ostris/ai-toolkit
1
0
Fork 0
mirror of https://github.com/ostris/ai-toolkit.git synced 2026-03-13 14:39:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Varous bug fixes

Browse Source
This commit is contained in:
Jaret Burkett
2024-07-10 15:20:04 -06:00
parent c008405480
commit c062b7716c
9 changed files with 294 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
testing/merge_in_text_encoder_adapter.py
View File

@@ -11,10 +11,10 @@ import json
# te_path = "google/flan-t5-xl"
# te_aug_path = "/mnt/Train/out/ip_adapter/t5xx_sd15_v1/t5xx_sd15_v1_000032000.safetensors"
# output_path = "/home/jaret/Dev/models/hf/kl-f16-d42_sd15_t5xl_raw"
model_path = "/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-1024-MS"
te_path = "google/flan-t5-base"
te_aug_path = "/home/jaret/Dev/models/tmp/pixart_sigma_t5base_000227500.safetensors"
output_path = "/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5base_raw"
model_path = "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS"
te_path = "google/flan-t5-large"
te_aug_path = "/home/jaret/Dev/models/tmp/pixart_sigma_t5l_000034000.safetensors"
output_path = "/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw"
print("Loading te adapter")

115
testing/shrink_pixart_sm.py
View File

@@ -2,62 +2,83 @@ import torch
from safetensors.torch import load_file, save_file
from collections import OrderedDict
model_path = "/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_tiny/transformer/diffusion_pytorch_model.orig.safetensors"
output_path = "/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_tiny/transformer/diffusion_pytorch_model.safetensors"
state_dict = load_file(model_path)
meta = OrderedDict()
meta["format"] = "pt"
meta['format'] = "pt"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def reduce_weight(weight, target_size):
weight = weight.to(device, torch.float32)
original_shape = weight.shape
flattened = weight.view(-1, original_shape[-1])
if flattened.shape[1] <= target_size:
return weight
U, S, V = torch.svd(flattened)
reduced = torch.mm(U[:, :target_size], torch.diag(S[:target_size]))
if reduced.shape[1] < target_size:
padding = torch.zeros(reduced.shape[0], target_size - reduced.shape[1], device=device)
reduced = torch.cat((reduced, padding), dim=1)
return reduced.view(original_shape[:-1] + (target_size,))
def reduce_bias(bias, target_size):
bias = bias.to(device, torch.float32)
original_size = bias.shape[0]
if original_size <= target_size:
return torch.nn.functional.pad(bias, (0, target_size - original_size))
else:
return bias.view(-1, original_size // target_size).mean(dim=1)[:target_size]
# Load your original state dict
state_dict = load_file(
"/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw/transformer/diffusion_pytorch_model.orig.safetensors")
# Create a new state dict for the reduced model
new_state_dict = {}
# Move non-blocks over
source_hidden_size = 1152
target_hidden_size = 1024
for key, value in state_dict.items():
if not key.startswith("transformer_blocks."):
new_state_dict[key] = value
value = value.to(device, torch.float32)
if 'weight' in key or 'scale_shift_table' in key:
if value.shape[0] == source_hidden_size:
value = value[:target_hidden_size]
elif value.shape[0] == source_hidden_size * 4:
value = value[:target_hidden_size * 4]
elif value.shape[0] == source_hidden_size * 6:
value = value[:target_hidden_size * 6]
block_names = ['transformer_blocks.{idx}.attn1.to_k.bias', 'transformer_blocks.{idx}.attn1.to_k.weight',
'transformer_blocks.{idx}.attn1.to_out.0.bias', 'transformer_blocks.{idx}.attn1.to_out.0.weight',
'transformer_blocks.{idx}.attn1.to_q.bias', 'transformer_blocks.{idx}.attn1.to_q.weight',
'transformer_blocks.{idx}.attn1.to_v.bias', 'transformer_blocks.{idx}.attn1.to_v.weight',
'transformer_blocks.{idx}.attn2.to_k.bias', 'transformer_blocks.{idx}.attn2.to_k.weight',
'transformer_blocks.{idx}.attn2.to_out.0.bias', 'transformer_blocks.{idx}.attn2.to_out.0.weight',
'transformer_blocks.{idx}.attn2.to_q.bias', 'transformer_blocks.{idx}.attn2.to_q.weight',
'transformer_blocks.{idx}.attn2.to_v.bias', 'transformer_blocks.{idx}.attn2.to_v.weight',
'transformer_blocks.{idx}.ff.net.0.proj.bias', 'transformer_blocks.{idx}.ff.net.0.proj.weight',
'transformer_blocks.{idx}.ff.net.2.bias', 'transformer_blocks.{idx}.ff.net.2.weight',
'transformer_blocks.{idx}.scale_shift_table']
if len(value.shape) > 1 and value.shape[
1] == source_hidden_size and 'attn2.to_k.weight' not in key and 'attn2.to_v.weight' not in key:
value = value[:, :target_hidden_size]
elif len(value.shape) > 1 and value.shape[1] == source_hidden_size * 4:
value = value[:, :target_hidden_size * 4]
# New block idx 0, 1, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 27
elif 'bias' in key:
if value.shape[0] == source_hidden_size:
value = value[:target_hidden_size]
elif value.shape[0] == source_hidden_size * 4:
value = value[:target_hidden_size * 4]
elif value.shape[0] == source_hidden_size * 6:
value = value[:target_hidden_size * 6]
current_idx = 0
for i in range(28):
if i not in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]:
# todo merge in with previous block
for name in block_names:
continue
# try:
# new_state_dict_key = name.format(idx=current_idx - 1)
# old_state_dict_key = name.format(idx=i)
# new_state_dict[new_state_dict_key] = (new_state_dict[new_state_dict_key] * 0.5) + (state_dict[old_state_dict_key] * 0.5)
# except KeyError:
# raise KeyError(f"KeyError: {name.format(idx=current_idx)}")
else:
for name in block_names:
new_state_dict[name.format(idx=current_idx)] = state_dict[name.format(idx=i)]
current_idx += 1
new_state_dict[key] = value
# make sure they are all fp16 and on cpu
# Move all to CPU and convert to float16
for key, value in new_state_dict.items():
new_state_dict[key] = value.to(torch.float16).cpu()
new_state_dict[key] = value.cpu().to(torch.float16)
# save the new state dict
save_file(new_state_dict, output_path, metadata=meta)
# Save the new state dict
save_file(new_state_dict,
"/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw/transformer/diffusion_pytorch_model.safetensors",
metadata=meta)
new_param_count = sum([v.numel() for v in new_state_dict.values()])
old_param_count = sum([v.numel() for v in state_dict.values()])
print(f"Old param count: {old_param_count:,}")
print(f"New param count: {new_param_count:,}")
print("Done!")

