Upload folder using huggingface_hub

0c17c2b verified 2 months ago

16.5 kB

"""
Flux2KleinBackend.py
FLUX.2 Klein 9B KV Nunchaku + NVFP4 Qwen3 Text Encoder Backend.
Optimized for 16GB GPUs (Blackwell sm_120 / sm_110):
- Distilled 4-step inference with KV-cache for multi-reference editing.
- Sequential GPU stage execution:
    Stage 1: Qwen3 NVFP4 Text Encoder on GPU -> offload to CPU.
    Stage 2: AutoencoderKLFlux2 VAE on GPU (if image conditioning) -> offload to CPU.
    Stage 3: NunchakuFlux2Transformer2DModel (FP4/INT4 SVDQuant) on GPU -> offload to CPU.
    Stage 4: AutoencoderKLFlux2 VAE on GPU (decode) -> offload to CPU.
- All models reside on CPU when idle, leaving the GPU with 0% VRAM allocation
  for other concurrent workloads.
"""

from __future__ import annotations

import gc
import os
import time
from typing import Any, List, Optional, Union

import numpy as np
import PIL.Image
import torch
from diffusers import AutoencoderKLFlux2, FlowMatchEulerDiscreteScheduler
from diffusers.pipelines.flux2.pipeline_flux2_klein_kv import (
    Flux2KleinKVPipeline,
    compute_empirical_mu,
    retrieve_timesteps,
)
from diffusers.pipelines.flux2.pipeline_output import Flux2PipelineOutput
from nunchaku import NunchakuFlux2Transformer2DModel
from nunchaku.utils import get_precision
from transformers import AutoTokenizer

from NVFP4TextEncoder import load_nvfp4_text_encoder


def _flush():
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()


class SequentialFlux2KleinPipeline:
    """Wrapper that runs Flux2KleinKVPipeline stages sequentially on CUDA,
    offloading each component to CPU immediately after its stage finishes.
    """

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKLFlux2,
        text_encoder: torch.nn.Module,
        tokenizer: AutoTokenizer,
        transformer: NunchakuFlux2Transformer2DModel,
        device: str | torch.device = "cuda",
    ):
        self.scheduler = scheduler
        self.vae = vae
        self.text_encoder = text_encoder
        self.tokenizer = tokenizer
        self.transformer = transformer
        self.device = torch.device(device)

        # Standard diffusers pipeline attributes for compatibility
        self.inner_pipe = Flux2KleinKVPipeline(
            scheduler=scheduler,
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            transformer=transformer,
            is_distilled=True,
        )
        self.image_processor = self.inner_pipe.image_processor
        self.vae_scale_factor = self.inner_pipe.vae_scale_factor

        # Expose components dictionary for server sleep / memory introspection
        self.components = {
            "text_encoder": self.text_encoder,
            "transformer": self.transformer,
            "vae": self.vae,
            "scheduler": self.scheduler,
            "tokenizer": self.tokenizer,
        }

        # Keep everything on CPU at initialization
        self.to("cpu")

    def to(self, device: str | torch.device):
        """Move components to specified device (typically 'cpu')."""
        target = str(device)
        if target == "cpu":
            for mod in [self.text_encoder, self.transformer, self.vae]:
                if isinstance(mod, torch.nn.Module):
                    try:
                        mod.to("cpu")
                    except Exception:
                        pass
            _flush()
        return self

    @torch.no_grad()
    def __call__(
        self,
        prompt: str | list[str] = "",
        image: Union[PIL.Image.Image, List[PIL.Image.Image], None] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 4,
        guidance_scale: float = 1.0,
        negative_prompt: Optional[str] = None,
        num_images_per_prompt: int = 1,
        generator: Optional[torch.Generator] = None,
        seed: Optional[int] = None,
        **kwargs,
    ) -> Flux2PipelineOutput:
        device = self.device

        if generator is None:
            generator = torch.Generator(device=device)
            if seed is not None:
                generator.manual_seed(seed)

        t_start = time.time()

