File size: 19,944 Bytes
bef115e 89cd62c bef115e 89cd62c bef115e 89cd62c bef115e 89cd62c bef115e 89cd62c bef115e 89cd62c bef115e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 | #!/usr/bin/env python3
"""
Supra2-IMG inference — standalone text-to-image (DiT + Flan-T5-Base + SD VAE).
Works on Linux and Windows. No imports from other project files.
Usage:
python inference.py --prompt "a sea jellyfish floating in the pitch-black ocean depths" \\
--seed 0 --cfg 3.0 --steps 50 --n 1 --out jellyfish.png
generate.bat "a red rose in daylight" rose.png
./generate.sh "a red rose in daylight" rose.png
Default checkpoint is Supra2-IMG_fp32.safetensors next to this script.
"""
from __future__ import annotations
import argparse
import math
import os
import sys
import time
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.nn.functional as F
# ---------------------------------------------------------------------------
# Architecture constants (must match the trained checkpoint)
# ---------------------------------------------------------------------------
IMG_SIZE = 256
LATENT_SIZE = 32 # 256 / 8 (f8 VAE)
LATENT_CH = 4
PATCH = 2
NUM_TOKENS = (LATENT_SIZE // PATCH) ** 2
D_MODEL = 576
DEPTH = 14
N_HEADS = 9
MLP_RATIO = 4.0
D_CTX = 768 # Flan-T5-Base
MAX_CTX_LEN = 128
LOCAL_DIR = os.path.dirname(os.path.abspath(__file__))
FLAN_WEIGHTS = os.path.join(LOCAL_DIR, "flan-t5-base.safetensors")
FLAN_CONFIG = os.path.join(LOCAL_DIR, "flan-t5-base.config.json")
FLAN_TOKENIZER = os.path.join(LOCAL_DIR, "flan-t5-base.tokenizer.json")
VAE_WEIGHTS = os.path.join(LOCAL_DIR, "sd-vae-ft-mse.safetensors")
VAE_CONFIG = os.path.join(LOCAL_DIR, "sd-vae-ft-mse.config.json")
VAE_SCALE = 0.18215
HF_REPO = "SupraLabs/Supra2-IMG"
DEFAULT_CKPT_SAFETENSORS = os.path.join(LOCAL_DIR, "Supra2-IMG_fp32.safetensors")
DEFAULT_CKPT_PT = os.path.join(LOCAL_DIR, "model_final_ema.pt")
DEFAULT_CKPT = DEFAULT_CKPT_SAFETENSORS if os.path.isfile(DEFAULT_CKPT_SAFETENSORS) else DEFAULT_CKPT_PT
def log(msg: str) -> None:
"""Print immediately so the user sees live progress."""
print(msg, flush=True)
def pick_device() -> torch.device:
if torch.cuda.is_available():
torch.cuda.set_device(0)
name = torch.cuda.get_device_name(0)
mem = torch.cuda.get_device_properties(0).total_memory / 1e9
log(f"[device] CUDA: {name} ({mem:.1f} GB)")
return torch.device("cuda:0")
if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
log("[device] Apple MPS")
return torch.device("mps")
log("[device] CPU (this will be slow)")
return torch.device("cpu")
# ===========================================================================
# MODEL
# ===========================================================================
def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
"""AdaLN: x * (1 + scale) + shift."""
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
class TimestepEmbedder(nn.Module):
"""Sinusoidal timestep embedding followed by an MLP."""
def __init__(self, hidden_size: int, freq_dim: int = 256) -> None:
super().__init__()
self.freq_dim = freq_dim
self.mlp = nn.Sequential(
nn.Linear(freq_dim, hidden_size),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size),
)
def _sinusoidal(self, t: torch.Tensor) -> torch.Tensor:
half = self.freq_dim // 2
freqs = torch.exp(
-math.log(10000.0) * torch.arange(half, device=t.device) / half
)
args = t[:, None].float() * freqs[None] * 1000.0
emb = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if self.freq_dim % 2:
emb = F.pad(emb, (0, 1))
return emb
def forward(self, t: torch.Tensor) -> torch.Tensor:
return self.mlp(self._sinusoidal(t))
class Attention(nn.Module):
"""Multi-head self- or cross-attention."""
