ParaAttention

FLUX.1-dev에 first block cache를 적용하려면 아래처럼 apply_cache_on_pipe를 호출하세요. 0.08은 FLUX 모델의 기본 residual difference 값입니다.

import time
import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    dtype=torch.bfloat16,
).to("cuda")  # or "mps", "xpu", "cpu"

from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe

apply_cache_on_pipe(pipe, residual_diff_threshold=0.08)

# Enable memory savings
# pipe.enable_model_cpu_offload()
# pipe.enable_sequential_cpu_offload()

begin = time.time()
image = pipe(
    "A cat holding a sign that says hello world",
    num_inference_steps=28,
).images[0]
end = time.time()
print(f"Time: {end - begin:.2f}s")

print("Saving image to flux.png")
image.save("flux.png")

최적화	원본	FBCache rdt=0.06	FBCache rdt=0.08	FBCache rdt=0.10	FBCache rdt=0.12
미리 보기
실제 시간 (초)	26.36	21.83	17.01	16.00	13.78

First Block Cache는 베이스라인 대비 추론 속도를 17.01초로 줄여 1.55배 더 빠르게 만들었으며, 거의 품질 손실이 없도록 유지했습니다.

import time
import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    dtype=torch.bfloat16,
).to("cuda")  # or "mps", "xpu", "cpu"

from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe

apply_cache_on_pipe(
    pipe,
    residual_diff_threshold=0.12,  # Use a larger value to make the cache take effect
)

from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, float8_weight_only

quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
pipe.transformer = torch.compile(
   pipe.transformer, mode="max-autotune-no-cudagraphs",
)

# Enable memory savings
# pipe.enable_model_cpu_offload()
# pipe.enable_sequential_cpu_offload()

for i in range(2):
    begin = time.time()
    image = pipe(
        "A cat holding a sign that says hello world",
        num_inference_steps=28,
    ).images[0]
    end = time.time()
    if i == 0:
        print(f"Warm up time: {end - begin:.2f}s")
    else:
        print(f"Time: {end - begin:.2f}s")

print("Saving image to flux.png")
image.save("flux.png")

fp8 동적 양자화 및 torch.compile을 적용하면 베이스라인 대비 추론 속도가 7.56초로 줄어 3.48배 빨라집니다.

import time
import torch
from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
from diffusers.utils import export_to_video

model_id = "tencent/HunyuanVideo"
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id,
    subfolder="transformer",
    dtype=torch.bfloat16,
    revision="refs/pr/18",
)
pipe = HunyuanVideoPipeline.from_pretrained(
    model_id,
    transformer=transformer,
    dtype=torch.float16,
    revision="refs/pr/18",
).to("cuda")  # or "mps", "xpu", "cpu"

from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe

apply_cache_on_pipe(pipe)

from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, float8_weight_only

quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
pipe.transformer = torch.compile(
   pipe.transformer, mode="max-autotune-no-cudagraphs",
)

# Enable memory savings
pipe.vae.enable_tiling()
# pipe.enable_model_cpu_offload()
# pipe.enable_sequential_cpu_offload()

for i in range(2):
    begin = time.time()
    output = pipe(
        prompt="A cat walks on the grass, realistic",
        height=720,
        width=1280,
        num_frames=129,
        num_inference_steps=1 if i == 0 else 30,
    ).frames[0]
    end = time.time()
    if i == 0:
        print(f"Warm up time: {end - begin:.2f}s")
    else:
        print(f"Time: {end - begin:.2f}s")

print("Saving video to hunyuan_video.mp4")
export_to_video(output, "hunyuan_video.mp4", fps=15)

NVIDIA L20 GPU는 메모리가 48GB에 불과하여, 컴파일 후 enable_model_cpu_offload를 호출하지 않으면 OOM(Out-Of-Memory) 오류가 발생할 수 있습니다. 이는 HunyuanVideo가 고해상도와 많은 프레임 수로 실행될 때 매우 큰 활성화 텐서를 생성하기 때문입니다. 80GB 미만의 메모리를 가진 GPU에서는 OOM 오류를 피하기 위해 해상도와 프레임 수를 줄여 보실 수 있습니다.

대형 비디오 생성 모델은 일반적으로 완전 연결 계층보다 어텐션 계산에 병목 현상이 발생합니다. 이러한 모델은 양자화 및 torch.compile에서 상당한 이점을 얻지 못합니다.

아래 코드 샘플은 First Block Cache, fp8 동적 양자화, torch.compile, Context Parallelism을 결합하여 가장 빠른 추론 속도를 제공합니다.

import time
import torch
import torch.distributed as dist
from diffusers import FluxPipeline

dist.init_process_group()

torch.cuda.set_device(dist.get_rank())

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    dtype=torch.bfloat16,
).to("cuda")  # or "mps", "xpu", "cpu"

from para_attn.context_parallel import init_context_parallel_mesh
from para_attn.context_parallel.diffusers_adapters import parallelize_pipe
from para_attn.parallel_vae.diffusers_adapters import parallelize_vae

mesh = init_context_parallel_mesh(
    pipe.device.type,
    max_ring_dim_size=2,
)
parallelize_pipe(
    pipe,
    mesh=mesh,
)
parallelize_vae(pipe.vae, mesh=mesh._flatten())

from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe

apply_cache_on_pipe(
    pipe,
    residual_diff_threshold=0.12,  # Use a larger value to make the cache take effect
)

from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, float8_weight_only

quantize_(pipe.text_encoder, float8_weight_only())
quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
torch._inductor.config.reorder_for_compute_comm_overlap = True
pipe.transformer = torch.compile(
   pipe.transformer, mode="max-autotune-no-cudagraphs",
)

