[core] Pyramid Attention Broadcast

[a-r-r-o-w]

What does this PR do?

Adds support for Pyramid Attention Broadcast.

• Paper: https://www.arxiv.org/abs/2408.12588
• Project Page: https://oahzxl.github.io/PAB/
• Code: https://github.com/NUS-HPC-AI-Lab/VideoSys

model_id	cache_method	time	model_memory	model_max_memory_reserved	inference_memory	inference_max_memory_reserved
latte	none	31.742	12.88	13.189	12.888	14.344
latte	fastercache	24.288	12.882	13.17	12.891	22.199
latte	pyramid_attention_broadcast	27.531	12.882	13.17	12.891	20.23
cogvideox-1.0	none	245.939	19.66	19.678	19.671	24.426
cogvideox-1.0	fastercache	159.031	19.661	19.678	19.672	40.721
cogvideox-1.0	pyramid_attention_broadcast	184.013	19.661	19.678	19.672	32.57
mochi	none	437.327	28.411	28.65	28.421	36.062
mochi	fastercache	358.871	28.411	28.648	28.422	36.062
mochi	pyramid_attention_broadcast	324.051	28.411	28.65	28.421	52.088
hunyuan_video	none	72.628	38.577	38.672	38.587	41.141
hunyuan_video	pyramid_attention_broadcast	63.892	38.578	38.672	38.587	44.785
flux	none	16.802	31.44	31.451	31.448	32.023
flux	pyramid_attention_broadcast	13.719	31.439	31.451	31.447	32.832

Usage

importtorchfromdiffusersimportCogVideoXPipeline, PyramidAttentionBroadcastConfig, apply_pyramid_attention_broadcastfromdiffusers.utilsimportexport_to_videopipe=CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
pipe.to("cuda")
config=PyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(100, 800),
current_timestep_callback=lambda: pipe._current_timestep,
)
apply_pyramid_attention_broadcast(pipe.transformer, config)
prompt= (
"A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. ""The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other ""pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, ""casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. ""The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical ""atmosphere of this unique musical performance."
)
video=pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
export_to_video(video, "output.mp4", fps=8)

Benchmark code

importargparseimportgcimportpathlibimporttracebackimportgitimportpandasaspdimporttorchfromdiffusersimport (
AllegroPipeline,
CogVideoXPipeline,
FluxPipeline,
HunyuanVideoPipeline,
LattePipeline,
MochiPipeline,
)
fromdiffusers.modelsimportHunyuanVideoTransformer3DModelfromdiffusers.utilsimportexport_to_videofromdiffusers.utils.loggingimportset_verbosity_info, set_verbosity_debugfromtabulateimporttabulaterepo=git.Repo(path="/home/aryan/work/diffusers")
branch=repo.active_branchifbranch.namein ["pyramid-attention-broadcast", "pyramid-attention-rewrite-2"]:
fromdiffusersimport (
apply_pyramid_attention_broadcast,
PyramidAttentionBroadcastConfig,
)
elifbranch.namein ["fastercache"]:
fromdiffusers.pipelines.fastercache_utilsimportapply_fastercache, FasterCacheConfigdefpretty_print_results(results, precision: int=3):
defformat_value(value):
ifisinstance(value, float):
returnf"{value:.{precision}f}"returnvaluefiltered_table= {k: format_value(v) fork, vinresults.items()}
print(tabulate([filtered_table], headers="keys", tablefmt="pipe", stralign="center"))
defbenchmark_fn(f, *args, **kwargs):
torch.cuda.synchronize()
start=torch.cuda.Event(enable_timing=True)
end=torch.cuda.Event(enable_timing=True)
start.record()
output=f(*args, **kwargs)
end.record()
torch.cuda.synchronize()
elapsed_time=round(start.elapsed_time(end) /1000, 3)
returnelapsed_time, outputdefprepare_allegro(dtype: torch.dtype, compile: bool=False, **kwargs):
model_id="rhymes-ai/Allegro"cache_dir=Nonepipe=AllegroPipeline.from_pretrained(model_id, torch_dtype=dtype, cache_dir=cache_dir)
pipe.to("cuda")
pipe.vae.enable_tiling()
ifcompile:
pipe.transformer=torch.compile(
pipe.transformer, mode="max-autotune-no-cudagraphs", fullgraph=True, dynamic=False
)
forkey, valueinlist(kwargs.items()):
iftorch.is_tensor(value):
kwargs[key] =value.to(device="cuda", dtype=dtype)
generation_kwargs= {
"prompt": "A seaside harbor with bright sunlight and sparkling seawater, with many boats in the water. From an aerial view, the boats vary in size and color, some moving and some stationary. Fishing boats in the water suggest that this location might be a popular spot for docking fishing boats.",
"height": 720,
"width": 1280,
"num_inference_steps": 50,
"guidance_scale": 5.0,
**kwargs,
}
returnpipe, generation_kwargsdefprepare_cogvideox_1_0(dtype: torch.dtype, compile: bool=False, **kwargs):
model_id="THUDM/CogVideoX-5b"cache_dir=Nonepipe=CogVideoXPipeline.from_pretrained(model_id, torch_dtype=dtype, cache_dir=cache_dir)
pipe.to("cuda")
ifcompile:
pipe.transformer=torch.compile(
pipe.transformer, mode="max-autotune-no-cudagraphs", fullgraph=True, dynamic=False
)
forkey, valueinlist(kwargs.items()):
iftorch.is_tensor(value):
kwargs[key] =value.to(device="cuda", dtype=dtype)
generation_kwargs= {
"prompt": (
"A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. ""The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other ""pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, ""casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. ""The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical ""atmosphere of this unique musical performance."
),
"height": 480,
"width": 720,
"num_frames": 49,
"num_inference_steps": 50,
"guidance_scale": 5.0,
**kwargs,
}
returnpipe, generation_kwargsdefprepare_flux(dtype: torch.dtype, compile: bool=False, **kwargs) ->None:
model_id="black-forest-labs/Flux.1-Dev"cache_dir="/raid/.cache/huggingface"pipe=FluxPipeline.from_pretrained(model_id, torch_dtype=dtype, cache_dir=cache_dir)
pipe.to("cuda")
ifcompile:
pipe.transformer=torch.compile(
pipe.transformer, mode="max-autotune-no-cudagraphs", fullgraph=True, dynamic=False
)
forkey, valueinlist(kwargs.items()):
iftorch.is_tensor(value):
kwargs[key] =value.to(device="cuda", dtype=dtype)
generation_kwargs= {
"prompt": "A cat holding a sign that says hello world",
"height": 768,
"width": 768,
"num_inference_steps": 50,
"guidance_scale": 5.0,
**kwargs,
}
returnpipe, generation_kwargsdefprepare_hunyuan_video(dtype: torch.dtype, compile: bool=False, **kwargs):
model_id="hunyuanvideo-community/HunyuanVideo"cache_dir=Nonetransformer=HunyuanVideoTransformer3DModel.from_pretrained(
model_id, subfolder="transformer", torch_dtype=torch.bfloat16
)
pipe=HunyuanVideoPipeline.from_pretrained(
model_id, transformer=transformer, torch_dtype=torch.float16, cache_dir=cache_dir
)
pipe.to("cuda")
ifcompile:
pipe.transformer=torch.compile(
