Flax support for Stable Diffusion 2 (#1423)

* Flax: start adapting to Stable Diffusion 2

* More changes.

* attention_head_dim can be a tuple.

* Fix typos

* Add simple SD 2 integration test.

Slice values taken from my Ampere GPU.

* Add simple UNet integration tests for Flax.

Note that the expected values are taken from the PyTorch results. This
ensures the Flax and PyTorch versions are not too far off.

* Apply suggestions from code review

Co-authored-by: Patrick von Platen <[email protected]>

* Typos and style

* Tests: verify jax is available.

* Style

* Make flake happy

* Remove typo.

* Simple Flax SD 2 pipeline tests.

* Import order

* Remove unused import.

Co-authored-by: Patrick von Platen <[email protected]>
Co-authored-by: @camenduru

Author: pcuenca

src/diffusers/models/attention_flax.py

@@ -104,17 +104,20 @@ class FlaxBasicTransformerBlock(nn.Module):
             Hidden states dimension inside each head
         dropout (:obj:`float`, *optional*, defaults to 0.0):
             Dropout rate
+        only_cross_attention (`bool`, defaults to `False`):
+            Whether to only apply cross attention.
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
     dim: int
     n_heads: int
     d_head: int
     dropout: float = 0.0
+    only_cross_attention: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
-        # self attention
+        # self attention (or cross_attention if only_cross_attention is True)
         self.attn1 = FlaxAttentionBlock(self.dim, self.n_heads, self.d_head, self.dropout, dtype=self.dtype)
         # cross attention
         self.attn2 = FlaxAttentionBlock(self.dim, self.n_heads, self.d_head, self.dropout, dtype=self.dtype)
@@ -126,7 +129,10 @@ def setup(self):
     def __call__(self, hidden_states, context, deterministic=True):
         # self attention
         residual = hidden_states
-        hidden_states = self.attn1(self.norm1(hidden_states), deterministic=deterministic)
+        if self.only_cross_attention:
+            hidden_states = self.attn1(self.norm1(hidden_states), context, deterministic=deterministic)
+        else:
+            hidden_states = self.attn1(self.norm1(hidden_states), deterministic=deterministic)
         hidden_states = hidden_states + residual
 
         # cross attention
@@ -159,6 +165,8 @@ class FlaxTransformer2DModel(nn.Module):
             Number of transformers block
         dropout (:obj:`float`, *optional*, defaults to 0.0):
             Dropout rate
+        use_linear_projection (`bool`, defaults to `False`): tbd
+        only_cross_attention (`bool`, defaults to `False`): tbd
         dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
             Parameters `dtype`
     """
@@ -167,49 +175,70 @@ class FlaxTransformer2DModel(nn.Module):
     d_head: int
     depth: int = 1
     dropout: float = 0.0
+    use_linear_projection: bool = False
+    only_cross_attention: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
         self.norm = nn.GroupNorm(num_groups=32, epsilon=1e-5)
 
         inner_dim = self.n_heads * self.d_head
-        self.proj_in = nn.Conv(
-            inner_dim,
-            kernel_size=(1, 1),
-            strides=(1, 1),
-            padding="VALID",
-            dtype=self.dtype,
-        )
+        if self.use_linear_projection:
+            self.proj_in = nn.Dense(inner_dim, dtype=self.dtype)
+        else:
+            self.proj_in = nn.Conv(
+                inner_dim,
+                kernel_size=(1, 1),
+                strides=(1, 1),
+                padding="VALID",
+                dtype=self.dtype,
+            )
 
         self.transformer_blocks = [
-            FlaxBasicTransformerBlock(inner_dim, self.n_heads, self.d_head, dropout=self.dropout, dtype=self.dtype)
+            FlaxBasicTransformerBlock(
+                inner_dim,
+                self.n_heads,
+                self.d_head,
+                dropout=self.dropout,
+                only_cross_attention=self.only_cross_attention,
+                dtype=self.dtype,
+            )
             for _ in range(self.depth)
         ]
 
-        self.proj_out = nn.Conv(
-            inner_dim,
-            kernel_size=(1, 1),
-            strides=(1, 1),
-            padding="VALID",
-            dtype=self.dtype,
-        )
+        if self.use_linear_projection:
+            self.proj_out = nn.Dense(inner_dim, dtype=self.dtype)
+        else:
+            self.proj_out = nn.Conv(
+                inner_dim,
+                kernel_size=(1, 1),
+                strides=(1, 1),
+                padding="VALID",
+                dtype=self.dtype,
+            )
 
     def __call__(self, hidden_states, context, deterministic=True):
         batch, height, width, channels = hidden_states.shape
         residual = hidden_states
         hidden_states = self.norm(hidden_states)
-        hidden_states = self.proj_in(hidden_states)
-
-        hidden_states = hidden_states.reshape(batch, height * width, channels)
+        if self.use_linear_projection:
+            hidden_states = hidden_states.reshape(batch, height * width, channels)
+            hidden_states = self.proj_in(hidden_states)
+        else:
+            hidden_states = self.proj_in(hidden_states)
+            hidden_states = hidden_states.reshape(batch, height * width, channels)
 
         for transformer_block in self.transformer_blocks:
             hidden_states = transformer_block(hidden_states, context, deterministic=deterministic)
 
-        hidden_states = hidden_states.reshape(batch, height, width, channels)
+        if self.use_linear_projection:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = hidden_states.reshape(batch, height, width, channels)
+        else:
+            hidden_states = hidden_states.reshape(batch, height, width, channels)
+            hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = self.proj_out(hidden_states)
         hidden_states = hidden_states + residual
-
         return hidden_states
 
 

src/diffusers/models/unet_2d_blocks_flax.py

@@ -46,6 +46,8 @@ class FlaxCrossAttnDownBlock2D(nn.Module):
     num_layers: int = 1
     attn_num_head_channels: int = 1
     add_downsample: bool = True
+    use_linear_projection: bool = False
+    only_cross_attention: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
@@ -68,6 +70,8 @@ def setup(self):
                 n_heads=self.attn_num_head_channels,
                 d_head=self.out_channels // self.attn_num_head_channels,
                 depth=1,
+                use_linear_projection=self.use_linear_projection,
+                only_cross_attention=self.only_cross_attention,
                 dtype=self.dtype,
             )
             attentions.append(attn_block)
@@ -178,6 +182,8 @@ class FlaxCrossAttnUpBlock2D(nn.Module):
     num_layers: int = 1
     attn_num_head_channels: int = 1
     add_upsample: bool = True
+    use_linear_projection: bool = False
+    only_cross_attention: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
@@ -201,6 +207,8 @@ def setup(self):
                 n_heads=self.attn_num_head_channels,
                 d_head=self.out_channels // self.attn_num_head_channels,
                 depth=1,
+                use_linear_projection=self.use_linear_projection,
+                only_cross_attention=self.only_cross_attention,
                 dtype=self.dtype,
             )
             attentions.append(attn_block)
@@ -310,6 +318,7 @@ class FlaxUNetMidBlock2DCrossAttn(nn.Module):
     dropout: float = 0.0
     num_layers: int = 1
     attn_num_head_channels: int = 1
+    use_linear_projection: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
@@ -331,6 +340,7 @@ def setup(self):
                 n_heads=self.attn_num_head_channels,
                 d_head=self.in_channels // self.attn_num_head_channels,
                 depth=1,
+                use_linear_projection=self.use_linear_projection,
                 dtype=self.dtype,
             )
             attentions.append(attn_block)

src/diffusers/models/unet_2d_condition_flax.py

@@ -79,7 +79,7 @@ class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
             The tuple of output channels for each block.
         layers_per_block (`int`, *optional*, defaults to 2):
             The number of layers per block.
-        attention_head_dim (`int`, *optional*, defaults to 8):
+        attention_head_dim (`int` or `Tuple[int]`, *optional*, defaults to 8):
             The dimension of the attention heads.
         cross_attention_dim (`int`, *optional*, defaults to 768):
             The dimension of the cross attention features.
@@ -97,11 +97,13 @@ class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
         "DownBlock2D",
     )
     up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")
+    only_cross_attention: Union[bool, Tuple[bool]] = False
     block_out_channels: Tuple[int] = (320, 640, 1280, 1280)
     layers_per_block: int = 2
-    attention_head_dim: int = 8
+    attention_head_dim: Union[int, Tuple[int]] = 8
     cross_attention_dim: int = 1280
     dropout: float = 0.0
+    use_linear_projection: bool = False
     dtype: jnp.dtype = jnp.float32
     freq_shift: int = 0
 
