xinference 1.9.0__py3-none-any.whl → 1.10.0__py3-none-any.whl

xinference/thirdparty/cosyvoice/transformer/upsample_encoder.py

@@ -56,16 +56,11 @@ class Upsample1D(nn.Module):
         # In this mode, first repeat interpolate, than conv with stride=1
         self.conv = nn.Conv1d(self.channels, self.out_channels, stride * 2 + 1, stride=1, padding=0)
 
-    def forward(self, inputs: torch.Tensor, input_lengths: torch.Tensor, conv_cache: torch.Tensor = torch.zeros(0, 0, 0)) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    def forward(self, inputs: torch.Tensor, input_lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
         outputs = F.interpolate(inputs, scale_factor=float(self.stride), mode="nearest")
-        if conv_cache.size(2) == 0:
-            outputs = F.pad(outputs, (self.stride * 2, 0), value=0.0)
-        else:
-            assert conv_cache.size(2) == self.stride * 2
-            outputs = torch.concat([conv_cache, outputs], dim=2)
-        conv_cache_new = outputs[:, :, -self.stride * 2:]
+        outputs = F.pad(outputs, (self.stride * 2, 0), value=0.0)
         outputs = self.conv(outputs)
-        return outputs, input_lengths * self.stride, conv_cache_new
+        return outputs, input_lengths * self.stride
 
 
 class PreLookaheadLayer(nn.Module):
@@ -83,7 +78,7 @@ class PreLookaheadLayer(nn.Module):
             kernel_size=3, stride=1, padding=0,
         )
 
-    def forward(self, inputs: torch.Tensor, context: torch.Tensor = torch.zeros(0, 0, 0), conv2_cache: torch.Tensor = torch.zeros(0, 0, 0)) -> Tuple[torch.Tensor, torch.Tensor]:
+    def forward(self, inputs: torch.Tensor, context: torch.Tensor = torch.zeros(0, 0, 0)) -> torch.Tensor:
         """
         inputs: (batch_size, seq_len, channels)
         """
@@ -93,22 +88,18 @@ class PreLookaheadLayer(nn.Module):
         if context.size(2) == 0:
             outputs = F.pad(outputs, (0, self.pre_lookahead_len), mode='constant', value=0.0)
         else:
+            assert self.training is False, 'you have passed context, make sure that you are running inference mode'
             assert context.size(2) == self.pre_lookahead_len
             outputs = F.pad(torch.concat([outputs, context], dim=2), (0, self.pre_lookahead_len - context.size(2)), mode='constant', value=0.0)
         outputs = F.leaky_relu(self.conv1(outputs))
         # outputs
-        if conv2_cache.size(2) == 0:
-            outputs = F.pad(outputs, (self.conv2.kernel_size[0] - 1, 0), mode='constant', value=0.0)
-        else:
-            assert conv2_cache.size(2) == self.conv2.kernel_size[0] - 1
-            outputs = torch.concat([conv2_cache, outputs], dim=2)
-        conv2_cache_new = outputs[:, :, -(self.conv2.kernel_size[0] - 1):]
+        outputs = F.pad(outputs, (self.conv2.kernel_size[0] - 1, 0), mode='constant', value=0.0)
         outputs = self.conv2(outputs)
         outputs = outputs.transpose(1, 2).contiguous()
 
         # residual connection
         outputs = outputs + inputs
-        return outputs, conv2_cache_new
+        return outputs
 