pipe.transformer, mode="max-autotune-no-cudagraphs", fullgraph=True, dynamic=False
)
forkey, valueinlist(kwargs.items()):
iftorch.is_tensor(value):
kwargs[key] =value.to(device="cuda", dtype=dtype)
generation_kwargs= {
"prompt": "A cat wearing sunglasses and working as a lifeguard at pool.",
"height": 320,
"width": 512,
"num_frames": 61,
"num_inference_steps": 30,
}
returnpipe, generation_kwargsdefprepare_latte(dtype: torch.dtype, compile: bool=False, **kwargs):
model_id="maxin-cn/Latte-1"cache_dir=Nonepipe=LattePipeline.from_pretrained(model_id, torch_dtype=dtype, cache_dir=cache_dir)
pipe.to("cuda")
ifcompile:
pipe.transformer=torch.compile(
pipe.transformer, mode="max-autotune-no-cudagraphs", fullgraph=True, dynamic=False
)
forkey, valueinlist(kwargs.items()):
iftorch.is_tensor(value):
kwargs[key] =value.to(device="cuda", dtype=dtype)
generation_kwargs= {
"prompt": "a cat wearing sunglasses and working as a lifeguard at pool.",
"height": 512,
"width": 512,
"video_length": 16,
"num_inference_steps": 50,
}
returnpipe, generation_kwargsdefprepare_mochi(dtype: torch.dtype, compile: bool=False, **kwargs):
model_id="genmo/mochi-1-preview"cache_dir=Nonepipe=MochiPipeline.from_pretrained(model_id, torch_dtype=dtype, cache_dir=cache_dir)
pipe.to("cuda")
pipe.vae.enable_tiling()
ifcompile:
pipe.transformer=torch.compile(
pipe.transformer, mode="max-autotune-no-cudagraphs", fullgraph=True, dynamic=False
)
forkey, valueinlist(kwargs.items()):
iftorch.is_tensor(value):
kwargs[key] =value.to(device="cuda", dtype=dtype)
generation_kwargs= {
"prompt": "Close-up of a chameleon's eye, with its scaly skin changing color. Ultra high resolution 4k.",
"height": 480,
"width": 848,
"num_frames": 85,
"num_inference_steps": 50,
}
returnpipe, generation_kwargsdefprepare_allegro_config(cache_method: str):
ifcache_method=="pyramid_attention_broadcast":
returnPyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
cross_attention_block_skip_range=6,
spatial_attention_timestep_skip_range=(100, 700),
cross_attention_timestep_skip_range=(100, 800),
spatial_attention_block_identifiers=["transformer_blocks"],
cross_attention_block_identifiers=["transformer_blocks"],
)
elifcache_method=="fastercache":
returnFasterCacheConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(-1, 681),
low_frequency_weight_update_timestep_range=(99, 641),
high_frequency_weight_update_timestep_range=(-1, 301),
spatial_attention_block_identifiers=["transformer_blocks"],
)
elifcache_method=="none":
returnNonedefprepare_cogvideox_1_0_config(cache_method: str):
ifcache_method=="pyramid_attention_broadcast":
returnPyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(100, 800),
spatial_attention_block_identifiers=["transformer_blocks"],
)
elifcache_method=="fastercache":
returnFasterCacheConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(-1, 681),
low_frequency_weight_update_timestep_range=(99, 641),
high_frequency_weight_update_timestep_range=(-1, 301),
spatial_attention_block_identifiers=["transformer_blocks"],
attention_weight_callback=lambda_: 0.3,
tensor_format="BFCHW",
)
elifcache_method=="none":
returnNonedefprepare_flux_config(cache_method: str):
ifcache_method=="pyramid_attention_broadcast":
returnPyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(100, 950),
spatial_attention_block_identifiers=["transformer_blocks", "single_transformer_blocks"],
)
elifcache_method=="fastercache":
returnFasterCacheConfig(
spatial_attention_block_skip_range=4,
spatial_attention_timestep_skip_range=(-1, 961),
spatial_attention_block_identifiers=["transformer_blocks", "single_transformer_blocks"],
tensor_format="BCHW",
)
elifcache_method=="none":
returnNonedefprepare_hunyuan_video_config(cache_method: str):
ifcache_method=="pyramid_attention_broadcast":
returnPyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(100, 800),
spatial_attention_block_identifiers=["transformer_blocks", "single_transformer_blocks"],
)
elifcache_method=="fastercache":
returnFasterCacheConfig(
spatial_attention_block_skip_range=4,
spatial_attention_timestep_skip_range=(99, 941),
spatial_attention_block_identifiers=["transformer_blocks", "single_transformer_blocks"],
tensor_format="BCFHW",
)
elifcache_method=="none":
returnNonedefprepare_latte_config(cache_method: str):
ifcache_method=="pyramid_attention_broadcast":
returnPyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
temporal_attention_block_skip_range=3,
cross_attention_block_skip_range=6,
spatial_attention_timestep_skip_range=(100, 700),