# Enable memory savings
# pipe.enable_model_cpu_offload(gpu_id=dist.get_rank())
# pipe.enable_sequential_cpu_offload(gpu_id=dist.get_rank())

for i in range(2):
    begin = time.time()
    image = pipe(
        "A cat holding a sign that says hello world",
        num_inference_steps=28,
        output_type="pil" if dist.get_rank() == 0 else "pt",
    ).images[0]
    end = time.time()
    if dist.get_rank() == 0:
        if i == 0:
            print(f"Warm up time: {end - begin:.2f}s")
        else:
            print(f"Time: {end - begin:.2f}s")

if dist.get_rank() == 0:
    print("Saving image to flux.png")
    image.save("flux.png")

dist.destroy_process_group()

run_flux.py에 저장하고 torchrun으로 실행하세요.

# Use --nproc_per_node to specify the number of GPUs
torchrun --nproc_per_node=2 run_flux.py

2개의 NVIDIA L20 GPU를 사용할 때, 기준값 대비 추론 속도가 8.20초로 감소하여 3.21배 빨라집니다. 4개의 L20을 사용할 경우, 추론 속도는 3.90초로, 6.75배 빨라집니다.

아래 코드 샘플은 First Block Cache와 Context Parallelism을 결합하여 가장 빠른 추론 속도를 제공합니다.

import time
import torch
import torch.distributed as dist
from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
from diffusers.utils import export_to_video

dist.init_process_group()

torch.cuda.set_device(dist.get_rank())

model_id = "tencent/HunyuanVideo"
transformer = HunyuanVideoTransformer3DModel.from_pretrained(
    model_id,
    subfolder="transformer",
    dtype=torch.bfloat16,
    revision="refs/pr/18",
)
pipe = HunyuanVideoPipeline.from_pretrained(
    model_id,
    transformer=transformer,
    dtype=torch.float16,
    revision="refs/pr/18",
).to("cuda")  # or "mps", "xpu", "cpu"

from para_attn.context_parallel import init_context_parallel_mesh
from para_attn.context_parallel.diffusers_adapters import parallelize_pipe
from para_attn.parallel_vae.diffusers_adapters import parallelize_vae

mesh = init_context_parallel_mesh(
    pipe.device.type,
)
parallelize_pipe(
    pipe,
    mesh=mesh,
)
parallelize_vae(pipe.vae, mesh=mesh._flatten())

from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe

apply_cache_on_pipe(pipe)

# from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight, float8_weight_only
#
# torch._inductor.config.reorder_for_compute_comm_overlap = True
#
# quantize_(pipe.text_encoder, float8_weight_only())
# quantize_(pipe.transformer, float8_dynamic_activation_float8_weight())
# pipe.transformer = torch.compile(
#    pipe.transformer, mode="max-autotune-no-cudagraphs",
# )

# Enable memory savings
pipe.vae.enable_tiling()
# pipe.enable_model_cpu_offload(gpu_id=dist.get_rank())
# pipe.enable_sequential_cpu_offload(gpu_id=dist.get_rank())

for i in range(2):
    begin = time.time()
    output = pipe(
        prompt="A cat walks on the grass, realistic",
        height=720,
        width=1280,
        num_frames=129,
        num_inference_steps=1 if i == 0 else 30,
        output_type="pil" if dist.get_rank() == 0 else "pt",
    ).frames[0]
    end = time.time()
    if dist.get_rank() == 0:
        if i == 0:
            print(f"Warm up time: {end - begin:.2f}s")
        else:
            print(f"Time: {end - begin:.2f}s")

if dist.get_rank() == 0:
    print("Saving video to hunyuan_video.mp4")
    export_to_video(output, "hunyuan_video.mp4", fps=15)

dist.destroy_process_group()

run_hunyuan_video.py에 저장하고 torchrun으로 실행하세요.

# Use --nproc_per_node to specify the number of GPUs
torchrun --nproc_per_node=8 run_hunyuan_video.py

8개의 NVIDIA L20 GPU를 사용할 때, 추론 속도는 베이스라인 대비 649.23초로 감소하여 5.66배 빨라집니다.

GPU 유형	GPU 수	최적화	실제 시간 (초)	속도 향상
NVIDIA L20	1	베이스라인	26.36	1.00x
NVIDIA L20	1	FBCache (rdt=0.08)	17.01	1.55x
NVIDIA L20	1	FP8 DQ	13.40	1.96x
NVIDIA L20	1	FBCache (rdt=0.12) + FP8 DQ	7.56	3.48x
NVIDIA L20	2	FBCache (rdt=0.12) + FP8 DQ + CP	4.92	5.35x
NVIDIA L20	4	FBCache (rdt=0.12) + FP8 DQ + CP	3.90	6.75x

GPU 유형	GPU 수	최적화	실제 시간 (초)	속도 향상
NVIDIA L20	1	베이스라인	3675.71	1.00x
NVIDIA L20	1	FBCache	2271.06	1.62x
NVIDIA L20	2	FBCache + CP	1132.90	3.24x
NVIDIA L20	4	FBCache + CP	718.15	5.12x
NVIDIA L20	8	FBCache + CP	649.23	5.66x

Diffusers

첫 번째 블록 캐시

fp8 양자화

컨텍스트 병렬

벤치마크