 
 class UpsampleConformerEncoder(torch.nn.Module):
@@ -253,6 +244,7 @@ class UpsampleConformerEncoder(torch.nn.Module):
         self,
         xs: torch.Tensor,
         xs_lens: torch.Tensor,
+        context: torch.Tensor = torch.zeros(0, 0, 0),
         decoding_chunk_size: int = 0,
         num_decoding_left_chunks: int = -1,
         streaming: bool = False,
@@ -285,15 +277,19 @@ class UpsampleConformerEncoder(torch.nn.Module):
         if self.global_cmvn is not None:
             xs = self.global_cmvn(xs)
         xs, pos_emb, masks = self.embed(xs, masks)
+        if context.size(1) != 0:
+            assert self.training is False, 'you have passed context, make sure that you are running inference mode'
+            context_masks = torch.ones(1, 1, context.size(1)).to(masks)
+            context, _, _ = self.embed(context, context_masks, offset=xs.size(1))
         mask_pad = masks  # (B, 1, T/subsample_rate)
         chunk_masks = add_optional_chunk_mask(xs, masks, False, False, 0, self.static_chunk_size if streaming is True else 0, -1)
         # lookahead + conformer encoder
-        xs, _ = self.pre_lookahead_layer(xs)
+        xs = self.pre_lookahead_layer(xs, context=context)
         xs = self.forward_layers(xs, chunk_masks, pos_emb, mask_pad)
 