def __init__(self, dim: int, n_heads: int, ctx_dim: int | None = None) -> None:
super().__init__()
self.n_heads = n_heads
self.head_dim = dim // n_heads
self.is_self = ctx_dim is None
if self.is_self:
self.qkv = nn.Linear(dim, dim * 3, bias=True)
else:
self.q = nn.Linear(dim, dim, bias=True)
self.kv = nn.Linear(ctx_dim, dim * 2, bias=True)
self.proj = nn.Linear(dim, dim, bias=True)
def forward(
self,
x: torch.Tensor,
ctx: torch.Tensor | None = None,
ctx_mask: torch.Tensor | None = None,
) -> torch.Tensor:
B, N, C = x.shape
if self.is_self:
qkv = self.qkv(x).view(B, N, 3, self.n_heads, self.head_dim)
q, k, v = (qkv[:, :, i].transpose(1, 2) for i in range(3))
else:
M = ctx.shape[1]
q = self.q(x).view(B, N, self.n_heads, self.head_dim).transpose(1, 2)
kv = self.kv(ctx).view(B, M, 2, self.n_heads, self.head_dim)
k, v = kv[:, :, 0].transpose(1, 2), kv[:, :, 1].transpose(1, 2)
attn_mask = None
if ctx_mask is not None:
attn_mask = ctx_mask.bool()[:, None, None, :]
out = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
out = out.transpose(1, 2).reshape(B, N, C)
return self.proj(out)
class DiTBlock(nn.Module):
"""DiT block: AdaLN-Zero self-attn + cross-attn + MLP."""
def __init__(self, dim: int, n_heads: int, ctx_dim: int, mlp_ratio: float) -> None:
super().__init__()
self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
self.self_attn = Attention(dim, n_heads)
self.norm_ca = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
self.cross_attn = Attention(dim, n_heads, ctx_dim=dim)
self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
hidden = int(dim * mlp_ratio)
self.mlp = nn.Sequential(
nn.Linear(dim, hidden),
nn.GELU(approximate="tanh"),
nn.Linear(hidden, dim),
)
self.adaln = nn.Sequential(nn.SiLU(), nn.Linear(dim, 6 * dim, bias=True))
def forward(
self,
x: torch.Tensor,
c: torch.Tensor,
ctx: torch.Tensor,
ctx_mask: torch.Tensor | None,
) -> torch.Tensor:
shift_sa, scale_sa, gate_sa, shift_mlp, scale_mlp, gate_mlp = self.adaln(c).chunk(6, dim=1)
x = x + gate_sa.unsqueeze(1) * self.self_attn(
modulate(self.norm1(x), shift_sa, scale_sa)
)
x = x + self.cross_attn(self.norm_ca(x), ctx=ctx, ctx_mask=ctx_mask)
x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
return x
class FinalLayer(nn.Module):
def __init__(self, dim: int, out_ch: int) -> None:
super().__init__()
self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
self.linear = nn.Linear(dim, out_ch, bias=True)
self.adaln = nn.Sequential(nn.SiLU(), nn.Linear(dim, 2 * dim, bias=True))
def forward(self, x: torch.Tensor, c: torch.Tensor) -> torch.Tensor:
shift, scale = self.adaln(c).chunk(2, dim=1)
return self.linear(modulate(self.norm(x), shift, scale))
class SupraDiT(nn.Module):
"""~100M parameter DiT for rectified-flow text-to-image."""
