xinference 0.16.0__py3-none-any.whl → 0.16.2__py3-none-any.whl

xinference/thirdparty/mlx/flux/trainer.py

@@ -0,0 +1,98 @@
+import mlx.core as mx
+import numpy as np
+from PIL import Image, ImageFile
+from tqdm import tqdm
+
+from .datasets import Dataset
+from .flux import FluxPipeline
+
+
+class Trainer:
+
+    def __init__(self, flux: FluxPipeline, dataset: Dataset, args):
+        self.flux = flux
+        self.dataset = dataset
+        self.args = args
+        self.latents = []
+        self.t5_features = []
+        self.clip_features = []
+
+    def _random_crop_resize(self, img):
+        resolution = self.args.resolution
+        width, height = img.size
+
+        a, b, c, d = mx.random.uniform(shape=(4,), stream=mx.cpu).tolist()
+
+        # Random crop the input image between 0.8 to 1.0 of its original dimensions
+        crop_size = (
+            max((0.8 + 0.2 * a) * width, resolution[0]),
+            max((0.8 + 0.2 * b) * height, resolution[1]),
+        )
+        pan = (width - crop_size[0], height - crop_size[1])
+        img = img.crop(
+            (
+                pan[0] * c,
+                pan[1] * d,
+                crop_size[0] + pan[0] * c,
+                crop_size[1] + pan[1] * d,
+            )
+        )
+
+        # Fit the largest rectangle with the ratio of resolution in the image
+        # rectangle.
+        width, height = crop_size
+        ratio = resolution[0] / resolution[1]
+        r1 = (height * ratio, height)
+        r2 = (width, width / ratio)
+        r = r1 if r1[0] <= width else r2
+        img = img.crop(
+            (
+                (width - r[0]) / 2,
+                (height - r[1]) / 2,
+                (width + r[0]) / 2,
+                (height + r[1]) / 2,
+            )
+        )
+
+        # Finally resize the image to resolution
+        img = img.resize(resolution, Image.LANCZOS)
+
+        return mx.array(np.array(img))
+
+    def _encode_image(self, input_img: ImageFile.ImageFile, num_augmentations: int):
+        for i in range(num_augmentations):
+            img = self._random_crop_resize(input_img)
+            img = (img[:, :, :3].astype(self.flux.dtype) / 255) * 2 - 1
+            x_0 = self.flux.ae.encode(img[None])
+            x_0 = x_0.astype(self.flux.dtype)
+            mx.eval(x_0)
+            self.latents.append(x_0)
+
+    def _encode_prompt(self, prompt):
+        t5_tok, clip_tok = self.flux.tokenize([prompt])
+        t5_feat = self.flux.t5(t5_tok)
+        clip_feat = self.flux.clip(clip_tok).pooled_output
+        mx.eval(t5_feat, clip_feat)
+        self.t5_features.append(t5_feat)
+        self.clip_features.append(clip_feat)
+
+    def encode_dataset(self):
+        """Encode the images & prompt in the latent space to prepare for training."""
+        self.flux.ae.eval()
+        for image, prompt in tqdm(self.dataset, desc="encode dataset"):
+            self._encode_image(image, self.args.num_augmentations)
+            self._encode_prompt(prompt)
+
+    def iterate(self, batch_size):
+        xs = mx.concatenate(self.latents)
+        t5 = mx.concatenate(self.t5_features)
+        clip = mx.concatenate(self.clip_features)
+        mx.eval(xs, t5, clip)
+        n_aug = self.args.num_augmentations
+        while True:
+            x_indices = mx.random.permutation(len(self.latents))
+            c_indices = x_indices // n_aug
+            for i in range(0, len(self.latents), batch_size):
+                x_i = x_indices[i : i + batch_size]
+                c_i = c_indices[i : i + batch_size]
+                yield xs[x_i], t5[c_i], clip[c_i]

