self.d_model,
+                dim=-1,
+            )
+            return (weight * self.norm(input) + bias, embedding)
+
+        weight, bias = torch.split(
+            self.project_layer(embedding),
+            split_size_or_sections=self.d_model,
+            dim=-1,
+        )
+        return weight * self.norm(input) + bias
+
+
+def _get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])

AR/text_processing/phonemizer.py

@@ -0,0 +1,78 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/text_processing/phonemizer.py
+import itertools
+import re
+from typing import Dict
+from typing import List
+
+import regex
+from gruut import sentences
+from gruut.const import Sentence
+from gruut.const import Word
+from AR.text_processing.symbols import SYMBOL_TO_ID
+
+
+class GruutPhonemizer:
+    def __init__(self, language: str):
+        self._phonemizer = sentences
+        self.lang = language
+        self.symbol_to_id = SYMBOL_TO_ID
+        self._special_cases_dict: Dict[str] = {
+            r"\.\.\.": "... ",
+            ";": "; ",
+            ":": ": ",
+            ",": ", ",
+            r"\.": ". ",
+            "!": "! ",
+            r"\?": "? ",
+            "—": "—",
+            "…": "… ",
+            "«": "«",
+            "»": "»",
+        }
+        self._punctuation_regexp: str = (
+            rf"([{''.join(self._special_cases_dict.keys())}])"
+        )
+
+    def _normalize_punctuation(self, text: str) -> str:
+        text = regex.sub(rf"\pZ+{self._punctuation_regexp}", r"\1", text)
+        text = regex.sub(rf"{self._punctuation_regexp}(\pL)", r"\1 \2", text)
+        text = regex.sub(r"\pZ+", r" ", text)
+        return text.strip()
+
+    def _convert_punctuation(self, word: Word) -> str:
+        if not word.phonemes:
+            return ""
+        if word.phonemes[0] in ["‖", "|"]:
+            return word.text.strip()
+
+        phonemes = "".join(word.phonemes)
+        # remove modifier characters ˈˌː with regex
+        phonemes = re.sub(r"[ˈˌː͡]", "", phonemes)
+        return phonemes.strip()
+
+    def phonemize(self, text: str, espeak: bool = False) -> str:
+        text_to_phonemize: str = self._normalize_punctuation(text)
+        sents: List[Sentence] = [
+            sent
+            for sent in self._phonemizer(text_to_phonemize, lang="en-us", espeak=espeak)
+        ]
+        words: List[str] = [
+            self._convert_punctuation(word) for word in itertools.chain(*sents)
+        ]
+        return " ".join(words)
+
+    def transform(self, phonemes):
+        # convert phonemes to ids
+        # dictionary is in symbols.py
+        return [self.symbol_to_id[p] for p in phonemes if p in self.symbol_to_id.keys()]
+
+
+if __name__ == "__main__":
+    phonemizer = GruutPhonemizer("en-us")
+    # text -> IPA
+    phonemes = phonemizer.phonemize("Hello, wor-ld ?")
+    print("phonemes:", phonemes)
+    print("len(phonemes):", len(phonemes))
+    phoneme_ids = phonemizer.transform(phonemes)
+    print("phoneme_ids:", phoneme_ids)
+    print("len(phoneme_ids):", len(phoneme_ids))

AR/text_processing/symbols.py

@@ -0,0 +1,9 @@
+# modified from https://github.com/feng-yufei/shared_debugging_code/blob/main/text_processing/symbols.py
+PAD = "_"
+PUNCTUATION = ';:,.!?¡¿—…"«»“” '
+LETTERS = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
+IPA_LETTERS = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
+SYMBOLS = [PAD] + list(PUNCTUATION) + list(LETTERS) + list(IPA_LETTERS)
+SPACE_ID = SYMBOLS.index(" ")
+SYMBOL_TO_ID = {s: i for i, s in enumerate(SYMBOLS)}
+ID_TO_SYMBOL = {i: s for i, s in enumerate(SYMBOLS)}

