implement DoubleTextEncoder for SDXL

src/refiners/foundationals/clip/text_encoder.py

@@ -255,7 +255,6 @@ class CLIPTextEncoderG(CLIPTextEncoder):
             num_layers=32,
             num_attention_heads=20,
             feedforward_dim=5120,
-            use_quick_gelu=True,
             tokenizer=tokenizer,
             device=device,
             dtype=dtype,

src/refiners/foundationals/latent_diffusion/sdxl_text_encoder.py

@@ -0,0 +1,88 @@
+from typing import cast
+from torch import device as Device, dtype as DType, Tensor, cat
+from refiners.adapters.adapter import Adapter
+from refiners.fluxion.context import Contexts
+import refiners.fluxion.layers as fl
+from refiners.foundationals.clip.text_encoder import CLIPTextEncoderG, CLIPTextEncoderL
+from jaxtyping import Float
+
+from refiners.foundationals.clip.tokenizer import CLIPTokenizer
+
+
+class TextEncoderWithPooling(fl.Chain, Adapter[CLIPTextEncoderG]):
+    def __init__(
+        self,
+        target: CLIPTextEncoderG,
+        projection: fl.Linear | None = None,
+    ) -> None:
+        with self.setup_adapter(target=target):
+            tokenizer = target.find(layer_type=CLIPTokenizer)
+            assert tokenizer is not None, "Tokenizer not found."
+            super().__init__(
+                tokenizer,
+                fl.SetContext(
+                    context="text_encoder_pooling", key="end_of_text_index", callback=self.set_end_of_text_index
+                ),
+                target[1:-2],
+                fl.Parallel(
+                    fl.Identity(),
+                    fl.Chain(
+                        target[-2:],
+                        projection
+                        or fl.Linear(
+                            in_features=1280, out_features=1280, bias=False, device=target.device, dtype=target.dtype
+                        ),
+                        fl.Lambda(func=self.pool),
+                    ),
+                ),
+            )
+
+    def init_context(self) -> Contexts:
+        return {"text_encoder_pooling": {"end_of_text_index": []}}
+
+    def __call__(self, text: str) -> tuple[Float[Tensor, "1 77 1280"], Float[Tensor, "1 1280"]]:
+        return super().__call__(text)
+
+    @property
+    def tokenizer(self) -> CLIPTokenizer:
+        tokenizer = self.find(layer_type=CLIPTokenizer)
+        assert tokenizer is not None, "Tokenizer not found."
+        return tokenizer
+
+    def set_end_of_text_index(self, end_of_text_index: list[int], tokens: Tensor) -> None:
+        position = (tokens == self.tokenizer.end_of_text_token_id).nonzero(as_tuple=True)[1].item()
+        end_of_text_index.append(cast(int, position))
+
+    def pool(self, x: Float[Tensor, "1 77 1280"]) -> Float[Tensor, "1 1280"]:
+        end_of_text_index = self.use_context(context_name="text_encoder_pooling").get("end_of_text_index", [])
+        assert len(end_of_text_index) == 1, "End of text index not found."
+        return x[:, end_of_text_index[0], :]
+
+
+class DoubleTextEncoder(fl.Chain):
+    def __init__(
+        self,
+        text_encoder_l: CLIPTextEncoderL | None = None,
+        text_encoder_g: CLIPTextEncoderG | None = None,
+        projection: fl.Linear | None = None,
+        device: Device | str | None = None,
+        dtype: DType | None = None,
+    ) -> None:
+        text_encoder_l = text_encoder_l or CLIPTextEncoderL(device=device, dtype=dtype)
+        text_encoder_g = text_encoder_g or CLIPTextEncoderG(device=device, dtype=dtype)
+        text_encoder_with_pooling = TextEncoderWithPooling(target=text_encoder_g, projection=projection)
+        super().__init__(
+            fl.Parallel(text_encoder_l[:-2], text_encoder_g),
+            fl.Lambda(func=self.concatenate_embeddings),
+        )
+        text_encoder_with_pooling.inject(parent=self.Parallel)
+
+    def __call__(self, text: str) -> tuple[Float[Tensor, "1 77 2048"], Float[Tensor, "1 1280"]]:
+        return super().__call__(text)
+
+    def concatenate_embeddings(
+        self, text_embedding_l: Tensor, text_embedding_with_pooling: tuple[Tensor, Tensor]
+    ) -> tuple[Tensor, Tensor]:
+        text_embedding_g, pooled_text_embedding = text_embedding_with_pooling
+        text_embedding = cat(tensors=[text_embedding_l, text_embedding_g], dim=-1)
+        return text_embedding, pooled_text_embedding

tests/foundationals/latent_diffusion/test_sdxl_double_encoder.py

@@ -0,0 +1,102 @@
+from typing import Any, Protocol, cast
+from pathlib import Path
+from warnings import warn
+import pytest
+import torch
+from torch import Tensor
+
+from refiners.fluxion.utils import manual_seed
+import refiners.fluxion.layers as fl
+from refiners.foundationals.clip.text_encoder import CLIPTextEncoderG, CLIPTextEncoderL
+from refiners.foundationals.latent_diffusion.sdxl_text_encoder import DoubleTextEncoder
+
+
+class DiffusersSDXL(Protocol):
+    unet: fl.Module
+    text_encoder: fl.Module
+    text_encoder_2: fl.Module
+    tokenizer: fl.Module
+    tokenizer_2: fl.Module
+    vae: fl.Module
+
+    def __call__(self, prompt: str, *args: Any, **kwargs: Any) -> Any:
+        ...
+
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: str | None = None,
+        negative_prompt: str | None = None,
+        negative_prompt_2: str | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        ...
+
+
+@pytest.fixture(scope="module")
+def stabilityai_sdxl_base_path(test_weights_path: Path) -> Path:
+    r = test_weights_path / "stabilityai" / "stable-diffusion-xl-base-1.0"
+    if not r.is_dir():
+        warn(message=f"could not find Stability SDXL base weights at {r}, skipping")
+        pytest.skip(allow_module_level=True)
+    return r
+
+
+@pytest.fixture(scope="module")
+def diffusers_sdxl(stabilityai_sdxl_base_path: Path) -> Any:
+    from diffusers import DiffusionPipeline  # type: ignore
+
+    return DiffusionPipeline.from_pretrained(pretrained_model_name_or_path=stabilityai_sdxl_base_path)  # type: ignore
+
+
+@pytest.fixture(scope="module")
+def double_text_encoder(test_weights_path: Path) -> DoubleTextEncoder:
+    text_encoder_l = CLIPTextEncoderL()
+    text_encoder_g_with_projection = CLIPTextEncoderG()
+    text_encoder_g_with_projection.append(module=fl.Linear(in_features=1280, out_features=1280, bias=False))
+
+    text_encoder_l_path = test_weights_path / "CLIPTextEncoderL.safetensors"
+    text_encdoer_g_path = test_weights_path / "CLIPTextEncoderGWithProjection.safetensors"
+
+    text_encoder_l.load_from_safetensors(tensors_path=text_encoder_l_path)
+    text_encoder_g_with_projection.load_from_safetensors(tensors_path=text_encdoer_g_path)
+
+    linear = text_encoder_g_with_projection.pop(index=-1)
+    assert isinstance(linear, fl.Linear)
+
+    double_text_encoder = DoubleTextEncoder(
+        text_encoder_l=text_encoder_l, text_encoder_g=text_encoder_g_with_projection, projection=linear
+    )
+
+    return double_text_encoder
+
+
+@torch.no_grad()
+def test_double_text_encoder(diffusers_sdxl: DiffusersSDXL, double_text_encoder: DoubleTextEncoder) -> None:
+    manual_seed(seed=0)
+    prompt = "A photo of a pizza."
+
+    prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = (
+        diffusers_sdxl.encode_prompt(prompt=prompt, negative_prompt="")
+    )
+
+    double_embedding, pooled_embedding = double_text_encoder(prompt)
+
+    assert double_embedding.shape == torch.Size([1, 77, 2048])
+    assert pooled_embedding.shape == torch.Size([1, 1280])
+
+    embedding_1, embedding_2 = cast(
+        tuple[Tensor, Tensor], prompt_embeds.split(split_size=[768, 1280], dim=-1)  # type: ignore
+    )
+
+    rembedding_1, rembedding_2 = cast(
+        tuple[Tensor, Tensor], double_embedding.split(split_size=[768, 1280], dim=-1)  # type: ignore
+    )
+
+    assert torch.allclose(input=embedding_1, other=rembedding_1, rtol=1e-3, atol=1e-3)
+    assert torch.allclose(input=embedding_2, other=rembedding_2, rtol=1e-3, atol=1e-3)
+    assert torch.allclose(input=pooled_embedding, other=pooled_prompt_embeds, rtol=1e-3, atol=1e-3)
+
+    negative_double_embedding, negative_pooled_embedding = double_text_encoder("")
+
+    assert torch.allclose(input=negative_double_embedding, other=negative_prompt_embeds, rtol=1e-3, atol=1e-3)
+    assert torch.allclose(input=negative_pooled_embedding, other=negative_pooled_prompt_embeds, rtol=1e-3, atol=1e-3)