From 231d70b65cc85347028327b0534ecd5d72695a3d Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 18:28:26 -0700
Subject: [PATCH 01/15] Debug

---
 cmake/onnxruntime_mlas.cmake | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
index 1a5eba9de4759..4058ee8cf8f0f 100644
--- a/cmake/onnxruntime_mlas.cmake
+++ b/cmake/onnxruntime_mlas.cmake
@@ -796,11 +796,6 @@ else()
           ${MLAS_SRC_DIR}/intrinsics/avx2/qdwconv_avx2.cpp
           ${MLAS_SRC_DIR}/intrinsics/avx2/saturation_check_avx2.cpp
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx2.cpp
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.h
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen64.h
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen128.h
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen32.h
           ${MLAS_SRC_DIR}/sqnbitgemm_lut_kernel_avx2.h
           ${MLAS_SRC_DIR}/sqnbitgemm_lut_kernel_avx2.cpp
           ${MLAS_SRC_DIR}/rotary_embedding_kernel_avx2.h
@@ -847,6 +842,11 @@ endif()
           ${MLAS_SRC_DIR}/x86_64/QgemmU8X8KernelAvx512Core.S
           ${MLAS_SRC_DIR}/x86_64/ConvSymKernelAvx512Core.S
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512.cpp
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.h
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen64.h
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen128.h
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen32.h
         )
         set_source_files_properties(${mlas_platform_srcs_avx512core} PROPERTIES COMPILE_FLAGS "-mfma -mavx512vnni -mavx512bw -mavx512dq -mavx512vl")
 

From 6bba3dd36a4926e2633ebde9d727b07fdca15fe7 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 18:36:17 -0700
Subject: [PATCH 02/15] Add debug file

---
 .../test/mlas/unittest/test_mlas_diag.cpp     | 251 ++++++++++++++++++
 1 file changed, 251 insertions(+)
 create mode 100644 onnxruntime/test/mlas/unittest/test_mlas_diag.cpp