def __init__(
self,
latent_ch: int = LATENT_CH,
d_model: int = D_MODEL,
depth: int = DEPTH,
n_heads: int = N_HEADS,
ctx_dim: int = D_CTX,
mlp_ratio: float = MLP_RATIO,
num_tokens: int = NUM_TOKENS,
) -> None:
super().__init__()
self.num_tokens = num_tokens
self.patch = PATCH
self.x_embed = nn.Linear(latent_ch * PATCH * PATCH, d_model)
self.pos_embed = nn.Parameter(torch.zeros(1, num_tokens, d_model))
self.t_embed = TimestepEmbedder(d_model)
self.ctx_proj = nn.Linear(ctx_dim, d_model)
self.blocks = nn.ModuleList(
[DiTBlock(d_model, n_heads, d_model, mlp_ratio) for _ in range(depth)]
)
self.final = FinalLayer(d_model, latent_ch * PATCH * PATCH)
def _position_embed(self, grid_h: int, grid_w: int) -> torch.Tensor:
"""Learned 16x16 positions, bicubic-resized when sampling above 256px."""
pos = self.pos_embed
side = int(math.isqrt(pos.shape[1]))
if grid_h == side and grid_w == side:
return pos
weight = pos.reshape(1, side, side, pos.shape[-1]).permute(0, 3, 1, 2).float()
weight = F.interpolate(weight, size=(grid_h, grid_w), mode="bicubic", align_corners=False)
return weight.permute(0, 2, 3, 1).reshape(1, grid_h * grid_w, pos.shape[-1]).to(dtype=pos.dtype)
def forward(
self,
z: torch.Tensor,
t: torch.Tensor,
ctx: torch.Tensor,
ctx_mask: torch.Tensor | None = None,
) -> torch.Tensor:
B, C, H, W = z.shape
P = self.patch
h, w = H // P, W // P
x = z.view(B, C, h, P, w, P).permute(0, 2, 4, 1, 3, 5).reshape(B, h * w, C * P * P)
x = self.x_embed(x) + self._position_embed(h, w)
c = self.t_embed(t)
ctx = self.ctx_proj(ctx)
for blk in self.blocks:
x = blk(x, c, ctx, ctx_mask)
x = self.final(x, c)
return x.view(B, h, w, C, P, P).permute(0, 3, 1, 4, 2, 5).reshape(B, C, H, W)
# ===========================================================================
# Checkpoint download
# ===========================================================================
def apply_lora(model: torch.nn.Module, path: str) -> None:
"""Merge a maltrainer Supra LoRA into the loaded weights. W = W + B @ A * (alpha / rank)."""
from safetensors.torch import load_file
tensors = load_file(path, device="cpu")
grouped: dict[str, dict[str, torch.Tensor]] = {}
for key, tensor in tensors.items():
if key.endswith(".alpha"):
grouped.setdefault(key[: -len(".alpha")], {})["alpha"] = tensor
elif key.endswith(".lora_A.weight"):
grouped.setdefault(key[: -len(".lora_A.weight")], {})["A"] = tensor
elif key.endswith(".lora_B.weight"):
grouped.setdefault(key[: -len(".lora_B.weight")], {})["B"] = tensor
applied = 0
for name, parts in grouped.items():
if "A" not in parts or "B" not in parts:
raise SystemExit(f"LoRA module {name} is missing lora_A or lora_B")
module = model.get_submodule(name)
down = parts["A"].to(device=module.weight.device, dtype=module.weight.dtype)
up = parts["B"].to(device=module.weight.device, dtype=module.weight.dtype)
rank = down.shape[0]
alpha = float(parts["alpha"].item()) if "alpha" in parts else float(rank)
module.weight.data.add_((up @ down) * (alpha / rank))
applied += 1
log(f"[lora] merged {applied} layers from {path} (alpha/rank scale)")
def ensure_checkpoint(path: str) -> str:
"""Use local checkpoint or download model_final_ema.pt from Hugging Face."""
if os.path.isfile(path):
log(f"[ckpt] found {path}")
return path
log(f"[ckpt] {path} not found — downloading from {HF_REPO} ...")
try:
from huggingface_hub import hf_hub_download
except ImportError:
log("[ckpt] installing huggingface_hub ...")
import subprocess
subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", "huggingface_hub"])
from huggingface_hub import hf_hub_download
filename = os.path.basename(path) or "model_final_ema.pt"
downloaded = hf_hub_download(
repo_id=HF_REPO,
filename=filename,
local_dir=".",
local_dir_use_symlinks=False,
)
# Prefer the expected local path if the hub placed it elsewhere
if os.path.isfile(filename) and os.path.abspath(filename) != os.path.abspath(path):
import shutil
shutil.copy2(filename, path)
log(f"[ckpt] copied to {path}")
return path
log(f"[ckpt] downloaded: {downloaded}")
return downloaded if os.path.isfile(downloaded) else path
# ===========================================================================
# Sampling (Euler integration of the flow ODE + optional CFG)
# ===========================================================================
@torch.no_grad()
def generate(args: argparse.Namespace, device: torch.device) -> None:
import json
from transformers import T5Config, T5EncoderModel, T5TokenizerFast
from diffusers import AutoencoderKL
import torchvision.utils as vutils
from safetensors import safe_open
from safetensors.torch import load_file
ckpt_path = ensure_checkpoint(args.ckpt or DEFAULT_CKPT)
log("[model] building SupraDiT ...")
model = SupraDiT().to(device).eval()
n_params = sum(p.numel() for p in model.parameters())
log(f"[model] {n_params / 1e6:.1f}M parameters")
log(f"[model] loading weights from {ckpt_path} ...")
t0 = time.perf_counter()
cfg = {}
if ckpt_path.endswith(".safetensors"):
weights = load_file(ckpt_path, device=str(device))
with safe_open(ckpt_path, framework="pt", device="cpu") as f:
meta = f.metadata() or {}
cfg = {k: int(v) if v.isdigit() else v for k, v in meta.items()}
if "__uncond_text__" in weights:
cfg["uncond_text"] = weights.pop("__uncond_text__")
if "__uncond_mask__" in weights:
cfg["uncond_mask"] = weights.pop("__uncond_mask__")
else:
state = torch.load(ckpt_path, map_location=device, weights_only=False)
cfg = state.get("config", {}) if isinstance(state, dict) else {}
if isinstance(state, dict):
if cfg.get("patch", PATCH) != PATCH:
raise SystemExit(f"Checkpoint PATCH={cfg['patch']} != script PATCH={PATCH}")
weights = state["ema"] if "ema" in state else state.get("model", state)
else:
weights = state
model.load_state_dict(weights, strict=True)
log(f"[model] weights loaded in {time.perf_counter() - t0:.1f}s")
if getattr(args, "lora", None):
apply_lora(model, args.lora)
ctx_len = int(cfg.get("ctx_len", MAX_CTX_LEN))
log(f"[text] ctx_len={ctx_len}")
log(f"[text] loading {os.path.basename(FLAN_WEIGHTS)} ...")
tokenizer = T5TokenizerFast(tokenizer_file=FLAN_TOKENIZER, extra_ids=100)
text_model = T5EncoderModel(T5Config.from_json_file(FLAN_CONFIG))
text_weights = load_file(FLAN_WEIGHTS)
text_incompat = text_model.load_state_dict(text_weights, strict=False)
bad_missing = [k for k in text_incompat.missing_keys if "embed_tokens" not in k]
bad_unexpected = [
k for k in text_incompat.unexpected_keys if not k.startswith("decoder.") and k != "lm_head.weight"
]
if bad_missing or bad_unexpected:
raise SystemExit(f"Flan-T5 load mismatch missing={bad_missing} unexpected={bad_unexpected}")
del text_weights
text_model = text_model.to(device).eval()
for p in text_model.parameters():
p.requires_grad = False
log(f"[vae] loading {os.path.basename(VAE_WEIGHTS)} ...")