        # ------------------------------------------------------------------
        # STAGE 1: Text Encoding (Qwen3 NVFP4 Text Encoder)
        # ------------------------------------------------------------------
        prompt_embeds = None
        text_ids = None
        try:
            print("[Flux2Klein] Stage 1: Moving Text Encoder to GPU...")
            self.text_encoder.to(device)
            _flush()

            prompt_embeds, text_ids = self.inner_pipe.encode_prompt(
                prompt=prompt,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=512,
            )
            # Store CPU copies until Stage 3 to minimize VRAM
            prompt_embeds = prompt_embeds.to("cpu")
            text_ids = text_ids.to("cpu")
        finally:
            self.text_encoder.to("cpu")
            _flush()
            print(f"[Flux2Klein] Stage 1 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

        # ------------------------------------------------------------------
        # STAGE 2: VAE Conditioning (if reference images provided)
        # ------------------------------------------------------------------
        condition_images = None
        if image is not None:
            if not isinstance(image, list):
                image = [image]
            condition_images = []
            for img in image:
                self.image_processor.check_image_input(img)
                img_w, img_h = img.size
                if img_w * img_h > 1024 * 1024:
                    img = self.image_processor._resize_to_target_area(img, 1024 * 1024)
                    img_w, img_h = img.size
                multiple_of = self.vae_scale_factor * 2
                img_w = (img_w // multiple_of) * multiple_of
                img_h = (img_h // multiple_of) * multiple_of
                img_pre = self.image_processor.preprocess(img, height=img_h, width=img_w, resize_mode="crop")
                condition_images.append(img_pre)
                height = height or img_h
                width = width or img_w

        height = height or 1024
        width = width or 1024

        image_latents = None
        image_latent_ids = None
        if condition_images is not None:
            try:
                print(f"[Flux2Klein] Stage 2: Moving VAE to GPU to encode {len(condition_images)} reference image(s)...")
                self.vae.to(device)
                _flush()

                image_latents, image_latent_ids = self.inner_pipe.prepare_image_latents(
                    images=condition_images,
                    batch_size=num_images_per_prompt,
                    generator=generator,
                    device=device,
                    dtype=self.vae.dtype,
                )
                image_latents = image_latents.to("cpu")
                image_latent_ids = image_latent_ids.to("cpu")
            finally:
                self.vae.to("cpu")
                _flush()
                print(f"[Flux2Klein] Stage 2 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

        # ------------------------------------------------------------------
        # STAGE 3: Transformer Denoising Loop (Nunchaku FP4/INT4 KV-Cache)
        # ------------------------------------------------------------------
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_ids = self.inner_pipe.prepare_latents(
            batch_size=num_images_per_prompt,
            num_latents_channels=num_channels_latents,
            height=height,
            width=width,
            dtype=self.transformer.dtype,
            device=device,
            generator=generator,
        )

        sigmas = np.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps)
        image_seq_len = latents.shape[1]
        mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps)
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            sigmas=sigmas,
            mu=mu,
        )
        self.scheduler.set_begin_index(0)

        kv_cache = None
        try:
            print("[Flux2Klein] Stage 3: Moving Nunchaku Transformer to GPU...")
            self.transformer.to(device)
            _flush()

            prompt_embeds = prompt_embeds.to(device)
            text_ids = text_ids.to(device)
            latents = latents.to(device)
            latent_ids = latent_ids.to(device)
            if image_latents is not None:
                image_latents = image_latents.to(device)
                image_latent_ids = image_latent_ids.to(device)

