Add Ascend NPU support for SDXL fine-tuning and fix the model saving bug when using DeepSpeed. (#7816)

* Add Ascend NPU support for SDXL fine-tuning and fix the model saving bug when using DeepSpeed.

* fix check code quality

* Decouple the NPU flash attention and make it an independent module.

* add doc and unit tests for npu flash attention.

---------

Co-authored-by: mhh001 <[email protected]>
Co-authored-by: Sayak Paul <[email protected]>

Author: 3 people

docs/source/en/api/attnprocessor.md

@@ -55,3 +55,6 @@ An attention processor is a class for applying different types of attention mech
 
 ## XFormersAttnProcessor
 [[autodoc]] models.attention_processor.XFormersAttnProcessor
+
+## AttnProcessorNPU
+[[autodoc]] models.attention_processor.AttnProcessorNPU

examples/controlnet/train_controlnet_sdxl.py

@@ -32,7 +32,7 @@
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
-from accelerate.utils import ProjectConfiguration, set_seed
+from accelerate.utils import DistributedType, ProjectConfiguration, set_seed
 from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
@@ -53,7 +53,7 @@
 from diffusers.optimization import get_scheduler
 from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
-from diffusers.utils.import_utils import is_xformers_available
+from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
@@ -64,6 +64,8 @@
 check_min_version("0.28.0.dev0")
 
 logger = get_logger(__name__)
+if is_torch_npu_available():
+    torch.npu.config.allow_internal_format = False
 
 
 def log_validation(vae, unet, controlnet, args, accelerator, weight_dtype, step, is_final_validation=False):
@@ -471,6 +473,9 @@ def parse_args(input_args=None):
     parser.add_argument(
         "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
     )
+    parser.add_argument(
+        "--enable_npu_flash_attention", action="store_true", help="Whether or not to use npu flash attention."
+    )
     parser.add_argument(
         "--set_grads_to_none",
         action="store_true",
@@ -936,6 +941,13 @@ def load_model_hook(models, input_dir):
     text_encoder_two.requires_grad_(False)
     controlnet.train()
 
+    if args.enable_npu_flash_attention:
+        if is_torch_npu_available():
+            logger.info("npu flash attention enabled.")
+            unet.enable_npu_flash_attention()
+        else:
+            raise ValueError("npu flash attention requires torch_npu extensions and is supported only on npu devices.")
+
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
             import xformers
@@ -1235,7 +1247,8 @@ def compute_embeddings(batch, proportion_empty_prompts, text_encoders, tokenizer
                 progress_bar.update(1)
                 global_step += 1
 
-                if accelerator.is_main_process:
+                # DeepSpeed requires saving weights on every device; saving weights only on the main process would cause issues.
+                if accelerator.distributed_type == DistributedType.DEEPSPEED or accelerator.is_main_process:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None:

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -32,7 +32,7 @@
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
-from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
+from accelerate.utils import DistributedDataParallelKwargs, DistributedType, ProjectConfiguration, set_seed
 from datasets import load_dataset
 from huggingface_hub import create_repo, upload_folder
 from packaging import version
@@ -60,14 +60,16 @@
     is_wandb_available,
 )
 from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
-from diffusers.utils.import_utils import is_xformers_available
+from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
 check_min_version("0.28.0.dev0")
 
 logger = get_logger(__name__)
+if is_torch_npu_available():
+    torch.npu.config.allow_internal_format = False
 
 
 def save_model_card(
@@ -419,6 +421,9 @@ def parse_args(input_args=None):
     parser.add_argument(
         "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
     )
+    parser.add_argument(
+        "--enable_npu_flash_attention", action="store_true", help="Whether or not to use npu flash attention."
+    )
     parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
     parser.add_argument(
         "--rank",
@@ -623,6 +628,13 @@ def main(args):
     text_encoder_one.to(accelerator.device, dtype=weight_dtype)
     text_encoder_two.to(accelerator.device, dtype=weight_dtype)
 
+    if args.enable_npu_flash_attention:
+        if is_torch_npu_available():
+            logger.info("npu flash attention enabled.")
+            unet.enable_npu_flash_attention()
+        else:
+            raise ValueError("npu flash attention requires torch_npu extensions and is supported only on npu devices.")
+
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
             import xformers
@@ -1149,7 +1161,8 @@ def compute_time_ids(original_size, crops_coords_top_left):
                 accelerator.log({"train_loss": train_loss}, step=global_step)
                 train_loss = 0.0
 
-                if accelerator.is_main_process:
+                # DeepSpeed requires saving weights on every device; saving weights only on the main process would cause issues.
+                if accelerator.distributed_type == DistributedType.DEEPSPEED or accelerator.is_main_process:
                     if global_step % args.checkpointing_steps == 0:
                         # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                         if args.checkpoints_total_limit is not None:

src/diffusers/models/activations.py

@@ -18,8 +18,12 @@
 from torch import nn
 
 from ..utils import deprecate
+from ..utils.import_utils import is_torch_npu_available
 
 
+if is_torch_npu_available():
+    import torch_npu
+
 ACTIVATION_FUNCTIONS = {
     "swish": nn.SiLU(),
     "silu": nn.SiLU(),
@@ -98,9 +102,13 @@ def forward(self, hidden_states, *args, **kwargs):
         if len(args) > 0 or kwargs.get("scale", None) is not None:
             deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
             deprecate("scale", "1.0.0", deprecation_message)
-
-        hidden_states, gate = self.proj(hidden_states).chunk(2, dim=-1)
-        return hidden_states * self.gelu(gate)
+        hidden_states = self.proj(hidden_states)
+        if is_torch_npu_available():
+            # using torch_npu.npu_geglu can run faster and save memory on NPU.
+            return torch_npu.npu_geglu(hidden_states, dim=-1, approximate=1)[0]
+        else:
+            hidden_states, gate = hidden_states.chunk(2, dim=-1)
+            return hidden_states * self.gelu(gate)
 
 
 class ApproximateGELU(nn.Module):

src/diffusers/models/attention_processor.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import inspect
+import math
 from importlib import import_module
 from typing import Callable, List, Optional, Union
 
@@ -21,13 +22,15 @@
 
 from ..image_processor import IPAdapterMaskProcessor
 from ..utils import deprecate, logging
-from ..utils.import_utils import is_xformers_available
+from ..utils.import_utils import is_torch_npu_available, is_xformers_available
 from ..utils.torch_utils import maybe_allow_in_graph
 from .lora import LoRALinearLayer
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+if is_torch_npu_available():
+    import torch_npu
 
 if is_xformers_available():
     import xformers
@@ -209,6 +212,23 @@ def __init__(
             )
         self.set_processor(processor)
 
+    def set_use_npu_flash_attention(self, use_npu_flash_attention: bool) -> None:
+        r"""
+        Set whether to use npu flash attention from `torch_npu` or not.
+
+        """
+        if use_npu_flash_attention:
+            processor = AttnProcessorNPU()
+        else:
+            # set attention processor
+            # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+            # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+            # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+            processor = (
+                AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
+            )
+        self.set_processor(processor)
+
     def set_use_memory_efficient_attention_xformers(
         self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
     ) -> None:
@@ -1207,6 +1227,116 @@ def __call__(
         return hidden_states
 
 
+class AttnProcessorNPU:
+
+    r"""
+    Processor for implementing flash attention using torch_npu. Torch_npu supports only fp16 and bf16 data types. If
+    fp32 is used, F.scaled_dot_product_attention will be used for computation, but the acceleration effect on NPU is
+    not significant.
+
+    """
+
+    def __init__(self):
+        if not is_torch_npu_available():
+            raise ImportError("AttnProcessorNPU requires torch_npu extensions and is supported only on npu devices.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        temb: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+
+        residual = hidden_states
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        if query.dtype in (torch.float16, torch.bfloat16):
+            hidden_states = torch_npu.npu_fusion_attention(
+                query,
+                key,
+                value,
+                attn.heads,
+                input_layout="BNSD",
+                pse=None,
+                atten_mask=attention_mask,
+                scale=1.0 / math.sqrt(query.shape[-1]),
+                pre_tockens=65536,
+                next_tockens=65536,
+                keep_prob=1.0,
+                sync=False,
+                inner_precise=0,
+            )[0]
+        else:
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
 class AttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).

src/diffusers/models/modeling_utils.py

@@ -272,6 +272,36 @@ def disable_gradient_checkpointing(self) -> None:
         if self._supports_gradient_checkpointing:
             self.apply(partial(self._set_gradient_checkpointing, value=False))
 
+    def set_use_npu_flash_attention(self, valid: bool) -> None:
+        r"""
+        Set the switch for the npu flash attention.
+        """
+
+        def fn_recursive_set_npu_flash_attention(module: torch.nn.Module):
+            if hasattr(module, "set_use_npu_flash_attention"):
+                module.set_use_npu_flash_attention(valid)
+
+            for child in module.children():
+                fn_recursive_set_npu_flash_attention(child)
+
+        for module in self.children():
+            if isinstance(module, torch.nn.Module):
+                fn_recursive_set_npu_flash_attention(module)
+
+    def enable_npu_flash_attention(self) -> None:
+        r"""
+        Enable npu flash attention from torch_npu
+
+        """
+        self.set_use_npu_flash_attention(True)
+
+    def disable_npu_flash_attention(self) -> None:
+        r"""
+        disable npu flash attention from torch_npu
+
+        """
+        self.set_use_npu_flash_attention(False)
+
     def set_use_memory_efficient_attention_xformers(
         self, valid: bool, attention_op: Optional[Callable] = None
     ) -> None: