English
Inference Endpoints
File size: 10,179 Bytes
d3841a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
b7d193e
 
 
 
 
 
 
 
 
9ae3cc2
 
d3841a2
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
 
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
 
 
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
 
 
 
 
6551d57
d3841a2
 
 
 
 
 
 
957bbbe
 
6551d57
d3841a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# 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 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 = 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)
    quantize_(vae, float8_dynamic_activation_float8_weight(granularity=PerRow()), device="cuda")
    transformer = FluxTransformer2DModel.from_pretrained(repo_id, subfolder="transformer", torch_dtype=dtype)
    quantize_(transformer, float8_dynamic_activation_float8_weight(granularity=PerRow()), device="cuda")
    pipe = FluxPipeline.from_pretrained(repo_id, vae=None, transformer=None, text_encoder_2=load_te2(repo_id, dtype), torch_dtype=dtype, quantization_config=int4_config)
    pipe.transformer = transformer
    pipe.vae = vae
    #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 = 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 = 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 = 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 = 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 = 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
        print("Loading pipeline...")
        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.pipeline.enable_vae_slicing()
        self.pipeline.enable_vae_tiling()
        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]