xinference 0.16.0__py3-none-any.whl → 0.16.1__py3-none-any.whl

xinference/thirdparty/mlx/flux/autoencoder.py

@@ -0,0 +1,357 @@
+# Copyright © 2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import List
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.upsample import upsample_nearest
+
+
+@dataclass
+class AutoEncoderParams:
+    resolution: int
+    in_channels: int
+    ch: int
+    out_ch: int
+    ch_mult: List[int]
+    num_res_blocks: int
+    z_channels: int
+    scale_factor: float
+    shift_factor: float
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = nn.GroupNorm(
+            num_groups=32,
+            dims=in_channels,
+            eps=1e-6,
+            affine=True,
+            pytorch_compatible=True,
+        )
+        self.q = nn.Linear(in_channels, in_channels)
+        self.k = nn.Linear(in_channels, in_channels)
+        self.v = nn.Linear(in_channels, in_channels)
+        self.proj_out = nn.Linear(in_channels, in_channels)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        B, H, W, C = x.shape
+
+        y = x.reshape(B, 1, -1, C)
+        y = self.norm(y)
+        q = self.q(y)
+        k = self.k(y)
+        v = self.v(y)
+        y = mx.fast.scaled_dot_product_attention(q, k, v, scale=C ** (-0.5))
+        y = self.proj_out(y)
+
+        return x + y.reshape(B, H, W, C)
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+
+        self.norm1 = nn.GroupNorm(
+            num_groups=32,
+            dims=in_channels,
+            eps=1e-6,
+            affine=True,
+            pytorch_compatible=True,
+        )
+        self.conv1 = nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.norm2 = nn.GroupNorm(
+            num_groups=32,
+            dims=out_channels,
+            eps=1e-6,
+            affine=True,
+            pytorch_compatible=True,
+        )
+        self.conv2 = nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if self.in_channels != self.out_channels:
+            self.nin_shortcut = nn.Linear(in_channels, out_channels)
+
+    def __call__(self, x):
+        h = x
+        h = self.norm1(h)
+        h = nn.silu(h)
+        h = self.conv1(h)
+
+        h = self.norm2(h)
+        h = nn.silu(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels, in_channels, kernel_size=3, stride=2, padding=0
+        )
+
+    def __call__(self, x: mx.array):
+        x = mx.pad(x, [(0, 0), (0, 1), (0, 1), (0, 0)])
+        x = self.conv(x)
+        return x
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels, in_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def __call__(self, x: mx.array):
+        x = upsample_nearest(x, (2, 2))
+        x = self.conv(x)
+        return x
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        ch: int,
+        ch_mult: list[int],
+        num_res_blocks: int,
+        z_channels: int,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        # downsampling
+        self.conv_in = nn.Conv2d(
+            in_channels, self.ch, kernel_size=3, stride=1, padding=1
+        )
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = []
+        block_in = self.ch
+        for i_level in range(self.num_resolutions):
+            block = []
+            attn = []  # TODO: Remove the attn, nobody appends anything to it
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            down = {}
+            down["block"] = block
+            down["attn"] = attn
+            if i_level != self.num_resolutions - 1:
+                down["downsample"] = Downsample(block_in)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = {}
+        self.mid["block_1"] = ResnetBlock(in_channels=block_in, out_channels=block_in)
+        self.mid["attn_1"] = AttnBlock(block_in)
+        self.mid["block_2"] = ResnetBlock(in_channels=block_in, out_channels=block_in)
+
+        # end
+        self.norm_out = nn.GroupNorm(
+            num_groups=32, dims=block_in, eps=1e-6, affine=True, pytorch_compatible=True
+        )
+        self.conv_out = nn.Conv2d(
+            block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def __call__(self, x: mx.array):
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level]["block"][i_block](hs[-1])
+
+                # TODO: Remove the attn
+                if len(self.down[i_level]["attn"]) > 0:
+                    h = self.down[i_level]["attn"][i_block](h)
+
+                hs.append(h)
+
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level]["downsample"](hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid["block_1"](h)
+        h = self.mid["attn_1"](h)
+        h = self.mid["block_2"](h)
+
+        # end
+        h = self.norm_out(h)
+        h = nn.silu(h)
+        h = self.conv_out(h)
+
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        ch: int,
+        out_ch: int,
+        ch_mult: list[int],
+        num_res_blocks: int,
+        in_channels: int,
+        resolution: int,
+        z_channels: int,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.ffactor = 2 ** (self.num_resolutions - 1)
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = nn.Conv2d(
+            z_channels, block_in, kernel_size=3, stride=1, padding=1
+        )
+
+        # middle
+        self.mid = {}
+        self.mid["block_1"] = ResnetBlock(in_channels=block_in, out_channels=block_in)
+        self.mid["attn_1"] = AttnBlock(block_in)
+        self.mid["block_2"] = ResnetBlock(in_channels=block_in, out_channels=block_in)
+
+        # upsampling
+        self.up = []
+        for i_level in reversed(range(self.num_resolutions)):
+            block = []
+            attn = []  # TODO: Remove the attn, nobody appends anything to it
+
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            up = {}
+            up["block"] = block
+            up["attn"] = attn
+            if i_level != 0:
+                up["upsample"] = Upsample(block_in)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = nn.GroupNorm(
+            num_groups=32, dims=block_in, eps=1e-6, affine=True, pytorch_compatible=True
+        )
+        self.conv_out = nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def __call__(self, z: mx.array):
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid["block_1"](h)
+        h = self.mid["attn_1"](h)
+        h = self.mid["block_2"](h)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level]["block"][i_block](h)
+
+                # TODO: Remove the attn
+                if len(self.up[i_level]["attn"]) > 0:
+                    h = self.up[i_level]["attn"][i_block](h)
+
+            if i_level != 0:
+                h = self.up[i_level]["upsample"](h)
+
+        # end
+        h = self.norm_out(h)
+        h = nn.silu(h)
+        h = self.conv_out(h)
+
+        return h
+
+
+class DiagonalGaussian(nn.Module):
+    def __call__(self, z: mx.array):
+        mean, logvar = mx.split(z, 2, axis=-1)
+        if self.training:
+            std = mx.exp(0.5 * logvar)
+            eps = mx.random.normal(shape=z.shape, dtype=z.dtype)
+            return mean + std * eps
+        else:
+            return mean
+
+
+class AutoEncoder(nn.Module):
+    def __init__(self, params: AutoEncoderParams):
+        super().__init__()
+        self.encoder = Encoder(
+            resolution=params.resolution,
+            in_channels=params.in_channels,
+            ch=params.ch,
+            ch_mult=params.ch_mult,
+            num_res_blocks=params.num_res_blocks,
+            z_channels=params.z_channels,
+        )
+        self.decoder = Decoder(
+            resolution=params.resolution,
+            in_channels=params.in_channels,
+            ch=params.ch,
+            out_ch=params.out_ch,
+            ch_mult=params.ch_mult,
+            num_res_blocks=params.num_res_blocks,
+            z_channels=params.z_channels,
+        )
+        self.reg = DiagonalGaussian()
+
+        self.scale_factor = params.scale_factor
+        self.shift_factor = params.shift_factor
+
+    def sanitize(self, weights):
+        new_weights = {}
+        for k, w in weights.items():
+            if w.ndim == 4:
+                w = w.transpose(0, 2, 3, 1)
+                w = w.reshape(-1).reshape(w.shape)
+                if w.shape[1:3] == (1, 1):
+                    w = w.squeeze((1, 2))
+            new_weights[k] = w
+        return new_weights
+
+    def encode(self, x: mx.array):
+        z = self.reg(self.encoder(x))
+        z = self.scale_factor * (z - self.shift_factor)
+        return z
+
+    def decode(self, z: mx.array):
+        z = z / self.scale_factor + self.shift_factor
+        return self.decoder(z)
+
+    def __call__(self, x: mx.array):
+        return self.decode(self.encode(x))

xinference/thirdparty/mlx/flux/clip.py

@@ -0,0 +1,154 @@
+# Copyright © 2024 Apple Inc.
+
+from dataclasses import dataclass
+from typing import List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+_ACTIVATIONS = {"quick_gelu": nn.gelu_fast_approx, "gelu": nn.gelu}
+
+
+@dataclass
+class CLIPTextModelConfig:
+    num_layers: int = 23
+    model_dims: int = 1024
+    num_heads: int = 16
+    max_length: int = 77
+    vocab_size: int = 49408
+    hidden_act: str = "quick_gelu"
+
+    @classmethod
+    def from_dict(cls, config):
+        return cls(
+            num_layers=config["num_hidden_layers"],
+            model_dims=config["hidden_size"],
+            num_heads=config["num_attention_heads"],
+            max_length=config["max_position_embeddings"],
+            vocab_size=config["vocab_size"],
+            hidden_act=config["hidden_act"],
+        )
+
+
+@dataclass
+class CLIPOutput:
+    # The last_hidden_state indexed at the EOS token and possibly projected if
+    # the model has a projection layer
+    pooled_output: Optional[mx.array] = None
+
+    # The full sequence output of the transformer after the final layernorm
+    last_hidden_state: Optional[mx.array] = None
+
+    # A list of hidden states corresponding to the outputs of the transformer layers
+    hidden_states: Optional[List[mx.array]] = None
+
+
+class CLIPEncoderLayer(nn.Module):
+    """The transformer encoder layer from CLIP."""
+
+    def __init__(self, model_dims: int, num_heads: int, activation: str):
+        super().__init__()
+
+        self.layer_norm1 = nn.LayerNorm(model_dims)
+        self.layer_norm2 = nn.LayerNorm(model_dims)
+
+        self.attention = nn.MultiHeadAttention(model_dims, num_heads, bias=True)
+
+        self.linear1 = nn.Linear(model_dims, 4 * model_dims)
+        self.linear2 = nn.Linear(4 * model_dims, model_dims)
+
+        self.act = _ACTIVATIONS[activation]
+
+    def __call__(self, x, attn_mask=None):
+        y = self.layer_norm1(x)
+        y = self.attention(y, y, y, attn_mask)
+        x = y + x
+
+        y = self.layer_norm2(x)
+        y = self.linear1(y)
+        y = self.act(y)
+        y = self.linear2(y)
+        x = y + x
+
+        return x
+
+
+class CLIPTextModel(nn.Module):
+    """Implements the text encoder transformer from CLIP."""
+
+    def __init__(self, config: CLIPTextModelConfig):
+        super().__init__()
+
+        self.token_embedding = nn.Embedding(config.vocab_size, config.model_dims)
+        self.position_embedding = nn.Embedding(config.max_length, config.model_dims)
+        self.layers = [
+            CLIPEncoderLayer(config.model_dims, config.num_heads, config.hidden_act)
+            for i in range(config.num_layers)
+        ]
+        self.final_layer_norm = nn.LayerNorm(config.model_dims)
+
+    def _get_mask(self, N, dtype):
+        indices = mx.arange(N)
+        mask = indices[:, None] < indices[None]
+        mask = mask.astype(dtype) * (-6e4 if dtype == mx.float16 else -1e9)
+        return mask
+
+    def sanitize(self, weights):
+        new_weights = {}
+        for key, w in weights.items():
+            # Remove prefixes
+            if key.startswith("text_model."):
+                key = key[11:]
+            if key.startswith("embeddings."):
+                key = key[11:]
+            if key.startswith("encoder."):
+                key = key[8:]
+
+            # Map attention layers
+            if "self_attn." in key:
+                key = key.replace("self_attn.", "attention.")
+            if "q_proj." in key:
+                key = key.replace("q_proj.", "query_proj.")
+            if "k_proj." in key:
+                key = key.replace("k_proj.", "key_proj.")
+            if "v_proj." in key:
+                key = key.replace("v_proj.", "value_proj.")
+
+            # Map ffn layers
+            if "mlp.fc1" in key:
+                key = key.replace("mlp.fc1", "linear1")
+            if "mlp.fc2" in key:
+                key = key.replace("mlp.fc2", "linear2")
+
+            new_weights[key] = w
+
+        return new_weights
+
+    def __call__(self, x):
+        # Extract some shapes
+        B, N = x.shape
+        eos_tokens = x.argmax(-1)
+
+        # Compute the embeddings
+        x = self.token_embedding(x)
+        x = x + self.position_embedding.weight[:N]
+
+        # Compute the features from the transformer
+        mask = self._get_mask(N, x.dtype)
+        hidden_states = []
+        for l in self.layers:
+            x = l(x, mask)
+            hidden_states.append(x)
+
+        # Apply the final layernorm and return
+        x = self.final_layer_norm(x)
+        last_hidden_state = x
+
+        # Select the EOS token
+        pooled_output = x[mx.arange(len(x)), eos_tokens]
+
+        return CLIPOutput(
+            pooled_output=pooled_output,
+            last_hidden_state=last_hidden_state,
+            hidden_states=hidden_states,
+        )

xinference/thirdparty/mlx/flux/datasets.py

@@ -0,0 +1,75 @@
+import json
+from pathlib import Path
+
+from PIL import Image
+
+
+class Dataset:
+    def __getitem__(self, index: int):
+        raise NotImplementedError()
+
+    def __len__(self):
+        raise NotImplementedError()
+
+
+class LocalDataset(Dataset):
+    prompt_key = "prompt"
+
+    def __init__(self, dataset: str, data_file):
+        self.dataset_base = Path(dataset)
+        with open(data_file, "r") as fid:
+            self._data = [json.loads(l) for l in fid]
+
+    def __len__(self):
+        return len(self._data)
+
+    def __getitem__(self, index: int):
+        item = self._data[index]
+        image = Image.open(self.dataset_base / item["image"])
+        return image, item[self.prompt_key]
+
+
+class LegacyDataset(LocalDataset):
+    prompt_key = "text"
+
+    def __init__(self, dataset: str):
+        self.dataset_base = Path(dataset)
+        with open(self.dataset_base / "index.json") as f:
+            self._data = json.load(f)["data"]
+
+
+class HuggingFaceDataset(Dataset):
+
+    def __init__(self, dataset: str):
+        from datasets import load_dataset as hf_load_dataset
+
+        self._df = hf_load_dataset(dataset)["train"]
+
+    def __len__(self):
+        return len(self._df)
+
+    def __getitem__(self, index: int):
+        item = self._df[index]
+        return item["image"], item["prompt"]
+
+
+def load_dataset(dataset: str):
+    dataset_base = Path(dataset)
+    data_file = dataset_base / "train.jsonl"
+    legacy_file = dataset_base / "index.json"
+
+    if data_file.exists():
+        print(f"Load the local dataset {data_file} .", flush=True)
+        dataset = LocalDataset(dataset, data_file)
+    elif legacy_file.exists():
+        print(f"Load the local dataset {legacy_file} .")
+        print()
+        print("     WARNING: 'index.json' is deprecated in favor of 'train.jsonl'.")
+        print("              See the README for details.")
+        print(flush=True)
+        dataset = LegacyDataset(dataset)
+    else:
+        print(f"Load the Hugging Face dataset {dataset} .", flush=True)
+        dataset = HuggingFaceDataset(dataset)
+
+    return dataset