ComfyUI/comfy/ldm/flipflop_transformer.py

421 lines
18 KiB
Python

from __future__ import annotations
import torch
import torch.cuda as cuda
import copy
from typing import List, Tuple
import comfy.model_management
class FlipFlopContext:
def __init__(self, holder: FlipFlopHolder):
self.holder = holder
self.reset()
def reset(self):
self.num_blocks = len(self.holder.blocks)
self.first_flip = True
self.first_flop = True
self.last_flip = False
self.last_flop = False
# TODO: the 'i' that's passed into func needs to be properly offset to do patches correctly
def __enter__(self):
self.reset()
return self
def __exit__(self, exc_type, exc_value, traceback):
self.holder.compute_stream.record_event(self.holder.cpy_end_event)
def do_flip(self, func, i: int, _, *args, **kwargs):
# flip
self.holder.compute_stream.wait_event(self.holder.cpy_end_event)
with torch.cuda.stream(self.holder.compute_stream):
out = func(i, self.holder.flip, *args, **kwargs)
self.holder.event_flip.record(self.holder.compute_stream)
# while flip executes, queue flop to copy to its next block
next_flop_i = i + 1
if next_flop_i >= self.num_blocks:
next_flop_i = next_flop_i - self.num_blocks
self.last_flip = True
if not self.first_flip:
self.holder._copy_state_dict(self.holder.flop.state_dict(), self.holder.blocks[next_flop_i].state_dict(), self.holder.event_flop, self.holder.cpy_end_event)
if self.last_flip:
self.holder._copy_state_dict(self.holder.flip.state_dict(), self.holder.blocks[0].state_dict(), cpy_start_event=self.holder.event_flip)
self.first_flip = False
return out
def do_flop(self, func, i: int, _, *args, **kwargs):
# flop
if not self.first_flop:
self.holder.compute_stream.wait_event(self.holder.cpy_end_event)
with torch.cuda.stream(self.holder.compute_stream):
out = func(i, self.holder.flop, *args, **kwargs)
self.holder.event_flop.record(self.holder.compute_stream)
# while flop executes, queue flip to copy to its next block
next_flip_i = i + 1
if next_flip_i >= self.num_blocks:
next_flip_i = next_flip_i - self.num_blocks
self.last_flop = True
self.holder._copy_state_dict(self.holder.flip.state_dict(), self.holder.blocks[next_flip_i].state_dict(), self.holder.event_flip, self.holder.cpy_end_event)
if self.last_flop:
self.holder._copy_state_dict(self.holder.flop.state_dict(), self.holder.blocks[1].state_dict(), cpy_start_event=self.holder.event_flop)
self.first_flop = False
return out
@torch.no_grad()
def __call__(self, func, i: int, block: torch.nn.Module, *args, **kwargs):
# flips are even indexes, flops are odd indexes
if i % 2 == 0:
return self.do_flip(func, i, block, *args, **kwargs)
else:
return self.do_flop(func, i, block, *args, **kwargs)
class FlipFlopHolder:
def __init__(self, blocks: List[torch.nn.Module], flip_amount: int, load_device="cuda", offload_device="cpu"):
self.load_device = torch.device(load_device)
self.offload_device = torch.device(offload_device)
self.blocks = blocks
self.flip_amount = flip_amount
self.block_module_size = 0
if len(self.blocks) > 0:
self.block_module_size = comfy.model_management.module_size(self.blocks[0])
self.flip: torch.nn.Module = None
self.flop: torch.nn.Module = None
# TODO: make initialization happen in model management code/model patcher, not here
self.init_flipflop_blocks(self.load_device)
self.compute_stream = cuda.default_stream(self.load_device)
self.cpy_stream = cuda.Stream(self.load_device)
self.event_flip = torch.cuda.Event(enable_timing=False)
self.event_flop = torch.cuda.Event(enable_timing=False)
self.cpy_end_event = torch.cuda.Event(enable_timing=False)