AR/utils/__init__.py

@@ -0,0 +1,37 @@
+import re
+
+
+def str2bool(str):
+    return True if str.lower() == 'true' else False
+
+
+def get_newest_ckpt(string_list):
+    # 定义一个正则表达式模式，用于匹配字符串中的数字
+    pattern = r'epoch=(\d+)-step=(\d+)\.ckpt'
+
+    # 使用正则表达式提取每个字符串中的数字信息，并创建一个包含元组的列表
+    extracted_info = []
+    for string in string_list:
+        match = re.match(pattern, string)
+        if match:
+            epoch = int(match.group(1))
+            step = int(match.group(2))
+            extracted_info.append((epoch, step, string))
+    # 按照 epoch 后面的数字和 step 后面的数字进行排序
+    sorted_info = sorted(
+        extracted_info, key=lambda x: (x[0], x[1]), reverse=True)
+    # 获取最新的 ckpt 文件名
+    newest_ckpt = sorted_info[0][2]
+    return newest_ckpt
+
+
+# 文本存在且不为空时 return True
+def check_txt_file(file_path):
+    try:
+        with open(file_path, 'r') as file:
+            text = file.readline().strip()
+        assert text.strip() != ''
+        return text
+    except Exception:
+        return False
+    return False

AR/utils/initialize.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python3
+"""Initialize modules for espnet2 neural networks."""
+import torch
+from typeguard import check_argument_types
+
+
+def initialize(model: torch.nn.Module, init: str):
+    """Initialize weights of a neural network module.
+
+    Parameters are initialized using the given method or distribution.
+
+    Custom initialization routines can be implemented into submodules
+    as function `espnet_initialization_fn` within the custom module.
+
+    Args:
+        model: Target.
+        init: Method of initialization.
+    """
+    assert check_argument_types()
+    print("init with", init)
+
+    # weight init
+    for p in model.parameters():
+        if p.dim() > 1:
+            if init == "xavier_uniform":
+                torch.nn.init.xavier_uniform_(p.data)
+            elif init == "xavier_normal":
+                torch.nn.init.xavier_normal_(p.data)
+            elif init == "kaiming_uniform":
+                torch.nn.init.kaiming_uniform_(p.data, nonlinearity="relu")
+            elif init == "kaiming_normal":
+                torch.nn.init.kaiming_normal_(p.data, nonlinearity="relu")
+            else:
+                raise ValueError("Unknown initialization: " + init)
+    # bias init
+    for name, p in model.named_parameters():
+        if ".bias" in name and p.dim() == 1:
+            p.data.zero_()

AR/utils/io.py

@@ -0,0 +1,34 @@
+import sys
+
+import torch
+import yaml
+
+
+def load_yaml_config(path):
+    with open(path) as f:
+        config = yaml.full_load(f)
+    return config
+
+
+def save_config_to_yaml(config, path):
+    assert path.endswith(".yaml")
+    with open(path, "w") as f:
+        f.write(yaml.dump(config))
+        f.close()
+
+
+def write_args(args, path):
+    args_dict = dict(
+        (name, getattr(args, name)) for name in dir(args) if not name.startswith("_")
+    )
+    with open(path, "a") as args_file:
+        args_file.write("==> torch version: {}\n".format(torch.__version__))
+        args_file.write(
+            "==> cudnn version: {}\n".format(torch.backends.cudnn.version())
+        )
+        args_file.write("==> Cmd:\n")
+        args_file.write(str(sys.argv))
+        args_file.write("\n==> args:\n")
+        for k, v in sorted(args_dict.items()):
+            args_file.write("  %s: %s\n" % (str(k), str(v)))
+        args_file.close()

MODELS/21/1.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60f90a8009bef17813c1902f9bff9724c669880827ab78e8b4ed354c7468a8e3
+size 335918

MODELS/21/11.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f2f64a37122fdf10c6244a6b7e9f08b74b5e9e157fe5056cccc05acb866970f
+size 303150

MODELS/21/191.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d89e44a5635a2aa9da893efb3db4fb8d0c5304a0a2830549c4cd6008258addff
+size 319534

MODELS/21/21.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4b29bb398a9dbed95c50489a2633f90a01c0c4ae1e4432f5d37d388401f9887
+size 155077753

MODELS/21/21.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfb359648e858765e9c1e3f7d51869aec9f607d18efd90d059cb83f1a7988141
+size 84927748

MODELS/22/22.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3632e3d1876f7a8e86850f346338c5e2390d09f382891277acf77a4e1a65a25
+size 155083315

MODELS/22/22.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dfe7fe2765b179db75d8e12bd2b32e1f8d624dcee9a3fecdecbc94904757c29
+size 84927982