scripts/inference_v4.py

import torch
from transformers import T5EncoderModel, T5Tokenizer, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

# Load text encoders
t5_tokenizer = T5Tokenizer.from_pretrained("google/flan-t5-base")
t5_model = T5EncoderModel.from_pretrained("google/flan-t5-base").to("cuda", torch.bfloat16)

clip_tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
clip_model = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14").to("cuda", torch.bfloat16)

# Load VAE
vae = AutoencoderKL.from_pretrained(
    "black-forest-labs/FLUX.1-schnell", 
    subfolder="vae",
    torch_dtype=torch.bfloat16
).to("cuda")

# Load TinyFlux-Deep
model_py = hf_hub_download("AbstractPhil/tiny-flux-deep", "scripts/model_v4.py")
exec(open(model_py).read())

config = TinyFluxConfig(
    use_sol_prior=False,  # Disabled until trained
    use_t5_vec=False,     # Disabled until trained
)
model = TinyFluxDeep(config).to("cuda", torch.bfloat16)
weights = load_file(hf_hub_download("AbstractPhil/tiny-flux-deep", "checkpoint_runs/v4_init/lailah_401434_v4_init.safetensors"))
model.load_state_dict(weights, strict=False)
model.eval()

def encode_prompt(prompt):
    """Encode prompt with both T5 and CLIP."""
    # T5
    t5_tokens = t5_tokenizer(prompt, return_tensors="pt", padding="max_length", 
                              max_length=77, truncation=True).to("cuda")
    with torch.no_grad():
        t5_emb = t5_model(**t5_tokens).last_hidden_state.to(torch.bfloat16)
    
    # CLIP
    clip_tokens = clip_tokenizer(prompt, return_tensors="pt", padding="max_length",
                                  max_length=77, truncation=True).to("cuda")
    with torch.no_grad():
        clip_out = clip_model(**clip_tokens)
        clip_pooled = clip_out.pooler_output.to(torch.bfloat16)
    
    return t5_emb, clip_pooled


def flux_shift(t, s=3.0):
    """Flux-style timestep shift."""
    return s * t / (1 + (s - 1) * t)


@torch.inference_mode()
def generate_image(prompt, num_steps=25, cfg_scale=4.0, seed=None):
    """
    Euler sampling for rectified flow.
    
    Flow matching formulation:
        x_t = (1 - t) * noise + t * data
        At t=0: pure noise
        At t=1: pure data  
        Velocity v = data - noise (constant)
        
    Sampling: Integrate from t=0 (noise) → t=1 (data)
    """
    if seed is not None:
        torch.manual_seed(seed)
    
    t5_emb, clip_pooled = encode_prompt(prompt)
    t5_null, clip_null = encode_prompt("")
    
    # Start from pure noise (t=0)
    x = torch.randn(1, 64*64, 16, device="cuda", dtype=torch.bfloat16)
    img_ids = TinyFluxDeep.create_img_ids(1, 64, 64, "cuda")
    
    # Timesteps: 0 → 1 with Flux shift
    t_linear = torch.linspace(0, 1, num_steps + 1, device="cuda", dtype=torch.float32)
    timesteps = flux_shift(t_linear, s=3.0)
    
    for i in range(num_steps):
        t_curr = timesteps[i]
        t_next = timesteps[i + 1]
        dt = t_next - t_curr  # Positive, moving toward data
        
        t_batch = t_curr.unsqueeze(0)
        
        # Predict velocity
        v_cond = model(x, t5_emb, clip_pooled, t_batch, img_ids)
        v_uncond = model(x, t5_null, clip_null, t_batch, img_ids)
        
        # Classifier-free guidance
        v = v_uncond + cfg_scale * (v_cond - v_uncond)
        
        # Euler step: x_{t+dt} = x_t + v * dt
        x = x + v * dt
    
    # Decode with VAE
    x = x.reshape(1, 64, 64, 16).permute(0, 3, 1, 2)  # [B, C, H, W]
    x = x / vae.config.scaling_factor
    image = vae.decode(x).sample
    
    # Convert to PIL
    image = (image / 2 + 0.5).clamp(0, 1)
    image = image[0].permute(1, 2, 0).cpu().float().numpy()
    image = (image * 255).astype("uint8")
    
    from PIL import Image
    return Image.fromarray(image)


# Generate
image = generate_image("a photograph of a tiger in natural habitat", seed=42)
image.save("tiger.png")
image