Added tpu flash attention.
Browse files
xora/models/transformers/attention.py
CHANGED
@@ -20,6 +20,13 @@ from diffusers.utils.torch_utils import maybe_allow_in_graph
|
|
20 |
from einops import rearrange
|
21 |
from torch import nn
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
# code adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
|
24 |
|
25 |
logger = logging.get_logger(__name__)
|
@@ -162,6 +169,15 @@ class BasicTransformerBlock(nn.Module):
|
|
162 |
self._chunk_size = None
|
163 |
self._chunk_dim = 0
|
164 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
|
166 |
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
|
167 |
# Sets chunk feed-forward
|
@@ -461,6 +477,13 @@ class Attention(nn.Module):
|
|
461 |
processor = AttnProcessor2_0()
|
462 |
self.set_processor(processor)
|
463 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
464 |
def set_processor(self, processor: "AttnProcessor") -> None:
|
465 |
r"""
|
466 |
Set the attention processor to use.
|
|
|
20 |
from einops import rearrange
|
21 |
from torch import nn
|
22 |
|
23 |
+
try:
|
24 |
+
from torch_xla.experimental.custom_kernel import flash_attention
|
25 |
+
except ImportError:
|
26 |
+
# workaround for automatic tests. Currently this function is manually patched
|
27 |
+
# to the torch_xla lib on setup of container
|
28 |
+
pass
|
29 |
+
|
30 |
# code adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
|
31 |
|
32 |
logger = logging.get_logger(__name__)
|
|
|
169 |
self._chunk_size = None
|
170 |
self._chunk_dim = 0
|
171 |
|
172 |
+
def set_use_tpu_flash_attention(self, device):
|
173 |
+
r"""
|
174 |
+
Function sets the flag in this object and propagates down the children. The flag will enforce the usage of TPU
|
175 |
+
attention kernel.
|
176 |
+
"""
|
177 |
+
if device == "xla":
|
178 |
+
self.use_tpu_flash_attention = True
|
179 |
+
self.attn1.set_use_tpu_flash_attention(device)
|
180 |
+
self.attn2.set_use_tpu_flash_attention(device)
|
181 |
|
182 |
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
|
183 |
# Sets chunk feed-forward
|
|
|
477 |
processor = AttnProcessor2_0()
|
478 |
self.set_processor(processor)
|
479 |
|
480 |
+
def set_use_tpu_flash_attention(self, device_type):
|
481 |
+
r"""
|
482 |
+
Function sets the flag in this object. The flag will enforce the usage of TPU attention kernel.
|
483 |
+
"""
|
484 |
+
if device_type == "xla":
|
485 |
+
self.use_tpu_flash_attention = True
|
486 |
+
|
487 |
def set_processor(self, processor: "AttnProcessor") -> None:
|
488 |
r"""
|
489 |
Set the attention processor to use.
|
xora/models/transformers/transformer3d.py
CHANGED
@@ -153,11 +153,11 @@ class Transformer3DModel(ModelMixin, ConfigMixin):
|
|
153 |
"""
|
154 |
logger.info(" ENABLE TPU FLASH ATTENTION -> TRUE")
|
155 |
# if using TPU -> configure components to use TPU flash attention
|
156 |
-
if
|
157 |
self.use_tpu_flash_attention = True
|
158 |
# push config down to the attention modules
|
159 |
for block in self.transformer_blocks:
|
160 |
-
block.set_use_tpu_flash_attention()
|
161 |
|
162 |
def initialize(self, embedding_std: float, mode: Literal["xora", "pixart"]):
|
163 |
def _basic_init(module):
|
|
|
153 |
"""
|
154 |
logger.info(" ENABLE TPU FLASH ATTENTION -> TRUE")
|
155 |
# if using TPU -> configure components to use TPU flash attention
|
156 |
+
if self.device.type == "xla":
|
157 |
self.use_tpu_flash_attention = True
|
158 |
# push config down to the attention modules
|
159 |
for block in self.transformer_blocks:
|
160 |
+
block.set_use_tpu_flash_attention(self.device.type)
|
161 |
|
162 |
def initialize(self, embedding_std: float, mode: Literal["xora", "pixart"]):
|
163 |
def _basic_init(module):
|
xora/utils/dist_util.py
CHANGED
@@ -1,11 +1,5 @@
|
|
1 |
from enum import Enum
|
2 |
|
3 |
-
class AccelerationType(Enum):
|
4 |
-
CPU = "cpu"
|
5 |
-
GPU = "gpu"
|
6 |
-
TPU = "tpu"
|
7 |
-
MPS = "mps"
|
8 |
-
|
9 |
def execute_graph() -> None:
|
10 |
if _acceleration_type == AccelerationType.TPU:
|
11 |
xm.mark_step()
|
|
|
1 |
from enum import Enum
|
2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
def execute_graph() -> None:
|
4 |
if _acceleration_type == AccelerationType.TPU:
|
5 |
xm.mark_step()
|