xinference 0.14.4.post1__py3-none-any.whl → 0.15.1__py3-none-any.whl

xinference/model/llm/transformers/llama_2.py

@@ -1,108 +0,0 @@
-# Copyright 2022-2023 XProbe Inc.
-#
-# 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.
-
-from typing import List, Optional
-
-from ....types import LoRA
-from ..llm_family import LLMFamilyV1, LLMSpecV1
-from .core import PytorchChatModel, PytorchModel, PytorchModelConfig
-
-
-class LlamaPytorchModel(PytorchModel):
-    def __init__(
-        self,
-        model_uid: str,
-        model_family: "LLMFamilyV1",
-        model_spec: "LLMSpecV1",
-        quantization: str,
-        model_path: str,
-        pytorch_model_config: Optional[PytorchModelConfig] = None,
-        peft_model: Optional[List[LoRA]] = None,
-    ):
-        super().__init__(
-            model_uid,
-            model_family,
-            model_spec,
-            quantization,
-            model_path,
-            pytorch_model_config=pytorch_model_config,
-            peft_model=peft_model,
-        )
-
-    def _load_model(self, **kwargs):
-        model, tokenizer = super()._load_model(**kwargs)
-        # Llama has no pad token by default
-        # https://github.com/huggingface/transformers/blob/07998ef39926b76d3f6667025535d0859eed61c3/docs/source/en/llm_tutorial.md?plain=1#L125
-        tokenizer.pad_token = tokenizer.eos_token
-        model.config.eos_token_id = tokenizer.eos_token_id
-        model.config.pad_token_id = tokenizer.pad_token_id
-        return model, tokenizer
-
-    @classmethod
-    def match(
-        cls, llm_family: "LLMFamilyV1", llm_spec: "LLMSpecV1", quantization: str
-    ) -> bool:
-        if llm_spec.model_format != "pytorch":
-            return False
-        model_family = llm_family.model_family or llm_family.model_name
-        if "llama-2" not in model_family:
-            return False
-        if "generate" not in llm_family.model_ability:
-            return False
-        return True
-
-
-class LlamaPytorchChatModel(PytorchChatModel):
-    def __init__(
-        self,
-        model_uid: str,
-        model_family: "LLMFamilyV1",
-        model_spec: "LLMSpecV1",
-        quantization: str,
-        model_path: str,
-        pytorch_model_config: Optional["PytorchModelConfig"] = None,
-        peft_model: Optional[List[LoRA]] = None,
-    ):
-        super().__init__(
-            model_uid,
-            model_family,
-            model_spec,
-            quantization,
-            model_path,
-            peft_model=peft_model,
-            pytorch_model_config=pytorch_model_config,
-        )
-        self._use_fast_tokenizer = False
-
-    def _load_model(self, **kwargs):
-        model, tokenizer = super()._load_model(**kwargs)
-        # Llama has no pad token by default
-        # https://github.com/huggingface/transformers/blob/07998ef39926b76d3f6667025535d0859eed61c3/docs/source/en/llm_tutorial.md?plain=1#L125
-        tokenizer.pad_token = tokenizer.eos_token
-        model.config.eos_token_id = tokenizer.eos_token_id
-        model.config.pad_token_id = tokenizer.pad_token_id
-        return model, tokenizer
-
-    @classmethod
-    def match(
-        cls, llm_family: "LLMFamilyV1", llm_spec: "LLMSpecV1", quantization: str
-    ) -> bool:
-        if llm_spec.model_format != "pytorch":
-            return False
-        model_family = llm_family.model_family or llm_family.model_name
-        if "llama-2" not in model_family:
-            return False
-        if "chat" not in llm_family.model_ability:
-            return False
-        return True

xinference/thirdparty/fish_speech/fish_speech/models/vqgan/lit_module.py

@@ -1,442 +0,0 @@
-import itertools
-import math
-from typing import Any, Callable
-
-import lightning as L
-import torch
-import torch.nn.functional as F
-# import wandb
-from lightning.pytorch.loggers import TensorBoardLogger, WandbLogger
-from matplotlib import pyplot as plt
-from torch import nn
-
-from fish_speech.models.vqgan.modules.discriminator import Discriminator
-from fish_speech.models.vqgan.modules.wavenet import WaveNet
-from fish_speech.models.vqgan.utils import avg_with_mask, plot_mel, sequence_mask
-
-
-class VQGAN(L.LightningModule):
-    def __init__(
-        self,
-        optimizer: Callable,
-        lr_scheduler: Callable,
-        encoder: WaveNet,
-        quantizer: nn.Module,
-        decoder: WaveNet,
-        discriminator: Discriminator,
-        vocoder: nn.Module,
-        encode_mel_transform: nn.Module,
-        gt_mel_transform: nn.Module,
-        weight_adv: float = 1.0,
-        weight_vq: float = 1.0,
-        weight_mel: float = 1.0,
-        sampling_rate: int = 44100,
-        freeze_encoder: bool = False,
-    ):
-        super().__init__()
-
-        # Model parameters
-        self.optimizer_builder = optimizer
-        self.lr_scheduler_builder = lr_scheduler
-
-        # Modules
-        self.encoder = encoder
-        self.quantizer = quantizer
-        self.decoder = decoder
-        self.vocoder = vocoder
-        self.discriminator = discriminator
-        self.encode_mel_transform = encode_mel_transform
-        self.gt_mel_transform = gt_mel_transform
-
-        # A simple linear layer to project quality to condition channels
-        self.quality_projection = nn.Linear(1, 768)
-
-        # Freeze vocoder
-        for param in self.vocoder.parameters():
-            param.requires_grad = False
-
-        # Loss weights
-        self.weight_adv = weight_adv
-        self.weight_vq = weight_vq
-        self.weight_mel = weight_mel
-
-        # Other parameters
-        self.sampling_rate = sampling_rate
-
-        # Disable strict loading
-        self.strict_loading = False
-
-        # If encoder is frozen
-        if freeze_encoder:
-            for param in self.encoder.parameters():
-                param.requires_grad = False
-
-            for param in self.quantizer.parameters():
-                param.requires_grad = False
-
-        self.automatic_optimization = False
-
-    def on_save_checkpoint(self, checkpoint):
-        # Do not save vocoder
-        state_dict = checkpoint["state_dict"]
-        for name in list(state_dict.keys()):
-            if "vocoder" in name:
-                state_dict.pop(name)
-
-    def configure_optimizers(self):
-        optimizer_generator = self.optimizer_builder(
-            itertools.chain(
-                self.encoder.parameters(),
-                self.quantizer.parameters(),
-                self.decoder.parameters(),
-                self.quality_projection.parameters(),
-            )
-        )
-        optimizer_discriminator = self.optimizer_builder(
-            self.discriminator.parameters()
-        )
-
-        lr_scheduler_generator = self.lr_scheduler_builder(optimizer_generator)
-        lr_scheduler_discriminator = self.lr_scheduler_builder(optimizer_discriminator)
-
-        return (
-            {
-                "optimizer": optimizer_generator,
-                "lr_scheduler": {
-                    "scheduler": lr_scheduler_generator,
-                    "interval": "step",
-                    "name": "optimizer/generator",
-                },
-            },
-            {
-                "optimizer": optimizer_discriminator,
-                "lr_scheduler": {
-                    "scheduler": lr_scheduler_discriminator,
-                    "interval": "step",
-                    "name": "optimizer/discriminator",
-                },
-            },
-        )
-
-    def training_step(self, batch, batch_idx):
-        optim_g, optim_d = self.optimizers()
-
-        audios, audio_lengths = batch["audios"], batch["audio_lengths"]
-
-        audios = audios.float()
-        audios = audios[:, None, :]
-
-        with torch.no_grad():
-            encoded_mels = self.encode_mel_transform(audios)
-            gt_mels = self.gt_mel_transform(audios)
-            quality = ((gt_mels.mean(-1) > -8).sum(-1) - 90) / 10
-            quality = quality.unsqueeze(-1)
-
-        mel_lengths = audio_lengths // self.gt_mel_transform.hop_length
-        mel_masks = sequence_mask(mel_lengths, gt_mels.shape[2])
-        mel_masks_float_conv = mel_masks[:, None, :].float()
-        gt_mels = gt_mels * mel_masks_float_conv
-        encoded_mels = encoded_mels * mel_masks_float_conv
-
-        # Encode
-        encoded_features = self.encoder(encoded_mels) * mel_masks_float_conv
-
-        # Quantize
-        vq_result = self.quantizer(encoded_features)
-        loss_vq = getattr("vq_result", "loss", 0.0)
-        vq_recon_features = vq_result.z * mel_masks_float_conv
-        vq_recon_features = (
-            vq_recon_features + self.quality_projection(quality)[:, :, None]
-        )
-
-        # VQ Decode
-        gen_mel = (
-            self.decoder(
-                torch.randn_like(vq_recon_features) * mel_masks_float_conv,
-                condition=vq_recon_features,
-            )
-            * mel_masks_float_conv
-        )
-
-        # Discriminator
-        real_logits = self.discriminator(gt_mels)
-        fake_logits = self.discriminator(gen_mel.detach())
-        d_mask = F.interpolate(
-            mel_masks_float_conv, size=(real_logits.shape[2],), mode="nearest"
-        )
-
-        loss_real = avg_with_mask((real_logits - 1) ** 2, d_mask)
-        loss_fake = avg_with_mask(fake_logits**2, d_mask)
-
-        loss_d = loss_real + loss_fake
-
-        self.log(
-            "train/discriminator/loss",
-            loss_d,
-            on_step=True,
-            on_epoch=False,
-            prog_bar=True,
-            logger=True,
-        )
-
-        # Discriminator backward
-        optim_d.zero_grad()
-        self.manual_backward(loss_d)
-        self.clip_gradients(
-            optim_d, gradient_clip_val=1000.0, gradient_clip_algorithm="norm"
-        )
-        optim_d.step()
-
-        # Mel Loss, applying l1, using a weighted sum
-        mel_distance = (
-            gen_mel - gt_mels
-        ).abs()  # * 0.5 + self.ssim(gen_mel, gt_mels) * 0.5
-        loss_mel_low_freq = avg_with_mask(mel_distance[:, :40, :], mel_masks_float_conv)
-        loss_mel_mid_freq = avg_with_mask(
-            mel_distance[:, 40:70, :], mel_masks_float_conv
-        )
-        loss_mel_high_freq = avg_with_mask(
-            mel_distance[:, 70:, :], mel_masks_float_conv
-        )
-        loss_mel = (
-            loss_mel_low_freq * 0.6 + loss_mel_mid_freq * 0.3 + loss_mel_high_freq * 0.1
-        )
-
-        # Adversarial Loss
-        fake_logits = self.discriminator(gen_mel)
-        loss_adv = avg_with_mask((fake_logits - 1) ** 2, d_mask)
-
-        # Total loss
-        loss = (
-            self.weight_vq * loss_vq
-            + self.weight_mel * loss_mel
-            + self.weight_adv * loss_adv
-        )
-
-        # Log losses
-        self.log(
-            "train/generator/loss",
-            loss,
-            on_step=True,
-            on_epoch=False,
-            prog_bar=True,
-            logger=True,
-        )
-        self.log(
-            "train/generator/loss_vq",
-            loss_vq,
-            on_step=True,
-            on_epoch=False,
-            prog_bar=False,
-            logger=True,
-        )
-        self.log(
-            "train/generator/loss_mel",
-            loss_mel,
-            on_step=True,
-            on_epoch=False,
-            prog_bar=False,
-            logger=True,
-        )
-        self.log(
-            "train/generator/loss_adv",
-            loss_adv,
-            on_step=True,
-            on_epoch=False,
-            prog_bar=False,
-            logger=True,
-        )
-
-        # Generator backward
-        optim_g.zero_grad()
-        self.manual_backward(loss)
-        self.clip_gradients(
-            optim_g, gradient_clip_val=1000.0, gradient_clip_algorithm="norm"
-        )
-        optim_g.step()
-
-        scheduler_g, scheduler_d = self.lr_schedulers()
-        scheduler_g.step()
-        scheduler_d.step()
-
-    def validation_step(self, batch: Any, batch_idx: int):
-        audios, audio_lengths = batch["audios"], batch["audio_lengths"]
-
-        audios = audios.float()
-        audios = audios[:, None, :]
-
-        encoded_mels = self.encode_mel_transform(audios)
-        gt_mels = self.gt_mel_transform(audios)
-
-        mel_lengths = audio_lengths // self.gt_mel_transform.hop_length
-        mel_masks = sequence_mask(mel_lengths, gt_mels.shape[2])
-        mel_masks_float_conv = mel_masks[:, None, :].float()
-        gt_mels = gt_mels * mel_masks_float_conv
-        encoded_mels = encoded_mels * mel_masks_float_conv
-
-        # Encode
-        encoded_features = self.encoder(encoded_mels) * mel_masks_float_conv
-
-        # Quantize
-        vq_recon_features = self.quantizer(encoded_features).z * mel_masks_float_conv
-        vq_recon_features = (
-            vq_recon_features
-            + self.quality_projection(
-                torch.ones(
-                    vq_recon_features.shape[0], 1, device=vq_recon_features.device
-                )
-                * 2
-            )[:, :, None]
-        )
-
-        # VQ Decode
-        gen_aux_mels = (
-            self.decoder(
-                torch.randn_like(vq_recon_features) * mel_masks_float_conv,
-                condition=vq_recon_features,
-            )
-            * mel_masks_float_conv
-        )
-        loss_mel = avg_with_mask((gen_aux_mels - gt_mels).abs(), mel_masks_float_conv)
-
-        self.log(
-            "val/loss_mel",
-            loss_mel,
-            on_step=False,
-            on_epoch=True,
-            prog_bar=False,
-            logger=True,
-            sync_dist=True,
-        )
-
-        recon_audios = self.vocoder(gt_mels)
-        gen_aux_audios = self.vocoder(gen_aux_mels)
-
-        # only log the first batch
-        if batch_idx != 0:
-            return
-
-        for idx, (
-            gt_mel,
-            gen_aux_mel,
-            audio,
-            gen_aux_audio,
-            recon_audio,
-            audio_len,
-        ) in enumerate(
-            zip(
-                gt_mels,
-                gen_aux_mels,
-                audios.cpu().float(),
-                gen_aux_audios.cpu().float(),
-                recon_audios.cpu().float(),
-                audio_lengths,
-            )
-        ):
-            if idx > 4:
-                break
-
-            mel_len = audio_len // self.gt_mel_transform.hop_length
-
-            image_mels = plot_mel(
-                [
-                    gt_mel[:, :mel_len],
-                    gen_aux_mel[:, :mel_len],
-                ],
-                [
-                    "Ground-Truth",
-                    "Auxiliary",
-                ],
-            )
-
-            if isinstance(self.logger, WandbLogger):
-                self.logger.experiment.log(
-                    {
-                        "reconstruction_mel": wandb.Image(image_mels, caption="mels"),
-                        "wavs": [
-                            wandb.Audio(
-                                audio[0, :audio_len],
-                                sample_rate=self.sampling_rate,
-                                caption="gt",
-                            ),
-                            wandb.Audio(
-                                gen_aux_audio[0, :audio_len],
-                                sample_rate=self.sampling_rate,
-                                caption="aux",
-                            ),
-                            wandb.Audio(
-                                recon_audio[0, :audio_len],
-                                sample_rate=self.sampling_rate,
-                                caption="recon",
-                            ),
-                        ],
-                    },
-                )
-
-            if isinstance(self.logger, TensorBoardLogger):
-                self.logger.experiment.add_figure(
-                    f"sample-{idx}/mels",
-                    image_mels,
-                    global_step=self.global_step,
-                )
-                self.logger.experiment.add_audio(
-                    f"sample-{idx}/wavs/gt",
-                    audio[0, :audio_len],
-                    self.global_step,
-                    sample_rate=self.sampling_rate,
-                )
-                self.logger.experiment.add_audio(
-                    f"sample-{idx}/wavs/gen",
-                    gen_aux_audio[0, :audio_len],
-                    self.global_step,
-                    sample_rate=self.sampling_rate,
-                )
-                self.logger.experiment.add_audio(
-                    f"sample-{idx}/wavs/recon",
-                    recon_audio[0, :audio_len],
-                    self.global_step,
-                    sample_rate=self.sampling_rate,
-                )
-
-            plt.close(image_mels)
-
-    def encode(self, audios, audio_lengths):
-        audios = audios.float()
-
-        mels = self.encode_mel_transform(audios)
-        mel_lengths = audio_lengths // self.encode_mel_transform.hop_length
-        mel_masks = sequence_mask(mel_lengths, mels.shape[2])
-        mel_masks_float_conv = mel_masks[:, None, :].float()
-        mels = mels * mel_masks_float_conv
-
-        # Encode
-        encoded_features = self.encoder(mels) * mel_masks_float_conv
-        feature_lengths = mel_lengths // math.prod(self.quantizer.downsample_factor)
-
-        return self.quantizer.encode(encoded_features), feature_lengths
-
-    def decode(self, indices, feature_lengths, return_audios=False):
-        factor = math.prod(self.quantizer.downsample_factor)
-        mel_masks = sequence_mask(feature_lengths * factor, indices.shape[2] * factor)
-        mel_masks_float_conv = mel_masks[:, None, :].float()
-
-        z = self.quantizer.decode(indices) * mel_masks_float_conv
-        z = (
-            z
-            + self.quality_projection(torch.ones(z.shape[0], 1, device=z.device) * 2)[
-                :, :, None
-            ]
-        )
-
-        gen_mel = (
-            self.decoder(
-                torch.randn_like(z) * mel_masks_float_conv,
-                condition=z,
-            )
-            * mel_masks_float_conv
-        )
-
-        if return_audios:
-            return self.vocoder(gen_mel)
-
-        return gen_mel

xinference/thirdparty/fish_speech/fish_speech/models/vqgan/modules/discriminator.py

@@ -1,44 +0,0 @@
-import torch
-from torch import nn
-from torch.nn.utils.parametrizations import weight_norm
-
-
-class Discriminator(nn.Module):
-    def __init__(self):
-        super().__init__()
-
-        blocks = []
-        convs = [
-            (1, 64, (3, 9), 1, (1, 4)),
-            (64, 128, (3, 9), (1, 2), (1, 4)),
-            (128, 256, (3, 9), (1, 2), (1, 4)),
-            (256, 512, (3, 9), (1, 2), (1, 4)),
-            (512, 1024, (3, 3), 1, (1, 1)),
-            (1024, 1, (3, 3), 1, (1, 1)),
-        ]
-
-        for idx, (in_channels, out_channels, kernel_size, stride, padding) in enumerate(
-            convs
-        ):
-            blocks.append(
-                weight_norm(
-                    nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding)
-                )
-            )
-
-            if idx != len(convs) - 1:
-                blocks.append(nn.SiLU(inplace=True))
-
-        self.blocks = nn.Sequential(*blocks)
-
-    def forward(self, x):
-        return self.blocks(x[:, None])[:, 0]
-
-
-if __name__ == "__main__":
-    model = Discriminator()
-    print(sum(p.numel() for p in model.parameters()) / 1_000_000)
