[llama-mm] Fix vision encoder model test

.github/workflows/pull.yml

@@ -72,6 +72,10 @@ jobs:
         conda activate "${CONDA_ENV}"
 
         MODEL_NAME=${{ matrix.model }}
+        # Install requirements for llama vision
+        if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" ]]; then
+          bash examples/models/llama3_2_vision/install_requirements.sh
+        fi
         BUILD_TOOL=${{ matrix.build-tool }}
         BACKEND=${{ matrix.backend }}
         DEMO_BACKEND_DELEGATION=${{ matrix.demo_backend_delegation }}

.github/workflows/trunk.yml

@@ -58,7 +58,11 @@ jobs:
         bash .ci/scripts/setup-conda.sh
         # Setup MacOS dependencies as there is no Docker support on MacOS atm
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/setup-macos.sh "${BUILD_TOOL}"
-        # Build and test xecutorch
+        # Install requirements for llama vision
+        if [[ "$MODEL_NAME" == "llama3_2_vision_encoder" ]]; then
+          ${CONDA_RUN} bash examples/models/llama3_2_vision/install_requirements.sh
+        fi
+        # Build and test executorch
         PYTHON_EXECUTABLE=python ${CONDA_RUN} bash .ci/scripts/test_model.sh "${MODEL_NAME}" "${BUILD_TOOL}" "${BACKEND}" "${DEMO_BACKEND_DELEGATION}"
 
   test-custom-ops-macos:

examples/models/llama3_2_vision/__init__.py

@@ -0,0 +1,12 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .vision_encoder import FlamingoVisionEncoderModel, VisionEncoderConfig
+
+__all__ = [
+    "FlamingoVisionEncoderModel",
+    "VisionEncoderConfig",
+]

examples/models/llama3_2_vision/vision_encoder/model.py

@@ -16,17 +16,6 @@
 )
 from torchtune.models.llama3_2_vision._component_builders import llama3_2_vision_encoder
 
-max_seq_len = 8192
-in_channels = 3
-tile_size = 560
-max_num_tiles = 4
-# how many tokens per image generated by the vision encoder
-tokens_per_image = 6404
-# how many images to cache in the kv cache in cross attention
-kv_cache_image_num = 1
-# maximum number of tokens generated by encoder and thus stored in the kv cache in cross attention
-encoder_max_seq_len = tokens_per_image * kv_cache_image_num
-
 
 @dataclass
 class VisionEncoderConfig:
@@ -42,11 +31,26 @@ class VisionEncoderConfig:
     in_channels: int = 3
 
 
+# 8 layers for CI testing purpose
+demo_config: VisionEncoderConfig = VisionEncoderConfig(
+    patch_size=14,
+    num_heads=8,
+    clip_embed_dim=768,
+    clip_num_layers=6,
+    clip_hidden_states=[1, 3, 5],
+    decoder_embed_dim=1024,
+    num_layers_projection=4,
+    tile_size=224,
+    max_num_tiles=4,
+    in_channels=3,
+)
+
+
 class FlamingoVisionEncoderModel(EagerModelBase):
     def __init__(self, config: Optional[VisionEncoderConfig] = None):
         super().__init__()
         if config is None:
-            config = VisionEncoderConfig()
+            config = demo_config
         self.config = config
         self.model = llama3_2_vision_encoder(
             patch_size=config.patch_size,

pytest.ini

@@ -17,7 +17,7 @@ addopts =
     # examples
     examples/models/llama/tests
     examples/models/llama3_2_vision/preprocess
-    # examples/models/llama3_2_vision/vision_encoder/test TODO: enable this
+    examples/models/llama3_2_vision/vision_encoder/test
     # examples/models/llava/test TODO: enable this
     # exir
     exir/_serialize/test