110
testing/shrink_pixart_sm2.py Normal file
View File

@@ -0,0 +1,110 @@
import torch
from safetensors.torch import load_file, save_file
from collections import OrderedDict
meta = OrderedDict()
meta['format'] = "pt"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def reduce_weight(weight, target_size):
weight = weight.to(device, torch.float32)
original_shape = weight.shape
if len(original_shape) == 1:
# For 1D tensors, simply truncate
return weight[:target_size]
if original_shape[0] <= target_size:
return weight
# Reshape the tensor to 2D
flattened = weight.reshape(original_shape[0], -1)
# Perform SVD
U, S, V = torch.svd(flattened)
# Reduce the dimensions
reduced = torch.mm(U[:target_size, :], torch.diag(S)).mm(V.t())
# Reshape back to the original shape with reduced first dimension
new_shape = (target_size,) + original_shape[1:]
return reduced.reshape(new_shape)
def reduce_bias(bias, target_size):
bias = bias.to(device, torch.float32)
return bias[:target_size]
# Load your original state dict
state_dict = load_file(
"/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw/transformer/diffusion_pytorch_model.orig.safetensors")
# Create a new state dict for the reduced model
new_state_dict = {}
for key, value in state_dict.items():
value = value.to(device, torch.float32)
if 'weight' in key or 'scale_shift_table' in key:
if value.shape[0] == 1152:
if len(value.shape) == 4:
orig_shape = value.shape
output_shape = (512, orig_shape[1], orig_shape[2], orig_shape[3]) # reshape to (1152, -1)
# reshape to (1152, -1)
value = value.view(value.shape[0], -1)
value = reduce_weight(value, 512)
value = value.view(output_shape)
else:
# value = reduce_weight(value.t(), 576).t().contiguous()
value = reduce_weight(value, 512)
pass
elif value.shape[0] == 4608:
if len(value.shape) == 4:
orig_shape = value.shape
output_shape = (2048, orig_shape[1], orig_shape[2], orig_shape[3])
value = value.view(value.shape[0], -1)
value = reduce_weight(value, 2048)
value = value.view(output_shape)
else:
value = reduce_weight(value, 2048)
elif value.shape[0] == 6912:
if len(value.shape) == 4:
orig_shape = value.shape
output_shape = (3072, orig_shape[1], orig_shape[2], orig_shape[3])
value = value.view(value.shape[0], -1)
value = reduce_weight(value, 3072)
value = value.view(output_shape)
else:
value = reduce_weight(value, 3072)
if len(value.shape) > 1 and value.shape[
1] == 1152 and 'attn2.to_k.weight' not in key and 'attn2.to_v.weight' not in key:
value = reduce_weight(value.t(), 512).t().contiguous() # Transpose before and after reduction
pass
elif len(value.shape) > 1 and value.shape[1] == 4608:
value = reduce_weight(value.t(), 2048).t().contiguous() # Transpose before and after reduction
pass
elif 'bias' in key:
if value.shape[0] == 1152:
value = reduce_bias(value, 512)
elif value.shape[0] == 4608:
value = reduce_bias(value, 2048)
elif value.shape[0] == 6912:
value = reduce_bias(value, 3072)
new_state_dict[key] = value
# Move all to CPU and convert to float16
for key, value in new_state_dict.items():
new_state_dict[key] = value.cpu().to(torch.float16)
# Save the new state dict
save_file(new_state_dict,
"/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw/transformer/diffusion_pytorch_model.safetensors",
metadata=meta)
print("Done!")

