feat(hstu): add native inference and history KV caching

Add model-owned HSTU conversion, TensorRT graphs and C++ ranking/retrieval
with verified history-prefix reuse, bounded native storage and independent
decode sessions. Keep optional original CUDA attention in a verified,
bundle-scoped provider with complete third-party notices.

Extend the shared runtime only for recommendation tasks, cache storage,
scoped plugins and stateful CUDA Graph execution. Document joint runtime
rebuild requirements and the external storage adapter boundary.

Signed-off-by: yifeif <277870278+yifeif-nv@users.noreply.github.com>

Author: yifeif-nv

ASSET_LICENSES.md

@@ -193,4 +193,29 @@ with the rest of the project:
 - `families/openfold3/tests/data/openfold3_features.npz`
 - `families/openfold3/tests/data/openfold3_structure.json`
 
+## HSTU native-provider license copies
+
+The [HSTU third-party notice](families/hstu/third_party/NOTICE.txt) records
+attribution for the optional native attention provider. Its accompanying
+license files preserve the full upstream terms. The kernel sources are
+obtained separately at build time and verified against
+`native_attention_source.json`.
+
+- `FBGEMM.LICENSE.txt` and `HSTU.LICENSE.txt`: verbatim BSD license files from
+  [FBGEMM revision 43791a0ade113a0ad5530c2a4948870dd0f7e417](https://github.com/pytorch/FBGEMM/tree/43791a0ade113a0ad5530c2a4948870dd0f7e417).
+- `CUTLASS.LICENSE.txt`: verbatim license from its pinned
+  [CUTLASS revision 571edeb2d0ac872a8392fc49285b156b07884b4e](https://github.com/jwfromm/cutlass/tree/571edeb2d0ac872a8392fc49285b156b07884b4e).
+  The compiled C++ headers use BSD-3-Clause; the separate Python CuTeDSL
+  exception in this license does not apply to those headers.
+- `CCCL.LICENSE.txt`: verbatim [CCCL 3.3.3 license](https://github.com/NVIDIA/cccl/blob/v3.3.3/LICENSE),
+  including CUB's BSD terms and Thrust/libcu++ notices. These headers come from
+  the installed CUDA development toolkit.
+- `NOTICE.txt`: source attribution and dependency boundaries. The builder
+  combines this text with the full licenses into `attention_native.NOTICE`
+  whenever it embeds the native attention library in a model bundle.
+
+This source distribution contains no third-party HSTU model weights or
+prebuilt native provider libraries. Users supply their own model checkpoints
+and the separately licensed CUDA, TensorRT and cuBLAS runtime dependencies.
+
 <!-- Collaborative review anchor: batch 2. -->

CMakeLists.txt

@@ -86,6 +86,7 @@ set(TRTMC_CUDA_INCLUDE_DIR ${CUDAToolkit_INCLUDE_DIRS})
 set(TRTMC_CUDART_LIBRARY CUDA::cudart)
 
 add_library(trtmc_core SHARED
+  core/runtime/cache/history_cache.cpp
   core/runtime/bundle/bundle_format.cpp
   core/runtime/primitives/cuda_common.cpp
   core/runtime/primitives/device_tensor.cpp
@@ -455,6 +456,12 @@ if(TRTMC_BUILD_EXAMPLES)
 endif()
 
 if(TRTMC_BUILD_TESTS)
+  add_executable(test_history_cache core/runtime/tests/test_history_cache.cpp)
+  target_link_libraries(test_history_cache PRIVATE trtmc_core)
+  target_compile_options(test_history_cache PRIVATE -Wall -Wextra -Wpedantic)
+  add_test(NAME history_cache COMMAND test_history_cache)
+  set_tests_properties(history_cache PROPERTIES LABELS cpu)
+
   add_executable(test_bundle_format_v1 core/runtime/tests/test_bundle_format_v1.cpp)
   target_include_directories(test_bundle_format_v1 PRIVATE ${PROJECT_SOURCE_DIR}/core)
   target_link_libraries(test_bundle_format_v1 PRIVATE trtmc_core)
@@ -552,6 +559,17 @@ if(TRTMC_BUILD_TESTS)
     LIBRARY_OUTPUT_DIRECTORY "${_trtmc_test_runtime_root}"
   )
 
