diff --git a/.clang-tidy b/.clang-tidy
index abcd9d7..497c2e3 100644
--- a/.clang-tidy
+++ b/.clang-tidy
@@ -1,4 +1,5 @@
 FormatStyle: file
+WarningsAsErrors: "*"
 Checks: "-*,\
           abseil-*,\
           -abseil-string-find-startswith,\
@@ -204,3 +205,6 @@ Checks: "-*,\
           -readability-uppercase-literal-suffix,\
           -readability-use-anyofallof
           "
+CheckOptions:
+    - { key: readability-identifier-naming.ConstexprVariableCase,    value: CamelCase }
+    - { key: readability-identifier-naming.ConstexprVariablePrefix,  value: k         }
diff --git a/BUILD.bazel b/BUILD.bazel
index 3019030..885aff5 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -69,6 +69,7 @@ cc_library(
     deps = [
         ":args",
         ":transformer_ops",
+        # "//base",
         "//compression:compress",
         "@hwy//:hwy",
         "@hwy//:matvec",
@@ -88,6 +89,7 @@ cc_binary(
         ":app",
         ":args",
         ":gemma_lib",
+        # "//base",
         "//compression:compress",
         "@hwy//:hwy",
         "@hwy//:nanobenchmark",
diff --git a/compression/blob_store.cc b/compression/blob_store.cc
index e088fc6..050dfbd 100644
--- a/compression/blob_store.cc
+++ b/compression/blob_store.cc
@@ -341,7 +341,7 @@ BlobError BlobReader::Open(const char* filename) {
 #endif
   if (fd_ < 0) return __LINE__;
 
-#if HWY_OS_LINUX
+#if HWY_OS_LINUX && (!defined(__ANDROID_API__) || __ANDROID_API__ >= 21)
   // Doubles the readahead window, which seems slightly faster when cached.
   (void)posix_fadvise(fd_, 0, 0, POSIX_FADV_SEQUENTIAL);
 #endif
diff --git a/gemma.cc b/gemma.cc
index c9066eb..4fe2782 100644
--- a/gemma.cc
+++ b/gemma.cc
@@ -19,18 +19,16 @@
 // which we pass the filename via macro 'argument'.
 #undef HWY_TARGET_INCLUDE
 #define HWY_TARGET_INCLUDE "gemma.cc"  // NOLINT
-#include "hwy/foreach_target.h"        // IWYU pragma: keep
+#include "hwy/foreach_target.h"  // IWYU pragma: keep
 // Must come after foreach_target.h to avoid redefinition errors.
 // copybara:import_next_line:gemma_cpp
 #include "compression/compress-inl.h"
 // copybara:import_next_line:gemma_cpp
 #include "ops.h"
-// copybara:import_next_line:gemma_cpp
 #include "hwy/contrib/matvec/matvec-inl.h"
 #include "hwy/highway.h"
 #include "hwy/profiler.h"
 #include "hwy/timer.h"
-#include "util/args.h"  // Path
 
 // Non-SIMD includes and types. Note that HWY_ONCE is only true on the last
 // compile pass, whereas we want this defined in the first.
@@ -766,10 +764,10 @@ GemmaImpl<Config>::GemmaImpl(
     std::unique_ptr<sentencepiece::SentencePieceProcessor>& tokenizer,
     hwy::AlignedFreeUniquePtr<uint8_t[]>& compressed_weights,
     hwy::ThreadPool& pool)
-    : compressed_weights(std::move(compressed_weights)),
+    : tokenizer(std::move(tokenizer)),
+      compressed_weights(std::move(compressed_weights)),
       prefill(hwy::MakeUniqueAligned<Activations<Config, kPrefillBatchSize>>()),
-      state(hwy::MakeUniqueAligned<Activations<Config, 1>>()),
-      tokenizer(std::move(tokenizer)) {}
+      state(hwy::MakeUniqueAligned<Activations<Config, 1>>()) {}
 
 template <>
 void GemmaImpl<ConfigGemma2B>::Generate(
diff --git a/ops.h b/ops.h
index 3725776..481e1d7 100644
--- a/ops.h
+++ b/ops.h
@@ -340,11 +340,21 @@ static HWY_NOINLINE HWY_MAYBE_UNUSED float Dot(const float* HWY_RESTRICT a,
 // = Dot(a, a, size), but that is not allowed due to HWY_RESTRICT.
 static HWY_NOINLINE HWY_MAYBE_UNUSED float SquaredL2(
     const float* HWY_RESTRICT a, size_t size) {
-  float total = 0.f;
-  for (size_t i = 0; i < size; ++i) {
-    total += a[i] * a[i];
+  const hn::ScalableTag<float> d;
+  const size_t N = hn::Lanes(d);
+  HWY_DASSERT(size >= 2 * N);
+  HWY_DASSERT(size % (2 * N) == 0);
+
+  auto sum0 = hn::Zero(d);
+  auto sum1 = hn::Zero(d);
+  for (size_t i = 0; i <= size - 2 * N; i += 2 * N) {
+    const auto a0 = LoadU(d, a + i);
+    sum0 = MulAdd(a0, a0, sum0);
+    const auto a1 = LoadU(d, a + i + N);
+    sum1 = MulAdd(a1, a1, sum1);
   }
-  return total;
+
+  return ReduceSum(d, Add(sum0, sum1));
 }
 
