scripts/inference_v3.py

# ============================================================================
# TinyFlux-Deep Inference Cell - With ExpertPredictor
# ============================================================================
# Run the model cell before this one (defines TinyFluxDeep, TinyFluxDeepConfig)
# Loads from: AbstractPhil/tiny-flux-deep or local checkpoint
#
# The ExpertPredictor runs standalone at inference - no SD1.5-flow needed.
# It predicts timestep expertise from (time_emb, clip_pooled).
# ============================================================================

import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from transformers import T5EncoderModel, T5Tokenizer, CLIPTextModel, CLIPTokenizer
from diffusers import AutoencoderKL
from PIL import Image
import numpy as np
import os

# ============================================================================
# CONFIG
# ============================================================================
DEVICE = "cuda"
DTYPE = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16

# Model loading
HF_REPO = "AbstractPhil/tiny-flux-deep"
# stable v3 step_316875
LOAD_FROM = "hub:step_346875"  # "hub", "hub:step_XXXXX", "hub:step_XXXXX_ema", "local:/path/to/weights.safetensors"

# Generation settings
NUM_STEPS = 50
GUIDANCE_SCALE = 5.0  # Note: this is now just for CFG, not the broken guidance_in
HEIGHT = 512
WIDTH = 512
SEED = None
SHIFT = 3.0

# Model architecture (must match training)
USE_EXPERT_PREDICTOR = True
EXPERT_DIM = 1280
EXPERT_HIDDEN_DIM = 512

# ============================================================================
# LOAD TEXT ENCODERS
# ============================================================================
print("Loading text encoders...")
t5_tok = T5Tokenizer.from_pretrained("google/flan-t5-base")
t5_enc = T5EncoderModel.from_pretrained("google/flan-t5-base", torch_dtype=DTYPE).to(DEVICE).eval()

clip_tok = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
clip_enc = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14", torch_dtype=DTYPE).to(DEVICE).eval()

# ============================================================================
# LOAD VAE
# ============================================================================
print("Loading Flux VAE...")
vae = AutoencoderKL.from_pretrained(
    "black-forest-labs/FLUX.1-schnell",
    subfolder="vae",
    torch_dtype=DTYPE
).to(DEVICE).eval()

# ============================================================================
# LOAD TINYFLUX-DEEP MODEL
# ============================================================================
print(f"Loading TinyFlux-Deep from: {LOAD_FROM}")

# Config with ExpertPredictor (no guidance_embeds)
config = TinyFluxDeepConfig(
    use_expert_predictor=USE_EXPERT_PREDICTOR,
    expert_dim=EXPERT_DIM,
    expert_hidden_dim=EXPERT_HIDDEN_DIM,
    guidance_embeds=False,  # Replaced by expert_predictor
)
model = TinyFluxDeep(config).to(DEVICE).to(DTYPE)

# Keys to handle during loading
DEPRECATED_KEYS = {
    'time_in.sin_basis',
    'guidance_in.sin_basis',
    'guidance_in.mlp.0.weight',
    'guidance_in.mlp.0.bias',
    'guidance_in.mlp.2.weight',
    'guidance_in.mlp.2.bias',
}


def load_weights(path):
    """Load weights from .safetensors or .pt file."""
    if path.endswith(".safetensors"):
        state_dict = load_file(path)
    elif path.endswith(".pt"):
        ckpt = torch.load(path, map_location=DEVICE, weights_only=False)
        if isinstance(ckpt, dict):
            if "model" in ckpt:
                state_dict = ckpt["model"]
            elif "state_dict" in ckpt:
                state_dict = ckpt["state_dict"]
            else:
                state_dict = ckpt
        else:
            state_dict = ckpt
    else:
        try:
            state_dict = load_file(path)
        except:
            state_dict = torch.load(path, map_location=DEVICE, weights_only=False)

    # Strip "_orig_mod." prefix from keys (added by torch.compile)
    if any(k.startswith("_orig_mod.") for k in state_dict.keys()):
        print("  Stripping torch.compile prefix...")
        state_dict = {k.replace("_orig_mod.", ""): v for k, v in state_dict.items()}

    return state_dict


def load_model_weights(model, weights, source_name):
    """Load weights with architecture upgrade support."""
    model_state = model.state_dict()

    loaded = []
    skipped_deprecated = []
    skipped_shape = []
    missing_new = []