            for i, t in enumerate(timesteps):
                t_step = time.time()
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                if i == 0 and image_latents is not None:
                    # Step 0: Extract KV cache for reference tokens
                    latent_model_input = torch.cat([image_latents, latents], dim=1).to(self.transformer.dtype)
                    latent_image_ids = torch.cat([image_latent_ids, latent_ids], dim=1)
                    noise_pred, kv_cache = self.transformer(
                        hidden_states=latent_model_input,
                        timestep=timestep / 1000,
                        guidance=None,
                        encoder_hidden_states=prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_image_ids,
                        joint_attention_kwargs=None,
                        return_dict=False,
                        kv_cache_mode="extract",
                        num_ref_tokens=image_latents.shape[1],
                    )
                elif kv_cache is not None:
                    # Steps 1+: Reuse cached reference KV
                    noise_pred = self.transformer(
                        hidden_states=latents.to(self.transformer.dtype),
                        timestep=timestep / 1000,
                        guidance=None,
                        encoder_hidden_states=prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_ids,
                        joint_attention_kwargs=None,
                        return_dict=False,
                        kv_cache=kv_cache,
                        kv_cache_mode="cached",
                    )[0]
                else:
                    # Standard text-to-image forward
                    noise_pred = self.transformer(
                        hidden_states=latents.to(self.transformer.dtype),
                        timestep=timestep / 1000,
                        guidance=None,
                        encoder_hidden_states=prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_ids,
                        joint_attention_kwargs=None,
                        return_dict=False,
                    )[0]

                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
                print(
                    f"[Flux2Klein] Step {i+1}/{len(timesteps)} in {time.time() - t_step:.2f}s | "
                    f"Peak VRAM: {torch.cuda.max_memory_allocated() / 1e9:.2f} GB"
                )

            if "noise_pred" in locals():
                del noise_pred
        finally:
            if kv_cache is not None:
                kv_cache.clear()
            self.transformer.to("cpu")
            del prompt_embeds, text_ids, image_latents, image_latent_ids
            _flush()
            print(f"[Flux2Klein] Stage 3 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

        # ------------------------------------------------------------------
        # STAGE 4: VAE Decoding
        # ------------------------------------------------------------------
        try:
            print("[Flux2Klein] Stage 4: Moving VAE to GPU to decode latents...")
            self.vae.to(device)
            _flush()

            latents = self.inner_pipe._unpack_latents_with_ids(latents, latent_ids)
            latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype)
            latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(
                latents.device, latents.dtype
            )
            latents = latents * latents_bn_std + latents_bn_mean
            latents = self.inner_pipe._unpatchify_latents(latents)
            decoded = self.vae.decode(latents, return_dict=False)[0]
            out_images = self.image_processor.postprocess(decoded, output_type="pil")
        finally:
            self.vae.to("cpu")
            del latents, latent_ids
            _flush()
            print(f"[Flux2Klein] Stage 4 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

        print(f"[Flux2Klein] Total generation finished in {time.time() - t_start:.2f}s!")
        return Flux2PipelineOutput(images=out_images)


class Flux2KleinBackend:
    """Backend factory for FLUX.2 Klein 9B KV Nunchaku."""

    def __init__(
        self,
        model_id: str,
        optimized_model_path: Optional[str] = None,
        nvfp4_text_encoder_path: Optional[str] = None,
    ):
        self.model_id = model_id
        self.optimized_model_path = optimized_model_path
        self.nvfp4_text_encoder_path = nvfp4_text_encoder_path
        self.pipeline = None

    def load(self):
        print(f"[Flux2KleinBackend] Initializing backend from base model: {self.model_id}...")

        # 1. Resolve optimized transformer path
        opt_path = self.optimized_model_path
        if not opt_path:
            precision = get_precision()
            opt_path = os.path.join(
                self.model_id,
                f"svdq-{precision}_r32-FLUX.2-klein-9b-kv-Nunchaku.safetensors",
            )
            if not os.path.exists(opt_path):
                # Fallback to fp4 if not found directly
                opt_path = os.path.join(
                    self.model_id,
                    "svdq-fp4_r32-FLUX.2-klein-9b-kv-Nunchaku.safetensors",
                )

        print(f"[Flux2KleinBackend] Loading Nunchaku transformer from {opt_path}...")
        transformer = NunchakuFlux2Transformer2DModel.from_pretrained(
            opt_path,
            torch_dtype=torch.bfloat16,
        )
        transformer.to("cpu")

        # 2. Resolve NVFP4 text encoder path
        enc_path = self.nvfp4_text_encoder_path
        if not enc_path:
            # Check if running directly from within the text encoder repository (extras/)
            candidate_parent = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
            if os.path.exists(os.path.join(candidate_parent, "config.json")):
                enc_path = candidate_parent
            elif os.path.exists(os.path.join(self.model_id, "text_encoder")):
                enc_path = os.path.join(self.model_id, "text_encoder")
            else:
                enc_path = "/path/to/FLUX.2-klein-text-encoder-NVFP4"

        print(f"[Flux2KleinBackend] Loading NVFP4 text encoder from {enc_path}...")
        text_encoder = load_nvfp4_text_encoder(enc_path, device="cpu", dtype=torch.bfloat16)
        tokenizer = AutoTokenizer.from_pretrained(enc_path)

        # 3. Load VAE and Scheduler
        print(f"[Flux2KleinBackend] Loading VAE and Scheduler from {self.model_id}...")
        vae = AutoencoderKLFlux2.from_pretrained(
            self.model_id,
            subfolder="vae",
            torch_dtype=torch.bfloat16,
        )
        vae.to("cpu")
        scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
            self.model_id,
            subfolder="scheduler",
        )

        # 4. Construct Sequential Pipeline
        self.pipeline = SequentialFlux2KleinPipeline(
            scheduler=scheduler,
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            transformer=transformer,
            device="cuda",
        )

        print("[Flux2KleinBackend] Pipeline successfully loaded and ready on CPU!")
        return self.pipeline, self.pipeline

	"""
	Flux2KleinBackend.py
	FLUX.2 Klein 9B KV Nunchaku + NVFP4 Qwen3 Text Encoder Backend.

	Optimized for 16GB GPUs (Blackwell sm_120 / sm_110):
	- Distilled 4-step inference with KV-cache for multi-reference editing.
	- Sequential GPU stage execution:
	Stage 1: Qwen3 NVFP4 Text Encoder on GPU -> offload to CPU.
	Stage 2: AutoencoderKLFlux2 VAE on GPU (if image conditioning) -> offload to CPU.
	Stage 3: NunchakuFlux2Transformer2DModel (FP4/INT4 SVDQuant) on GPU -> offload to CPU.
	Stage 4: AutoencoderKLFlux2 VAE on GPU (decode) -> offload to CPU.
	- All models reside on CPU when idle, leaving the GPU with 0% VRAM allocation
	for other concurrent workloads.
	"""

	from __future__ import annotations

	import gc
	import os
	import time
	from typing import Any, List, Optional, Union

	import numpy as np
	import PIL.Image
	import torch
	from diffusers import AutoencoderKLFlux2, FlowMatchEulerDiscreteScheduler
	from diffusers.pipelines.flux2.pipeline_flux2_klein_kv import (
	Flux2KleinKVPipeline,
	compute_empirical_mu,
	retrieve_timesteps,
	)
	from diffusers.pipelines.flux2.pipeline_output import Flux2PipelineOutput
	from nunchaku import NunchakuFlux2Transformer2DModel
	from nunchaku.utils import get_precision
	from transformers import AutoTokenizer

	from NVFP4TextEncoder import load_nvfp4_text_encoder


	def _flush():
	gc.collect()
	if torch.cuda.is_available():
	torch.cuda.empty_cache()


	class SequentialFlux2KleinPipeline:
	"""Wrapper that runs Flux2KleinKVPipeline stages sequentially on CUDA,
	offloading each component to CPU immediately after its stage finishes.
	"""

	def __init__(
	self,
	scheduler: FlowMatchEulerDiscreteScheduler,
	vae: AutoencoderKLFlux2,
	text_encoder: torch.nn.Module,
	tokenizer: AutoTokenizer,
	transformer: NunchakuFlux2Transformer2DModel,
	device: str \| torch.device = "cuda",
	):
	self.scheduler = scheduler
	self.vae = vae
	self.text_encoder = text_encoder
	self.tokenizer = tokenizer
	self.transformer = transformer
	self.device = torch.device(device)