temporal_attention_timestep_skip_range=(100, 800),
cross_attention_timestep_skip_range=(100, 800),
spatial_attention_block_identifiers=["transformer_blocks"],
temporal_attention_block_identifiers=["temporal_transformer_blocks"],
cross_attention_block_identifiers=["transformer_blocks"],
)
elifcache_method=="fastercache":
returnFasterCacheConfig(
spatial_attention_block_skip_range=2,
temporal_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(-1, 681),
temporal_attention_timestep_skip_range=(-1, 681),
low_frequency_weight_update_timestep_range=(99, 641),
high_frequency_weight_update_timestep_range=(-1, 301),
spatial_attention_block_identifiers=["transformer_blocks.*attn1"],
temporal_attention_block_identifiers=["temporal_transformer_blocks"],
)
elifcache_method=="none":
returnNonedefprepare_mochi_config(cache_method: str):
ifcache_method=="pyramid_attention_broadcast":
returnPyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(400, 987),
spatial_attention_block_identifiers=["transformer_blocks"],
)
elifcache_method=="fastercache":
returnFasterCacheConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(-1, 981),
low_frequency_weight_update_timestep_range=(301, 961),
high_frequency_weight_update_timestep_range=(-1, 851),
unconditional_batch_skip_range=4,
unconditional_batch_timestep_skip_range=(-1, 975),
spatial_attention_block_identifiers=["transformer_blocks"],
attention_weight_callback=lambda_: 0.6,
)
elifcache_method=="none":
returnNonedefdecode_allegro(pipe: AllegroPipeline, latents: torch.Tensor, filename: pathlib.Path, **kwargs):
filename=f"{filename.as_posix()}.mp4"video=pipe.decode_latents(latents)
video=pipe.video_processor.postprocess_video(video=video, output_type="pil")[0]
export_to_video(video, filename, fps=8)
returnfilenamedefdecode_cogvideox_1_0(pipe: CogVideoXPipeline, latents: torch.Tensor, filename: pathlib.Path, **kwargs):
filename=f"{filename.as_posix()}.mp4"video=pipe.decode_latents(latents)
video=pipe.video_processor.postprocess_video(video=video, output_type="pil")[0]
export_to_video(video, filename, fps=8)
returnfilenamedefdecode_flux(pipe: FluxPipeline, latents: torch.Tensor, filename: pathlib.Path, **kwargs):
height=kwargs["height"]
width=kwargs["width"]
filename=f"{filename.as_posix()}.png"latents=pipe._unpack_latents(latents, height, width, pipe.vae_scale_factor)
latents= (latents/pipe.vae.config.scaling_factor) +pipe.vae.config.shift_factorimage=pipe.vae.decode(latents, return_dict=False)[0]
image=pipe.image_processor.postprocess(image, output_type="pil")[0]
image.save(filename)
returnfilenamedefdecode_hunyuan_video(pipe: HunyuanVideoPipeline, latents: torch.Tensor, filename: pathlib.Path, **kwargs):
filename=f"{filename.as_posix()}.mp4"latents=latents.to(pipe.vae.dtype) /pipe.vae.config.scaling_factorvideo=pipe.vae.decode(latents, return_dict=False)[0]
video=pipe.video_processor.postprocess_video(video, output_type="pil")[0]
export_to_video(video, filename, fps=8)
returnfilenamedefdecode_latte(pipe: LattePipeline, latents: torch.Tensor, filename: pathlib.Path, **kwargs):
filename=f"{filename.as_posix()}.mp4"video=pipe.decode_latents(latents, video_length=kwargs["video_length"])
video=pipe.video_processor.postprocess_video(video=video, output_type="pil")[0]
export_to_video(video, filename, fps=8)
returnfilenamedefdecode_mochi(pipe: MochiPipeline, latents: torch.Tensor, filename: pathlib.Path, **kwargs):
filename=f"{filename.as_posix()}.mp4"latents_mean=torch.tensor(pipe.vae.config.latents_mean).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
latents_std=torch.tensor(pipe.vae.config.latents_std).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
latents=latents*latents_std/pipe.vae.config.scaling_factor+latents_meanvideo=pipe.vae.decode(latents, return_dict=False)[0]
video=pipe.video_processor.postprocess_video(video=video, output_type="pil")[0]
export_to_video(video, filename, fps=8)
returnfilenameMODEL_MAPPING= {
"allegro": {
"prepare": prepare_allegro,
"config": prepare_allegro_config,
"decode": decode_allegro,
},
"cogvideox-1.0": {
"prepare": prepare_cogvideox_1_0,
"config": prepare_cogvideox_1_0_config,
"decode": decode_cogvideox_1_0,
},
"flux": {
"prepare": prepare_flux,
"config": prepare_flux_config,
"decode": decode_flux,
},
"hunyuan_video": {
"prepare": prepare_hunyuan_video,
"config": prepare_hunyuan_video_config,
"decode": decode_hunyuan_video,
},
"latte": {
"prepare": prepare_latte,
"config": prepare_latte_config,