xinference/thirdparty/mlx/flux/utils.py

@@ -0,0 +1,179 @@
+# Copyright © 2024 Apple Inc.
+
+import json
+import os
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+
+from .autoencoder import AutoEncoder, AutoEncoderParams
+from .clip import CLIPTextModel, CLIPTextModelConfig
+from .model import Flux, FluxParams
+from .t5 import T5Config, T5Encoder
+from .tokenizers import CLIPTokenizer, T5Tokenizer
+
+
+@dataclass
+class ModelSpec:
+    params: FluxParams
+    ae_params: AutoEncoderParams
+    ckpt_path: Optional[str]
+    ae_path: Optional[str]
+    repo_id: Optional[str]
+    repo_flow: Optional[str]
+    repo_ae: Optional[str]
+
+
+configs = {
+    "flux-dev": ModelSpec(
+        repo_id="black-forest-labs/FLUX.1-dev",
+        repo_flow="flux1-dev.safetensors",
+        repo_ae="ae.safetensors",
+        ckpt_path=os.getenv("FLUX_DEV"),
+        params=FluxParams(
+            in_channels=64,
+            vec_in_dim=768,
+            context_in_dim=4096,
+            hidden_size=3072,
+            mlp_ratio=4.0,
+            num_heads=24,
+            depth=19,
+            depth_single_blocks=38,
+            axes_dim=[16, 56, 56],
+            theta=10_000,
+            qkv_bias=True,
+            guidance_embed=True,
+        ),
+        ae_path=os.getenv("AE"),
+        ae_params=AutoEncoderParams(
+            resolution=256,
+            in_channels=3,
+            ch=128,
+            out_ch=3,
+            ch_mult=[1, 2, 4, 4],
+            num_res_blocks=2,
+            z_channels=16,
+            scale_factor=0.3611,
+            shift_factor=0.1159,
+        ),
+    ),
+    "flux-schnell": ModelSpec(
+        repo_id="black-forest-labs/FLUX.1-schnell",
+        repo_flow="flux1-schnell.safetensors",
+        repo_ae="ae.safetensors",
+        ckpt_path=os.getenv("FLUX_SCHNELL"),
+        params=FluxParams(
+            in_channels=64,
+            vec_in_dim=768,
+            context_in_dim=4096,
+            hidden_size=3072,
+            mlp_ratio=4.0,
+            num_heads=24,
+            depth=19,
+            depth_single_blocks=38,
+            axes_dim=[16, 56, 56],
+            theta=10_000,
+            qkv_bias=True,
+            guidance_embed=False,
+        ),
+        ae_path=os.getenv("AE"),
+        ae_params=AutoEncoderParams(
+            resolution=256,
+            in_channels=3,
+            ch=128,
+            out_ch=3,
+            ch_mult=[1, 2, 4, 4],
+            num_res_blocks=2,
+            z_channels=16,
+            scale_factor=0.3611,
+            shift_factor=0.1159,
+        ),
+    ),
+}
+
+
+def load_flow_model(name: str, ckpt_path: str):
+    # Make the model
+    model = Flux(configs[name].params)
+
+    # Load the checkpoint if needed
+    if os.path.isdir(ckpt_path):
+        ckpt_path = os.path.join(ckpt_path, configs[name].repo_flow)
+    weights = mx.load(ckpt_path)
+    weights = model.sanitize(weights)
+    model.load_weights(list(weights.items()))
+
+    return model
+
+
+def load_ae(name: str, ckpt_path: str):
+    # Make the autoencoder
+    ae = AutoEncoder(configs[name].ae_params)
+
+    # Load the checkpoint if needed
+    ckpt_path = os.path.join(ckpt_path, "ae.safetensors")
+    weights = mx.load(ckpt_path)
+    weights = ae.sanitize(weights)
+    ae.load_weights(list(weights.items()))
+
+    return ae
+
+
+def load_clip(name: str, ckpt_path: str):
+    config_path = os.path.join(ckpt_path, "text_encoder/config.json")
+    with open(config_path) as f:
+        config = CLIPTextModelConfig.from_dict(json.load(f))
+
+    # Make the clip text encoder
+    clip = CLIPTextModel(config)
+
+    ckpt_path = os.path.join(ckpt_path, "text_encoder/model.safetensors")
+    weights = mx.load(ckpt_path)
+    weights = clip.sanitize(weights)
+    clip.load_weights(list(weights.items()))
+
+    return clip
+
+
+def load_t5(name: str, ckpt_path: str):
+    config_path = os.path.join(ckpt_path, "text_encoder_2/config.json")
+    with open(config_path) as f:
+        config = T5Config.from_dict(json.load(f))
+
+    # Make the T5 model
+    t5 = T5Encoder(config)
+
+    model_index = os.path.join(ckpt_path, "text_encoder_2/model.safetensors.index.json")
+    weight_files = set()
+    with open(model_index) as f:
+        for _, w in json.load(f)["weight_map"].items():
+            weight_files.add(w)
+    weights = {}
+    for w in weight_files:
+        w = f"text_encoder_2/{w}"
+        w = os.path.join(ckpt_path, w)
+        weights.update(mx.load(w))
+    weights = t5.sanitize(weights)
+    t5.load_weights(list(weights.items()))
+
+    return t5
+
+
+def load_clip_tokenizer(name: str, ckpt_path: str):
+    vocab_file = os.path.join(ckpt_path, "tokenizer/vocab.json")
+    with open(vocab_file, encoding="utf-8") as f:
+        vocab = json.load(f)
+
+    merges_file = os.path.join(ckpt_path, "tokenizer/merges.txt")
+    with open(merges_file, encoding="utf-8") as f:
+        bpe_merges = f.read().strip().split("\n")[1 : 49152 - 256 - 2 + 1]
+    bpe_merges = [tuple(m.split()) for m in bpe_merges]
+    bpe_ranks = dict(map(reversed, enumerate(bpe_merges)))
+
+    return CLIPTokenizer(bpe_ranks, vocab, max_length=77)
+
+
+def load_t5_tokenizer(name: str, ckpt_path: str, pad: bool = True):
+    model_file = os.path.join(ckpt_path, "tokenizer_2/spiece.model")
+    return T5Tokenizer(model_file, 256 if "schnell" in name else 512)