diff --git a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
new file mode 100644
index 0000000000000..f707ce0fc1644
--- /dev/null
+++ b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
@@ -0,0 +1,251 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+//
+// MLAS diagnostic helper. Always-on (no env var) so it works in CI
+// environments where setting environment variables is not possible.
+//
+// At process startup (before any test runs) this dumps:
+//   * Build-time ISA macros (__AVX2__, __AVX512F__, ...)
+//   * Runtime MlasPlatform feature flags (Avx2Supported_, Avx512Supported_)
+//   * Active QNBitGemmDispatch table identity (Avx2 / Avx2vnni / Avx512 /
+//     Avx512vnni / other)
+//   * On Linux, raw CPUID leaves 1 and 7 sub-leaf 0 so we can confirm what
+//     the kernel/hypervisor actually exposes vs. what MlasPlatform decided.
+//
+// When this file is linked into the test binary it also installs a SIGILL
+// handler (Linux only) that prints the faulting RIP, the module/symbol
+// nearest to it (dladdr), and the first 16 bytes at RIP so we can
+// disassemble the offending instruction offline. The handler is
+// async-signal-safe (write(2) only). It re-raises SIGILL with the default
+// disposition afterwards so the process still terminates and gtest still
+// reports the failing test.
+//
+// All output is tagged with the "MLAS-DIAG:" prefix so it can be grep'd
+// out of CI logs.
+//
+
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+#include "gtest/gtest.h"
+
+#include "core/mlas/lib/mlasi.h"
+
+#if defined(__linux__)
+#include <unistd.h>
+#include <signal.h>
+#include <dlfcn.h>
+#endif
+
+namespace {
+
+constexpr const char* kTag = "MLAS-DIAG: ";
+
+void PrintBuildMacros() {
+  std::fprintf(stderr, "%sbuild macros:", kTag);
+#if defined(__AVX__)
+  std::fprintf(stderr, " __AVX__");
+#endif
+#if defined(__AVX2__)
+  std::fprintf(stderr, " __AVX2__");
+#endif
+#if defined(__FMA__)
+  std::fprintf(stderr, " __FMA__");
+#endif
+#if defined(__AVX512F__)
+  std::fprintf(stderr, " __AVX512F__");
+#endif
+#if defined(__AVX512BW__)
+  std::fprintf(stderr, " __AVX512BW__");
+#endif
+#if defined(__AVX512DQ__)
+  std::fprintf(stderr, " __AVX512DQ__");
+#endif
+#if defined(__AVX512VL__)
+  std::fprintf(stderr, " __AVX512VL__");
+#endif
+#if defined(__AVX512VNNI__)
+  std::fprintf(stderr, " __AVX512VNNI__");
+#endif
+#if defined(__GNUC__)
+  std::fprintf(stderr, " __GNUC__=%d", __GNUC__);
+#endif
+#if defined(_MSC_VER)
+  std::fprintf(stderr, " _MSC_VER=%d", _MSC_VER);
+#endif
+  std::fprintf(stderr, "\n");
+}
+
+void PrintMlasPlatform() {
+  const auto& p = GetMlasPlatform();
+  std::fprintf(stderr,
+               "%sMlasPlatform: Avx2Supported_=%d Avx512Supported_=%d\n",
+               kTag,
+               static_cast<int>(p.Avx2Supported_),
+               static_cast<int>(p.Avx512Supported_));
+
+  const void* dispatch = static_cast<const void*>(p.QNBitGemmDispatch);
+  const char* name = "unknown/null";
+#if defined(MLAS_TARGET_AMD64) || defined(MLAS_TARGET_IX86)
+  if (dispatch == &MlasSQNBitGemmDispatchAvx2) {
+    name = "Avx2";
+  } else if (dispatch == &MlasSQNBitGemmDispatchAvx2vnni) {
+    name = "Avx2vnni";
+  } else if (dispatch == &MlasSQNBitGemmDispatchAvx512) {
+    name = "Avx512";
+  } else if (dispatch == &MlasSQNBitGemmDispatchAvx512vnni) {
+    name = "Avx512vnni";
+  } else if (dispatch == nullptr) {
+    name = "null";
+  }
+#endif
+  std::fprintf(stderr,
+               "%sQNBitGemmDispatch=%p (%s)\n",
+               kTag, dispatch, name);
+}
+
+#if defined(__linux__) && (defined(__x86_64__) || defined(__i386__))
+
+void PrintRawCpuid() {
+  auto cpuid = [](uint32_t leaf, uint32_t subleaf,
+                  uint32_t& eax, uint32_t& ebx, uint32_t& ecx, uint32_t& edx) {
+    __asm__ volatile(
+        "cpuid"
+        : "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx)
+        : "a"(leaf), "c"(subleaf));
+  };
+  uint32_t a, b, c, d;
+  cpuid(1, 0, a, b, c, d);
+  std::fprintf(stderr, "%sCPUID(1,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
+               kTag, a, b, c, d);
+  cpuid(7, 0, a, b, c, d);
+  std::fprintf(stderr, "%sCPUID(7,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
+               kTag, a, b, c, d);
+}
+
+// Async-signal-safe writer: integer -> hex into a small buffer.
+size_t HexWrite(uintptr_t v, char* out, size_t out_len) {
+  static const char kHex[] = "0123456789abcdef";
+  // Reserve room for "0x" + up to 16 hex digits + null.
+  char tmp[20];
+  size_t n = 0;
+  if (v == 0) {
+    tmp[n++] = '0';
+  } else {
+    while (v != 0 && n < sizeof(tmp)) {
+      tmp[n++] = kHex[v & 0xF];
+      v >>= 4;
+    }
+  }
+  // Now reverse into out, prefixed with "0x".
+  size_t w = 0;
+  if (out_len > 0) out[w++] = '0';
+  if (out_len > 1) out[w++] = 'x';
+  while (n > 0 && w < out_len) {
+    out[w++] = tmp[--n];
+  }
+  return w;
+}
+
+void SafeWriteStr(const char* s) {
+  if (s == nullptr) return;
+  size_t n = 0;
+  while (s[n] != '\0') ++n;
+  (void)write(STDERR_FILENO, s, n);
+}
+
+void SafeWriteHex(uintptr_t v) {
+  char buf[20];
+  size_t n = HexWrite(v, buf, sizeof(buf));
+  (void)write(STDERR_FILENO, buf, n);
+}
+
+void SigillHandler(int sig, siginfo_t* info, void* /*ucontext*/) {
+  SafeWriteStr(kTag);
+  SafeWriteStr("SIGILL caught. si_addr=");
+  uintptr_t addr = reinterpret_cast<uintptr_t>(info ? info->si_addr : nullptr);
+  SafeWriteHex(addr);
+  SafeWriteStr("\n");
+
+  // Try to resolve module + nearest symbol. dladdr is not strictly
+  // async-signal-safe, but in practice on glibc it is safe enough for a
+  // best-effort crash dump; if it deadlocks we still get the si_addr above.
+  Dl_info dlinfo{};
+  if (info && dladdr(info->si_addr, &dlinfo) != 0) {
+    SafeWriteStr(kTag);
+    SafeWriteStr("dladdr: fname=");
+    SafeWriteStr(dlinfo.dli_fname ? dlinfo.dli_fname : "(null)");
+    SafeWriteStr(" fbase=");
+    SafeWriteHex(reinterpret_cast<uintptr_t>(dlinfo.dli_fbase));
+    SafeWriteStr(" sname=");
+    SafeWriteStr(dlinfo.dli_sname ? dlinfo.dli_sname : "(null)");
+    SafeWriteStr(" saddr=");
+    SafeWriteHex(reinterpret_cast<uintptr_t>(dlinfo.dli_saddr));
+    SafeWriteStr("\n");
+  }
+
+  // Dump first 16 bytes at si_addr so we can decode the instruction offline.
+  if (info && info->si_addr != nullptr) {
+    SafeWriteStr(kTag);
+    SafeWriteStr("bytes@si_addr:");
+    const unsigned char* p = static_cast<const unsigned char*>(info->si_addr);
+    for (int i = 0; i < 16; ++i) {
+      char two[3];
+      static const char kHex[] = "0123456789abcdef";
+      two[0] = kHex[(p[i] >> 4) & 0xF];
+      two[1] = kHex[p[i] & 0xF];
+      two[2] = ' ';
+      (void)write(STDERR_FILENO, " ", 1);
+      (void)write(STDERR_FILENO, two, 3);
+    }
+    SafeWriteStr("\n");
+  }
+
+  // Re-raise with default disposition so gtest still records the crash.
+  struct sigaction dfl{};
+  dfl.sa_handler = SIG_DFL;
+  sigemptyset(&dfl.sa_mask);
+  sigaction(sig, &dfl, nullptr);
+  raise(sig);
+}
+
+void InstallSigillHandler() {
+  struct sigaction sa{};
+  sa.sa_sigaction = &SigillHandler;
+  sa.sa_flags = SA_SIGINFO | SA_RESETHAND;
+  sigemptyset(&sa.sa_mask);
+  if (sigaction(SIGILL, &sa, nullptr) != 0) {
+    std::fprintf(stderr, "%sfailed to install SIGILL handler\n", kTag);
+  } else {
+    std::fprintf(stderr, "%sSIGILL handler installed\n", kTag);
+  }
+}
+
+#else  // !linux x86
+
+void PrintRawCpuid() {}
+void InstallSigillHandler() {}
+
+#endif
+
+class MlasDiagEnvironment : public ::testing::Environment {
+ public:
+  void SetUp() override {
+    std::fprintf(stderr, "%s---- MLAS diagnostic dump ----\n", kTag);
+    PrintBuildMacros();
+    PrintRawCpuid();
+    PrintMlasPlatform();
+    InstallSigillHandler();
+    std::fprintf(stderr, "%s---- end MLAS diagnostic dump ----\n", kTag);
+    std::fflush(stderr);
+  }
+};
+
+// Register at program startup so it runs before any TEST() body executes.
+const ::testing::Environment* const kMlasDiagEnv =
+    ::testing::AddGlobalTestEnvironment(new MlasDiagEnvironment());
+
+}  // namespace

From d5bab99a29a4b8e0ffd6d8d1fc43ad7abed8b1c4 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 19:26:38 -0700
Subject: [PATCH 03/15] More debug

---
 onnxruntime/test/mlas/unittest/test_main.cpp  |  8 ++++++++
 .../test/mlas/unittest/test_mlas_diag.cpp     | 20 +++++++++++++++----
 2 files changed, 24 insertions(+), 4 deletions(-)

diff --git a/onnxruntime/test/mlas/unittest/test_main.cpp b/onnxruntime/test/mlas/unittest/test_main.cpp
index 505c0c01dfa90..edda089a6b75d 100644
--- a/onnxruntime/test/mlas/unittest/test_main.cpp
+++ b/onnxruntime/test/mlas/unittest/test_main.cpp
@@ -56,6 +56,12 @@ bool AddTestRegister(TestRegister test_register) {
   return true;
 }
 
+// Defined in test_mlas_diag.cpp. Forces that TU to be linked in even when
+// the linker would otherwise discard it (it has no externally-referenced
+// symbols of its own). The diagnostic environment is registered as a side
+// effect of this call.
+void RegisterMlasDiagEnvironment();
+
 int main(int argc, char** argv) {
   bool is_short_execute = (argc <= 1 || strcmp("--long", argv[1]) != 0);
   std::cout << "-------------------------------------------------------" << std::endl;
@@ -67,6 +73,8 @@ int main(int argc, char** argv) {
   std::cout << "----Total " << test_count << " tests registered programmably!" << std::endl;
   std::cout << "-------------------------------------------------------" << std::endl;
 
+  RegisterMlasDiagEnvironment();
+
   ::testing::InitGoogleTest(&argc, argv);
 
   return RUN_ALL_TESTS();
diff --git a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
index f707ce0fc1644..dcbecd9dd3912 100644
--- a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
+++ b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
@@ -244,8 +244,20 @@ class MlasDiagEnvironment : public ::testing::Environment {
   }
 };
 
-// Register at program startup so it runs before any TEST() body executes.
-const ::testing::Environment* const kMlasDiagEnv =
-    ::testing::AddGlobalTestEnvironment(new MlasDiagEnvironment());
-
 }  // namespace
+
+// Externally-visible entry point so test_main.cpp can force this TU to link
+// and register the diagnostic GTest environment. Without an externally
+// referenced symbol, the MSVC linker will discard the anonymous-namespace
+// static initializer that registers the environment, which is exactly what
+// happened in CI: the local test process showed MLAS-DIAG: output but the
+// CI test binary did not, because nothing referenced this TU.
+void RegisterMlasDiagEnvironment() {
+  // ::testing::AddGlobalTestEnvironment takes ownership of the pointer.
+  static bool registered = false;
+  if (registered) {
+    return;
+  }
+  registered = true;
+  ::testing::AddGlobalTestEnvironment(new MlasDiagEnvironment());
+}

From 8d1fccd63802f36a8b20db5f8363dfe8081e960e Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 20:05:32 -0700
Subject: [PATCH 04/15] Add right test exclusion when the test runs on
 non-AVX512 hosts

---
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    | 30 +++++++++++++++++++
 1 file changed, 30 insertions(+)

diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 7690b32c3c6a8..906e045df3a7a 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -310,6 +310,16 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 // handler in a follow-up.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64) {
+  // The scalar entry point lives in sqnbitgemm_kernel_avx512_2bit.cpp, which is
+  // compiled with AVX-512 flags so its SIMD siblings can use intrinsics. The
+  // compiler may autovectorize this "scalar" function to AVX-512 instructions.
+  // Calling it directly here bypasses MlasIsQNBitGemmAvailable, so on hosts
+  // that do not safely expose AVX-512 this can SIGILL. Mirror the AVX-512
+  // SIMD test guard. Long-term fix: split the TU so the scalar entry has
+  // flag-neutral codegen.
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
+  }
   struct Shape {
     size_t M, N, K;
   };
@@ -351,6 +361,10 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64) {
 // Same coverage with per-block non-default zero points.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
+  // See Scalar_BlkLen64 comment.
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
+  }
   struct Shape {
     size_t M, N, K;
   };
@@ -806,6 +820,10 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128) {
+  // See Scalar_BlkLen64 comment.
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -819,6 +837,10 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
+  // See Scalar_BlkLen64 comment.
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -1160,6 +1182,10 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32) {
+  // See Scalar_BlkLen64 comment.
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1173,6 +1199,10 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
+  // See Scalar_BlkLen64 comment.
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {

From 86302ddc738a6290e615ccba74173527af326be2 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 20:44:40 -0700
Subject: [PATCH 05/15] Fixes

---
 cmake/onnxruntime_mlas.cmake | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)

diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
index 4058ee8cf8f0f..acb5988956611 100644
--- a/cmake/onnxruntime_mlas.cmake
+++ b/cmake/onnxruntime_mlas.cmake
@@ -842,14 +842,30 @@ endif()
           ${MLAS_SRC_DIR}/x86_64/QgemmU8X8KernelAvx512Core.S
           ${MLAS_SRC_DIR}/x86_64/ConvSymKernelAvx512Core.S
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512.cpp
+          # NOTE: sqnbitgemm_kernel_avx512_2bit_blklen{32,64,128}.h are also AVX-512
+          # only, but they are header files included by sqnbitgemm_kernel_avx512.cpp
+          # and sqnbitgemm_kernel_avx512vnni.cpp, so they pick up those TUs'
+          # AVX-512 compile flags automatically. Listing them here just adds
+          # them to IDE project files for navigation.
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.h
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen64.h
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen128.h
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen32.h
         )
         set_source_files_properties(${mlas_platform_srcs_avx512core} PROPERTIES COMPILE_FLAGS "-mfma -mavx512vnni -mavx512bw -mavx512dq -mavx512vl")
 
+        # sqnbitgemm_kernel_avx512_2bit.cpp is named "avx512" but is intentionally
+        # pure scalar C++ -- it implements the W2 pack helper and a scalar
+        # reference oracle (see file header comment). It must NOT be compiled
+        # with AVX-512 flags, otherwise the compiler may autovectorize it to
+        # AVX-512 (e.g. VNNI vpdpbusd) instructions that some virtualized
+        # hosts do not accept, producing SIGILL when tests call those symbols
+        # directly.
+        set(mlas_platform_srcs_w2_scalar
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp
+        )
+        set_source_files_properties(${mlas_platform_srcs_w2_scalar} PROPERTIES COMPILE_FLAGS "-mavx2 -mfma -mf16c")
+
         set(mlas_platform_srcs_avx512vnni
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512vnni.cpp
           ${MLAS_SRC_DIR}/qkv_quant_kernel_avx512vnni.cpp
@@ -867,6 +883,7 @@ endif()
           ${mlas_platform_srcs_avx2}
           ${mlas_platform_srcs_avx512f}
           ${mlas_platform_srcs_avx512core}
+          ${mlas_platform_srcs_w2_scalar}
           ${mlas_platform_srcs_avx512vnni}
         )
 

From f26fe93f91b0bd53e1eb3bb674bd5a6348dabd29 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 21:15:03 -0700
Subject: [PATCH 06/15] More debug

---
 .../test/mlas/unittest/test_mlas_diag.cpp     | 109 ++++++++++++------
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    |  30 -----
 2 files changed, 74 insertions(+), 65 deletions(-)

diff --git a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
index dcbecd9dd3912..96e99022fed23 100644
--- a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
+++ b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
@@ -25,6 +25,7 @@
 // out of CI logs.
 //
 
+#include <cstdarg>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
@@ -44,48 +45,63 @@ namespace {
 
 constexpr const char* kTag = "MLAS-DIAG: ";
 
+// Print a formatted message to both stdout and stderr. CI infrastructures
+// often capture only one of the two; doing both maximizes the chance the
+// diagnostic shows up.
+void PrintBoth(const char* fmt, ...) {
+  va_list ap;
+  va_start(ap, fmt);
+  va_list ap2;
+  va_copy(ap2, ap);
+  std::vfprintf(stderr, fmt, ap);
+  std::vfprintf(stdout, fmt, ap2);
+  va_end(ap);
+  va_end(ap2);
+  std::fflush(stderr);
+  std::fflush(stdout);
+}
+
 void PrintBuildMacros() {
-  std::fprintf(stderr, "%sbuild macros:", kTag);
+  PrintBoth("%sbuild macros:", kTag);
 #if defined(__AVX__)
-  std::fprintf(stderr, " __AVX__");
+  PrintBoth(" __AVX__");
 #endif
 #if defined(__AVX2__)
-  std::fprintf(stderr, " __AVX2__");
+  PrintBoth(" __AVX2__");
 #endif
 #if defined(__FMA__)
-  std::fprintf(stderr, " __FMA__");
+  PrintBoth(" __FMA__");
 #endif
 #if defined(__AVX512F__)
-  std::fprintf(stderr, " __AVX512F__");
+  PrintBoth(" __AVX512F__");
 #endif
 #if defined(__AVX512BW__)
-  std::fprintf(stderr, " __AVX512BW__");
+  PrintBoth(" __AVX512BW__");
 #endif
 #if defined(__AVX512DQ__)
-  std::fprintf(stderr, " __AVX512DQ__");
+  PrintBoth(" __AVX512DQ__");
 #endif
 #if defined(__AVX512VL__)
-  std::fprintf(stderr, " __AVX512VL__");
+  PrintBoth(" __AVX512VL__");
 #endif
 #if defined(__AVX512VNNI__)
-  std::fprintf(stderr, " __AVX512VNNI__");
+  PrintBoth(" __AVX512VNNI__");
 #endif
 #if defined(__GNUC__)
-  std::fprintf(stderr, " __GNUC__=%d", __GNUC__);
+  PrintBoth(" __GNUC__=%d", __GNUC__);
 #endif
 #if defined(_MSC_VER)
-  std::fprintf(stderr, " _MSC_VER=%d", _MSC_VER);
+  PrintBoth(" _MSC_VER=%d", _MSC_VER);
 #endif
-  std::fprintf(stderr, "\n");
+  PrintBoth("\n");
 }
 
 void PrintMlasPlatform() {
   const auto& p = GetMlasPlatform();
-  std::fprintf(stderr,
-               "%sMlasPlatform: Avx2Supported_=%d Avx512Supported_=%d\n",
-               kTag,
-               static_cast<int>(p.Avx2Supported_),
-               static_cast<int>(p.Avx512Supported_));
+  PrintBoth("%sMlasPlatform: Avx2Supported_=%d Avx512Supported_=%d\n",
+            kTag,
+            static_cast<int>(p.Avx2Supported_),
+            static_cast<int>(p.Avx512Supported_));
 
   const void* dispatch = static_cast<const void*>(p.QNBitGemmDispatch);
   const char* name = "unknown/null";
@@ -102,9 +118,8 @@ void PrintMlasPlatform() {
     name = "null";
   }
 #endif
-  std::fprintf(stderr,
-               "%sQNBitGemmDispatch=%p (%s)\n",
-               kTag, dispatch, name);
+  PrintBoth("%sQNBitGemmDispatch=%p (%s)\n",
+            kTag, dispatch, name);
 }
 
 #if defined(__linux__) && (defined(__x86_64__) || defined(__i386__))
@@ -119,11 +134,11 @@ void PrintRawCpuid() {
   };
   uint32_t a, b, c, d;
   cpuid(1, 0, a, b, c, d);
-  std::fprintf(stderr, "%sCPUID(1,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
-               kTag, a, b, c, d);
+  PrintBoth("%sCPUID(1,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
+            kTag, a, b, c, d);
   cpuid(7, 0, a, b, c, d);
-  std::fprintf(stderr, "%sCPUID(7,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
-               kTag, a, b, c, d);
+  PrintBoth("%sCPUID(7,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
+            kTag, a, b, c, d);
 }
 
