Include audio preprocessing for raw audio tensor

examples/models/llava/export_llava.py

@@ -226,11 +226,11 @@ def export_all(llava_model: LlavaModel):
         {
             "image_encoder": image_encoder_ep,
             "token_embedding": token_embedding_ep,
-            "text_model": text_model_ep,
+            "text_decoder": text_model_ep,
         },
         partitioner={
             "image_encoder": [XnnpackPartitioner()],
-            "text_model": [
+            "text_decoder": [
                 # First partition the DQLinear nodes, then partition the rest of the nodes,
                 # to avoid multiple DQLinear nodes in the same partition,
                 # to avoid holding multiple unpacked and packed weight buffers in memory,
@@ -254,7 +254,7 @@ def export_all(llava_model: LlavaModel):
             memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
             sym_shape_eval_pass={
                 "image_encoder": ConstraintBasedSymShapeEvalPass(),
-                "text_model": ConstraintBasedSymShapeEvalPass(),
+                "text_decoder": ConstraintBasedSymShapeEvalPass(),
                 "token_embedding": HintBasedSymShapeEvalPass(),
             },
         )

examples/models/llava/runner/llava_text_decoder_runner.h

@@ -89,7 +89,7 @@ class ET_EXPERIMENTAL LlavaTextDecoderRunner
   }
 
   inline static const std::string kTokenEmbeddingMethod = "token_embedding";
-  inline static const std::string kTextModelMethod = "text_model";
+  inline static const std::string kTextModelMethod = "text_decoder";
 };
 
 } // namespace example

examples/models/llava/test/test_llava.py

@@ -96,7 +96,7 @@ def test_llava_export(self):
             "token_embedding", (prompt_before_image,)
         )[0]
         llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_before_img),
         )
 
@@ -107,7 +107,7 @@ def test_llava_export(self):
         # pte prefill image
         pte_embeds_img = llava_module.run_method("image_encoder", (resized,))[0]
         llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (
                 torch.tensor([start_pos], dtype=torch.int64),
                 pte_embeds_img,
@@ -122,7 +122,7 @@ def test_llava_export(self):
             "token_embedding", (prompt_after_image,)
         )[0]
         pte_prefill_after_img = llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_after_img),
         )[0]
 
@@ -139,7 +139,7 @@ def test_llava_export(self):
                 "token_embedding", (torch.tensor([[new_tokens[i]]], dtype=torch.int64),)
             )[0]
             logits = llava_module.run_method(
-                "text_model",
+                "text_decoder",
                 (torch.tensor([start_pos + i], dtype=torch.int64), token_embeds),
             )[0]
             new_tokens.append(torch.argmax(logits).item())

examples/models/llava/test/test_pte.py

@@ -47,7 +47,7 @@ def main():
         "token_embedding", (prompt_before_image,)
     )[0]
     pte_prefill_before_img = llava_module.run_method(
-        "text_model",
+        "text_decoder",
         (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_before_img),
     )[0]
     print(pte_prefill_before_img)
@@ -60,7 +60,7 @@ def main():
     logging.warning("Image encoder finished")
     logging.warning("Image token prefill started")
     pte_prefill_img = llava_module.run_method(
-        "text_model",
+        "text_decoder",
         (
             torch.tensor([start_pos], dtype=torch.int64),
             pte_embeds_img,
@@ -77,7 +77,7 @@ def main():
         "token_embedding", (prompt_after_image,)
     )[0]
     pte_prefill_after_img = llava_module.run_method(
-        "text_model",
+        "text_decoder",
         (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_after_img),
     )[0]
     logging.warning("Text token prefill finished")
@@ -91,7 +91,7 @@ def main():
             "token_embedding", (torch.tensor([[new_tokens[i]]], dtype=torch.int64),)
         )[0]
         logits = llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (torch.tensor([start_pos + i], dtype=torch.int64), token_embeds),
         )[0]
         new_tokens.append(torch.argmax(logits[..., -1, :]).item())

examples/models/voxtral/CMakeLists.txt

@@ -0,0 +1,98 @@
+# 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.
+
+#
+# Simple CMake build system for voxtral runner.
+#
+cmake_minimum_required(VERSION 3.24)
+project(voxtral)
+
+set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
+
+include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
+
+if(CMAKE_TOOLCHAIN_FILE MATCHES ".*(iOS|ios\.toolchain)\.cmake$")
+  set(CMAKE_TOOLCHAIN_IOS ON)
+else()
+  set(CMAKE_TOOLCHAIN_IOS OFF)
+endif()
+
+# Let files say "include <executorch/path/to/header.h>"
+set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+
+# Need this for gflags for some reason
+set(gflags_DIR ${CMAKE_CURRENT_BINARY_DIR}/../../../third-party/gflags)
+find_package(gflags REQUIRED)
+
+# Find `executorch` libraries, same as for gflags
+list(APPEND CMAKE_FIND_ROOT_PATH ${CMAKE_CURRENT_BINARY_DIR}/../../..)
+find_package(executorch CONFIG REQUIRED FIND_ROOT_PATH_BOTH)
+executorch_target_link_options_shared_lib(executorch)
+
+set(link_libraries executorch gflags)
+set(_srcs multimodal.cpp)
+
+list(
+  APPEND
+  link_libraries
+  optimized_native_cpu_ops_lib
+  quantized_ops_lib
+  custom_ops
+  cpublas
+  eigen_blas
+)
+executorch_target_link_options_shared_lib(optimized_native_cpu_ops_lib)
+executorch_target_link_options_shared_lib(quantized_ops_lib)
+executorch_target_link_options_shared_lib(custom_ops)
+
+# XNNPACK
+if(TARGET xnnpack_backend)
+  set(xnnpack_backend_libs xnnpack_backend XNNPACK xnnpack-microkernels-prod)
+  if(TARGET kleidiai)
+    list(APPEND xnnpack_backend_libs kleidiai)
+  endif()
+  list(APPEND link_libraries ${xnnpack_backend_libs})
+  executorch_target_link_options_shared_lib(xnnpack_backend)
+endif()
+
+# Add LLM runner and extension module
+if(NOT TARGET extension_llm_runner)
+  message(
+    FATAL_ERROR
+      "ExecuTorch must be installed with EXECUTORCH_BUILD_EXTENSION_LLM_RUNNER enabled."
+  )
+endif()
+
+# Needed for cpuinfo where it uses android specific log lib
+if(ANDROID)
+  list(APPEND link_libraries log)
+endif()
+
+# Add the required ExecuTorch extensions for multimodal LLM runner
+list(
+  APPEND
+  link_libraries
+  extension_llm_runner
+  extension_module
+  extension_data_loader
+  extension_tensor
+  extension_flat_tensor
+)
+
+# Add tokenizers
+list(APPEND link_libraries tokenizers::tokenizers)
+
+add_executable(voxtral_runner ${_srcs})
+if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
+  target_link_options_gc_sections(voxtral_runner)
+  if(NOT APPLE)
+    target_link_options(voxtral_runner PRIVATE "LINKER:-s")
+  endif()
+endif()
+
+target_include_directories(voxtral_runner PUBLIC ${_common_include_directories})
+target_link_libraries(voxtral_runner PUBLIC ${link_libraries})
+target_compile_options(voxtral_runner PUBLIC ${_common_compile_options})