"decode": decode_latte,
},
"mochi": {
"prepare": prepare_mochi,
"config": prepare_mochi_config,
"decode": decode_mochi,
},
}
STR_TO_COMPUTE_DTYPE= {
"bf16": torch.bfloat16,
"fp16": torch.float16,
"fp32": torch.float32,
}
defrun_inference(pipe, generation_kwargs):
generator=torch.Generator("cuda").manual_seed(181201)
output=pipe(generator=generator, output_type="latent", **generation_kwargs)[0]
torch.cuda.synchronize()
returnoutput@torch.no_grad()defmain(model_id: str, cache_method: str, output_dir: str, dtype: str, compile: bool=False):
ifmodel_idnotinMODEL_MAPPING.keys():
raiseValueError("Unsupported `model_id` specified.")
output_dir=pathlib.Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
csv_filename=output_dir/f"{model_id}.csv"compute_dtype=STR_TO_COMPUTE_DTYPE[dtype]
model=MODEL_MAPPING[model_id]
try:
torch.cuda.reset_peak_memory_stats()
torch.cuda.reset_accumulated_memory_stats()
gc.collect()
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
torch.cuda.synchronize()
# 1. Prepare inputs and generation kwargspipe, generation_kwargs=model["prepare"](dtype=compute_dtype, compile=compile)
model_memory=round(torch.cuda.memory_allocated() /1024**3, 3)
model_max_memory_reserved=round(torch.cuda.max_memory_reserved() /1024**3, 3)
# 2. Apply attention approximation techniqueconfig=model["config"](cache_method)
ifcache_method=="pyramid_attention_broadcast":
config.current_timestep_callback=lambda: pipe._current_timestepifcache_method=="pyramid_attention_broadcast":
apply_pyramid_attention_broadcast(pipe.transformer, config)
elifcache_method=="fastercache":
apply_fastercache(pipe, config)
elifcache_method=="none":
passelse:
raiseValueError(f"Invalid {cache_method=} provided.")
# 3. Warmupnum_warmups=1original_num_inference_steps=generation_kwargs["num_inference_steps"]
generation_kwargs["num_inference_steps"] =2for_inrange(num_warmups):
run_inference(pipe, generation_kwargs)
generation_kwargs["num_inference_steps"] =original_num_inference_steps# 4. Benchmarktime, latents=benchmark_fn(run_inference, pipe, generation_kwargs)
inference_memory=round(torch.cuda.memory_allocated() /1024**3, 3)
inference_max_memory_reserved=round(torch.cuda.max_memory_reserved() /1024**3, 3)
# 5. Decode latentsfilename=output_dir/f"{model_id}---dtype-{dtype}---cache_method-{cache_method}---compile-{compile}"filename=model["decode"](
pipe,
latents,
filename,
height=generation_kwargs["height"],
width=generation_kwargs["width"],
video_length=generation_kwargs.get("video_length", None),
)
# 6. Save artifactsinfo= {
"model_id": model_id,
"cache_method": cache_method,
"compute_dtype": dtype,
"compile": compile,
"time": time,
"model_memory": model_memory,
"model_max_memory_reserved": model_max_memory_reserved,
"inference_memory": inference_memory,
"inference_max_memory_reserved": inference_max_memory_reserved,
"branch": branch,
"filename": filename,
"exception": None,
}
exceptExceptionase:
print(f"An error occurred: {e}")
traceback.print_exc()
# 6. Save artifactsinfo= {
"model_id": model_id,
"cache_method": cache_method,
"compute_dtype": dtype,
"compile": compile,
"time": None,
"model_memory": None,
"model_max_memory_reserved": None,
"inference_memory": None,
"inference_max_memory_reserved": None,
"branch": branch,
"filename": None,
"exception": str(e),
}
pretty_print_results(info, precision=3)
df=pd.DataFrame([info])
df.to_csv(csv_filename.as_posix(), mode="a", index=False, header=notcsv_filename.is_file())
if__name__=="__main__":
parser=argparse.ArgumentParser()
parser.add_argument(
"--model_id",
type=str,
default="flux",
choices=["flux", "cogvideox-1.0", "latte", "allegro", "hunyuan_video", "mochi"],
help="Model to run benchmark for.",
)
parser.add_argument(
"--cache_method",
type=str,
default="pyramid_attention_broadcast",
choices=["pyramid_attention_broadcast", "fastercache", "none"],
help="Cache method to use.",
)
parser.add_argument(
"--output_dir", type=str, help="Path where the benchmark artifacts and outputs are the be saved."
)
parser.add_argument("--dtype", type=str, help="torch.dtype to use for inference")
parser.add_argument(
"--compile",
action="store_true",
default=False,
help="Whether to torch.compile the denoiser.",
)
parser.add_argument("-v", "--verbose", action="store_true", help="Enable verbose logging.")
args=parser.parse_args()
ifargs.verbose:
set_verbosity_debug()
else:
set_verbosity_info()
main(args.model_id, args.cache_method, args.output_dir, args.dtype, args.compile)

Flux
Mochi
mochi---dtype-bf16---cache_method-none---compile-False.mp4	mochi---dtype-bf16---cache_method-pyramid_attention_broadcast---compile-False.mp4
Hunyuan Video
hunyuan_video---dtype-bf16---cache_method-none---compile-False.mp4	hunyuan_video---dtype-bf16---cache_method-pyramid_attention_broadcast---compile-False.mp4
CogVideoX-2b T2V
cogvideox_2b.mp4	cogvideox_pab_2b.mp4
CogVideoX-5b T2V
cogvideox_5b.mp4	cogvideox_pab_5b.mp4
CogVideoX-5b I2V
cogvideox_5b_i2v.mp4	cogvideox_pab_5b_i2v.mp4
Latte
latte---dtype-fp16---cache_method-none---compile-False.mp4	latte---dtype-fp16---cache_method-pyramid_attention_broadcast---compile-False.mp4

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.

@yiyixuxu@sayakpaul

@oahzxl for PAB, @zRzRzRzRzRzRzR for CogVideoX related changes, @maxin-cn for Latte related changes

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[a-r-r-o-w]

I can't seem to replicate the results for PAB on CogVideoX-5b T2V or I2V. This is what I get:

CogVideoX-5b T2V
cogvideox_5b.mp4	cogvideox_pab_5b.mp4
CogVideoX-5b I2V
cogvideox_5b_i2v.mp4	cogvideox_pab_5b_i2v.mp4

@oahzxl Would you be able to give this a review when free? I'm unable to figure out what I'm doing wrong that's causing poor results in these cases. Thank you!