         # upsample + conformer encoder
         xs = xs.transpose(1, 2).contiguous()
-        xs, xs_lens, _ = self.up_layer(xs, xs_lens)
+        xs, xs_lens = self.up_layer(xs, xs_lens)
         xs = xs.transpose(1, 2).contiguous()
         T = xs.size(1)
         masks = ~make_pad_mask(xs_lens, T).unsqueeze(1)  # (B, 1, T)
@@ -322,100 +318,3 @@ class UpsampleConformerEncoder(torch.nn.Module):
         for layer in self.up_encoders:
             xs, chunk_masks, _, _ = layer(xs, chunk_masks, pos_emb, mask_pad)
         return xs
-
-    @torch.jit.export
-    def forward_chunk(
-        self,
-        xs: torch.Tensor,
-        xs_lens: torch.Tensor,
-        offset: int = 0,
-        context: torch.Tensor = torch.zeros(0, 0, 0),
-        pre_lookahead_layer_conv2_cache: torch.Tensor = torch.zeros(0, 0, 0),
-        encoders_kv_cache: torch.Tensor = torch.zeros(0, 0, 0, 0, 0),
-        upsample_offset: int = 0,
-        upsample_conv_cache: torch.Tensor = torch.zeros(0, 0, 0),
-        upsample_kv_cache: torch.Tensor = torch.zeros(0, 0, 0, 0, 0)
-    ) -> Tuple[torch.Tensor, torch.Tensor, Tuple[int, torch.Tensor, torch.Tensor, int, torch.Tensor, torch.Tensor]]:
-        """Embed positions in tensor.
-
-        Args:
-            xs: padded input tensor (B, T, D)
-            xs_lens: input length (B)
-            decoding_chunk_size: decoding chunk size for dynamic chunk
-                0: default for training, use random dynamic chunk.
-                <0: for decoding, use full chunk.
-                >0: for decoding, use fixed chunk size as set.
-            num_decoding_left_chunks: number of left chunks, this is for decoding,
-            the chunk size is decoding_chunk_size.
-                >=0: use num_decoding_left_chunks
-                <0: use all left chunks
-        Returns:
-            encoder output tensor xs, and subsampled masks
-            xs: padded output tensor (B, T' ~= T/subsample_rate, D)
-            masks: torch.Tensor batch padding mask after subsample
-                (B, 1, T' ~= T/subsample_rate)
-        NOTE(xcsong):
-            We pass the `__call__` method of the modules instead of `forward` to the
-            checkpointing API because `__call__` attaches all the hooks of the module.
-            https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2
-        """
-        assert xs.size(0) == 1
-        # tmp_masks is just for interface compatibility
-        tmp_masks = torch.ones(1,
-                               xs.size(1),
-                               device=xs.device,
-                               dtype=torch.bool)
-        tmp_masks = tmp_masks.unsqueeze(1)
-        if self.global_cmvn is not None:
-            xs = self.global_cmvn(xs)
-        # NOTE(xcsong): Before embed, shape(xs) is (b=1, time, mel-dim)
-        xs, pos_emb, _ = self.embed(xs, tmp_masks, offset)
-        offset += xs.size(1)
-        tmp_masks = torch.ones(1,
-                               context.size(1),
-                               device=context.device,
-                               dtype=torch.bool)
-        tmp_masks = tmp_masks.unsqueeze(1)
-        if context.size(1) != 0:
-            context, _, _ = self.embed(context, tmp_masks, offset)
-
-        # lookahead + conformer encoder
-        xs, pre_lookahead_layer_conv2_cache = self.pre_lookahead_layer(xs, context, pre_lookahead_layer_conv2_cache)
-        # NOTE in cache mode we do not need to call add_optional_chunk_mask
-        chunk_masks = torch.ones((1, xs.size(1), offset), dtype=torch.bool, device=xs.device)
-        mask_pad = torch.ones((0, 0, 0), dtype=torch.bool, device=xs.device)
-        encoders_kv_cache_list = []
-        for index, layer in enumerate(self.encoders):
-            xs, chunk_masks, encoders_kv_cache_new, _ = layer(xs, chunk_masks, pos_emb, mask_pad, encoders_kv_cache[index])
-            encoders_kv_cache_list.append(encoders_kv_cache_new)
-        encoders_kv_cache = torch.stack(encoders_kv_cache_list, dim=0)
-
-        # upsample
-        xs = xs.transpose(1, 2).contiguous()
-        xs, xs_lens, upsample_conv_cache = self.up_layer(xs, xs_lens, upsample_conv_cache)
-        xs = xs.transpose(1, 2).contiguous()
-
-        # tmp_masks is just for interface compatibility
-        tmp_masks = torch.ones(1,
-                               xs.size(1),
-                               device=xs.device,
-                               dtype=torch.bool)
-        tmp_masks = tmp_masks.unsqueeze(1)
-        xs, pos_emb, masks = self.up_embed(xs, tmp_masks, upsample_offset)
-        upsample_offset += xs.size(1)
-
-        # conformer encoder
-        chunk_masks = torch.ones((1, xs.size(1), upsample_offset), dtype=torch.bool, device=xs.device)
-        mask_pad = torch.ones((0, 0, 0), dtype=torch.bool, device=xs.device)
-        upsample_kv_cache_list = []
-        for index, layer in enumerate(self.up_encoders):
-            xs, chunk_masks, upsample_kv_cache_new, _ = layer(xs, chunk_masks, pos_emb, mask_pad, upsample_kv_cache[index])
-            upsample_kv_cache_list.append(upsample_kv_cache_new)
-        upsample_kv_cache = torch.stack(upsample_kv_cache_list, dim=0)
-
-        if self.normalize_before:
-            xs = self.after_norm(xs)
-        # Here we assume the mask is not changed in encoder layers, so just
-        # return the masks before encoder layers, and the masks will be used
-        # for cross attention with decoder later
-        return xs, masks, (offset, pre_lookahead_layer_conv2_cache, encoders_kv_cache, upsample_offset, upsample_conv_cache, upsample_kv_cache)

xinference/thirdparty/cosyvoice/utils/common.py

@@ -1,5 +1,6 @@
 # Copyright (c) 2020 Mobvoi Inc (Binbin Zhang)
 #               2024 Alibaba Inc (authors: Xiang Lyu)
+#               2025 Alibaba Inc (authors: Xiang Lyu, Bofan Zhou)
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -15,6 +16,7 @@
 # Modified from ESPnet(https://github.com/espnet/espnet)
 """Unility functions for Transformer."""
 