# INIT - is this actually needed?
self.compute_stream.record_event(self.cpy_end_event)
def _copy_state_dict(self, dst, src, cpy_start_event: torch.cuda.Event=None, cpy_end_event: torch.cuda.Event=None):
if cpy_start_event:
self.cpy_stream.wait_event(cpy_start_event)
with torch.cuda.stream(self.cpy_stream):
for k, v in src.items():
dst[k].copy_(v, non_blocking=True)
if cpy_end_event:
cpy_end_event.record(self.cpy_stream)
def context(self):
return FlipFlopContext(self)
def init_flipflop_blocks(self, load_device: torch.device):
self.flip = copy.deepcopy(self.blocks[0]).to(device=load_device)
self.flop = copy.deepcopy(self.blocks[1]).to(device=load_device)
def clean_flipflop_blocks(self):
del self.flip
del self.flop
self.flip = None
self.flop = None
class FlopFlopModule(torch.nn.Module):
def __init__(self, block_types: tuple[str, ...]):
super().__init__()
self.block_types = block_types
self.flipflop: dict[str, FlipFlopHolder] = {}
def setup_flipflop_holders(self, block_percentage: float):
for block_type in self.block_types:
if block_type in self.flipflop:
continue
num_blocks = int(len(self.transformer_blocks) * (1.0-block_percentage))
self.flipflop["transformer_blocks"] = FlipFlopHolder(self.transformer_blocks[num_blocks:], num_blocks)
def clean_flipflop_holders(self):
for block_type in self.flipflop.keys():
self.flipflop[block_type].clean_flipflop_blocks()
del self.flipflop[block_type]
def get_blocks(self, block_type: str) -> torch.nn.ModuleList:
if block_type not in self.block_types:
raise ValueError(f"Block type {block_type} not found in {self.block_types}")
if block_type in self.flipflop:
return getattr(self, block_type)[:self.flipflop[block_type].flip_amount]
return getattr(self, block_type)
def get_all_block_module_sizes(self, sort_by_size: bool = False) -> list[tuple[str, int]]:
'''
Returns a list of (block_type, size).
If sort_by_size is True, the list is sorted by size.
'''
sizes = [(block_type, self.get_block_module_size(block_type)) for block_type in self.block_types]
if sort_by_size:
sizes.sort(key=lambda x: x[1])
return sizes
def get_block_module_size(self, block_type: str) -> int:
return comfy.model_management.module_size(getattr(self, block_type)[0])
# Below is the implementation from contentis' prototype flip flop
class FlipFlopTransformer:
def __init__(self, transformer_blocks: List[torch.nn.Module], block_wrap_fn, out_names: Tuple[str], pinned_staging: bool = False, inference_device="cuda", offloading_device="cpu"):
self.transformer_blocks = transformer_blocks
self.offloading_device = torch.device(offloading_device)
self.inference_device = torch.device(inference_device)
self.staging = pinned_staging
self.flip = copy.deepcopy(self.transformer_blocks[0]).to(device=self.inference_device)
self.flop = copy.deepcopy(self.transformer_blocks[1]).to(device=self.inference_device)
self._cpy_fn = self._copy_state_dict
if self.staging:
self.staging_buffer = self._pin_module(self.transformer_blocks[0]).state_dict()
self._cpy_fn = self._copy_state_dict_with_staging
self.compute_stream = cuda.default_stream(self.inference_device)
self.cpy_stream = cuda.Stream(self.inference_device)
self.event_flip = torch.cuda.Event(enable_timing=False)
self.event_flop = torch.cuda.Event(enable_timing=False)
self.cpy_end_event = torch.cuda.Event(enable_timing=False)
self.block_wrap_fn = block_wrap_fn
self.out_names = out_names
self.num_blocks = len(self.transformer_blocks)
self.extra_run = self.num_blocks % 2
# INIT
self.compute_stream.record_event(self.cpy_end_event)
def _copy_state_dict(self, dst, src, cpy_start_event=None, cpy_end_event=None):
if cpy_start_event:
self.cpy_stream.wait_event(cpy_start_event)
with torch.cuda.stream(self.cpy_stream):
for k, v in src.items():
dst[k].copy_(v, non_blocking=True)
if cpy_end_event:
cpy_end_event.record(self.cpy_stream)
def _copy_state_dict_with_staging(self, dst, src, cpy_start_event=None, cpy_end_event=None):
if cpy_start_event:
self.cpy_stream.wait_event(cpy_start_event)
with torch.cuda.stream(self.cpy_stream):
for k, v in src.items():
self.staging_buffer[k].copy_(v, non_blocking=True)
dst[k].copy_(self.staging_buffer[k], non_blocking=True)
if cpy_end_event:
cpy_end_event.record(self.cpy_stream)
def _pin_module(self, module):
pinned_module = copy.deepcopy(module)
for param in pinned_module.parameters():
param.data = param.data.pin_memory()
# Pin all buffers (if any)
for buffer in pinned_module.buffers():
buffer.data = buffer.data.pin_memory()
return pinned_module
def _reset(self):
if self.extra_run:
self._copy_state_dict(self.flop.state_dict(), self.transformer_blocks[1].state_dict(), cpy_start_event=self.event_flop)
self._copy_state_dict(self.flip.state_dict(), self.transformer_blocks[0].state_dict(), cpy_start_event=self.event_flip)
else:
self._copy_state_dict(self.flip.state_dict(), self.transformer_blocks[0].state_dict(), cpy_start_event=self.event_flip)
self._copy_state_dict(self.flop.state_dict(), self.transformer_blocks[1].state_dict(), cpy_start_event=self.event_flop)
self.compute_stream.record_event(self.cpy_end_event)
@torch.no_grad()
def __call__(self, **feed_dict):
'''
Flip accounts for even blocks (0 is first block), flop accounts for odd blocks.
'''
# separated flip flop refactor
num_blocks = len(self.transformer_blocks)
first_flip = True
first_flop = True
last_flip = False
last_flop = False
for i, block in enumerate(self.transformer_blocks):
is_flip = i % 2 == 0
if is_flip:
# flip
self.compute_stream.wait_event(self.cpy_end_event)
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flip, **feed_dict)
self.event_flip.record(self.compute_stream)
# while flip executes, queue flop to copy to its next block
next_flop_i = i + 1
if next_flop_i >= num_blocks:
next_flop_i = next_flop_i - num_blocks
last_flip = True
if not first_flip:
self._copy_state_dict(self.flop.state_dict(), self.transformer_blocks[next_flop_i].state_dict(), self.event_flop, self.cpy_end_event)
if last_flip:
self._copy_state_dict(self.flip.state_dict(), self.transformer_blocks[0].state_dict(), cpy_start_event=self.event_flip)
first_flip = False
else:
# flop
if not first_flop:
self.compute_stream.wait_event(self.cpy_end_event)
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flop, **feed_dict)
self.event_flop.record(self.compute_stream)
# while flop executes, queue flip to copy to its next block
next_flip_i = i + 1
if next_flip_i >= num_blocks:
next_flip_i = next_flip_i - num_blocks
last_flop = True
self._copy_state_dict(self.flip.state_dict(), self.transformer_blocks[next_flip_i].state_dict(), self.event_flip, self.cpy_end_event)
if last_flop:
self._copy_state_dict(self.flop.state_dict(), self.transformer_blocks[1].state_dict(), cpy_start_event=self.event_flop)
first_flop = False
self.compute_stream.record_event(self.cpy_end_event)
outputs = [feed_dict[name] for name in self.out_names]
if len(outputs) == 1:
return outputs[0]
return tuple(outputs)
@torch.no_grad()
def __call__old(self, **feed_dict):
# contentis' prototype flip flop
# Wait for reset
self.compute_stream.wait_event(self.cpy_end_event)
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flip, **feed_dict)
self.event_flip.record(self.compute_stream)
for i in range(self.num_blocks // 2 - 1):
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flop, **feed_dict)
self.event_flop.record(self.compute_stream)
self._cpy_fn(self.flip.state_dict(), self.transformer_blocks[(i + 1) * 2].state_dict(), self.event_flip,
self.cpy_end_event)
self.compute_stream.wait_event(self.cpy_end_event)
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flip, **feed_dict)
self.event_flip.record(self.compute_stream)
self._cpy_fn(self.flop.state_dict(), self.transformer_blocks[(i + 1) * 2 + 1].state_dict(), self.event_flop,
self.cpy_end_event)
self.compute_stream.wait_event(self.cpy_end_event)
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flop, **feed_dict)
self.event_flop.record(self.compute_stream)
if self.extra_run:
self._cpy_fn(self.flip.state_dict(), self.transformer_blocks[-1].state_dict(), self.event_flip,
self.cpy_end_event)
self.compute_stream.wait_event(self.cpy_end_event)
with torch.cuda.stream(self.compute_stream):
feed_dict = self.block_wrap_fn(self.flip, **feed_dict)
self.event_flip.record(self.compute_stream)
self._reset()
outputs = [feed_dict[name] for name in self.out_names]
if len(outputs) == 1:
return outputs[0]
return tuple(outputs)
# @register("Flux")
# class Flux:
# @staticmethod
# def double_block_wrap(block, **kwargs):
# kwargs["img"], kwargs["txt"] = block(img=kwargs["img"],
# txt=kwargs["txt"],
# vec=kwargs["vec"],
# pe=kwargs["pe"],
# attn_mask=kwargs.get("attn_mask"))
# return kwargs
# @staticmethod
# def single_block_wrap(block, **kwargs):
# kwargs["img"] = block(kwargs["img"],
# vec=kwargs["vec"],
# pe=kwargs["pe"],
# attn_mask=kwargs.get("attn_mask"))
# return kwargs
# double_config = FlipFlopConfig(block_name="double_blocks",
# block_wrap_fn=double_block_wrap,
# out_names=("img", "txt"),
# overwrite_forward="double_transformer_fwd",
# pinned_staging=False)
# single_config = FlipFlopConfig(block_name="single_blocks",
# block_wrap_fn=single_block_wrap,
# out_names=("img",),
# overwrite_forward="single_transformer_fwd",
# pinned_staging=False)
# @staticmethod
# def patch(model):
# patch_model_from_config(model, Flux.double_config)
# patch_model_from_config(model, Flux.single_config)
# return model
# @register("WanModel")
# class Wan:
# @staticmethod
# def wan_blocks_wrap(block, **kwargs):
# kwargs["x"] = block(x=kwargs["x"],
# context=kwargs["context"],
# e=kwargs["e"],
# freqs=kwargs["freqs"],
# context_img_len=kwargs.get("context_img_len"))
# return kwargs
# blocks_config = FlipFlopConfig(block_name="blocks",
# block_wrap_fn=wan_blocks_wrap,
# out_names=("x",),
# overwrite_forward="block_fwd",
# pinned_staging=False)
# @staticmethod
# def patch(model):
# patch_model_from_config(model, Wan.blocks_config)
# return model
# @register("QwenImageTransformer2DModel")
# class QwenImage:
# @staticmethod
# def qwen_blocks_wrap(block, **kwargs):
# kwargs["encoder_hidden_states"], kwargs["hidden_states"] = block(hidden_states=kwargs["hidden_states"],
# encoder_hidden_states=kwargs["encoder_hidden_states"],
# encoder_hidden_states_mask=kwargs["encoder_hidden_states_mask"],
# temb=kwargs["temb"],
# image_rotary_emb=kwargs["image_rotary_emb"],
# transformer_options=kwargs["transformer_options"])
# return kwargs
# blocks_config = FlipFlopConfig(block_name="transformer_blocks",
# block_wrap_fn=qwen_blocks_wrap,
# out_names=("encoder_hidden_states", "hidden_states"),
# overwrite_forward="blocks_fwd",
# pinned_staging=False)
# @staticmethod
# def patch(model):
# patch_model_from_config(model, QwenImage.blocks_config)
# return model