兼容了flash_attention的批量推理,并修复了一些bug GPT_SoVITS/AR/models/t2s_model.py

批量推理备份文件:   GPT_SoVITS/AR/models/t2s_model_batch_only.py

GPT_SoVITS/AR/models/t2s_model.py

@@ -99,7 +99,8 @@ class T2SBlock:
 
         attn = F.scaled_dot_product_attention(q, k, v, ~attn_mask)
 
-        attn = attn.permute(2, 0, 1, 3).reshape(batch_size, -1, self.hidden_dim)
+        attn = attn.permute(2, 0, 1, 3).reshape(batch_size*q_len, self.hidden_dim)
+        attn = attn.view(q_len, batch_size, self.hidden_dim).transpose(1, 0)
         attn = F.linear(attn, self.out_w, self.out_b)
 
         x = F.layer_norm(
@@ -114,15 +115,15 @@ class T2SBlock:
         )
         return x, k_cache, v_cache
 
-    def decode_next_token(self, x, k_cache, v_cache):
+    def decode_next_token(self, x, k_cache, v_cache, attn_mask : torch.Tensor):
         q, k, v = F.linear(x, self.qkv_w, self.qkv_b).chunk(3, dim=-1)
 
         k_cache = torch.cat([k_cache, k], dim=1)
         v_cache = torch.cat([v_cache, v], dim=1)
-        kv_len = k_cache.shape[1]
+        
-
         batch_size = q.shape[0]
         q_len = q.shape[1]
+        kv_len = k_cache.shape[1]
 
         q = q.view(batch_size, q_len, self.num_heads, -1).transpose(1, 2)
         k = k_cache.view(batch_size, kv_len, self.num_heads, -1).transpose(1, 2)
@@ -131,7 +132,8 @@ class T2SBlock:
 
         attn = F.scaled_dot_product_attention(q, k, v)
 
-        attn = attn.permute(2, 0, 1, 3).reshape(batch_size, -1, self.hidden_dim)
+        attn = attn.permute(2, 0, 1, 3).reshape(batch_size*q_len, self.hidden_dim)
+        attn = attn.view(q_len, batch_size, self.hidden_dim).transpose(1, 0)
         attn = F.linear(attn, self.out_w, self.out_b)
 
         x = F.layer_norm(
@@ -164,10 +166,10 @@ class T2STransformer:
         return x, k_cache, v_cache
 
     def decode_next_token(
-        self, x, k_cache: List[torch.Tensor], v_cache: List[torch.Tensor]
+        self, x, k_cache: List[torch.Tensor], v_cache: List[torch.Tensor], attn_mask : torch.Tensor
     ):
         for i in range(self.num_blocks):
-            x, k_cache[i], v_cache[i] = self.blocks[i].decode_next_token(x, k_cache[i], v_cache[i])
+            x, k_cache[i], v_cache[i] = self.blocks[i].decode_next_token(x, k_cache[i], v_cache[i], attn_mask)
         return x, k_cache, v_cache
 
 
@@ -543,12 +545,16 @@ class Text2SemanticDecoder(nn.Module):
         xy_attn_mask = torch.concat([x_attn_mask_pad, y_attn_mask], dim=0).to(
             x.device
         )
-
+        
+        y_list = [None]*y.shape[0]
+        batch_idx_map = list(range(y.shape[0]))
+        idx_list = [None]*y.shape[0]
+        cache_y_emb = y_emb
         for idx in tqdm(range(1500)):
-            if xy_attn_mask is not None:
+            if idx == 0:
                 xy_dec, k_cache, v_cache = self.t2s_transformer.process_prompt(xy_pos, xy_attn_mask)
             else:
-                xy_dec, k_cache, v_cache = self.t2s_transformer.decode_next_token(xy_pos, k_cache, v_cache)
+                xy_dec, k_cache, v_cache = self.t2s_transformer.decode_next_token(xy_pos, k_cache, v_cache, xy_attn_mask)
 
             logits = self.ar_predict_layer(
                 xy_dec[:, -1]
@@ -557,18 +563,51 @@ class Text2SemanticDecoder(nn.Module):
             if idx == 0:
                 xy_attn_mask = None
                 logits = logits[:, :-1]
+                
             samples = sample(
-                logits[0], y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
+                logits, y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
-            )[0].unsqueeze(0)
+            )[0]
 
             y = torch.concat([y, samples], dim=1)
-
+            
-            if early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num:
+            ####### 移除batch中已经生成完毕的序列,进一步优化计算量
+            reserved_idx_of_batch_for_y = None
+            if (self.EOS in samples[:, 0]) or \
+                (self.EOS in torch.argmax(logits, dim=-1)):  ###如果生成到EOS，则停止
+                    l = samples[:, 0]==self.EOS
+                    removed_idx_of_batch_for_y = torch.where(l==True)[0].tolist()
+                    reserved_idx_of_batch_for_y = torch.where(l==False)[0]
+                    # batch_indexs = torch.tensor(batch_idx_map, device=y.device)[removed_idx_of_batch_for_y]
+                    for i in removed_idx_of_batch_for_y:
+                        batch_index = batch_idx_map[i]
+                        idx_list[batch_index] = idx - 1
+                        y_list[batch_index] = y[i, :-1]
+                
+                    batch_idx_map = [batch_idx_map[i] for i in reserved_idx_of_batch_for_y.tolist()]
+                
+            # 只保留batch中未生成完毕的序列 
+            if reserved_idx_of_batch_for_y is not None:
+                # index = torch.LongTensor(batch_idx_map).to(y.device)
+                y = torch.index_select(y, dim=0, index=reserved_idx_of_batch_for_y)
+                if cache_y_emb is not None:
+                    cache_y_emb = torch.index_select(cache_y_emb, dim=0, index=reserved_idx_of_batch_for_y)
+                if k_cache is not None :
+                    for i in range(len(k_cache)):
+                        k_cache[i] = torch.index_select(k_cache[i], dim=0, index=reserved_idx_of_batch_for_y)
+                        v_cache[i] = torch.index_select(v_cache[i], dim=0, index=reserved_idx_of_batch_for_y)
+                
+                
+            if (early_stop_num != -1 and (y.shape[1] - prefix_len) > early_stop_num) or idx==1499:
                 print("use early stop num:", early_stop_num)
                 stop = True
-
+                for i, batch_index in enumerate(batch_idx_map):
-            if torch.argmax(logits, dim=-1)[0] == self.EOS or samples[0, 0] == self.EOS:
+                    batch_index = batch_idx_map[i]
+                    idx_list[batch_index] = idx
+                    y_list[batch_index] = y[i, :-1]
+                
+            if not (None in idx_list):
                 stop = True
+                
             if stop:
                 if y.shape[1]==0:
                     y = torch.concat([y, torch.zeros_like(samples)], dim=1)
@@ -580,6 +619,11 @@ class Text2SemanticDecoder(nn.Module):
             y_emb = self.ar_audio_embedding(y[:, -1:])
             xy_pos = y_emb * self.ar_audio_position.x_scale + self.ar_audio_position.alpha * self.ar_audio_position.pe[:, y_len + idx]
 
+        if (None in idx_list):
+            for i in range(x.shape[0]):
+                if idx_list[i] is None:
+                    idx_list[i] = 1500-1  ###如果没有生成到EOS，就用最大长度代替
+                    
         if ref_free:
-            return y[:, :-1], 0
+            return y_list, [0]*x.shape[0]
-        return y[:, :-1], idx - 1
+        return y_list, idx_list

GPT_SoVITS/AR/models/t2s_model_batch_only.py

@@ -0,0 +1,483 @@
+# 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,
+    dpo_loss,
+    make_reject_y, 
+    get_batch_logps
+)
+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 make_input_data(self, x, x_lens, y, y_lens, bert_feature):
+        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)
+
+        return xy_pos, xy_attn_mask, targets
+
+    def forward(self, x, x_lens, y, y_lens, bert_feature):
+        """
+        x: phoneme_ids
+        y: semantic_ids
+        """
+
+        reject_y, reject_y_lens = make_reject_y(y, y_lens)
+
+        xy_pos, xy_attn_mask, targets = self.make_input_data(x, x_lens, y, y_lens, bert_feature)
+
+        xy_dec, _ = self.h(
+            (xy_pos, None),
+            mask=xy_attn_mask,
+        )
+        x_len = x_lens.max()
+        logits = self.ar_predict_layer(xy_dec[:, x_len:])
+
+        ###### DPO #############
+        reject_xy_pos, reject_xy_attn_mask, reject_targets = self.make_input_data(x, x_lens, reject_y, reject_y_lens, bert_feature)
+
+        reject_xy_dec, _ = self.h(
+            (reject_xy_pos, None),
+            mask=reject_xy_attn_mask,
+        )
+        x_len = x_lens.max()
+        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len:])
+
+        # loss
+        # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
+
+        loss_1 = F.cross_entropy(logits.permute(0, 2, 1), targets, reduction="sum")
+        acc = self.ar_accuracy_metric(logits.permute(0, 2, 1).detach(), targets).item()
+
+        A_logits, R_logits = get_batch_logps(logits, reject_logits, targets, reject_targets)
+        loss_2, _, _ = dpo_loss(A_logits, R_logits, 0, 0, 0.2, reference_free=True)
+        
+        loss = loss_1 + loss_2
+
+        return loss, acc
+
+    def forward_old(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,
+        top_p: 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
+        
+        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,
+        }
+        ###################  first step ##########################
+        if y is not None:
+            y_emb = self.ar_audio_embedding(y)
+            y_len = y_emb.shape[1]
+            prefix_len = y.shape[1]
+            y_pos = self.ar_audio_position(y_emb)
+            xy_pos = torch.concat([x, y_pos], dim=1)
+            cache["y_emb"] = y_emb
+            ref_free = False
+        else:
+            y_emb = None
+            y_len = 0
+            prefix_len = 0
+            y_pos = None
+            xy_pos = x
+            y = torch.zeros(x.shape[0], 0, dtype=torch.int, device=x.device)
+            ref_free = True
+
+        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(
+            x.device
+        )
+        
+        y_list = [None]*y.shape[0]
+        batch_idx_map = list(range(y.shape[0]))
+        idx_list = [None]*y.shape[0]
+        for idx in tqdm(range(1500)):
+            
+            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, y, top_k=top_k, top_p=top_p, repetition_penalty=1.35, temperature=temperature
+            )[0]
+            # 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) 
+
+            # 移除已经生成完毕的序列
+            reserved_idx_of_batch_for_y = None
+            if (self.EOS in torch.argmax(logits, dim=-1)) or \
+                (self.EOS in samples[:, 0]):  ###如果生成到EOS，则停止
+                    l = samples[:, 0]==self.EOS
+                    removed_idx_of_batch_for_y = torch.where(l==True)[0].tolist()
+                    reserved_idx_of_batch_for_y = torch.where(l==False)[0]
+                    # batch_indexs = torch.tensor(batch_idx_map, device=y.device)[removed_idx_of_batch_for_y]
+                    for i in removed_idx_of_batch_for_y:
+                        batch_index = batch_idx_map[i]
+                        idx_list[batch_index] = idx - 1
+                        y_list[batch_index] = y[i, :-1]
+                
+                    batch_idx_map = [batch_idx_map[i] for i in reserved_idx_of_batch_for_y.tolist()]
+                
+            # 只保留未生成完毕的序列 
+            if reserved_idx_of_batch_for_y is not None:
+                # index = torch.LongTensor(batch_idx_map).to(y.device)
+                y = torch.index_select(y, dim=0, index=reserved_idx_of_batch_for_y)
+                if cache["y_emb"] is not None:
+                    cache["y_emb"] = torch.index_select(cache["y_emb"], dim=0, index=reserved_idx_of_batch_for_y)
+                if cache["k"] is not None:
+                    for i in range(self.num_layers):
+                        # 因为kv转置了，所以batch dim是1
+                        cache["k"][i] = torch.index_select(cache["k"][i], dim=1, index=reserved_idx_of_batch_for_y)
+                        cache["v"][i] = torch.index_select(cache["v"][i], dim=1, index=reserved_idx_of_batch_for_y)
+                
+                
+            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 not (None in idx_list):
+                # 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")
+                if y.shape[1]==0:
+                    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
+            
+            ####################### update next step ###################################
+            cache["first_infer"] = 0
+            if cache["y_emb"] is not None:
+                y_emb = torch.cat(
+                    [cache["y_emb"], self.ar_audio_embedding(y[:, -1:])], dim = 1
+                )
+                cache["y_emb"] = y_emb
+                y_pos = self.ar_audio_position(y_emb)
+                xy_pos = y_pos[:, -1:]
+            else:
+                y_emb = self.ar_audio_embedding(y[:, -1:])
+                cache["y_emb"] = y_emb
+                y_pos = self.ar_audio_position(y_emb)
+                xy_pos = y_pos
+            y_len = y_pos.shape[1]
+
+            ###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
+            )
+            
+        if (None in idx_list):
+            for i in range(x.shape[0]):
+                if idx_list[i] is None:
+                    idx_list[i] = 1500-1  ###如果没有生成到EOS，就用最大长度代替
+                    
+        if ref_free:
+            return y_list, [0]*x.shape[0]
+        return y_list, idx_list