+import queue
 import random
 from typing import List
 
@@ -164,3 +166,21 @@ def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
     #     chunk_masks = (1.0 - chunk_masks) * torch.finfo(dtype).min
     mask = (1.0 - mask) * -1.0e+10
     return mask
+
+
+class TrtContextWrapper:
+    def __init__(self, trt_engine, trt_concurrent=1, device='cuda:0'):
+        self.trt_context_pool = queue.Queue(maxsize=trt_concurrent)
+        self.trt_engine = trt_engine
+        for _ in range(trt_concurrent):
+            trt_context = trt_engine.create_execution_context()
+            trt_stream = torch.cuda.stream(torch.cuda.Stream(device))
+            assert trt_context is not None, 'failed to create trt context, maybe not enough CUDA memory, try reduce current trt concurrent {}'.format(trt_concurrent)
+            self.trt_context_pool.put([trt_context, trt_stream])
+        assert self.trt_context_pool.empty() is False, 'no avaialbe estimator context'
+
+    def acquire_estimator(self):
+        return self.trt_context_pool.get(), self.trt_engine
+
+    def release_estimator(self, context, stream):
+        self.trt_context_pool.put([context, stream])

xinference/thirdparty/cosyvoice/utils/executor.py

@@ -25,14 +25,16 @@ from cosyvoice.utils.train_utils import update_parameter_and_lr, log_per_step, l
 
 class Executor:
 
-    def __init__(self, gan: bool = False):
+    def __init__(self, gan: bool = False, ref_model: torch.nn.Module = None, dpo_loss: torch.nn.Module = None):
         self.gan = gan
+        self.ref_model = ref_model
+        self.dpo_loss = dpo_loss
         self.step = 0
         self.epoch = 0
         self.rank = int(os.environ.get('RANK', 0))
         self.device = torch.device('cuda:{}'.format(self.rank))
 
-    def train_one_epoc(self, model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, scaler, group_join):
+    def train_one_epoc(self, model, optimizer, scheduler, train_data_loader, cv_data_loader, writer, info_dict, scaler, group_join, ref_model=None):
         ''' Train one epoch
         '''
 
@@ -44,6 +46,8 @@ class Executor:
         # torch.nn.parallel.DistributedDataParallel to be able to train
         # with uneven inputs across participating processes.
         model.train()
+        if self.ref_model is not None:
+            self.ref_model.eval()
         model_context = model.join if info_dict['train_engine'] == 'torch_ddp' else nullcontext
         with model_context():
             for batch_idx, batch_dict in enumerate(train_data_loader):
@@ -65,7 +69,7 @@ class Executor:
                     context = nullcontext
 
                 with context():
-                    info_dict = batch_forward(model, batch_dict, scaler, info_dict)
+                    info_dict = batch_forward(model, batch_dict, scaler, info_dict, ref_model=self.ref_model, dpo_loss=self.dpo_loss)
                     info_dict = batch_backward(model, scaler, info_dict)
 
                 info_dict = update_parameter_and_lr(model, optimizer, scheduler, scaler, info_dict)
@@ -162,7 +166,7 @@ class Executor:
             for k, v in info_dict['loss_dict'].items():
                 if k not in total_loss_dict:
                     total_loss_dict[k] = []
-                total_loss_dict[k].append(v.item() * num_utts)
+                total_loss_dict[k].append(v.mean().item() * num_utts)
             log_per_step(None, info_dict)
         for k, v in total_loss_dict.items():
             total_loss_dict[k] = sum(v) / total_num_utts