xinference/web/ui/build/asset-manifest.json

@@ -1,14 +1,14 @@
 {
   "files": {
     "main.css": "./static/css/main.5061c4c3.css",
-    "main.js": "./static/js/main.f7da0140.js",
+    "main.js": "./static/js/main.2f269bb3.js",
     "static/media/icon.webp": "./static/media/icon.4603d52c63041e5dfbfd.webp",
     "index.html": "./index.html",
     "main.5061c4c3.css.map": "./static/css/main.5061c4c3.css.map",
-    "main.f7da0140.js.map": "./static/js/main.f7da0140.js.map"
+    "main.2f269bb3.js.map": "./static/js/main.2f269bb3.js.map"
   },
   "entrypoints": [
     "static/css/main.5061c4c3.css",
-    "static/js/main.f7da0140.js"
+    "static/js/main.2f269bb3.js"
   ]
 }

xinference/web/ui/build/index.html

@@ -1 +1 @@
-<!doctype html><html lang="en"><head><meta charset="utf-8"/><link rel="icon" href="./favicon.svg"/><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="theme-color" content="#000000"/><meta name="description" content="Web site created using create-react-app"/><link rel="apple-touch-icon" href="./logo192.png"/><link rel="manifest" href="./manifest.json"/><title>Xinference</title><script defer="defer" src="./static/js/main.f7da0140.js"></script><link href="./static/css/main.5061c4c3.css" rel="stylesheet"></head><body><noscript>You need to enable JavaScript to run this app.</noscript><div id="root"></div></body></html>
+<!doctype html><html lang="en"><head><meta charset="utf-8"/><link rel="icon" href="./favicon.svg"/><meta name="viewport" content="width=device-width,initial-scale=1"/><meta name="theme-color" content="#000000"/><meta name="description" content="Web site created using create-react-app"/><link rel="apple-touch-icon" href="./logo192.png"/><link rel="manifest" href="./manifest.json"/><title>Xinference</title><script defer="defer" src="./static/js/main.2f269bb3.js"></script><link href="./static/css/main.5061c4c3.css" rel="stylesheet"></head><body><noscript>You need to enable JavaScript to run this app.</noscript><div id="root"></div></body></html>