Upload 3 files

README.md

@@ -0,0 +1,8 @@
+---
+license: other
+license_name: flux-1-dev-non-commercial-license
+license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.
+language:
+- en
+inference: true
+---

handler.py

@@ -0,0 +1,203 @@
+# https://github.com/sayakpaul/diffusers-torchao
+# https://github.com/pytorch/ao/releases
+# https://developer.nvidia.com/cuda-gpus
+
+import os
+from typing import Any, Dict
+import gc
+from PIL import Image
+from huggingface_hub import hf_hub_download
+import torch
+from torchao.quantization import quantize_, autoquant, int8_dynamic_activation_int8_weight, int8_dynamic_activation_int4_weight, float8_dynamic_activation_float8_weight
+from torchao.quantization.quant_api import PerRow
+from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, TorchAoConfig
+from transformers import T5EncoderModel, BitsAndBytesConfig
+from optimum.quanto import freeze, qfloat8, quantize
+from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
+from huggingface_inference_toolkit.logging import logger
+
+import subprocess
+subprocess.run("pip list", shell=True)
+
+print(torch.cuda.get_device_name())
+print(torch.cuda.get_device_capability())
+print(torch.cuda.get_arch_list())
+
+IS_NEW_GPU = False if torch.cuda.get_device_capability() < (8, 9) else True
+IS_TURBO = False
+IS_4BIT = True
+IS_COMPILE = False
+IS_AUTOQ = False
+IS_PARA = True
+IS_LVRAM = True
+
+# Set high precision for float32 matrix multiplications.
+# This setting optimizes performance on NVIDIA GPUs with Ampere architecture (e.g., A100, RTX 30 series) or newer.
+torch.set_float32_matmul_precision("high")
+
+if IS_COMPILE:
+    import torch._dynamo
+    torch._dynamo.config.suppress_errors = True
+
+def offload_pipe(pipe) -> Any:
+    if IS_LVRAM: pipe.enable_model_cpu_offload()
+    return pipe
+
+def load_te2(repo_id: str, dtype: torch.dtype) -> Any:
+    if IS_4BIT:
+        nf4_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.bfloat16)
+        text_encoder_2 = T5EncoderModel.from_pretrained(repo_id, subfolder="text_encoder_2", torch_dtype=dtype, quantization_config=nf4_config)
+    else:
+        text_encoder_2 = T5EncoderModel.from_pretrained(repo_id, subfolder="text_encoder_2", torch_dtype=dtype)
+        quantize(text_encoder_2, weights=qfloat8)
+        freeze(text_encoder_2)
+    return text_encoder_2
+
+def load_pipeline_stable(repo_id: str, dtype: torch.dtype) -> Any:
+    quantization_config = TorchAoConfig("int4dq" if IS_4BIT else "float8dq" if IS_NEW_GPU else "int8wo")
+    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
+    pipe = offload_pipe(FluxPipeline.from_pretrained(repo_id, vae=vae, text_encoder_2=load_te2(repo_id, dtype), torch_dtype=dtype, quantization_config=quantization_config))
+    pipe.transformer.fuse_qkv_projections()
+    pipe.vae.fuse_qkv_projections()
+    return pipe
+
+def load_pipeline_lowvram(repo_id: str, dtype: torch.dtype) -> Any:
+    int4_config = TorchAoConfig("int4dq")
+    float8_config = TorchAoConfig("float8dq")
+    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
+    transformer = AutoencoderKL.from_pretrained(repo_id, subfolder="transformer", torch_dtype=dtype, quantization_config=float8_config)
+    pipe = FluxPipeline.from_pretrained(repo_id, vae=vae, transformer=transformer, text_encoder_2=load_te2(repo_id, dtype), torch_dtype=dtype, quantization_config=int4_config)
+    pipe.transformer.fuse_qkv_projections()
+    pipe.vae.fuse_qkv_projections()
+    pipe.to("cuda")
+    return pipe
+