    # Load matching weights
    for k, v in weights.items():
        if k in DEPRECATED_KEYS or k.startswith('guidance_in.'):
            skipped_deprecated.append(k)
        elif k in model_state:
            if v.shape == model_state[k].shape:
                model_state[k] = v
                loaded.append(k)
            else:
                skipped_shape.append((k, v.shape, model_state[k].shape))
        else:
            # Key not in model (maybe old architecture)
            skipped_deprecated.append(k)

    # Find new keys not in checkpoint
    for k in model_state:
        if k not in weights and not any(k.startswith(d.split('.')[0]) for d in DEPRECATED_KEYS if '.' in d):
            missing_new.append(k)

    # Apply loaded weights
    model.load_state_dict(model_state, strict=False)

    # Report
    print(f"  ✓ Loaded: {len(loaded)} weights")
    if skipped_deprecated:
        print(f"  ✓ Skipped deprecated: {len(skipped_deprecated)} (guidance_in, etc)")
    if skipped_shape:
        print(f"  ⚠ Shape mismatch: {len(skipped_shape)}")
        for k, old, new in skipped_shape[:3]:
            print(f"      {k}: {old} vs {new}")
    if missing_new:
        # Group by module
        modules = set(k.split('.')[0] for k in missing_new)
        print(f"  ℹ New modules (fresh init): {modules}")

    print(f"✓ Loaded from {source_name}")


if LOAD_FROM == "hub":
    try:
        weights_path = hf_hub_download(repo_id=HF_REPO, filename="model.safetensors")
    except:
        weights_path = hf_hub_download(repo_id=HF_REPO, filename="model.pt")
    weights = load_weights(weights_path)
    load_model_weights(model, weights, HF_REPO)

elif LOAD_FROM.startswith("hub:"):
    ckpt_name = LOAD_FROM[4:]
    for ext in [".safetensors", ".pt", ""]:
        try:
            if ckpt_name.endswith((".safetensors", ".pt")):
                filename = ckpt_name if "/" in ckpt_name else f"checkpoints/{ckpt_name}"
            else:
                filename = f"checkpoints/{ckpt_name}{ext}"
            weights_path = hf_hub_download(repo_id=HF_REPO, filename=filename)
            weights = load_weights(weights_path)
            load_model_weights(model, weights, f"{HF_REPO}/{filename}")
            break
        except Exception as e:
            continue
    else:
        raise ValueError(f"Could not find checkpoint: {ckpt_name}")

elif LOAD_FROM.startswith("local:"):
    weights_path = LOAD_FROM[6:]
    weights = load_weights(weights_path)
    load_model_weights(model, weights, weights_path)

else:
    raise ValueError(f"Unknown LOAD_FROM: {LOAD_FROM}")

model.eval()

# Count parameters
total_params = sum(p.numel() for p in model.parameters())
expert_params = sum(p.numel() for p in model.expert_predictor.parameters()) if model.expert_predictor else 0
print(f"Model params: {total_params:,} (expert_predictor: {expert_params:,})")

# ============================================================================
# ENCODING FUNCTIONS
# ============================================================================
@torch.inference_mode()
def encode_prompt(prompt: str, max_length: int = 128):
    """Encode prompt with flan-t5-base and CLIP-L."""
    t5_in = t5_tok(
        prompt,
        max_length=max_length,
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    ).to(DEVICE)
    t5_out = t5_enc(
        input_ids=t5_in.input_ids,
        attention_mask=t5_in.attention_mask
    ).last_hidden_state

    clip_in = clip_tok(
        prompt,
        max_length=77,
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    ).to(DEVICE)
    clip_out = clip_enc(
        input_ids=clip_in.input_ids,
        attention_mask=clip_in.attention_mask
    )
    clip_pooled = clip_out.pooler_output

    return t5_out.to(DTYPE), clip_pooled.to(DTYPE)


# ============================================================================
# FLOW MATCHING HELPERS
# ============================================================================
def flux_shift(t, s=SHIFT):
    """Flux timestep shift - biases towards higher t (closer to data)."""
    return s * t / (1 + (s - 1) * t)


# ============================================================================
# EULER DISCRETE FLOW MATCHING SAMPLER
# ============================================================================
@torch.inference_mode()
def euler_sample(
    model,
    prompt: str,
    negative_prompt: str = "",
    num_steps: int = 28,
    guidance_scale: float = 3.5,
    height: int = 512,
    width: int = 512,
    seed: int = None,
):
    """
    Euler discrete sampler for rectified flow matching.

    Flow Matching formulation:
        x_t = (1 - t) * noise + t * data
        At t=0: noise, At t=1: data
        Velocity v = data - noise (constant)

    Sampling: Integrate from t=0 (noise) to t=1 (data)

    With ExpertPredictor:
        - No guidance embedding needed
        - Expert predictor runs internally from (time_emb, clip_pooled)
        - CFG still works via positive/negative prompt difference
    """
    if seed is not None:
        torch.manual_seed(seed)
        generator = torch.Generator(device=DEVICE).manual_seed(seed)
    else:
        generator = None

    H_lat = height // 8
    W_lat = width // 8
    C_lat = 16

    # Encode prompts
    t5_cond, clip_cond = encode_prompt(prompt)
    if guidance_scale > 1.0 and negative_prompt is not None:
        t5_uncond, clip_uncond = encode_prompt(negative_prompt)
    else:
        t5_uncond, clip_uncond = None, None

    # Start from pure noise (t=0)
    x = torch.randn(1, H_lat * W_lat, C_lat, device=DEVICE, dtype=DTYPE, generator=generator)

    # Create image position IDs
    img_ids = TinyFluxDeep.create_img_ids(1, H_lat, W_lat, DEVICE)

    # Timesteps: 0 → 1 with flux shift
    t_linear = torch.linspace(0, 1, num_steps + 1, device=DEVICE, dtype=DTYPE)
    timesteps = flux_shift(t_linear, s=SHIFT)

    print(f"Sampling with {num_steps} Euler steps (t: 0→1, shifted)...")

    for i in range(num_steps):
        t_curr = timesteps[i]
        t_next = timesteps[i + 1]
        dt = t_next - t_curr

        t_batch = t_curr.unsqueeze(0)

        # Predict velocity (no guidance embedding, expert_predictor runs internally)
        v_cond = model(
            hidden_states=x,
            encoder_hidden_states=t5_cond,
            pooled_projections=clip_cond,
            timestep=t_batch,
            img_ids=img_ids,
            # No guidance parameter - ExpertPredictor handles timestep awareness
            # No expert_features - predictor runs standalone at inference
        )

        # Classifier-free guidance (true CFG via prompt difference)
        if guidance_scale > 1.0 and t5_uncond is not None:
            v_uncond = model(
                hidden_states=x,
                encoder_hidden_states=t5_uncond,
                pooled_projections=clip_uncond,
                timestep=t_batch,
                img_ids=img_ids,
            )
            v = v_uncond + guidance_scale * (v_cond - v_uncond)
        else:
            v = v_cond

        # Euler step: x_{t+dt} = x_t + v * dt
        x = x + v * dt

        if (i + 1) % max(1, num_steps // 5) == 0 or i == num_steps - 1:
            print(f"  Step {i+1}/{num_steps}, t={t_next.item():.3f}")

    # Reshape: (1, H*W, C) -> (1, C, H, W)
    latents = x.reshape(1, H_lat, W_lat, C_lat).permute(0, 3, 1, 2)

    return latents


# ============================================================================
# DECODE LATENTS TO IMAGE
# ============================================================================
@torch.inference_mode()
def decode_latents(latents):
    """Decode VAE latents to PIL Image."""
    latents = latents / vae.config.scaling_factor
    image = vae.decode(latents.to(vae.dtype)).sample
    image = (image / 2 + 0.5).clamp(0, 1)
    image = image[0].float().permute(1, 2, 0).cpu().numpy()
    image = (image * 255).astype(np.uint8)
    return Image.fromarray(image)


# ============================================================================
# MAIN GENERATION FUNCTION
# ============================================================================
def generate(
    prompt: str,
    negative_prompt: str = "",
    num_steps: int = NUM_STEPS,
    guidance_scale: float = GUIDANCE_SCALE,
    height: int = HEIGHT,
    width: int = WIDTH,
    seed: int = SEED,
    save_path: str = None,
):
    """
    Generate an image from a text prompt.