xinference/thirdparty/cosyvoice/utils/file_utils.py

@@ -1,5 +1,6 @@
 # Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang)
 #               2024 Alibaba Inc (authors: Xiang Lyu, Zetao Hu)
+#               2025 Alibaba Inc (authors: Xiang Lyu, Yabin Li)
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -13,7 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
 import json
+import torch
 import torchaudio
 import logging
 logging.getLogger('matplotlib').setLevel(logging.WARNING)
@@ -56,7 +59,7 @@ def convert_onnx_to_trt(trt_model, trt_kwargs, onnx_model, fp16):
     network = builder.create_network(network_flags)
     parser = trt.OnnxParser(network, logger)
     config = builder.create_builder_config()
-    config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 33)  # 8GB
+    config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, 1 << 32)  # 4GB
     if fp16:
         config.set_flag(trt.BuilderFlag.FP16)
     profile = builder.create_optimization_profile()
@@ -83,3 +86,44 @@ def convert_onnx_to_trt(trt_model, trt_kwargs, onnx_model, fp16):
     with open(trt_model, "wb") as f:
         f.write(engine_bytes)
     logging.info("Succesfully convert onnx to trt...")
+
+
+def export_cosyvoice2_vllm(model, model_path, device):
+    if os.path.exists(model_path):
+        return
+    pad_to = DEFAULT_VOCAB_PADDING_SIZE = 64
+    vocab_size = model.speech_embedding.num_embeddings
+    feature_size = model.speech_embedding.embedding_dim
+    pad_vocab_size = ((vocab_size + pad_to - 1) // pad_to) * pad_to
+
+    dtype = torch.bfloat16
+    # lm_head
+    new_lm_head = torch.nn.Linear(in_features=feature_size, out_features=pad_vocab_size, bias=True)
+    with torch.no_grad():
+        new_lm_head.weight[:vocab_size] = model.llm_decoder.weight
+        new_lm_head.bias[:vocab_size] = model.llm_decoder.bias
+        new_lm_head.weight[vocab_size:] = 0
+        new_lm_head.bias[vocab_size:] = 0
+    model.llm.model.lm_head = new_lm_head
+    new_codec_embed = torch.nn.Linear(in_features=feature_size, out_features=pad_vocab_size)
+    # embed_tokens
+    embed_tokens = model.llm.model.model.embed_tokens
+    with torch.no_grad():
+        new_codec_embed.weight[:vocab_size] = model.speech_embedding.weight
+        new_codec_embed.weight[vocab_size:] = 0
+    model.llm.model.set_input_embeddings(new_codec_embed)
+    model.llm.model.to(device)
+    model.llm.model.to(dtype)
+    tmp_vocab_size = model.llm.model.config.vocab_size
+    tmp_tie_embedding = model.llm.model.config.tie_word_embeddings
+    del model.llm.model.generation_config.eos_token_id
+    del model.llm.model.config.bos_token_id
+    del model.llm.model.config.eos_token_id
+    model.llm.model.config.vocab_size = pad_vocab_size
+    model.llm.model.config.tie_word_embeddings = False
+    model.llm.model.config.use_bias = True
+    model.llm.model.save_pretrained(model_path)
+    os.system('sed -i s@Qwen2ForCausalLM@CosyVoice2ForCausalLM@g {}/config.json'.format(os.path.abspath(model_path)))
+    model.llm.model.config.vocab_size = tmp_vocab_size
+    model.llm.model.config.tie_word_embeddings = tmp_tie_embedding
+    model.llm.model.set_input_embeddings(embed_tokens)

xinference/thirdparty/cosyvoice/utils/losses.py

@@ -1,5 +1,6 @@
 import torch
 import torch.nn.functional as F
+from typing import Tuple
 
