diff --git a/mlir/CMakeLists.txt b/mlir/CMakeLists.txt
index a8e493e2397e..13bcf0c05310 100644
--- a/mlir/CMakeLists.txt
+++ b/mlir/CMakeLists.txt
@@ -53,6 +53,11 @@ add_subdirectory(lib)
 # C API needs all dialects for registration, but should be built before tests.
 add_subdirectory(lib/CAPI)
 
+# Process utils before test so that targets defined there (e.g.
+# ci-performance-scripts, which copies performance scripts onto the lit PATH)
+# exist when the test suite registers its build dependencies.
+add_subdirectory(utils)
+
 if (MLIR_INCLUDE_TESTS)
   add_definitions(-DMLIR_INCLUDE_TESTS)
   add_custom_target(RocMLIRUnitTests)
@@ -64,4 +69,3 @@ if (MLIR_INCLUDE_TESTS)
   add_subdirectory(test)
 endif()
 add_subdirectory(tools)
-add_subdirectory(utils)
diff --git a/mlir/test/CMakeLists.txt b/mlir/test/CMakeLists.txt
index a11fbf077298..7a47173f84d6 100644
--- a/mlir/test/CMakeLists.txt
+++ b/mlir/test/CMakeLists.txt
@@ -114,6 +114,13 @@ if (TARGET amd_arch_db)
   list(APPEND ROCMLIR_TEST_DEPENDS amd_arch_db)
 endif()
 
+# Performance scripts (e.g. tuningRunner.py) are copied into ROCMLIR_BIN_DIR,
+# which is on the lit PATH, by the ci-performance-scripts target. Tests that
+# invoke these scripts directly need them copied before the test suite runs.
+if (TARGET ci-performance-scripts)
+  list(APPEND ROCMLIR_TEST_DEPENDS ci-performance-scripts)
+endif()
+
 if(MLIR_ENABLE_ROCM_RUNNER)
   list(APPEND ROCMLIR_TEST_DEPENDS
     mlir_runner_utils
diff --git a/mlir/test/perf-scripts/runtime/tuningRunner-gemm.py b/mlir/test/perf-scripts/runtime/tuningRunner-gemm.py
new file mode 100644
index 000000000000..f92f25a96a44
--- /dev/null
+++ b/mlir/test/perf-scripts/runtime/tuningRunner-gemm.py
@@ -0,0 +1,17 @@
+# Using a tiny GEMM with --debug-quick-tune-data. This emits a `.debug`
+# TSV of the per-config table entries (PerfConfig + TFlops) but, unlike the
+# full ``--debug`` flag, omits the heavy per-iteration ``MeasurementsMs``
+# arrays. Verify the debug file is produced, has the expected header and
+# per-config rows, and that the measurements column is absent.
+#
+# tuningRunner.py drives real GPU tuning, so it needs the ROCm runner / GPU
+# runtime.
+# REQUIRES: rocm-runner
+# RUN: rm -f %t2.tsv %t2.tsv.state %t2.tsv.debug
+# RUN: tuningRunner.py --op gemm --tuning-space=quick --debug-quick-tune-data \
+# RUN:     --config='-g 1 -m 64 -n 64 -k 64 -t f32 -out_datatype f32 -transA 0 -transB 0' \
+# RUN:     -q -o %t2.tsv
+# RUN: FileCheck %s --check-prefix=DEBUG --implicit-check-not=MeasurementsMs < %t2.tsv.debug
+#
+# DEBUG: PerfConfig{{.*}}TFlops
+# DEBUG: {{v[0-9]+:}}
diff --git a/mlir/utils/performance/tests/test_tuningRunner.py b/mlir/utils/performance/tests/test_tuningRunner.py
index f74284f8d52c..de94136554a2 100644
--- a/mlir/utils/performance/tests/test_tuningRunner.py
+++ b/mlir/utils/performance/tests/test_tuningRunner.py
@@ -241,6 +241,7 @@ def _options(self, output_path, arch="gfx900", num_cu=64, num_chiplets=1, tuning
             num_cu=num_cu,
             num_chiplets=num_chiplets,
             debug=False,
+            debug_quick_tune_data=False,
             quiet=False,
             verbose=False,
             tuning_space_kind=tuning_space,
@@ -452,6 +453,7 @@ def test_cache_loaded_with_canonical_key(self):
                 num_cu=64,
                 num_chiplets=1,
                 debug=False,
+                debug_quick_tune_data=False,
                 quiet=False,
                 verbose=False,
                 tuning_space_kind="full",
diff --git a/mlir/utils/performance/tuningRunner.py b/mlir/utils/performance/tuningRunner.py
index dc8c11e547a5..e792aa7fda45 100755
--- a/mlir/utils/performance/tuningRunner.py
+++ b/mlir/utils/performance/tuningRunner.py
@@ -169,6 +169,7 @@ def get_gpu_logger(gpu_id: int) -> logging.LoggerAdapter:
 class Options:
     """Configuration options for the tuning process."""
     debug: bool
+    debug_quick_tune_data: bool
     tuning_space_kind: str
     quiet: bool
     verbose: bool
@@ -1640,8 +1641,9 @@ def tune_configs(ctx: TuningContext, status_only: bool) -> bool:
 
     has_errors = False
 
-    debug_enabled = ctx.options.debug and ctx.options.output != '-'
-    if ctx.options.debug and not debug_enabled:
+    debug_requested = ctx.options.debug or ctx.options.debug_quick_tune_data
+    debug_enabled = debug_requested and ctx.options.output != '-'
+    if debug_requested and not debug_enabled:
         logger.warning("Debug output disabled when writing to stdout")
 
     with (OutputFileWriter(ctx.options.output, ctx.options) as results_writer,
@@ -1939,7 +1941,13 @@ def parse_arguments(gpu_topology: GpuTopology,
                         "--debug",
                         action='store_true',
                         default=False,
-                        help="Enable debug output including detailed measurements")
+                        help="Enable debug output including detailed per-iteration measurements")
+
+    parser.add_argument("--debug-quick-tune-data",
+                        action='store_true',
+                        default=False,
+                        help="Enable debug output for quick tuning data generation without the "
+                        "detailed per-iteration measurement arrays")
 
     parser.add_argument("--tuning-space",
                         default="full",
@@ -2125,6 +2133,7 @@ def main(args=None):
                       num_cu=num_cu,
                       num_chiplets=num_chiplets,
                       debug=parsed_args.debug,
+                      debug_quick_tune_data=parsed_args.debug_quick_tune_data,
                       quiet=parsed_args.quiet,
                       verbose=parsed_args.verbose,
                       tuning_space_kind=parsed_args.tuning_space,