with open(VAE_CONFIG, encoding="utf-8") as f:
vae = AutoencoderKL.from_config(json.load(f))
vae_weights = load_file(VAE_WEIGHTS)
vae._fix_state_dict_keys_on_load(vae_weights)
vae.load_state_dict(vae_weights, strict=True)
del vae_weights
vae = vae.to(device).eval()
torch.manual_seed(args.seed)
if device.type == "cuda":
torch.cuda.manual_seed_all(args.seed)
prompts = [args.prompt] * args.n
n_tok = len(tokenizer(args.prompt)["input_ids"])
log(f"[text] prompt tokens={n_tok} n={args.n} seed={args.seed} cfg={args.cfg} steps={args.steps}")
if n_tok > ctx_len:
log(f"[text] WARNING: prompt has {n_tok} tokens, truncated to ctx_len={ctx_len}")
tok = tokenizer(
prompts,
padding="max_length",
truncation=True,
max_length=ctx_len,
return_tensors="pt",
).to(device)
use_amp = device.type == "cuda"
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=use_amp):
ctx = text_model(**tok).last_hidden_state.float()
cmask = tok["attention_mask"].float()
use_cfg = args.cfg > 1.0
if use_cfg:
if isinstance(cfg, dict) and "uncond_text" in cfg:
uncond_ctx = cfg["uncond_text"].to(device).float().unsqueeze(0).expand(args.n, -1, -1)
uncond_mask = cfg["uncond_mask"].to(device).float().unsqueeze(0).expand(args.n, -1)
log("[cfg] using stored unconditional embeddings")
else:
u_tok = tokenizer(
[""] * args.n,
padding="max_length",
truncation=True,
max_length=ctx_len,
return_tensors="pt",
).to(device)
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=use_amp):
uncond_ctx = text_model(**u_tok).last_hidden_state.float()
uncond_mask = u_tok["attention_mask"].float()
log("[cfg] encoded empty string as unconditional")
ctx_all = torch.cat([ctx, uncond_ctx], 0)
mask_all = torch.cat([cmask, uncond_mask], 0)
image_size = int(getattr(args, "size", IMG_SIZE))
if image_size % 16 != 0:
raise SystemExit(f"--size must be a multiple of 16, got {image_size}")
latent_size = image_size // 8
log(f"[sample] image {image_size}px, latent {latent_size}")
z = torch.randn(args.n, LATENT_CH, latent_size, latent_size, device=device)
dt = 1.0 / args.steps
log(f"[sample] Euler flow, {args.steps} steps ...")
t_sample = time.perf_counter()
for i in tqdm(range(args.steps), desc="Generating", unit="step", dynamic_ncols=True):
t = torch.full((args.n,), i * dt, device=device)
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=use_amp):
if use_cfg:
v_both = model(torch.cat([z, z], 0), torch.cat([t, t], 0), ctx_all, mask_all)
v_cond, v_uncond = v_both.float().chunk(2, 0)
v = v_uncond + args.cfg * (v_cond - v_uncond)
else:
v = model(z, t, ctx, cmask).float()
z = z + dt * v
log(f"[sample] denoising done in {time.perf_counter() - t_sample:.1f}s")
log("[vae] decoding latents ...")
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=use_amp):
imgs = vae.decode(z / VAE_SCALE).sample
imgs = (imgs.clamp(-1, 1) + 1) / 2
out_dir = os.path.dirname(os.path.abspath(args.out))
if out_dir:
os.makedirs(out_dir, exist_ok=True)
vutils.save_image(imgs, args.out, nrow=int(math.ceil(math.sqrt(args.n))))
log(f"[done] saved {args.n} image(s) -> {args.out}")
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description="Supra2-IMG standalone inference")
p.add_argument(
"--ckpt",
default=None,
help="Path to checkpoint (.safetensors or .pt)",
)
p.add_argument(
"--prompt",
default="a sea jellyfish floating in the pitch-black ocean depths",
help="Text prompt",
)
p.add_argument("--seed", type=int, default=0, help="RNG seed")
p.add_argument("--cfg", type=float, default=3.0, help="Classifier-free guidance scale")
p.add_argument("--steps", type=int, default=50, help="Euler ODE steps")
p.add_argument("--n", type=int, default=1, help="Number of images")
p.add_argument("--out", default="jellyfish.png", help="Output image path")
p.add_argument("--size", type=int, default=IMG_SIZE, help="Output side in pixels. 256 is native; 512 interpolates positions.")
p.add_argument("--lora", default=None, help="Optional Supra LoRA .safetensors to merge before sampling")
return p.parse_args()
def main() -> None:
args = parse_args()
log("=== Supra2-IMG inference ===")
device = pick_device()
generate(args, device)
if __name__ == "__main__":
main() |