-    x = torch.randn(1, 128, 1024)
-    y = model(x)
-    print(y.shape)
-    print(y)

xinference/thirdparty/fish_speech/fish_speech/models/vqgan/modules/reference.py

@@ -1,115 +0,0 @@
-from typing import Optional
-
-import torch
-import torch.nn.functional as F
-from torch import nn
-
-from fish_speech.utils import autocast_exclude_mps
-
-from .wavenet import WaveNet
-
-
-class ReferenceEncoder(WaveNet):
-    def __init__(
-        self,
-        input_channels: Optional[int] = None,
-        output_channels: Optional[int] = None,
-        residual_channels: int = 512,
-        residual_layers: int = 20,
-        dilation_cycle: Optional[int] = 4,
-        num_heads: int = 8,
-        latent_len: int = 4,
-    ):
-        super().__init__(
-            input_channels=input_channels,
-            residual_channels=residual_channels,
-            residual_layers=residual_layers,
-            dilation_cycle=dilation_cycle,
-        )
-
-        self.head_dim = residual_channels // num_heads
-        self.num_heads = num_heads
-
-        self.latent_len = latent_len
-        self.latent = nn.Parameter(torch.zeros(1, self.latent_len, residual_channels))
-
-        self.q = nn.Linear(residual_channels, residual_channels, bias=True)
-        self.kv = nn.Linear(residual_channels, residual_channels * 2, bias=True)
-        self.q_norm = nn.LayerNorm(self.head_dim)
-        self.k_norm = nn.LayerNorm(self.head_dim)
-        self.proj = nn.Linear(residual_channels, residual_channels)
-        self.proj_drop = nn.Dropout(0.1)
-
-        self.norm = nn.LayerNorm(residual_channels)
-        self.mlp = nn.Sequential(
-            nn.Linear(residual_channels, residual_channels * 4),
-            nn.SiLU(),
-            nn.Linear(residual_channels * 4, residual_channels),
-        )
-        self.output_projection_attn = nn.Linear(residual_channels, output_channels)
-
-        torch.nn.init.trunc_normal_(self.latent, std=0.02)
-        self.apply(self.init_weights)
-
-    def init_weights(self, m):
-        if isinstance(m, nn.Linear):
-            torch.nn.init.trunc_normal_(m.weight, std=0.02)
-            if m.bias is not None:
-                torch.nn.init.constant_(m.bias, 0)
-
-    def forward(self, x, attn_mask=None):
-        x = super().forward(x).mT
-        B, N, C = x.shape
-
-        # Calculate mask
-        if attn_mask is not None:
-            assert attn_mask.shape == (B, N) and attn_mask.dtype == torch.bool
-
-            attn_mask = attn_mask[:, None, None, :].expand(
-                B, self.num_heads, self.latent_len, N
-            )
-
-        q_latent = self.latent.expand(B, -1, -1)
-        q = (
-            self.q(q_latent)
-            .reshape(B, self.latent_len, self.num_heads, self.head_dim)
-            .transpose(1, 2)
-        )
-
-        kv = (
-            self.kv(x)
-            .reshape(B, N, 2, self.num_heads, self.head_dim)
-            .permute(2, 0, 3, 1, 4)
-        )
-        k, v = kv.unbind(0)
-
-        q, k = self.q_norm(q), self.k_norm(k)
-        x = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
-
-        x = x.transpose(1, 2).reshape(B, self.latent_len, C)
-        x = self.proj(x)
-        x = self.proj_drop(x)
-
-        x = x + self.mlp(self.norm(x))
-        x = self.output_projection_attn(x)
-        x = x.mean(1)
-
-        return x
-
-
-if __name__ == "__main__":
-    with autocast_exclude_mps(device_type="cpu", dtype=torch.bfloat16):
-        model = ReferenceEncoder(
-            input_channels=128,
-            output_channels=64,
-            residual_channels=384,
-            residual_layers=20,
-            dilation_cycle=4,
-            num_heads=8,
-        )
-        x = torch.randn(4, 128, 64)
-        mask = torch.ones(4, 64, dtype=torch.bool)
-        y = model(x, mask)
-        print(y.shape)
-        loss = F.mse_loss(y, torch.randn(4, 64))
-        loss.backward()