 // Async-signal-safe writer: integer -> hex into a small buffer.
@@ -155,12 +170,14 @@ void SafeWriteStr(const char* s) {
   size_t n = 0;
   while (s[n] != '\0') ++n;
   (void)write(STDERR_FILENO, s, n);
+  (void)write(STDOUT_FILENO, s, n);
 }
 
 void SafeWriteHex(uintptr_t v) {
   char buf[20];
   size_t n = HexWrite(v, buf, sizeof(buf));
   (void)write(STDERR_FILENO, buf, n);
+  (void)write(STDOUT_FILENO, buf, n);
 }
 
 void SigillHandler(int sig, siginfo_t* info, void* /*ucontext*/) {
@@ -193,13 +210,14 @@ void SigillHandler(int sig, siginfo_t* info, void* /*ucontext*/) {
     SafeWriteStr("bytes@si_addr:");
     const unsigned char* p = static_cast<const unsigned char*>(info->si_addr);
     for (int i = 0; i < 16; ++i) {
-      char two[3];
+      char two[4];
       static const char kHex[] = "0123456789abcdef";
-      two[0] = kHex[(p[i] >> 4) & 0xF];
-      two[1] = kHex[p[i] & 0xF];
-      two[2] = ' ';
-      (void)write(STDERR_FILENO, " ", 1);
+      two[0] = ' ';
+      two[1] = kHex[(p[i] >> 4) & 0xF];
+      two[2] = kHex[p[i] & 0xF];
+      two[3] = '\0';
       (void)write(STDERR_FILENO, two, 3);
+      (void)write(STDOUT_FILENO, two, 3);
     }
     SafeWriteStr("\n");
   }
@@ -213,15 +231,22 @@ void SigillHandler(int sig, siginfo_t* info, void* /*ucontext*/) {
 }
 
 void InstallSigillHandler() {
+  static bool installed = false;
+  if (installed) return;
+  installed = true;
   struct sigaction sa{};
   sa.sa_sigaction = &SigillHandler;
   sa.sa_flags = SA_SIGINFO | SA_RESETHAND;
   sigemptyset(&sa.sa_mask);
   if (sigaction(SIGILL, &sa, nullptr) != 0) {
     std::fprintf(stderr, "%sfailed to install SIGILL handler\n", kTag);
+    std::fprintf(stdout, "%sfailed to install SIGILL handler\n", kTag);
   } else {
     std::fprintf(stderr, "%sSIGILL handler installed\n", kTag);
+    std::fprintf(stdout, "%sSIGILL handler installed\n", kTag);
   }
+  std::fflush(stderr);
+  std::fflush(stdout);
 }
 
 #else  // !linux x86
@@ -234,13 +259,14 @@ void InstallSigillHandler() {}
 class MlasDiagEnvironment : public ::testing::Environment {
  public:
   void SetUp() override {
-    std::fprintf(stderr, "%s---- MLAS diagnostic dump ----\n", kTag);
+    PrintBoth("%s---- MLAS diagnostic dump ----\n", kTag);
     PrintBuildMacros();
     PrintRawCpuid();
     PrintMlasPlatform();
     InstallSigillHandler();
-    std::fprintf(stderr, "%s---- end MLAS diagnostic dump ----\n", kTag);
+    PrintBoth("%s---- end MLAS diagnostic dump ----\n", kTag);
     std::fflush(stderr);
+    std::fflush(stdout);
   }
 };
 
@@ -248,11 +274,24 @@ class MlasDiagEnvironment : public ::testing::Environment {
 
 // Externally-visible entry point so test_main.cpp can force this TU to link
 // and register the diagnostic GTest environment. Without an externally
-// referenced symbol, the MSVC linker will discard the anonymous-namespace
-// static initializer that registers the environment, which is exactly what
-// happened in CI: the local test process showed MLAS-DIAG: output but the
-// CI test binary did not, because nothing referenced this TU.
+// referenced symbol, the linker may discard the anonymous-namespace static
+// initializer that registers the environment.
 void RegisterMlasDiagEnvironment() {
+  // Install the SIGILL handler as soon as we are called -- before GTest
+  // even sets up the environment. This way a fault during static init or
+  // very early test setup is still captured.
+  InstallSigillHandler();
+
+  // Also print the basic platform dump immediately so CI logs have it
+  // even if a crash kills us before the Environment::SetUp() runs.
+  PrintBoth("%s---- MLAS early diagnostic dump ----\n", kTag);
+  PrintBuildMacros();
+  PrintRawCpuid();
+  PrintMlasPlatform();
+  PrintBoth("%s---- end MLAS early diagnostic dump ----\n", kTag);
+  std::fflush(stderr);
+  std::fflush(stdout);
+
   // ::testing::AddGlobalTestEnvironment takes ownership of the pointer.
   static bool registered = false;
   if (registered) {
diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 906e045df3a7a..7690b32c3c6a8 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -310,16 +310,6 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 // handler in a follow-up.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64) {
-  // The scalar entry point lives in sqnbitgemm_kernel_avx512_2bit.cpp, which is
-  // compiled with AVX-512 flags so its SIMD siblings can use intrinsics. The
-  // compiler may autovectorize this "scalar" function to AVX-512 instructions.
-  // Calling it directly here bypasses MlasIsQNBitGemmAvailable, so on hosts
-  // that do not safely expose AVX-512 this can SIGILL. Mirror the AVX-512
-  // SIMD test guard. Long-term fix: split the TU so the scalar entry has
-  // flag-neutral codegen.
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
-  }
   struct Shape {
     size_t M, N, K;
   };
@@ -361,10 +351,6 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64) {
 // Same coverage with per-block non-default zero points.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
-  // See Scalar_BlkLen64 comment.
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
-  }
   struct Shape {
     size_t M, N, K;
   };
@@ -820,10 +806,6 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128) {
-  // See Scalar_BlkLen64 comment.
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
-  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -837,10 +819,6 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
-  // See Scalar_BlkLen64 comment.
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
-  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -1182,10 +1160,6 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32) {
-  // See Scalar_BlkLen64 comment.
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
-  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1199,10 +1173,6 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
-  // See Scalar_BlkLen64 comment.
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "W2 scalar reference TU is compiled with AVX-512 flags; skip on non-AVX-512 hosts.";
-  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {

From 8b8dc5f71d289bc82c1bee1bed86dc71083f690f Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 22:02:23 -0700
Subject: [PATCH 07/15] More debug 2

---
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    | 28 +++++++++++++++----
 1 file changed, 22 insertions(+), 6 deletions(-)

diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 7690b32c3c6a8..d9cbf62f354bf 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -309,7 +309,18 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 // NOT a multiple of 256 so it's excluded; that shape will need a tail
 // handler in a follow-up.
 //
-TEST(MlasSq2BitTest, Scalar_BlkLen64) {
+// TODO(W2-scalar-disabled): re-enable once the W2 scalar reference TU codegen
+// is pinned away from AVX-512 autovectorization (LTO can override per-file
+// flags). Tracked in follow-up. These tests bypass MlasIsQNBitGemmAvailable
+// and call internal scalar symbols directly, so they execute the same TU's
+// instructions regardless of dispatch gating. On some virtualized hosts
+// (e.g. Azure Standard_D32ads_v6, Zen 4) certain AVX-512 forms still SIGILL
+// even when CPUID reports support, killing the whole test binary and
+// blocking unrelated suites. Disable them as a CI unblock; production paths
+// are unaffected (the scalar GEMM kernel is not installed in any dispatch
+// table; the scalar pack helper only runs on AVX-512 hosts where the same
+// codegen has been observed to be safe in practice).
+TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen64) {
   struct Shape {
     size_t M, N, K;
   };
@@ -350,7 +361,8 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64) {
 //
 // Same coverage with per-block non-default zero points.
 //
-TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
+// See DISABLED_Scalar_BlkLen64 comment.
+TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen64_WithZeroPoints) {
   struct Shape {
     size_t M, N, K;
   };
@@ -805,7 +817,8 @@ constexpr struct {
 
 }  // namespace
 
-TEST(MlasSq2BitTest, Scalar_BlkLen128) {
+// See DISABLED_Scalar_BlkLen64 comment.
+TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen128) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -818,7 +831,8 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128) {
   }
 }
 
-TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
+// See DISABLED_Scalar_BlkLen64 comment.
+TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen128_WithZeroPoints) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -1159,7 +1173,8 @@ constexpr struct {
 
 }  // namespace
 
-TEST(MlasSq2BitTest, Scalar_BlkLen32) {
+// See DISABLED_Scalar_BlkLen64 comment.
+TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen32) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1172,7 +1187,8 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32) {
   }
 }
 
-TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
+// See DISABLED_Scalar_BlkLen64 comment.
+TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen32_WithZeroPoints) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {

From 118ba2b6394f966834753973ade3da107902f5aa Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 22:43:23 -0700
Subject: [PATCH 08/15] Debug 3

---
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    | 69 +++++++++++++------
 1 file changed, 47 insertions(+), 22 deletions(-)

diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index d9cbf62f354bf..7b753489e9e3f 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -302,25 +302,45 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 
 }  // namespace
 
+// =============================================================================
+// W2 scalar reference tests -- temporarily disabled via #if 0.
+//
+// These tests call internal scalar W2 symbols directly, bypassing
+// MlasIsQNBitGemmAvailable. The scalar reference TU
+// (sqnbitgemm_kernel_avx512_2bit.cpp) is compiled with AVX-512-capable
+// flags and the build uses LTO, so the compiler may autovectorize these
+// "scalar" functions into AVX-512 instructions. On some virtualized Linux
+// hosts (e.g. Azure Standard_D32ads_v6 / Zen 4 CI image) certain AVX-512
+// VNNI encodings still SIGILL even when CPUID reports support, killing the
+// whole onnxruntime_mlas_test binary and blocking all downstream test
+// suites including the production-path MatMul2Bits op tests in
+// onnxruntime_test_all.
+//
+// Note: GTest's DISABLED_ prefix did not skip these in CI in practice
+// (binary still showed [ RUN ] for the original name even after rename),
+// so we wrap them out of the binary completely via #if 0.
+//
+// Production paths are unaffected:
+//   * SQ2BitGemmKernel_BlkSum_CompInt8_Scalar is not installed in any
+//     dispatch table -- it is only used as an oracle in these tests.
+//   * SQ2BitGemmPackQuantBDataAndBlkSum_Scalar runs at model load time
+//     under the AVX-512 dispatch path, which is only selected on hosts
+//     where it has been observed to be safe in practice.
+//
+// TODO(W2-scalar-disabled): re-enable once the W2 scalar reference TU
+// codegen is pinned away from AVX-512 autovectorization, e.g. by adding
+// __attribute__((target("..."))) on the scalar entry points or by
+// excluding the TU from LTO.
+// =============================================================================
+
+#if 0
 //
 // Scalar block-group test, no zero-points. Covers the same small synthetic
 // shapes + representative prefill sizes used by the production W2 tests. All
 // shapes have K as a multiple of (kBlkLen * kBlockGroupBlks) = 256. K=384 is
 // NOT a multiple of 256 so it's excluded; that shape will need a tail
 // handler in a follow-up.
-//
-// TODO(W2-scalar-disabled): re-enable once the W2 scalar reference TU codegen
-// is pinned away from AVX-512 autovectorization (LTO can override per-file
-// flags). Tracked in follow-up. These tests bypass MlasIsQNBitGemmAvailable
-// and call internal scalar symbols directly, so they execute the same TU's
-// instructions regardless of dispatch gating. On some virtualized hosts
-// (e.g. Azure Standard_D32ads_v6, Zen 4) certain AVX-512 forms still SIGILL
-// even when CPUID reports support, killing the whole test binary and
-// blocking unrelated suites. Disable them as a CI unblock; production paths
-// are unaffected (the scalar GEMM kernel is not installed in any dispatch
-// table; the scalar pack helper only runs on AVX-512 hosts where the same
-// codegen has been observed to be safe in practice).
-TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen64) {
+TEST(MlasSq2BitTest, Scalar_BlkLen64) {
   struct Shape {
     size_t M, N, K;
   };
@@ -362,7 +382,7 @@ TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen64) {
 // Same coverage with per-block non-default zero points.
 //
 // See DISABLED_Scalar_BlkLen64 comment.
-TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen64_WithZeroPoints) {
+TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
   struct Shape {
     size_t M, N, K;
   };
@@ -398,6 +418,7 @@ TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen64_WithZeroPoints) {
     }
   }
 }
+#endif  // W2 scalar reference tests (BlkLen64 pair)
 
 //
 // SIMD block-group shape coverage. Phase-3+K-tail kernel requires:
@@ -817,8 +838,9 @@ constexpr struct {
 
 }  // namespace
 
-// See DISABLED_Scalar_BlkLen64 comment.
-TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen128) {
+#if 0  // See top-of-file W2 scalar reference tests note.
+// See top-of-file note for why these are disabled.
+TEST(MlasSq2BitTest, Scalar_BlkLen128) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -831,8 +853,8 @@ TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen128) {
   }
 }
 
-// See DISABLED_Scalar_BlkLen64 comment.
-TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen128_WithZeroPoints) {
+// See top-of-file note for why these are disabled.
+TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -844,6 +866,7 @@ TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen128_WithZeroPoints) {
     }
   }
 }
+#endif  // W2 scalar reference tests (BlkLen128 pair)
 
 TEST(MlasSq2BitTest, BlkLen128_Avx512) {
   if (!GetMlasPlatform().Avx512Supported_) {
@@ -1173,8 +1196,9 @@ constexpr struct {
 
 }  // namespace
 
-// See DISABLED_Scalar_BlkLen64 comment.
-TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen32) {
+#if 0  // See top-of-file W2 scalar reference tests note.
+// See top-of-file note for why these are disabled.
+TEST(MlasSq2BitTest, Scalar_BlkLen32) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1187,8 +1211,8 @@ TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen32) {
   }
 }
 
-// See DISABLED_Scalar_BlkLen64 comment.
-TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen32_WithZeroPoints) {
+// See top-of-file note for why these are disabled.
+TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1200,6 +1224,7 @@ TEST(MlasSq2BitTest, DISABLED_Scalar_BlkLen32_WithZeroPoints) {
     }
   }
 }
+#endif  // W2 scalar reference tests (BlkLen32 pair)
 
 TEST(MlasSq2BitTest, BlkLen32_Avx512) {
   if (!GetMlasPlatform().Avx512Supported_) {

From dd744c93a58b611233115c6e54366deb2082b045 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Fri, 19 Jun 2026 23:23:21 -0700
Subject: [PATCH 09/15] Debug 5

---
 onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h    | 4 ++++
 onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp | 4 ++++
 2 files changed, 8 insertions(+)

diff --git a/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h b/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h
index 16e23b3993d57..8a1927130750f 100644
--- a/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h
+++ b/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h
@@ -4,6 +4,10 @@ Copyright (c) Microsoft Corporation. All rights reserved.
 
 Licensed under the MIT License.
 
+// CI-CACHE-BUSTER 2026-06-19T22:50: force ccache/sccache miss after wrapping
+// W2 scalar reference tests in #if 0. Safe to delete once cache invalidation
+// is confirmed in the next CI run.
+
 Module Name:
 
     sqnbitgemm_kernel_avx512_2bit.h
diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 7b753489e9e3f..10e751367b152 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -4,6 +4,10 @@ Copyright (c) Microsoft Corporation. All rights reserved.
 
 Licensed under the MIT License.
 
+// CI-CACHE-BUSTER 2026-06-19T22:50: pair with sqnbitgemm_kernel_avx512_2bit.h
+// to force ccache/sccache invalidation after wrapping W2 scalar reference
+// tests in #if 0.
+
 Module Name:
 
     test_sqnbitgemm_2bit.cpp

From f2ea65b2724ec46d58032b60040dfc8f63d71a29 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Sat, 20 Jun 2026 00:07:16 -0700
Subject: [PATCH 10/15] Debug 6

---
 cmake/onnxruntime_unittests.cmake                           | 6 ++++++
 ...tgemm_2bit_gemm.cpp => test_sqnbitgemm_2bit_gemm_v2.cpp} | 0
 2 files changed, 6 insertions(+)
 rename onnxruntime/test/mlas/unittest/{test_sqnbitgemm_2bit_gemm.cpp => test_sqnbitgemm_2bit_gemm_v2.cpp} (100%)

diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
index 23eccb22476df..0842ef992c624 100644
--- a/cmake/onnxruntime_unittests.cmake
+++ b/cmake/onnxruntime_unittests.cmake
@@ -1847,6 +1847,12 @@ endif()
       )
     endif()
     target_compile_definitions(onnxruntime_mlas_test PRIVATE ${mlas_private_compile_definitions})
+    # CI-CACHE-BUSTER: force rebuild of every TU in onnxruntime_mlas_test when
+    # a persistent /build directory would otherwise reuse stale .obj files.
+    # Bump the timestamp to invalidate. Safe to remove once the
+    # W2 scalar-test #if 0 wrapping is confirmed in CI.
+    target_compile_definitions(onnxruntime_mlas_test PRIVATE
+        MLAS_TEST_REBUILD_TAG=20260619T2310)
     target_include_directories(onnxruntime_mlas_test PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc ${ONNXRUNTIME_ROOT}
             ${CMAKE_CURRENT_BINARY_DIR})
     target_link_libraries(onnxruntime_mlas_test PRIVATE GTest::gtest GTest::gmock ${ONNXRUNTIME_MLAS_LIBS} onnxruntime_common)
diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm_v2.cpp
similarity index 100%
rename from onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
rename to onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm_v2.cpp

