[executorch][muse] Add run_solo_batching

.github/workflows/cuda.yml

@@ -480,8 +480,11 @@ jobs:
         pip install gguf
         python -m pytest examples/models/gemma4_31b/tests/ --ignore=examples/models/gemma4_31b/tests/test_mlx_pipeline.py -v -o "addopts="
 
-        # Muse Glimmer batched CUDA export on a tiny model
+        # Muse Glimmer batched CUDA export on a tiny model, and the batched
+        # runner end to end on it
         python -m pytest examples/models/muse-glimmer/tests/test_cuda_batching_pipeline.py -v -o "addopts="
+        make muse-glimmer-cuda
+        python -m pytest examples/models/muse-glimmer/tests/test_run_solo_batching.py -v -o "addopts="
 
   unittest-cuda-runtime:
     name: unittest-cuda-runtime

Makefile

@@ -492,6 +492,7 @@ muse-glimmer-cuda:
 	@echo ""
 	@echo "✓ Build complete!"
 	@echo "  Solo runner:   cmake-out/examples/models/muse-glimmer/solo_runner"
+	@echo "  Batched runner: cmake-out/examples/models/muse-glimmer/run_solo_batching"
 	@echo "  DFlash runner: cmake-out/examples/models/muse-glimmer/dflash_runner"
 	@echo "  Worker:        cmake-out/examples/models/muse-glimmer/muse_glimmer_worker"
 

examples/models/muse-glimmer/CMakeLists.txt

@@ -189,3 +189,19 @@ if(TARGET mlxdelegate)
   executorch_target_copy_mlx_metallib(dflash_runner)
   executorch_target_copy_mlx_metallib(muse_glimmer_worker)
 endif()
+
+# Batched text generation on the batching extension (CUDA only): runs what
+# export/export_solo_batching.py writes.
+if(EXECUTORCH_BUILD_CUDA)
+  add_executable(run_solo_batching runtime/runners/run_solo_batching.cpp)
+  target_include_directories(
+    run_solo_batching PUBLIC ${_common_include_directories} ${_json_include}
+  )
+  target_link_libraries(
+    run_solo_batching PUBLIC ${link_libraries} cuda_batching Threads::Threads
+  )
+  # CUDA AOTI blobs resolve symbols from the host executable.
+  if(NOT APPLE AND NOT MSVC)
+    target_link_options(run_solo_batching PRIVATE "LINKER:--export-dynamic")
+  endif()
+endif()

examples/models/muse-glimmer/CMakePresets.json

@@ -43,7 +43,7 @@
             "name": "muse-glimmer-cuda",
             "displayName": "Build Muse Glimmer runner (CUDA)",
             "configurePreset": "muse-glimmer-cuda",
-            "targets": ["solo_runner", "dflash_runner", "muse_glimmer_worker"]
+            "targets": ["solo_runner", "run_solo_batching", "dflash_runner", "muse_glimmer_worker"]
         },
         {
             "name": "muse-glimmer-mlx",