From 743c720b4feaa2f61c19575230937ee208ea5f2b Mon Sep 17 00:00:00 2001
From: Callum Mitchell <callumm@amd.com>
Date: Tue, 27 Oct 2026 21:37:29 -0600
Subject: [PATCH] server: reset roofline profiler after warmup

Add ggml_cuda_roofline_reset() call at the end of load_model(), after
all initialization completes (model load, MTP/speculative setup, mmproj
load) but before any user requests. This excludes warmup/initialization
overhead from TTFT measurements.

Following the pattern from PR #42 (llama-mtmd-cli) and llama-bench:
- Add roofline header include with #ifdef GGML_HIP_ROOFLINE guard
- Call ggml_cuda_roofline_reset() after params propagation
- Update CMakeLists.txt with compile definitions and include dirs

Co-Authored-By: Claude <noreply@anthropic.com>
---
 tools/server/CMakeLists.txt     |  7 +++++++
 tools/server/server-context.cpp | 13 +++++++++++++
 2 files changed, 20 insertions(+)

diff --git a/tools/server/CMakeLists.txt b/tools/server/CMakeLists.txt
index 70b8d7da8eab..7218d76847c5 100644
--- a/tools/server/CMakeLists.txt
+++ b/tools/server/CMakeLists.txt
@@ -35,6 +35,13 @@ target_include_directories(${TARGET} PRIVATE ../mtmd)
 target_include_directories(${TARGET} PRIVATE ${PROJECT_SOURCE_DIR})
 target_link_libraries(${TARGET} PUBLIC llama-common mtmd ${CMAKE_THREAD_LIBS_INIT})
 
+# Compile in the post-warmup roofline reset only when the HIP profiler is built
+# (ggml-cuda-roofline.cpp, linked via ggml-hip). No-op in every other build.
+if (GGML_HIP_ROOFLINE)
+    target_compile_definitions(${TARGET} PRIVATE GGML_HIP_ROOFLINE)
+    target_include_directories(${TARGET} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../../ggml/src/ggml-cuda)
+endif()
+
 # llama-server-impl: server logic, reusable by app
 
 set(TARGET llama-server-impl)
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 60c563345399..5833eebaf151 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -18,6 +18,13 @@
 #include "mtmd.h"
 #include "mtmd-helper.h"
 
+// Optional per-op roofline profiler (HIP build with GGML_HIP_ROOFLINE). Used to
+// discard warmup so the report covers only measured inference. Compiled out
+// otherwise. See ggml/src/ggml-cuda/ggml-cuda-roofline.cpp.
+#ifdef GGML_HIP_ROOFLINE
+#include "ggml-cuda-roofline.h"
+#endif
+
 #include <algorithm>
 #include <cstddef>
 #include <cinttypes>
@@ -1527,6 +1534,12 @@ struct server_context_impl {
         // propagate new defaults back to caller
         params = params_base;
 
+        // Discard warmup from the per-op roofline profiler so the report covers
+        // only measured inference. Compiled out unless GGML_HIP_ROOFLINE.
+#ifdef GGML_HIP_ROOFLINE
+        ggml_cuda_roofline_reset();
+#endif
+
         if (!is_resume) {
             return init();
         }