[oahzxl]

sure, thanks for your code! i guess it may be related with pos embed or encoder concat of 5b model. i can have a look at the code soon!

[oahzxl]

hi, i have done some experiments and here are my conclusions:

i first try a simple implementation

the org attention is:

 attn_hidden_states, attn_encoder_hidden_states = self.attn1(
hidden_states=norm_hidden_states,
encoder_hidden_states=norm_encoder_hidden_states,
image_rotary_emb=image_rotary_emb,
)

for simplicty, i just add pab's logic here:

 # in init
self.attn_count = 0
self.last_attn = None
...
# in forward
if (10 < self.attn_count < 45) and (self.attn_count % 2 != 0):
attn_hidden_states, attn_encoder_hidden_states = self.last_attn
else:
attn_hidden_states, attn_encoder_hidden_states = self.attn1(
hidden_states=norm_hidden_states,
encoder_hidden_states=norm_encoder_hidden_states,
image_rotary_emb=image_rotary_emb,
)
self.last_attn = attn_hidden_states, attn_encoder_hidden_states

this should be exactly the same as the logic in pab processor.

then i find pab will be numerically unstable with fp16 for cogvideox-5b. so i change to bfloat16, and it works!

output-bf16-new.mp4

->> so the first problem is float16!

2. then i test pab processor

but fail even if i use bfloat16

i find even i set spatial_attn_skip_range to 1 (which means no broadcast), it will also generate random noise.

->> so i think the second problem is in processor, but no clue for now

hope it can help you!

[a-r-r-o-w]

Thank you so much for the investigation! I think I found the bug. This line checks if the processor signature supports a specific keyword arguments before passing them. In this case, since we replace the attention processor with PyramidAttentionBroadcastAttentionProcessor, which only has args and kwargs, it drops the image_rotary_emb kwargs necessary for generation. So, RoPE embeddings are not passed at all causing bad video.

[oahzxl]

glad i can help :) !

[sayakpaul]

Thanks for the PR! The numbers are extremely promising!

I want to brainstorm a bit about how we should incorporate PAB design-wise.

IIUC, PAB can be applied at a model-level and it rejigs the attention computation of the concerned model. IMO, this is a bit similar to how we do QKV fusion. Entry point to QKV fusion can either be from a pipeline or from a model (if there's support).

If this is correct, then I wonder if supporting PAB through a Mixin class makes the most elegant design as opposed to enabling it via set_attn_processor().

Or are we relying on a Mixin because we need to depend on pipeline-level attributes which may not be suitable for a model?

[stevhliu]

Very nice, thanks for adding docs for this method! Same comments apply to latte.md :)

[yiyixuxu]

Thanks! I think it will be a nice feature, but not very sure about the design.

• the Attention processor wrapper is not aligned with our design, we should just make custom attention processors (even though we might have to make one for each model that has non default attention processor)
• For another thing, I think this would also won't be compatible with torch.compile, no? I think we should consider a design similar to https://github.com/huggingface/diffusers/pull/9524/files. We can probably store the attention output cache (a dict) on pipeline and pass as cross_attention_kwargs on each iterations (just putting the ideas here. not something I have already carefully thought through, so it might not work. feel free to brainstorm)

[DN6]

Might be misunderstanding here, but shouldn't we just return should_compute_attention directly here? Why use return not should_compute_attention?

[a-r-r-o-w]

This function is used to determine whether attention computation should be skipped or not. So, if skip_callback were to return True, it means that should_compute_attention had to have been False, and vice versa, so this is correct.

[a-r-r-o-w]

@DN6 Addressed the review comments. Could you give this another look?