+  add_library(trtmc_test_backend_fake_trt SHARED core/runtime/tests/fake_backend.cpp)
+  target_include_directories(trtmc_test_backend_fake_trt PRIVATE
+    ${PROJECT_SOURCE_DIR}/core/runtime/include
+  )
+  target_link_libraries(trtmc_test_backend_fake_trt PRIVATE CUDA::cudart)
+  target_compile_definitions(trtmc_test_backend_fake_trt PRIVATE TRTMC_FAKE_BACKEND_NAME="trt")
+  set_target_properties(trtmc_test_backend_fake_trt PROPERTIES
+    OUTPUT_NAME trtmc_backend_trt
+    LIBRARY_OUTPUT_DIRECTORY "${_trtmc_test_runtime_root}"
+  )
+
   add_library(trtmc_test_family_fake SHARED core/runtime/tests/fake_family.cpp)
   target_include_directories(trtmc_test_family_fake PRIVATE ${PROJECT_SOURCE_DIR}/core/runtime/include)
   target_link_libraries(trtmc_test_family_fake PRIVATE trtmc_core)
@@ -571,6 +589,7 @@ if(TRTMC_BUILD_TESTS)
   add_dependencies(test_family_loader
     trtmc_test_backend_fake
     trtmc_test_backend_fake_rtx
+    trtmc_test_backend_fake_trt
     trtmc_test_family_fake
   )
   add_custom_command(TARGET test_family_loader POST_BUILD

apps/task_runtime.h

@@ -33,6 +33,7 @@ inline bool uses_existing_task_runtime(std::string_view primary_task) noexcept {
         ISpeechToolSessionProvider::kTask,
         IEmbedding::kTask,
         IEncoding::kTask,
+        IRecommendation::kTask,
         IReranking::kTask,
         ISegmentation::kTask,
         IPointPromptedSegmentation::kTask,

conanfile.py

@@ -5,6 +5,7 @@
 
 import json
 import os
+import shutil
 import stat
 import subprocess
 from pathlib import Path
@@ -44,6 +45,34 @@ def _make_executable(path: Path) -> None:
     path.chmod(path.stat().st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH)
 
 
+def _package_native_commands(build: Path, destinations: tuple[Path, ...]) -> None:
+    """Ship root-level native commands, including family-owned trtmc-* binaries."""
+    commands = [build / "trtmc"]
+    commands.extend(
+        path for path in sorted(build.glob("trtmc-*")) if not path.suffix and not path.is_dir()
+    )
+    for command in commands:
+        try:
+            mode = command.lstat().st_mode
+            if not stat.S_ISREG(mode) or not mode & (stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH):
+                raise ConanException(f"native command is not a regular executable: {command}")
+            with command.open("rb") as source:
+                if source.read(4) != b"\x7fELF":
+                    raise ConanException(f"native command is not an ELF executable: {command}")
+            for directory in destinations:
+                directory.mkdir(parents=True, exist_ok=True)
+                target = directory / command.name
+                if target.exists() or target.is_symlink():
+                    raise ConanException(f"duplicate native command destination: {target}")
+                # Copy this exact build output; recursive basename matching can
+                # otherwise replace it with a nested target of the same name.
+                shutil.copy2(command, target)
+                _make_executable(target)
+                _set_runpath(target, "$ORIGIN")
+        except OSError as error:
+            raise ConanException(f"cannot package native command {command}: {error}") from error
+
+
 class TensorRTModelConnectConan(ConanFile):
     name = "tensorrt-model-connect"
     version = "0.1.0"
@@ -91,8 +120,7 @@ def package(self) -> None:
             ],
             check=True,
         )
-        copy(self, "trtmc", src=str(build), dst=str(module_bin), keep_path=False)
-        copy(self, "trtmc-server", src=str(build), dst=str(module_bin), keep_path=False)
+        _package_native_commands(build, (module_bin,))
         for library in ("libtrtmc_core.so", "libtrtmc_runtime.so"):
             copy(self, library, src=str(build), dst=str(module_bin), keep_path=False)
         copy(
@@ -196,7 +224,7 @@ def package(self) -> None:
         ):
             raise ConanException("native runtime package is incomplete")
 
-        for executable in (native, native_server, benchmark_worker, dataset_benchmark):
+        for executable in (benchmark_worker, dataset_benchmark):
             _make_executable(executable)
             _set_runpath(executable, "$ORIGIN")
         for library in shared_runtime: