Kohaku-Blueleaf eb2f78b4e0
[Training Node] algo support, grad acc, optional grad ckpt (#9015)
* Add factorization utils for lokr

* Add lokr train impl

* Add loha train impl

* Add adapter map for algo selection

* Add optional grad ckpt and algo selection

* Update __init__.py

* correct key name for loha

* Use custom fwd/bwd func and better init for loha

* Support gradient accumulation

* Fix bugs of loha

* use more stable init

* Add OFT training

* linting
2025-07-23 20:57:27 -04:00

162 lines
5.3 KiB
Python

import logging
from typing import Optional
import torch
import comfy.model_management
from .base import WeightAdapterBase, WeightAdapterTrainBase, weight_decompose, factorization
class OFTDiff(WeightAdapterTrainBase):
def __init__(self, weights):
super().__init__()
# Unpack weights tuple from LoHaAdapter
blocks, rescale, alpha, _ = weights
# Create trainable parameters
self.oft_blocks = torch.nn.Parameter(blocks)
if rescale is not None:
self.rescale = torch.nn.Parameter(rescale)
self.rescaled = True
else:
self.rescaled = False
self.block_num, self.block_size, _ = blocks.shape
self.constraint = float(alpha)
self.alpha = torch.nn.Parameter(torch.tensor(alpha), requires_grad=False)
def __call__(self, w):
org_dtype = w.dtype
I = torch.eye(self.block_size, device=self.oft_blocks.device)
## generate r
# for Q = -Q^T
q = self.oft_blocks - self.oft_blocks.transpose(1, 2)
normed_q = q
if self.constraint:
q_norm = torch.norm(q) + 1e-8
if q_norm > self.constraint:
normed_q = q * self.constraint / q_norm
# use float() to prevent unsupported type
r = (I + normed_q) @ (I - normed_q).float().inverse()
## Apply chunked matmul on weight
_, *shape = w.shape
org_weight = w.to(dtype=r.dtype)
org_weight = org_weight.unflatten(0, (self.block_num, self.block_size))
# Init R=0, so add I on it to ensure the output of step0 is original model output
weight = torch.einsum(
"k n m, k n ... -> k m ...",
r,
org_weight,
).flatten(0, 1)
if self.rescaled:
weight = self.rescale * weight
return weight.to(org_dtype)
def passive_memory_usage(self):
"""Calculates memory usage of the trainable parameters."""
return sum(param.numel() * param.element_size() for param in self.parameters())
class OFTAdapter(WeightAdapterBase):
name = "oft"
def __init__(self, loaded_keys, weights):
self.loaded_keys = loaded_keys
self.weights = weights
@classmethod
def create_train(cls, weight, rank=1, alpha=1.0):
out_dim = weight.shape[0]
block_size, block_num = factorization(out_dim, rank)
block = torch.zeros(block_num, block_size, block_size, device=weight.device, dtype=weight.dtype)
return OFTDiff(
(block, None, alpha, None)
)
def to_train(self):
return OFTDiff(self.weights)
@classmethod
def load(
cls,
x: str,
lora: dict[str, torch.Tensor],
alpha: float,
dora_scale: torch.Tensor,
loaded_keys: set[str] = None,
) -> Optional["OFTAdapter"]:
if loaded_keys is None:
loaded_keys = set()
blocks_name = "{}.oft_blocks".format(x)
rescale_name = "{}.rescale".format(x)
blocks = None
if blocks_name in lora.keys():
blocks = lora[blocks_name]
if blocks.ndim == 3:
loaded_keys.add(blocks_name)
else:
blocks = None
if blocks is None:
return None
rescale = None
if rescale_name in lora.keys():
rescale = lora[rescale_name]
loaded_keys.add(rescale_name)
weights = (blocks, rescale, alpha, dora_scale)
return cls(loaded_keys, weights)
def calculate_weight(
self,
weight,
key,
strength,
strength_model,
offset,
function,
intermediate_dtype=torch.float32,
original_weight=None,
):
v = self.weights
blocks = v[0]
rescale = v[1]
alpha = v[2]
if alpha is None:
alpha = 0
dora_scale = v[3]
blocks = comfy.model_management.cast_to_device(blocks, weight.device, intermediate_dtype)
if rescale is not None:
rescale = comfy.model_management.cast_to_device(rescale, weight.device, intermediate_dtype)
block_num, block_size, *_ = blocks.shape
try:
# Get r
I = torch.eye(block_size, device=blocks.device, dtype=blocks.dtype)
# for Q = -Q^T
q = blocks - blocks.transpose(1, 2)
normed_q = q
if alpha > 0: # alpha in oft/boft is for constraint
q_norm = torch.norm(q) + 1e-8
if q_norm > alpha:
normed_q = q * alpha / q_norm
# use float() to prevent unsupported type in .inverse()
r = (I + normed_q) @ (I - normed_q).float().inverse()
r = r.to(weight)
_, *shape = weight.shape
lora_diff = torch.einsum(
"k n m, k n ... -> k m ...",
(r * strength) - strength * I,
weight.view(block_num, block_size, *shape),
).view(-1, *shape)
if dora_scale is not None:
weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
else:
weight += function((strength * lora_diff).type(weight.dtype))
except Exception as e:
logging.error("ERROR {} {} {}".format(self.name, key, e))
return weight