@@ -134,6 +136,14 @@ def setup(self):
         self.time_proj = FlaxTimesteps(block_out_channels[0], freq_shift=self.config.freq_shift)
         self.time_embedding = FlaxTimestepEmbedding(time_embed_dim, dtype=self.dtype)
 
+        only_cross_attention = self.only_cross_attention
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = (only_cross_attention,) * len(self.down_block_types)
+
+        attention_head_dim = self.attention_head_dim
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(self.down_block_types)
+
         # down
         down_blocks = []
         output_channel = block_out_channels[0]
@@ -148,8 +158,10 @@ def setup(self):
                     out_channels=output_channel,
                     dropout=self.dropout,
                     num_layers=self.layers_per_block,
-                    attn_num_head_channels=self.attention_head_dim,
+                    attn_num_head_channels=attention_head_dim[i],
                     add_downsample=not is_final_block,
+                    use_linear_projection=self.use_linear_projection,
+                    only_cross_attention=only_cross_attention[i],
                     dtype=self.dtype,
                 )
             else:
@@ -169,13 +181,16 @@ def setup(self):
         self.mid_block = FlaxUNetMidBlock2DCrossAttn(
             in_channels=block_out_channels[-1],
             dropout=self.dropout,
-            attn_num_head_channels=self.attention_head_dim,
+            attn_num_head_channels=attention_head_dim[-1],
+            use_linear_projection=self.use_linear_projection,
             dtype=self.dtype,
         )
 
         # up
         up_blocks = []
         reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_attention_head_dim = list(reversed(attention_head_dim))
+        only_cross_attention = list(reversed(only_cross_attention))
         output_channel = reversed_block_out_channels[0]
         for i, up_block_type in enumerate(self.up_block_types):
             prev_output_channel = output_channel
@@ -190,9 +205,11 @@ def setup(self):
                     out_channels=output_channel,
                     prev_output_channel=prev_output_channel,
                     num_layers=self.layers_per_block + 1,
-                    attn_num_head_channels=self.attention_head_dim,
+                    attn_num_head_channels=reversed_attention_head_dim[i],
                     add_upsample=not is_final_block,
                     dropout=self.dropout,
+                    use_linear_projection=self.use_linear_projection,
+                    only_cross_attention=only_cross_attention[i],
                     dtype=self.dtype,
                 )
             else:

tests/models/test_models_unet_2d.py

@@ -639,3 +639,29 @@ def test_compvis_sd_inpaint_fp16(self, seed, timestep, expected_slice):
         expected_output_slice = torch.tensor(expected_slice)
 
         assert torch_all_close(output_slice, expected_output_slice, atol=5e-3)
+
+    @parameterized.expand(
+        [
+            # fmt: off
+            [83, 4, [0.1514, 0.0807, 0.1624, 0.1016, -0.1896, 0.0263, 0.0677, 0.2310]],
+            [17, 0.55, [0.1164, -0.0216, 0.0170, 0.1589, -0.3120, 0.1005, -0.0581, -0.1458]],
+            [8, 0.89, [-0.1758, -0.0169, 0.1004, -0.1411, 0.1312, 0.1103, -0.1996, 0.2139]],
+            [3, 1000, [0.1214, 0.0352, -0.0731, -0.1562, -0.0994, -0.0906, -0.2340, -0.0539]],
+            # fmt: on
+        ]
+    )
+    @require_torch_gpu
+    def test_stabilityai_sd_v2_fp16(self, seed, timestep, expected_slice):
+        model = self.get_unet_model(model_id="stabilityai/stable-diffusion-2", fp16=True)
+        latents = self.get_latents(seed, shape=(4, 4, 96, 96), fp16=True)
+        encoder_hidden_states = self.get_encoder_hidden_states(seed, shape=(4, 77, 1024), fp16=True)
+
+        with torch.no_grad():
+            sample = model(latents, timestep=timestep, encoder_hidden_states=encoder_hidden_states).sample
+
+        assert sample.shape == latents.shape
+
+        output_slice = sample[-1, -2:, -2:, :2].flatten().float().cpu()
+        expected_output_slice = torch.tensor(expected_slice)
+
+        assert torch_all_close(output_slice, expected_output_slice, atol=5e-3)