    Args:
        prompt: Text description of desired image
        negative_prompt: What to avoid (empty string for none)
        num_steps: Number of Euler steps (20-50 recommended)
        guidance_scale: CFG scale (1.0=none, 3-7 typical)
        height: Output height in pixels (divisible by 8)
        width: Output width in pixels (divisible by 8)
        seed: Random seed (None for random)
        save_path: Path to save image (None to skip)

    Returns:
        PIL.Image
    """
    print(f"\nGenerating: '{prompt}'")
    print(f"Settings: {num_steps} steps, cfg={guidance_scale}, {width}x{height}, seed={seed}")

    latents = euler_sample(
        model=model,
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_steps=num_steps,
        guidance_scale=guidance_scale,
        height=height,
        width=width,
        seed=seed,
    )

    print("Decoding latents...")
    image = decode_latents(latents)

    if save_path:
        image.save(save_path)
        print(f"✓ Saved to {save_path}")

    print("✓ Done!")
    return image


# ============================================================================
# BATCH GENERATION
# ============================================================================
def generate_batch(
    prompts: list,
    negative_prompt: str = "",
    num_steps: int = NUM_STEPS,
    guidance_scale: float = GUIDANCE_SCALE,
    height: int = HEIGHT,
    width: int = WIDTH,
    seed: int = SEED,
    output_dir: str = "./outputs",
):
    """Generate multiple images."""
    os.makedirs(output_dir, exist_ok=True)
    images = []

    for i, prompt in enumerate(prompts):
        img_seed = seed + i if seed is not None else None
        image = generate(
            prompt=prompt,
            negative_prompt=negative_prompt,
            num_steps=num_steps,
            guidance_scale=guidance_scale,
            height=height,
            width=width,
            seed=img_seed,
            save_path=os.path.join(output_dir, f"{i:03d}.png"),
        )
        images.append(image)

    return images


# ============================================================================
# COMPARISON FUNCTION (old vs new model)
# ============================================================================
def compare_with_without_expert(
    prompt: str,
    negative_prompt: str = "",
    num_steps: int = 30,
    guidance_scale: float = 5.0,
    seed: int = 42,
    save_prefix: str = "compare",
):
    """
    Generate same prompt with expert_predictor enabled vs disabled.
    Useful for A/B testing the effect of the distilled expert.
    """
    # With expert
    image_with = generate(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_steps=num_steps,
        guidance_scale=guidance_scale,
        seed=seed,
        save_path=f"{save_prefix}_with_expert.png",
    )

    # Without expert (temporarily disable)
    old_predictor = model.expert_predictor
    model.expert_predictor = None

    image_without = generate(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_steps=num_steps,
        guidance_scale=guidance_scale,
        seed=seed,
        save_path=f"{save_prefix}_without_expert.png",
    )

    # Restore
    model.expert_predictor = old_predictor

    # Side by side
    combined = Image.new('RGB', (image_with.width * 2, image_with.height))
    combined.paste(image_without, (0, 0))
    combined.paste(image_with, (image_with.width, 0))
    combined.save(f"{save_prefix}_comparison.png")

    print(f"\n✓ Comparison saved: {save_prefix}_comparison.png")
    print(f"  Left: without expert | Right: with expert")

    return image_without, image_with, combined


# ============================================================================
# QUICK TEST
# ============================================================================
print("\n" + "="*60)
print("TinyFlux-Deep + ExpertPredictor Inference Ready!")
print("="*60)
print(f"Config: {config.hidden_size} hidden, {config.num_attention_heads} heads")
print(f"        {config.num_double_layers} double, {config.num_single_layers} single layers")
print(f"        ExpertPredictor: {config.use_expert_predictor} (dim={config.expert_dim})")
print(f"Total:  {total_params:,} parameters")

# Example usage:
image = generate(
    prompt="subject, animal, feline, lion, natural habitat",
    negative_prompt="",
    num_steps=50,
    guidance_scale=5.0,
    seed=4545,
    width=512,
    height=512,
)
image