[a-r-r-o-w]

We need to make some more updates here before merging to address the case of using multiple hooks at once. The current implementation does not really work, if say both FP8 and PAB are enabled together. I will take it up in this PR before merging after layerwise upcasting is merged: #10347

This has already been addressed in group offloading PR but that will take some more time to complete: #10503

[a-r-r-o-w]

With the latest changes, it is now possible to use multiple forward-modifying hooks now. Here's an example with FP8 layerwise-upcasting and PAB:

importtorchfromdiffusersimportCogVideoXPipeline, PyramidAttentionBroadcastConfigfromdiffusers.utilsimportexport_to_videofromdiffusers.utils.loggingimportset_verbosity_debugset_verbosity_debug()
pipe=CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
config=PyramidAttentionBroadcastConfig(
spatial_attention_block_skip_range=2,
spatial_attention_timestep_skip_range=(150, 700),
current_timestep_callback=lambda: pipe.current_timestep,
)
pipe.transformer.enable_cache(config)
pipe.transformer.enable_layerwise_casting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
pipe.to("cuda")
prompt= (
"A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. ""The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other ""pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, ""casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. ""The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical ""atmosphere of this unique musical performance."
)
video=pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
export_to_video(video, "output.mp4", fps=8)

output.mp4

[a-r-r-o-w]

@DN6 Some major changes made to the hooks addition/removal process to be able to support:

• adding multiple hooks to affect the forward pass
• remove hooks arbitrarily (out-of-order is supported as well)

Please take a look when you can. Happy to answer any questions and iterate further if needed

[a-r-r-o-w]

This may look a bit weird - why are we assigning the function reference of next index to the old forward of the hook being removed?

TLDR; Hook invocation order is the reverse of the order in which they are added. Please take a look at the tests related to invocation for better understanding.

When we add hook A followed by hook B, the execution order of methods looks like:

forward:
pre_forward of B
pre_forward of A
actual_module_orginal_forward
post_forward of A
post_forward of B

In this example, the function references would have A-related ones at index 0, and B related ones at index 1. Removing hook A requires pointing the B->old_forward to actual_module_forward. Removing hook B requires pointing module->forward to new_forward(A). We handle both cases here.

Let's take a more complex example to understand better. We add hook A that only has pre/post-forward. We add hook B that has a new_forward implementation. We add hook C that only has pre/post-forward. The invocation order would be

forward:
pre_forward of C pre_forward of B /> pre_forward of A
new_forward of B --/ actual_module_original_forward
post_forward of B \</ post_forward of A
post_forward of C

In this example, the function references would have A-related ones at index 0, and B related ones at index 1, and C related ones at index 2. Removing hook A requires pointing B->old_forward (since we overwrote the original forward implementation by making use of a new_forward method) to actual_module_original_forward. Removing hook B requires pointing C->old_forward to new_forward(A). Removing hook C requires pointing module->forward to new_forward(B)

On a separate note from the explanation, the invocation design here is very friendly with parellism too IMO, so we can eventually introduce different parallel methods utilizing the same hook design, without being invasive in the actual modeling implementations themselves.

There are a few tests added to make sure execution order is correct.

[a-r-r-o-w]

If a hook implements a new_forward method, it can choose to make a call to the forward method it overwrote. The overwritten function is always stored in the overwritten_forward attribute of FunctionReference objects

[a-r-r-o-w]

Some basic fast tests to check simple functionality of the hooks on dummy model. This is necessary to make sure all methods are being invoked correctly, and that the hooks are behaving in a predictable manner when added, or arbitrarily removed out-of-order.

We can also support out-of-order hook addition, but currently there is no such use case, so support has not been added.

[a-r-r-o-w]

I think we're good to merge now and also got the approval from Dhruv after working together on latest changes! Thanks for the patience and the reviews everyone 🤗 Will merge once CI is green and wrap up the open cache PRs

@oahzxl Congratulations on the success of your new work - Data centric parallel! I also really liked reading about the pyramid activation checkpointing that was introduced in VideoSys. Thanks for your patience and help, and also for your work that inspired multiple other papers researching caching mechanism specific to video models. We will be sure to integrate as much as possible to make the methods more easily accessible :)