+def load_pipeline_compile(repo_id: str, dtype: torch.dtype) -> Any:
+    quantization_config = TorchAoConfig("int4dq" if IS_4BIT else "float8dq" if IS_NEW_GPU else "int8wo")
+    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
+    pipe = offload_pipe(FluxPipeline.from_pretrained(repo_id, vae=vae, text_encoder_2=load_te2(repo_id, dtype), torch_dtype=dtype, quantization_config=quantization_config))
+    pipe.transformer.fuse_qkv_projections()
+    pipe.vae.fuse_qkv_projections()
+    pipe.transformer.to(memory_format=torch.channels_last)
+    pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
+    pipe.vae.to(memory_format=torch.channels_last)
+    pipe.vae = torch.compile(pipe.vae, mode="max-autotune", fullgraph=True)
+    return pipe
+
+def load_pipeline_autoquant(repo_id: str, dtype: torch.dtype) -> Any:
+    pipe = offload_pipe(FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype))
+    pipe.transformer.fuse_qkv_projections()
+    pipe.vae.fuse_qkv_projections()
+    pipe.transformer.to(memory_format=torch.channels_last)
+    pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
+    pipe.vae.to(memory_format=torch.channels_last)
+    pipe.vae = torch.compile(pipe.vae, mode="max-autotune", fullgraph=True)
+    pipe.transformer = autoquant(pipe.transformer, error_on_unseen=False)
+    pipe.vae = autoquant(pipe.vae, error_on_unseen=False)
+    return pipe
+
+def load_pipeline_turbo(repo_id: str, dtype: torch.dtype) -> Any:
+    pipe = offload_pipe(FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype))
+    pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"), adapter_name="hyper-sd")
+    pipe.set_adapters(["hyper-sd"], adapter_weights=[0.125])
+    pipe.fuse_lora()
+    pipe.unload_lora_weights()
+    pipe.transformer.fuse_qkv_projections()
+    pipe.vae.fuse_qkv_projections()
+    weight = int8_dynamic_activation_int4_weight() if IS_4BIT else int8_dynamic_activation_int8_weight()
+    quantize_(pipe.transformer, weight, device="cuda")
+    quantize_(pipe.vae, weight, device="cuda")
+    return pipe
+
+def load_pipeline_turbo_compile(repo_id: str, dtype: torch.dtype) -> Any:
+    pipe = offload_pipe(FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype))
+    pipe.load_lora_weights(hf_hub_download("ByteDance/Hyper-SD", "Hyper-FLUX.1-dev-8steps-lora.safetensors"), adapter_name="hyper-sd")
+    pipe.set_adapters(["hyper-sd"], adapter_weights=[0.125])
+    pipe.fuse_lora()
+    pipe.unload_lora_weights()
+    pipe.transformer.fuse_qkv_projections()
+    pipe.vae.fuse_qkv_projections()
+    weight = int8_dynamic_activation_int4_weight() if IS_4BIT else int8_dynamic_activation_int8_weight()
+    quantize_(pipe.transformer, weight, device="cuda")
+    quantize_(pipe.vae, weight, device="cuda")
+    pipe.transformer.to(memory_format=torch.channels_last)
+    pipe.transformer = torch.compile(pipe.transformer, mode="max-autotune", fullgraph=True)
+    pipe.vae.to(memory_format=torch.channels_last)
+    pipe.vae = torch.compile(pipe.vae, mode="max-autotune", fullgraph=True)
+    return pipe
+
+def load_pipeline_opt(repo_id: str, dtype: torch.dtype) -> Any:
+    quantization_config = TorchAoConfig("int4dq" if IS_4BIT else "float8dq" if IS_NEW_GPU else "int8wo")
+    weight = int8_dynamic_activation_int4_weight() if IS_4BIT else int8_dynamic_activation_int8_weight()
+    transformer = FluxTransformer2DModel.from_pretrained(repo_id, subfolder="transformer", torch_dtype=dtype)
+    transformer.fuse_qkv_projections()