 static HWY_NOINLINE HWY_MAYBE_UNUSED void RMSNorm(
@@ -362,12 +372,30 @@ static HWY_NOINLINE HWY_MAYBE_UNUSED void RMSNorm(
 static HWY_NOINLINE HWY_MAYBE_UNUSED void RMSNorm(
     const float* HWY_RESTRICT x, const hwy::bfloat16_t* HWY_RESTRICT weight,
     float* HWY_RESTRICT out, size_t size) {
-  constexpr float eps = 1e-6f;
-  float ss = SquaredL2(x, size);
-  ss = 1.0f / sqrtf(ss / StaticCast<float>(size) + eps);
-  for (size_t j = 0; j < size; j++) {
-    // Note 1.0f centering here
-    out[j] = (1.0f + hwy::F32FromBF16(weight[j])) * (ss * x[j]);
+  namespace hn = hwy::HWY_NAMESPACE;
+
+  constexpr float kEps = 1e-6f;
+  constexpr size_t kUnrollSize = 2;
+
+  const hn::ScalableTag<hwy::bfloat16_t> dbf;
+  const hn::Repartition<float, decltype(dbf)> df32;
+  const size_t N32 = hn::Lanes(df32);
+
+  const float ss = SquaredL2(x, size);
+  const auto vss =
+      hn::Set(df32, 1.0f / sqrtf(ss / StaticCast<float>(size) + kEps));
+
+  HWY_DASSERT(size % (kUnrollSize * MaxLanes(df32)) == 0);
+  for (size_t i = 0; i < size; i += kUnrollSize * N32) {
+    const hn::Vec<decltype(dbf)> w16 = hn::LoadU(dbf, weight + i);
+    const auto w0 = hn::PromoteLowerTo(df32, w16);
+    const auto w1 = hn::PromoteUpperTo(df32, w16);
+    const auto m0 = hn::Mul(vss, hn::LoadU(df32, x + i));
+    const auto m1 = hn::Mul(vss, hn::LoadU(df32, x + i + N32));
+
+    // (1+weight) * m = m + weight*m = one FMA.
+    hn::StoreU(hn::MulAdd(m0, w0, m0), df32, out + i);
+    hn::StoreU(hn::MulAdd(m1, w1, m1), df32, out + i + N32);
   }
 }
 
diff --git a/run.cc b/run.cc
index 1487a8a..8bc0910 100644
--- a/run.cc
+++ b/run.cc
@@ -66,8 +66,8 @@ void ShowConfig(LoaderArgs& loader, InferenceArgs& inference, AppArgs& app) {
               << std::thread::hardware_concurrency() << std::endl
               << "Instruction set               : "
               << hwy::TargetName(hwy::DispatchedTarget()) << " ("
-              << hwy::VectorBytes() * 8 << " bits)"
-              << "\n"
+              << hwy::VectorBytes() * 8 << " bits)" << "\n"
+              << "Compiled config               : " << CompiledConfig() << "\n"
               << "Weight Type                   : "
               << gcpp::TypeName(gcpp::WeightT()) << "\n"
               << "EmbedderInput Type            : "
@@ -119,7 +119,7 @@ void ReplGemma(gcpp::Gemma& model, gcpp::KVCache& kv_cache,
                        verbosity](int token, float) {
     ++abs_pos;
     ++current_pos;
-    if (current_pos < prompt_size) {
+    if (current_pos <= prompt_size) {
       std::cerr << "." << std::flush;
     } else if (token == gcpp::EOS_ID) {
       if (!args.multiturn) {
diff --git a/util/app.h b/util/app.h
index 8a59e09..ed97525 100644
--- a/util/app.h
+++ b/util/app.h
@@ -36,6 +36,24 @@
 
 namespace gcpp {
 
+static inline const char* CompiledConfig() {
+  if (HWY_IS_ASAN) {
+    return "asan";
+  } else if (HWY_IS_MSAN) {
+    return "msan";
+  } else if (HWY_IS_TSAN) {
+    return "tsan";
+#if defined(HWY_IS_UBSAN)
+  } else if (HWY_IS_UBSAN) {
+    return "ubsan";
+#endif
+  } else if (HWY_IS_DEBUG_BUILD) {
+    return "dbg";
+  } else {
+    return "opt";
+  }
+}
+
 static inline void PinThreadToCore(size_t cpu_index) {
 #if HWY_OS_LINUX
   // Forces the thread to run on the logical processor with the same number.