	# Standard diffusers pipeline attributes for compatibility
	self.inner_pipe = Flux2KleinKVPipeline(
	scheduler=scheduler,
	vae=vae,
	text_encoder=text_encoder,
	tokenizer=tokenizer,
	transformer=transformer,
	is_distilled=True,
	)
	self.image_processor = self.inner_pipe.image_processor
	self.vae_scale_factor = self.inner_pipe.vae_scale_factor

	# Expose components dictionary for server sleep / memory introspection
	self.components = {
	"text_encoder": self.text_encoder,
	"transformer": self.transformer,
	"vae": self.vae,
	"scheduler": self.scheduler,
	"tokenizer": self.tokenizer,
	}

	# Keep everything on CPU at initialization
	self.to("cpu")

	def to(self, device: str \| torch.device):
	"""Move components to specified device (typically 'cpu')."""
	target = str(device)
	if target == "cpu":
	for mod in [self.text_encoder, self.transformer, self.vae]:
	if isinstance(mod, torch.nn.Module):
	try:
	mod.to("cpu")
	except Exception:
	pass
	_flush()
	return self

	@torch.no_grad()
	def __call__(
	self,
	prompt: str \| list[str] = "",
	image: Union[PIL.Image.Image, List[PIL.Image.Image], None] = None,
	height: Optional[int] = None,
	width: Optional[int] = None,
	num_inference_steps: int = 4,
	guidance_scale: float = 1.0,
	negative_prompt: Optional[str] = None,
	num_images_per_prompt: int = 1,
	generator: Optional[torch.Generator] = None,
	seed: Optional[int] = None,
	**kwargs,
	) -> Flux2PipelineOutput:
	device = self.device

	if generator is None:
	generator = torch.Generator(device=device)
	if seed is not None:
	generator.manual_seed(seed)

	t_start = time.time()

	# ------------------------------------------------------------------
	# STAGE 1: Text Encoding (Qwen3 NVFP4 Text Encoder)
	# ------------------------------------------------------------------
	prompt_embeds = None
	text_ids = None
	try:
	print("[Flux2Klein] Stage 1: Moving Text Encoder to GPU...")
	self.text_encoder.to(device)
	_flush()

	prompt_embeds, text_ids = self.inner_pipe.encode_prompt(
	prompt=prompt,
	device=device,
	num_images_per_prompt=num_images_per_prompt,
	max_sequence_length=512,
	)
	# Store CPU copies until Stage 3 to minimize VRAM
	prompt_embeds = prompt_embeds.to("cpu")
	text_ids = text_ids.to("cpu")
	finally:
	self.text_encoder.to("cpu")
	_flush()
	print(f"[Flux2Klein] Stage 1 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

	# ------------------------------------------------------------------
	# STAGE 2: VAE Conditioning (if reference images provided)
	# ------------------------------------------------------------------
	condition_images = None
	if image is not None:
	if not isinstance(image, list):
	image = [image]
	condition_images = []
	for img in image:
	self.image_processor.check_image_input(img)
	img_w, img_h = img.size
	if img_w * img_h > 1024 * 1024:
	img = self.image_processor._resize_to_target_area(img, 1024 * 1024)
	img_w, img_h = img.size
	multiple_of = self.vae_scale_factor * 2
	img_w = (img_w // multiple_of) * multiple_of
	img_h = (img_h // multiple_of) * multiple_of
	img_pre = self.image_processor.preprocess(img, height=img_h, width=img_w, resize_mode="crop")
	condition_images.append(img_pre)
	height = height or img_h
	width = width or img_w

	height = height or 1024
	width = width or 1024

	image_latents = None
	image_latent_ids = None
	if condition_images is not None:
	try:
	print(f"[Flux2Klein] Stage 2: Moving VAE to GPU to encode {len(condition_images)} reference image(s)...")
	self.vae.to(device)
	_flush()

	image_latents, image_latent_ids = self.inner_pipe.prepare_image_latents(
	images=condition_images,
	batch_size=num_images_per_prompt,
	generator=generator,
	device=device,
	dtype=self.vae.dtype,
	)
	image_latents = image_latents.to("cpu")
	image_latent_ids = image_latent_ids.to("cpu")
	finally:
	self.vae.to("cpu")
	_flush()
	print(f"[Flux2Klein] Stage 2 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

	# ------------------------------------------------------------------
	# STAGE 3: Transformer Denoising Loop (Nunchaku FP4/INT4 KV-Cache)
	# ------------------------------------------------------------------
	num_channels_latents = self.transformer.config.in_channels // 4
	latents, latent_ids = self.inner_pipe.prepare_latents(
	batch_size=num_images_per_prompt,
	num_latents_channels=num_channels_latents,
	height=height,
	width=width,
	dtype=self.transformer.dtype,
	device=device,
	generator=generator,
	)

	sigmas = np.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps)
	image_seq_len = latents.shape[1]
	mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps)
	timesteps, num_inference_steps = retrieve_timesteps(
	self.scheduler,
	num_inference_steps,
	device,
	sigmas=sigmas,
	mu=mu,
	)
	self.scheduler.set_begin_index(0)

	kv_cache = None
	try:
	print("[Flux2Klein] Stage 3: Moving Nunchaku Transformer to GPU...")
	self.transformer.to(device)
	_flush()

	prompt_embeds = prompt_embeds.to(device)
	text_ids = text_ids.to(device)
	latents = latents.to(device)
	latent_ids = latent_ids.to(device)
	if image_latents is not None:
	image_latents = image_latents.to(device)
	image_latent_ids = image_latent_ids.to(device)

	for i, t in enumerate(timesteps):
	t_step = time.time()
	timestep = t.expand(latents.shape[0]).to(latents.dtype)

	if i == 0 and image_latents is not None:
	# Step 0: Extract KV cache for reference tokens
	latent_model_input = torch.cat([image_latents, latents], dim=1).to(self.transformer.dtype)
	latent_image_ids = torch.cat([image_latent_ids, latent_ids], dim=1)
	noise_pred, kv_cache = self.transformer(
	hidden_states=latent_model_input,
	timestep=timestep / 1000,
	guidance=None,
	encoder_hidden_states=prompt_embeds,
	txt_ids=text_ids,
	img_ids=latent_image_ids,
	joint_attention_kwargs=None,
	return_dict=False,
	kv_cache_mode="extract",
	num_ref_tokens=image_latents.shape[1],
	)
	elif kv_cache is not None:
	# Steps 1+: Reuse cached reference KV
	noise_pred = self.transformer(
	hidden_states=latents.to(self.transformer.dtype),
	timestep=timestep / 1000,
	guidance=None,
	encoder_hidden_states=prompt_embeds,
	txt_ids=text_ids,
	img_ids=latent_ids,
	joint_attention_kwargs=None,
	return_dict=False,
	kv_cache=kv_cache,
	kv_cache_mode="cached",
	)[0]
	else:
	# Standard text-to-image forward
	noise_pred = self.transformer(
	hidden_states=latents.to(self.transformer.dtype),
	timestep=timestep / 1000,
	guidance=None,
	encoder_hidden_states=prompt_embeds,
	txt_ids=text_ids,
	img_ids=latent_ids,
	joint_attention_kwargs=None,
	return_dict=False,
	)[0]

	latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
	print(
	f"[Flux2Klein] Step {i+1}/{len(timesteps)} in {time.time() - t_step:.2f}s \| "
	f"Peak VRAM: {torch.cuda.max_memory_allocated() / 1e9:.2f} GB"
	)

	if "noise_pred" in locals():
	del noise_pred
	finally:
	if kv_cache is not None:
	kv_cache.clear()
	self.transformer.to("cpu")
	del prompt_embeds, text_ids, image_latents, image_latent_ids
	_flush()
	print(f"[Flux2Klein] Stage 3 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

	# ------------------------------------------------------------------
	# STAGE 4: VAE Decoding
	# ------------------------------------------------------------------
	try:
	print("[Flux2Klein] Stage 4: Moving VAE to GPU to decode latents...")
	self.vae.to(device)
	_flush()

	latents = self.inner_pipe._unpack_latents_with_ids(latents, latent_ids)
	latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.device, latents.dtype)
	latents_bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(
	latents.device, latents.dtype
	)
	latents = latents * latents_bn_std + latents_bn_mean
	latents = self.inner_pipe._unpatchify_latents(latents)
	decoded = self.vae.decode(latents, return_dict=False)[0]
	out_images = self.image_processor.postprocess(decoded, output_type="pil")
	finally:
	self.vae.to("cpu")
	del latents, latent_ids
	_flush()
	print(f"[Flux2Klein] Stage 4 complete. Allocated VRAM: {torch.cuda.memory_allocated() / 1e9:.3f} GB")

	print(f"[Flux2Klein] Total generation finished in {time.time() - t_start:.2f}s!")
	return Flux2PipelineOutput(images=out_images)


	class Flux2KleinBackend:
	"""Backend factory for FLUX.2 Klein 9B KV Nunchaku."""

	def __init__(
	self,
	model_id: str,
	optimized_model_path: Optional[str] = None,
	nvfp4_text_encoder_path: Optional[str] = None,
	):
	self.model_id = model_id
	self.optimized_model_path = optimized_model_path
	self.nvfp4_text_encoder_path = nvfp4_text_encoder_path
	self.pipeline = None

	def load(self):
	print(f"[Flux2KleinBackend] Initializing backend from base model: {self.model_id}...")

	# 1. Resolve optimized transformer path
	opt_path = self.optimized_model_path
	if not opt_path:
	precision = get_precision()
	opt_path = os.path.join(
	self.model_id,
	f"svdq-{precision}_r32-FLUX.2-klein-9b-kv-Nunchaku.safetensors",
	)
	if not os.path.exists(opt_path):
	# Fallback to fp4 if not found directly
	opt_path = os.path.join(
	self.model_id,
	"svdq-fp4_r32-FLUX.2-klein-9b-kv-Nunchaku.safetensors",
	)

	print(f"[Flux2KleinBackend] Loading Nunchaku transformer from {opt_path}...")
	transformer = NunchakuFlux2Transformer2DModel.from_pretrained(
	opt_path,
	torch_dtype=torch.bfloat16,
	)
	transformer.to("cpu")

	# 2. Resolve NVFP4 text encoder path
	enc_path = self.nvfp4_text_encoder_path
	if not enc_path:
	# Check if running directly from within the text encoder repository (extras/)
	candidate_parent = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
	if os.path.exists(os.path.join(candidate_parent, "config.json")):
	enc_path = candidate_parent
	elif os.path.exists(os.path.join(self.model_id, "text_encoder")):
	enc_path = os.path.join(self.model_id, "text_encoder")
	else:
	enc_path = "/path/to/FLUX.2-klein-text-encoder-NVFP4"

	print(f"[Flux2KleinBackend] Loading NVFP4 text encoder from {enc_path}...")
	text_encoder = load_nvfp4_text_encoder(enc_path, device="cpu", dtype=torch.bfloat16)
	tokenizer = AutoTokenizer.from_pretrained(enc_path)

	# 3. Load VAE and Scheduler
	print(f"[Flux2KleinBackend] Loading VAE and Scheduler from {self.model_id}...")
	vae = AutoencoderKLFlux2.from_pretrained(
	self.model_id,
	subfolder="vae",
	torch_dtype=torch.bfloat16,
	)
	vae.to("cpu")
	scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
	self.model_id,
	subfolder="scheduler",
	)

	# 4. Construct Sequential Pipeline
	self.pipeline = SequentialFlux2KleinPipeline(
	scheduler=scheduler,
	vae=vae,
	text_encoder=text_encoder,
	tokenizer=tokenizer,
	transformer=transformer,
	device="cuda",
	)

	print("[Flux2KleinBackend] Pipeline successfully loaded and ready on CPU!")
	return self.pipeline, self.pipeline