VictorMylle/Kohya-ss-sd-scripts
1
0
Fork 1
mirror of https://github.com/kohya-ss/sd-scripts.git synced 2026-04-08 14:34:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

IPEX fix SDPA

Browse Source
This commit is contained in:
Disty0
2023-12-19 22:59:06 +03:00
parent aff05e043f
commit 8556b9d7f5
2 changed files with 64 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
library/ipex/attention.py
View File

@@ -20,22 +20,21 @@ def torch_bmm_32_bit(input, mat2, *, out=None):
if split_slice_size <= 1:
split_slice_size = 1
break
split_2_slice_size = input_tokens
if split_slice_size * slice_block_size > 4:
slice_block_size_2 = split_slice_size * mat2_shape / 1024 / 1024 * block_multiply
do_split_2 = True
# Find something divisible with the input_tokens
while (split_2_slice_size * slice_block_size_2) > 4:
split_2_slice_size = split_2_slice_size // 2
if split_2_slice_size <= 1:
split_2_slice_size = 1
break
else:
do_split_2 = False
else:
do_split = False
split_2_slice_size = input_tokens
if split_slice_size * slice_block_size > 4:
slice_block_size2 = split_slice_size * mat2_shape / 1024 / 1024 * block_multiply
do_split_2 = True
# Find something divisible with the input_tokens
while (split_2_slice_size * slice_block_size2) > 4:
split_2_slice_size = split_2_slice_size // 2
if split_2_slice_size <= 1:
split_2_slice_size = 1
break
else:
do_split_2 = False
if do_split:
hidden_states = torch.zeros(input.shape[0], input.shape[1], mat2.shape[2], device=input.device, dtype=input.dtype)
for i in range(batch_size_attention // split_slice_size):
@@ -64,42 +63,51 @@ original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_a
def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False):
# ARC GPUs can't allocate more than 4GB to a single block, Slice it:
if len(query.shape) == 3:
batch_size_attention, query_tokens, shape_four = query.shape
shape_one = 1
no_shape_one = True
batch_size_attention, query_tokens, shape_three = query.shape
shape_four = 1
else:
shape_one, batch_size_attention, query_tokens, shape_four = query.shape
no_shape_one = False
batch_size_attention, query_tokens, shape_three, shape_four = query.shape
block_multiply = query.element_size()
slice_block_size = shape_one * query_tokens * shape_four / 1024 / 1024 * block_multiply
slice_block_size = query_tokens * shape_three * shape_four / 1024 / 1024 * block_multiply
block_size = batch_size_attention * slice_block_size
split_slice_size = batch_size_attention
if block_size > 4:
do_split = True
# Find something divisible with the shape_one
# Find something divisible with the batch_size_attention
while (split_slice_size * slice_block_size) > 4:
split_slice_size = split_slice_size // 2
if split_slice_size <= 1:
split_slice_size = 1
break
split_2_slice_size = query_tokens
if split_slice_size * slice_block_size > 4:
slice_block_size_2 = split_slice_size * shape_three * shape_four / 1024 / 1024 * block_multiply
do_split_2 = True
# Find something divisible with the query_tokens
while (split_2_slice_size * slice_block_size_2) > 4:
split_2_slice_size = split_2_slice_size // 2
if split_2_slice_size <= 1:
split_2_slice_size = 1
break
split_3_slice_size = shape_three
if split_2_slice_size * slice_block_size_2 > 4:
slice_block_size_3 = split_slice_size * split_2_slice_size * shape_four / 1024 / 1024 * block_multiply
do_split_3 = True
# Find something divisible with the shape_three
while (split_3_slice_size * slice_block_size_3) > 4:
split_3_slice_size = split_3_slice_size // 2
if split_3_slice_size <= 1:
split_3_slice_size = 1
break
else:
do_split_3 = False
else:
do_split_2 = False
else:
do_split = False
split_2_slice_size = query_tokens
if split_slice_size * slice_block_size > 4:
slice_block_size2 = shape_one * split_slice_size * shape_four / 1024 / 1024 * block_multiply
do_split_2 = True
# Find something divisible with the batch_size_attention
while (split_2_slice_size * slice_block_size2) > 4:
split_2_slice_size = split_2_slice_size // 2
if split_2_slice_size <= 1:
split_2_slice_size = 1
break
else:
do_split_2 = False
if do_split:
hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
for i in range(batch_size_attention // split_slice_size):
@@ -109,7 +117,18 @@ def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropo
for i2 in range(query_tokens // split_2_slice_size): # pylint: disable=invalid-name
start_idx_2 = i2 * split_2_slice_size
end_idx_2 = (i2 + 1) * split_2_slice_size
if no_shape_one:
if do_split_3:
for i3 in range(shape_three // split_3_slice_size): # pylint: disable=invalid-name
start_idx_3 = i3 * split_3_slice_size
end_idx_3 = (i3 + 1) * split_3_slice_size
hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = original_scaled_dot_product_attention(
query[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
key[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
value[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] if attn_mask is not None else attn_mask,
dropout_p=dropout_p, is_causal=is_causal
)
else:
hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = original_scaled_dot_product_attention(
query[start_idx:end_idx, start_idx_2:end_idx_2],
key[start_idx:end_idx, start_idx_2:end_idx_2],
@@ -117,31 +136,14 @@ def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropo
attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2] if attn_mask is not None else attn_mask,
dropout_p=dropout_p, is_causal=is_causal
)
else:
hidden_states[:, start_idx:end_idx, start_idx_2:end_idx_2] = original_scaled_dot_product_attention(
query[:, start_idx:end_idx, start_idx_2:end_idx_2],
key[:, start_idx:end_idx, start_idx_2:end_idx_2],
value[:, start_idx:end_idx, start_idx_2:end_idx_2],
attn_mask=attn_mask[:, start_idx:end_idx, start_idx_2:end_idx_2] if attn_mask is not None else attn_mask,
dropout_p=dropout_p, is_causal=is_causal
)
else:
if no_shape_one:
hidden_states[start_idx:end_idx] = original_scaled_dot_product_attention(
query[start_idx:end_idx],
key[start_idx:end_idx],
value[start_idx:end_idx],
attn_mask=attn_mask[start_idx:end_idx] if attn_mask is not None else attn_mask,
dropout_p=dropout_p, is_causal=is_causal
)
else:
hidden_states[:, start_idx:end_idx] = original_scaled_dot_product_attention(
query[:, start_idx:end_idx],
key[:, start_idx:end_idx],
value[:, start_idx:end_idx],
attn_mask=attn_mask[:, start_idx:end_idx] if attn_mask is not None else attn_mask,
dropout_p=dropout_p, is_causal=is_causal
)
hidden_states[start_idx:end_idx] = original_scaled_dot_product_attention(
query[start_idx:end_idx],
key[start_idx:end_idx],
value[start_idx:end_idx],
attn_mask=attn_mask[start_idx:end_idx] if attn_mask is not None else attn_mask,
dropout_p=dropout_p, is_causal=is_causal
)
else:
return original_scaled_dot_product_attention(
query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal

4
library/ipex/hijacks.py
View File

@@ -185,6 +185,10 @@ def ipex_hijacks():
CondFunc('torch.Generator',
lambda orig_func, device=None: torch.xpu.Generator(return_xpu(device)),
lambda orig_func, device=None: device is not None and device != torch.device("cpu") and device != "cpu")
else:
CondFunc('torch.Generator',
lambda orig_func, device=None: orig_func(return_xpu(device)),
lambda orig_func, device=None: check_device(device))
# TiledVAE and ControlNet:
CondFunc('torch.batch_norm',