Add BUCK files for llava python and C++ libs (#8297)

Summary:

Add BUCK targets so it can be used in fbcode, bento etc.

Reviewed By: luyich

Differential Revision: D69278781

Author: larryliu0820

examples/models/llava/export_llava.py

@@ -67,6 +67,7 @@ def export(self) -> "LlavaEdgeManager":
                 dynamic_shapes=dynamic_shape,
                 strict=False,
             )
+            # pyre-ignore: Incompatible attribute type [8]: Attribute `pre_autograd_graph_module` declared in class `LLMEdgeManager` has type `Optional[GraphModule]` but is used as type `Module`.
             self.pre_autograd_graph_module = self.export_program.module()
         return self
 

examples/models/llava/image_util.py

@@ -21,6 +21,7 @@
 logging.basicConfig(level=logging.INFO, format=FORMAT)
 
 
+# pyre-ignore: Undefined or invalid type [11]: Annotation `Image` is not defined as a type.
 def prepare_image(image: Image, target_h: int, target_w: int) -> torch.Tensor:
     """Read image into a tensor and resize the image so that it fits in
     a target_h x target_w canvas.

examples/models/llava/model.py

@@ -48,6 +48,7 @@ def __init__(
         self.use_sdpa_with_kv_cache_op = use_sdpa_with_kv_cache_op
         self.model_ = llava_model
         self.image_processor = image_processor
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `config`.
         self.vision_feature_layer = self.model_.config.vision_feature_layer
         self.vision_feature_select_strategy = (
             self.model_.config.vision_feature_select_strategy
@@ -76,6 +77,7 @@ def __init__(
         )
 
     def _translate_state_dict_for_text_model(self) -> Dict[str, Any]:
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `language_model`.
         state_dict = self.model_.language_model.state_dict()
         key_map = {
             # fmt: off
@@ -128,9 +130,11 @@ def get_model(self):
         return self.model_.get_model()
 
     def embed_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `language_model`.
         return self.model_.language_model.model.embed_tokens(tokens)
 
     def encode_images(self, images: torch.Tensor) -> torch.Tensor:
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `dtype`.
         images = images.to(dtype=self.model_.dtype)
         if type(images) is list:
             image_features = []
@@ -144,15 +148,19 @@ def encode_images(self, images: torch.Tensor) -> torch.Tensor:
                 image_feature = self._feature_select(image_forward_out).to(image.dtype)
                 image_features.append(image_feature)
         else:
+            # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `vision_tower`.
             image_forward_outs = self.model_.vision_tower(
+                # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `device`.
                 images.to(device=self.model_.device, dtype=self.model_.dtype),
                 output_hidden_states=True,
             )
             image_features = self._feature_select(image_forward_outs).to(images.dtype)
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `multi_modal_projector`.
         image_features = self.model_.multi_modal_projector(image_features)
         return image_features
 
     def image_preprocess(self, img: torch.Tensor) -> torch.Tensor:
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_vision_objects.CLIPImageProcessor` has no attribute `crop_size`.
         target_h = self.image_processor.crop_size["height"]
         target_w = self.image_processor.crop_size["width"]
         # pad the image with median rgb value, to make a square
@@ -195,10 +203,14 @@ def image_preprocess(self, img: torch.Tensor) -> torch.Tensor:
         # print(resized.shape)
         # cropped = F.center_crop(img, output_size=[w, w])
         # print(cropped.shape)
+        # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_vision_objects.CLIPImageProcessor` has no attribute `rescale_factor`.
         scaled = resized * self.image_processor.rescale_factor
         # print(scaled)
         normed = F.normalize(
-            scaled, self.image_processor.image_mean, self.image_processor.image_std
+            # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_vision_objects.CLIPImageProcessor` has no attribute `image_std`.
+            scaled,
+            self.image_processor.image_mean,
+            self.image_processor.image_std,
         )
         # print(normed)
         return normed.unsqueeze(0)
@@ -249,7 +261,9 @@ def prefill_ref(
     ) -> torch.Tensor:
         """Avoiding the torch.where() call to find <image> placeholder and insert image embedding. Taking 3 inputs instead."""
         embeds = self.prefill_embedding(prompt_before_image, images, prompt_after_image)
+        # pyre-ignore: Undefined attribute [16]: Module `transformers` has no attribute `LlamaForCausalLM`.
         return LlamaForCausalLM.forward(
+            # pyre-ignore: Undefined attribute [16]: `transformers.utils.dummy_pt_objects.LlavaForConditionalGeneration` has no attribute `language_model`.
             self.model_.language_model,
             inputs_embeds=embeds,
             return_dict=False,
@@ -268,12 +282,16 @@ class LlavaModel(EagerModelBase):
     def __init__(self, use_sdpa_with_kv_cache_op=True, max_seq_len=768):
         self.use_sdpa_with_kv_cache_op = use_sdpa_with_kv_cache_op
         self.max_seq_len = max_seq_len
-        self.processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
+        self.processor = AutoProcessor.from_pretrained(
+            "llava-hf/llava-1.5-7b-hf",
+            revision="a272c74b2481d8aff3aa6fc2c4bf891fe57334fb",  # Need this for transformers >= 4.44.2
+        )
         self.tokenizer = self.processor.tokenizer
         self.image_processor = self.processor.image_processor
         self.model = LlavaForConditionalGeneration.from_pretrained(
             "llava-hf/llava-1.5-7b-hf",
             device_map="cpu",
+            revision="a272c74b2481d8aff3aa6fc2c4bf891fe57334fb",  # Need this for transformers >= 4.44.2
         )
         self.image = Image.open(
             requests.get(

examples/models/llava/targets.bzl

@@ -0,0 +1,24 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "get_oss_build_kwargs", "runtime")
+
+def define_common_targets():
+    runtime.cxx_binary(
+        name = "main",
+        srcs = [
+            "main.cpp",
+        ],
+        compiler_flags = ["-Wno-global-constructors"],
+        preprocessor_flags = [
+            "-DET_USE_THREADPOOL",
+        ],
+        deps = [
+            "//executorch/examples/models/llava/runner:runner",
+            "//executorch/extension/evalue_util:print_evalue",
+            "//executorch/extension/threadpool:cpuinfo_utils",
+            "//executorch/extension/threadpool:threadpool",
+        ],
+        external_deps = [
+            "gflags",
+            "torch-core-cpp",
+        ],
+        **get_oss_build_kwargs()
+    )