+    if IS_NEW_GPU: quantize_(transformer, float8_dynamic_activation_float8_weight(granularity=PerRow()), device="cuda")
+    else: quantize_(transformer, weight, device="cuda")
+    transformer.to(memory_format=torch.channels_last)
+    transformer = torch.compile(transformer, mode="max-autotune", fullgraph=True)
+    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae", torch_dtype=dtype)
+    vae.fuse_qkv_projections()
+    if IS_NEW_GPU: quantize_(vae, float8_dynamic_activation_float8_weight(granularity=PerRow()), device="cuda")
+    else: quantize_(vae, weight, device="cuda")
+    vae.to(memory_format=torch.channels_last)
+    vae = torch.compile(vae, mode="max-autotune", fullgraph=True)
+    pipe = offload_pipe(FluxPipeline.from_pretrained(repo_id, transformer=None, vae=None, text_encoder_2=load_te2(repo_id, dtype), torch_dtype=dtype, quantization_config=quantization_config))
+    pipe.transformer = transformer
+    pipe.vae = vae
+    return pipe
+
+class EndpointHandler:
+    def __init__(self, path=""):
+        repo_id = "NoMoreCopyrightOrg/flux-dev-8step" if IS_TURBO else "NoMoreCopyrightOrg/flux-dev"
+        dtype = torch.bfloat16
+        #dtype = torch.float16 # for older nVidia GPUs
+        if IS_AUTOQ: self.pipeline = load_pipeline_autoquant(repo_id, dtype)
+        elif IS_COMPILE: self.pipeline = load_pipeline_opt(repo_id, dtype)
+        elif IS_LVRAM and IS_NEW_GPU: self.pipeline = load_pipeline_lowvram(repo_id, dtype)
+        else: self.pipeline = load_pipeline_stable(repo_id, dtype)
+        if IS_PARA: apply_cache_on_pipe(self.pipeline, residual_diff_threshold=0.12)
+        gc.collect()
+        torch.cuda.empty_cache()
+        self.enable_vae_slicing()
+        self.enable_vae_tiling()
+        if IS_LVRAM:
+            self.pipeline.transformer.to("cuda")
+            self.pipeline.vae.to("cuda")
+        else: self.pipeline.to("cuda")
+        print(self.pipeline)
+
+    def __call__(self, data: Dict[str, Any]) -> Image.Image:
+        logger.info(f"Received incoming request with {data=}")
+
+        if "inputs" in data and isinstance(data["inputs"], str):
+            prompt = data.pop("inputs")
+        elif "prompt" in data and isinstance(data["prompt"], str):
+            prompt = data.pop("prompt")
+        else:
+            raise ValueError(
+                "Provided input body must contain either the key `inputs` or `prompt` with the"
+                " prompt to use for the image generation, and it needs to be a non-empty string."
+            )
+
+        parameters = data.pop("parameters", {})
+
+        num_inference_steps = parameters.get("num_inference_steps", 8 if IS_TURBO else 28)
+        width = parameters.get("width", 1024)
+        height = parameters.get("height", 1024)
+        guidance_scale = parameters.get("guidance_scale", 3.5)
+
+        # seed generator (seed cannot be provided as is but via a generator)
+        seed = parameters.get("seed", 0)
+        generator = torch.manual_seed(seed)
+
+        return self.pipeline(  # type: ignore
+            prompt,
+            height=height,
+            width=width,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            output_type="pil",
+        ).images[0]
+
+

requirements.txt

@@ -0,0 +1,20 @@
+--extra-index-url https://download.pytorch.org/whl/cu126
+torch>=2.6.0
+torchvision
+torchaudio
+huggingface_hub
+torchao>=0.9.0
+diffusers>=0.32.2
+peft
+transformers==4.48.3
+accelerate
+numpy
+scipy
+Pillow
+sentencepiece
+protobuf
+triton
+gemlite
+para-attn
+bitsandbytes
+optimum-quanto