 
 def tpr_loss(disc_real_outputs, disc_generated_outputs, tau):
@@ -18,3 +19,39 @@ def mel_loss(real_speech, generated_speech, mel_transforms):
         mel_g = transform(generated_speech)
         loss += F.l1_loss(mel_g, mel_r)
     return loss
+
+
+class DPOLoss(torch.nn.Module):
+    """
+    DPO Loss
+    """
+
+    def __init__(self, beta: float, label_smoothing: float = 0.0, ipo: bool = False) -> None:
+        super().__init__()
+        self.beta = beta
+        self.label_smoothing = label_smoothing
+        self.ipo = ipo
+
+    def forward(
+        self,
+        policy_chosen_logps: torch.Tensor,
+        policy_rejected_logps: torch.Tensor,
+        reference_chosen_logps: torch.Tensor,
+        reference_rejected_logps: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        pi_logratios = policy_chosen_logps - policy_rejected_logps
+        ref_logratios = reference_chosen_logps - reference_rejected_logps
+        logits = pi_logratios - ref_logratios
+        if self.ipo:
+            losses = (logits - 1 / (2 * self.beta)) ** 2  # Eq. 17 of https://arxiv.org/pdf/2310.12036v2.pdf
+        else:
+            # Eq. 3 https://ericmitchell.ai/cdpo.pdf; label_smoothing=0 gives original DPO (Eq. 7 of https://arxiv.org/pdf/2305.18290.pdf)
+            losses = (
+                -F.logsigmoid(self.beta * logits) * (1 - self.label_smoothing)
+                - F.logsigmoid(-self.beta * logits) * self.label_smoothing
+            )
+        loss = losses.mean()
+        chosen_rewards = self.beta * (policy_chosen_logps - reference_chosen_logps).detach()
+        rejected_rewards = self.beta * (policy_rejected_logps - reference_rejected_logps).detach()
+
+        return loss, chosen_rewards, rejected_rewards

xinference/thirdparty/cosyvoice/utils/mask.py

@@ -86,7 +86,7 @@ def subsequent_mask(
     return mask
 
 
-def subsequent_chunk_mask(
+def subsequent_chunk_mask_deprecated(
         size: int,
         chunk_size: int,
         num_left_chunks: int = -1,
@@ -124,6 +124,40 @@ def subsequent_chunk_mask(
     return ret
 
 
+def subsequent_chunk_mask(
+        size: int,
+        chunk_size: int,
+        num_left_chunks: int = -1,
+        device: torch.device = torch.device("cpu"),
+) -> torch.Tensor:
+    """Create mask for subsequent steps (size, size) with chunk size,
+       this is for streaming encoder
+
+    Args:
+        size (int): size of mask
+        chunk_size (int): size of chunk
+        num_left_chunks (int): number of left chunks
+            <0: use full chunk
+            >=0: use num_left_chunks
+        device (torch.device): "cpu" or "cuda" or torch.Tensor.device
+
+    Returns:
+        torch.Tensor: mask
+
+    Examples:
+        >>> subsequent_chunk_mask(4, 2)
+        [[1, 1, 0, 0],
+         [1, 1, 0, 0],
+         [1, 1, 1, 1],
+         [1, 1, 1, 1]]
+    """
+    # NOTE this modified implementation meets onnx export requirements, but it doesn't support num_left_chunks
+    pos_idx = torch.arange(size, device=device)
+    block_value = (torch.div(pos_idx, chunk_size, rounding_mode='trunc') + 1) * chunk_size
+    ret = pos_idx.unsqueeze(0) < block_value.unsqueeze(1)
+    return ret
+
+
 def add_optional_chunk_mask(xs: torch.Tensor,
                             masks: torch.Tensor,
                             use_dynamic_chunk: bool,

xinference/thirdparty/cosyvoice/utils/train_utils.py

@@ -50,10 +50,10 @@ def init_distributed(args):
     return world_size, local_rank, rank
 
 
-def init_dataset_and_dataloader(args, configs, gan):
+def init_dataset_and_dataloader(args, configs, gan, dpo):
     data_pipeline = configs['data_pipeline_gan'] if gan is True else configs['data_pipeline']
-    train_dataset = Dataset(args.train_data, data_pipeline=data_pipeline, mode='train', gan=gan, shuffle=True, partition=True)
-    cv_dataset = Dataset(args.cv_data, data_pipeline=data_pipeline, mode='train', gan=gan, shuffle=False, partition=False)
+    train_dataset = Dataset(args.train_data, data_pipeline=data_pipeline, mode='train', gan=gan, dpo=dpo, shuffle=True, partition=True)
+    cv_dataset = Dataset(args.cv_data, data_pipeline=data_pipeline, mode='train', gan=gan, dpo=dpo, shuffle=False, partition=False)
 
     # do not use persistent_workers=True, as whisper tokenizer opens tiktoken file each time when the for loop starts
     train_data_loader = DataLoader(train_dataset,
@@ -71,7 +71,7 @@ def init_dataset_and_dataloader(args, configs, gan):
 
 def check_modify_and_save_config(args, configs):
     if args.train_engine == "torch_ddp":
-        configs['train_conf']["dtype"] = 'fp32'
+        configs['train_conf']["dtype"] = 'bf16' if args.use_amp is True else 'fp32'
     else:
         with open(args.deepspeed_config, 'r') as fin:
             ds_configs = json.load(fin)
@@ -235,7 +235,7 @@ def cosyvoice_join(group_join, info_dict):
         return False
 
 
-def batch_forward(model, batch, scaler, info_dict):
+def batch_forward(model, batch, scaler, info_dict, ref_model=None, dpo_loss=None):
     device = int(os.environ.get('LOCAL_RANK', 0))
 
     dtype = info_dict["dtype"]
@@ -247,12 +247,30 @@ def batch_forward(model, batch, scaler, info_dict):
         dtype = torch.float32
 
     if info_dict['train_engine'] == 'torch_ddp':
-        autocast = torch.cuda.amp.autocast(enabled=scaler is not None)
+        autocast = torch.cuda.amp.autocast(enabled=scaler is not None, dtype=dtype)
     else:
         autocast = torch.cuda.amp.autocast(enabled=True, dtype=dtype, cache_enabled=False)
 
     with autocast:
         info_dict['loss_dict'] = model(batch, device)
+        if ref_model is not None and dpo_loss is not None:
+            chosen_logps = info_dict['loss_dict']["chosen_logps"]
+            rejected_logps = info_dict['loss_dict']["rejected_logps"]
+            sft_loss = info_dict['loss_dict']['loss']
+            with torch.no_grad():
+                ref_loss_dict = ref_model(batch, device)
+            reference_chosen_logps = ref_loss_dict["chosen_logps"]
+            reference_rejected_logps = ref_loss_dict["rejected_logps"]
+            preference_loss, chosen_reward, reject_reward = dpo_loss(
+                chosen_logps, rejected_logps, reference_chosen_logps, reference_rejected_logps
+            )
+            dpo_acc = (chosen_reward > reject_reward).float().mean()
+            info_dict['loss_dict']["loss"] = preference_loss + sft_loss
+            info_dict['loss_dict']["sft_loss"] = sft_loss
+            info_dict['loss_dict']["dpo_loss"] = preference_loss
+            info_dict['loss_dict']["dpo_acc"] = dpo_acc
+            info_dict['loss_dict']["chosen_reward"] = chosen_reward.mean()
+            info_dict['loss_dict']["reject_reward"] = reject_reward.mean()
     return info_dict
 
 

xinference/thirdparty/cosyvoice/vllm/cosyvoice2.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/qwen2/modeling_qwen2.py
+# Copyright 2024 The Qwen team.
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# 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.
+"""Inference-only Qwen2 model compatible with HuggingFace weights."""
+from vllm.model_executor.models.qwen2 import *
+
+
+class CosyVoice2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.lora_config = lora_config
+
+        self.quant_config = quant_config
+        self.model = Qwen2Model(vllm_config=vllm_config,
+                                prefix=maybe_prefix(prefix, "model"))
+
+        if get_pp_group().is_last_rank:
+            if config.tie_word_embeddings:
+                self.lm_head = self.model.embed_tokens
+            else:
+                self.lm_head = ParallelLMHead(config.vocab_size,
+                                              config.hidden_size,
+                                              True,
+                                              quant_config=quant_config,
+                                              prefix=maybe_prefix(
+                                                  prefix, "lm_head"))
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata, self.lm_head.bias)
+        return logits
+
+    def load_weights(self, weights: Iterable[tuple[str,
+                                                   torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head."]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights)