diff --git a/backprop/backward.cc b/backprop/backward.cc
index 7f06fd4..da63e3a 100644
--- a/backprop/backward.cc
+++ b/backprop/backward.cc
@@ -24,8 +24,10 @@
 
 // Compiles this file for multiple architectures via "foreach_target.h", to
 // which we pass the filename via macro 'argument'.
+// clang-format off
 #undef HWY_TARGET_INCLUDE
 #define HWY_TARGET_INCLUDE "backprop/backward.cc"  // NOLINT
+// clang-format on
 #include "hwy/foreach_target.h"        // IWYU pragma: keep
 
 #include "hwy/highway.h"
diff --git a/gemma/weights.cc b/gemma/weights.cc
index a4d800e..405f409 100644
--- a/gemma/weights.cc
+++ b/gemma/weights.cc
@@ -15,13 +15,15 @@
 
 #include "gemma/weights.h"
 
-#include <cstdio>
+#include <stdio.h>
+
 #include <cstdlib>
 
 #include "compression/compress.h"
 #include "compression/io.h"  // Path
 #include "gemma/common.h"
 #include "gemma/configs.h"
+#include "util/allocator.h"
 #include "hwy/base.h"  // HWY_ABORT
 #include "hwy/contrib/thread_pool/thread_pool.h"
 #include "hwy/profiler.h"
@@ -72,7 +74,10 @@ struct LoadCompressedWeightsT {
       }
       HWY_ASSERT(scale_pos == TConfig::kNumTensorScales);
     }
-    c_weights->Reshape();
+    {
+      PROFILER_ZONE("Startup.Reshape");
+      c_weights->Reshape();
+    }
     return c_weights_u8;
   }
 };
diff --git a/gemma/weights.h b/gemma/weights.h
index 818ae8e..42fc23b 100644
--- a/gemma/weights.h
+++ b/gemma/weights.h
@@ -25,7 +25,6 @@
 #include "hwy/aligned_allocator.h"
 #include "hwy/base.h"
 #include "hwy/contrib/thread_pool/thread_pool.h"
-#include "hwy/profiler.h"
 
 namespace gcpp {
 
@@ -122,8 +121,6 @@ struct CompressedLayer {
   // after loading weights via ForEachTensor.
   // TODO: update compression/convert_weights to bake this in.
   void Reshape() {
-    PROFILER_ZONE("Startup.Reshape");
-
     constexpr size_t kModelDim = TConfig::kModelDim;
     constexpr size_t kHeads = TConfig::kHeads;
     constexpr size_t kQKVDim = TConfig::kQKVDim;
diff --git a/ops/matvec-inl.h b/ops/matvec-inl.h
index ffc8677..5a56413 100644
--- a/ops/matvec-inl.h
+++ b/ops/matvec-inl.h
@@ -23,7 +23,6 @@
 
 #include "hwy/base.h"
 #include "hwy/contrib/thread_pool/thread_pool.h"
-#include "hwy/profiler.h"
 
 #endif  // THIRD_PARTY_GEMMA_CPP_OPS_MATVEC_INL_H_
 
@@ -39,6 +38,7 @@
 #include "ops/dot-inl.h"
 #include "hwy/contrib/math/math-inl.h"
 #include "hwy/contrib/matvec/matvec-inl.h"
+#include "hwy/profiler.h"  // also uses SIMD
 
 HWY_BEFORE_NAMESPACE();
 namespace gcpp {