File size: 13,583 Bytes
4f03424 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | # Adapted from MLX-VLM 95b01ccad2d9f65a9e87f6a87bd1c5df69626261.
# Copyright (c) 2025 Prince Canuma. MIT: see MLX_VLM_LICENSE.txt.
# Modifications: text-to-image prefix KV cache and sequential component loading.
"""Generation and reference-image editing for Qwen-Image-2.1."""
from __future__ import annotations
import json
import math
import gc
from collections.abc import Sequence
from pathlib import Path
import mlx.core as mx
import numpy as np
from PIL import Image
from mlx_vlm.models.qwen_image.config import QwenImageVariant, get_variant
from mlx_vlm.models.qwen_image.download import download_model, validate_model_layout
from mlx_vlm.models.qwen_image.kv_cache import QwenImageKVCache
from mlx_vlm.models.qwen_image.scheduler import FlowMatchEulerDiscreteScheduler
from mlx_vlm.models.qwen_image.text_encoder import QwenImageTextEncoder
from mlx_vlm.models.qwen_image.weights import _read_quant, load_text_encoder, load_transformer, load_vae
def _image_dimensions(resolution: int, ratio: float) -> tuple[int, int]:
width = math.sqrt(resolution * resolution * ratio)
return max(32, round(width / 32) * 32), max(32, round(width / ratio / 32) * 32)
class QwenImagePipeline:
def __init__(
self,
*,
variant: QwenImageVariant,
model_path: str | Path,
text_encoder,
transformer,
vae,
phase_offload: bool = True,
) -> None:
self.variant = variant
self.model_path = Path(model_path)
self.text_encoder = QwenImageTextEncoder(
model=text_encoder, model_path=self.model_path
)
self.transformer = transformer
self.vae = vae
self.phase_offload = phase_offload
cfg = json.loads((self.model_path / "vae" / "config.json").read_text())
self.z_dim = cfg["z_dim"]
self.latents_mean = mx.array(cfg["latents_mean"]).reshape(
1, self.z_dim, 1, 1, 1
)
self.latents_std = mx.array(cfg["latents_std"]).reshape(1, self.z_dim, 1, 1, 1)
self.quantization_config = _read_quant(self.model_path / "transformer")
scheduler_path = self.model_path / "scheduler" / "scheduler_config.json"
scheduler_config = (
json.loads(scheduler_path.read_text()) if scheduler_path.exists() else {}
)
self.scheduler_config = {
key: scheduler_config[key]
for key in (
"base_shift",
"max_shift",
"base_image_seq_len",
"max_image_seq_len",
"num_train_timesteps",
"shift_terminal",
)
if key in scheduler_config
}
@classmethod
def from_pretrained(
cls,
variant: str | QwenImageVariant = "qwen-image-2.1",
*,
model_path: str | Path | None = None,
download: bool = True,
token: str | None = None,
revision: str | None = None,
force_download: bool = False,
phase_offload: bool = True,
) -> "QwenImagePipeline":
variant = get_variant(variant)
if model_path is None:
if not download:
raise ValueError("model_path is required when download=False")
model_path = download_model(
variant, token=token, revision=revision, force_download=force_download
)
model_path = validate_model_layout(model_path)
return cls(
variant=variant,
model_path=model_path,
text_encoder=None if phase_offload else load_text_encoder(model_path),
transformer=None if phase_offload else load_transformer(model_path, variant),
vae=None if phase_offload else load_vae(model_path),
phase_offload=phase_offload,
)
def activate(self, component: str) -> None:
"""Keep only the currently needed component resident in unified memory.
Callers materialize activations before switching, so no lazy graph keeps
the preceding model alive. Loading happens before resetting sample RNG.
"""
if self.phase_offload:
if component != 'text_encoder': self.text_encoder.model = None
if component != 'transformer': self.transformer = None
if component != 'vae': self.vae = None
gc.collect()
mx.clear_cache()
if component == 'text_encoder':
if self.text_encoder.model is None:
self.text_encoder.model = load_text_encoder(self.model_path)
model = self.text_encoder.model
elif component == 'transformer':
if self.transformer is None:
self.transformer = load_transformer(self.model_path, self.variant)
model = self.transformer
elif component == 'vae':
if self.vae is None: self.vae = load_vae(self.model_path)
model = self.vae
else: raise ValueError(component)
mx.eval(model.parameters())
def release(self) -> None:
if self.phase_offload:
self.text_encoder.model = self.transformer = self.vae = None
gc.collect()
mx.clear_cache()
def count_prompt_tokens(self, prompt: str) -> int:
return len(self.text_encoder.tokenizer(prompt)["input_ids"])
def generate_array(
self,
prompt: str,
*,
seed: int = 0,
steps: int = 30,
width: int = 512,
height: int = 512,
guidance: float = 1.0,
negative_prompt: str = " ",
num_images: int = 1,
) -> mx.array:
"""Generate one ``[H, W, 3]`` image, or ``[N, H, W, 3]`` when num_images > 1."""
self.activate('text_encoder')
emb = self.text_encoder.encode(prompt).astype(mx.bfloat16)
do_cfg = guidance is not None and guidance > 1.0
neg = (
self.text_encoder.encode(negative_prompt).astype(mx.bfloat16)
if do_cfg
else None
)
return self._sample(
emb,
neg,
seed=seed,
steps=steps,
width=width,
height=height,
guidance=guidance,
num_images=num_images,
)[..., :3]
def edit_array(
self,
prompt: str,
image_paths: Sequence[str | Path],
*,
seed: int = 0,
steps: int = 40,
width: int | None = None,
height: int | None = None,
guidance: float = 1.0,
negative_prompt: str = " ",
output_resolution: int = 1024,
use_kv_cache: bool = True,
) -> mx.array:
"""Edit one or more references; return an [H, W, 4] uint8 RGBA image."""
if not image_paths:
raise ValueError("At least one reference image is required")
if output_resolution < 256:
raise ValueError("output_resolution must be at least 256")
self.activate('vae')
references, reference_latents, reference_shapes = [], [], []
for path in image_paths:
with Image.open(Path(path).expanduser()) as source:
image = source.convert("RGBA")
size = _image_dimensions(output_resolution, image.width / image.height)
image = image.resize(size, Image.Resampling.LANCZOS)
references.append(image)
pixels = mx.array(np.asarray(image).astype(np.float32) / 127.5 - 1.0)
pixels = pixels.transpose(2, 0, 1)[None, :, None].astype(mx.bfloat16)
mean, variance = self.vae.encode(pixels)
mx.eval(mean, variance)
normalized = (
mean - self.latents_mean.astype(mean.dtype)
) / self.latents_std.astype(mean.dtype)
h, w = normalized.shape[-2:]
reference_shapes.append((1, h, w))
reference_latents.append(
normalized.reshape(1, self.z_dim, h * w).transpose(0, 2, 1)
)
mx.eval(reference_latents[-1])
default_width, default_height = references[-1].size
width = (default_width if width is None else width) // 32 * 32
height = (default_height if height is None else height) // 32 * 32
if width < 32 or height < 32:
raise ValueError("Output width and height must be at least 32")
self.activate('text_encoder')
emb, image_pad_mask = self.text_encoder.encode_edit(prompt, references)
neg = negative_image_pad_mask = None
if guidance is not None and guidance > 1.0:
neg, negative_image_pad_mask = self.text_encoder.encode_edit(
negative_prompt, references
)
neg = neg.astype(mx.bfloat16)
return self._sample(
emb.astype(mx.bfloat16),
neg,
seed=seed,
steps=steps,
width=width,
height=height,
guidance=guidance,
reference_latents=mx.concatenate(reference_latents, axis=1).astype(
mx.bfloat16
),
reference_image_shapes=reference_shapes,
image_pad_mask=image_pad_mask,
negative_image_pad_mask=negative_image_pad_mask,
use_kv_cache=use_kv_cache,
)
def _sample(
self,
emb,
neg,
*,
seed,
steps,
width,
height,
guidance,
num_images=1,
reference_latents=None,
reference_image_shapes=None,
image_pad_mask=None,
negative_image_pad_mask=None,
use_kv_cache=False,
) -> mx.array:
mx.eval(emb, [] if neg is None else neg,
[] if reference_latents is None else reference_latents)
self.activate('transformer')
z = self.z_dim
h_lat, w_lat = height // 16, width // 16
tokens = h_lat * w_lat
mx.random.seed(seed)
latents = mx.random.normal((num_images, 1, z, h_lat, w_lat)).astype(mx.bfloat16)
latents = latents.reshape(num_images, z, tokens).transpose(0, 2, 1)
if num_images > 1:
emb = mx.broadcast_to(emb, (num_images, *emb.shape[1:]))
if neg is not None:
neg = mx.broadcast_to(neg, (num_images, *neg.shape[1:]))
scheduler = FlowMatchEulerDiscreteScheduler(
image_seq_len=tokens, num_inference_steps=steps, **self.scheduler_config
)
cache = negative_cache = None
if (
use_kv_cache
and steps > 1
and self.transformer.causal_condition
):
num_layers = len(self.transformer.transformer_blocks)
cache = QwenImageKVCache(num_layers)
if neg is not None:
negative_cache = QwenImageKVCache(num_layers)
for i in range(steps):
t = scheduler.timesteps[i : i + 1].astype(latents.dtype) / 1000
model_input = latents
edit_kwargs = {}
if reference_latents is not None:
if cache is None or i == 0:
model_input = mx.concatenate([reference_latents, latents], axis=1)
edit_kwargs = dict(
reference_image_shapes=reference_image_shapes,
image_pad_mask=image_pad_mask,
)
if cache is not None:
edit_kwargs.update(
kv_cache=cache, kv_cache_mode="extract" if i == 0 else "cached"
)
pred = self.transformer(
hidden_states=model_input,
encoder_hidden_states=emb,
timestep=t,
img_shape=(1, h_lat, w_lat),
**edit_kwargs,
)
if neg is not None:
if reference_latents is not None:
edit_kwargs["image_pad_mask"] = negative_image_pad_mask
if negative_cache is not None:
edit_kwargs["kv_cache"] = negative_cache
neg_pred = self.transformer(
hidden_states=model_input,
encoder_hidden_states=neg,
timestep=t,
img_shape=(1, h_lat, w_lat),
**edit_kwargs,
)
pred = neg_pred + guidance * (pred - neg_pred)
latents = scheduler.step(noise=pred, step_index=i, latents=latents)
if cache is not None and i == 0:
# Materialize compact prefix storage at the existing step boundary,
# releasing the extraction graph before the next denoising step.
mx.eval(
latents,
cache.arrays(),
[] if negative_cache is None else negative_cache.arrays(),
)
else:
mx.eval(latents)
# The VAE decoder does not need the per-layer prefix buffers.
if cache is not None:
cache.clear()
if negative_cache is not None:
negative_cache.clear()
z_lat = (
latents.transpose(0, 2, 1)
.reshape(num_images, z, 1, h_lat, w_lat)
.astype(mx.float32)
)
z_lat = z_lat * self.latents_std + self.latents_mean
mx.eval(z_lat)
self.activate('vae')
image = self.vae.decode(z_lat)[:, :, 0]
image = ((mx.clip(image, -1.0, 1.0) + 1.0) / 2.0 * 255).astype(mx.uint8)
images = image.transpose(0, 2, 3, 1)
return images[0] if num_images == 1 else images
__all__ = ["QwenImagePipeline"]
|