100
testing/shrink_pixart_sm3.py Normal file
View File

@@ -0,0 +1,100 @@
import torch
from safetensors.torch import load_file, save_file
from collections import OrderedDict
meta = OrderedDict()
meta['format'] = "pt"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def reduce_weight(weight, target_size):
weight = weight.to(device, torch.float32)
# resize so target_size is the first dimension
tmp_weight = weight.view(1, 1, weight.shape[0], weight.shape[1])
# use interpolate to resize the tensor
new_weight = torch.nn.functional.interpolate(tmp_weight, size=(target_size, weight.shape[1]), mode='bicubic', align_corners=True)
# reshape back to original shape
return new_weight.view(target_size, weight.shape[1])
def reduce_bias(bias, target_size):
bias = bias.view(1, 1, bias.shape[0], 1)
new_bias = torch.nn.functional.interpolate(bias, size=(target_size, 1), mode='bicubic', align_corners=True)
return new_bias.view(target_size)
# Load your original state dict
state_dict = load_file(
"/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw/transformer/diffusion_pytorch_model.orig.safetensors")
# Create a new state dict for the reduced model
new_state_dict = {}
for key, value in state_dict.items():
value = value.to(device, torch.float32)
if 'weight' in key or 'scale_shift_table' in key:
if value.shape[0] == 1152:
if len(value.shape) == 4:
orig_shape = value.shape
output_shape = (512, orig_shape[1], orig_shape[2], orig_shape[3]) # reshape to (1152, -1)
# reshape to (1152, -1)
value = value.view(value.shape[0], -1)
value = reduce_weight(value, 512)
value = value.view(output_shape)
else:
# value = reduce_weight(value.t(), 576).t().contiguous()
value = reduce_weight(value, 512)
pass
elif value.shape[0] == 4608:
if len(value.shape) == 4:
orig_shape = value.shape
output_shape = (2048, orig_shape[1], orig_shape[2], orig_shape[3])
value = value.view(value.shape[0], -1)
value = reduce_weight(value, 2048)
value = value.view(output_shape)
else:
value = reduce_weight(value, 2048)
elif value.shape[0] == 6912:
if len(value.shape) == 4:
orig_shape = value.shape
output_shape = (3072, orig_shape[1], orig_shape[2], orig_shape[3])
value = value.view(value.shape[0], -1)
value = reduce_weight(value, 3072)
value = value.view(output_shape)
else:
value = reduce_weight(value, 3072)
if len(value.shape) > 1 and value.shape[
1] == 1152 and 'attn2.to_k.weight' not in key and 'attn2.to_v.weight' not in key:
value = reduce_weight(value.t(), 512).t().contiguous() # Transpose before and after reduction
pass
elif len(value.shape) > 1 and value.shape[1] == 4608:
value = reduce_weight(value.t(), 2048).t().contiguous() # Transpose before and after reduction
pass
elif 'bias' in key:
if value.shape[0] == 1152:
value = reduce_bias(value, 512)
elif value.shape[0] == 4608:
value = reduce_bias(value, 2048)
elif value.shape[0] == 6912:
value = reduce_bias(value, 3072)
new_state_dict[key] = value
# Move all to CPU and convert to float16
for key, value in new_state_dict.items():
new_state_dict[key] = value.cpu().to(torch.float16)
# Save the new state dict
save_file(new_state_dict,
"/home/jaret/Dev/models/hf/PixArt-Sigma-XL-2-512_MS_t5large_raw/transformer/diffusion_pytorch_model.safetensors",
metadata=meta)
print("Done!")