From 412a5b4e62cd45f1fe735b41a919ac402319ba4d Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Sat, 20 Jun 2026 02:47:35 -0700
Subject: [PATCH 11/15] Debug 7

---
 cmake/onnxruntime_unittests.cmake             |   6 -
 .../mlas/lib/sqnbitgemm_kernel_avx512_2bit.h  |   4 -
 onnxruntime/test/mlas/unittest/test_main.cpp  |   8 -
 .../test/mlas/unittest/test_mlas_diag.cpp     | 302 ------------------
 ...m_v2.cpp => test_sqnbitgemm_2bit_gemm.cpp} |  70 ++--
 5 files changed, 26 insertions(+), 364 deletions(-)
 delete mode 100644 onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
 rename onnxruntime/test/mlas/unittest/{test_sqnbitgemm_2bit_gemm_v2.cpp => test_sqnbitgemm_2bit_gemm.cpp} (95%)

diff --git a/cmake/onnxruntime_unittests.cmake b/cmake/onnxruntime_unittests.cmake
index 0842ef992c624..23eccb22476df 100644
--- a/cmake/onnxruntime_unittests.cmake
+++ b/cmake/onnxruntime_unittests.cmake
@@ -1847,12 +1847,6 @@ endif()
       )
     endif()
     target_compile_definitions(onnxruntime_mlas_test PRIVATE ${mlas_private_compile_definitions})
-    # CI-CACHE-BUSTER: force rebuild of every TU in onnxruntime_mlas_test when
-    # a persistent /build directory would otherwise reuse stale .obj files.
-    # Bump the timestamp to invalidate. Safe to remove once the
-    # W2 scalar-test #if 0 wrapping is confirmed in CI.
-    target_compile_definitions(onnxruntime_mlas_test PRIVATE
-        MLAS_TEST_REBUILD_TAG=20260619T2310)
     target_include_directories(onnxruntime_mlas_test PRIVATE ${ONNXRUNTIME_ROOT}/core/mlas/inc ${ONNXRUNTIME_ROOT}
             ${CMAKE_CURRENT_BINARY_DIR})
     target_link_libraries(onnxruntime_mlas_test PRIVATE GTest::gtest GTest::gmock ${ONNXRUNTIME_MLAS_LIBS} onnxruntime_common)
diff --git a/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h b/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h
index 8a1927130750f..16e23b3993d57 100644
--- a/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h
+++ b/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx512_2bit.h
@@ -4,10 +4,6 @@ Copyright (c) Microsoft Corporation. All rights reserved.
 
 Licensed under the MIT License.
 
-// CI-CACHE-BUSTER 2026-06-19T22:50: force ccache/sccache miss after wrapping
-// W2 scalar reference tests in #if 0. Safe to delete once cache invalidation
-// is confirmed in the next CI run.
-
 Module Name:
 
     sqnbitgemm_kernel_avx512_2bit.h
diff --git a/onnxruntime/test/mlas/unittest/test_main.cpp b/onnxruntime/test/mlas/unittest/test_main.cpp
index edda089a6b75d..505c0c01dfa90 100644
--- a/onnxruntime/test/mlas/unittest/test_main.cpp
+++ b/onnxruntime/test/mlas/unittest/test_main.cpp
@@ -56,12 +56,6 @@ bool AddTestRegister(TestRegister test_register) {
   return true;
 }
 
-// Defined in test_mlas_diag.cpp. Forces that TU to be linked in even when
-// the linker would otherwise discard it (it has no externally-referenced
-// symbols of its own). The diagnostic environment is registered as a side
-// effect of this call.
-void RegisterMlasDiagEnvironment();
-
 int main(int argc, char** argv) {
   bool is_short_execute = (argc <= 1 || strcmp("--long", argv[1]) != 0);
   std::cout << "-------------------------------------------------------" << std::endl;
@@ -73,8 +67,6 @@ int main(int argc, char** argv) {
   std::cout << "----Total " << test_count << " tests registered programmably!" << std::endl;
   std::cout << "-------------------------------------------------------" << std::endl;
 
-  RegisterMlasDiagEnvironment();
-
   ::testing::InitGoogleTest(&argc, argv);
 
   return RUN_ALL_TESTS();
diff --git a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp b/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
deleted file mode 100644
index 96e99022fed23..0000000000000
--- a/onnxruntime/test/mlas/unittest/test_mlas_diag.cpp
+++ /dev/null
@@ -1,302 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-//
-// MLAS diagnostic helper. Always-on (no env var) so it works in CI
-// environments where setting environment variables is not possible.
-//
-// At process startup (before any test runs) this dumps:
-//   * Build-time ISA macros (__AVX2__, __AVX512F__, ...)
-//   * Runtime MlasPlatform feature flags (Avx2Supported_, Avx512Supported_)
-//   * Active QNBitGemmDispatch table identity (Avx2 / Avx2vnni / Avx512 /
-//     Avx512vnni / other)
-//   * On Linux, raw CPUID leaves 1 and 7 sub-leaf 0 so we can confirm what
-//     the kernel/hypervisor actually exposes vs. what MlasPlatform decided.
-//
-// When this file is linked into the test binary it also installs a SIGILL
-// handler (Linux only) that prints the faulting RIP, the module/symbol
-// nearest to it (dladdr), and the first 16 bytes at RIP so we can
-// disassemble the offending instruction offline. The handler is
-// async-signal-safe (write(2) only). It re-raises SIGILL with the default
-// disposition afterwards so the process still terminates and gtest still
-// reports the failing test.
-//
-// All output is tagged with the "MLAS-DIAG:" prefix so it can be grep'd
-// out of CI logs.
-//
-
-#include <cstdarg>
-#include <cstdint>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-
-#include "gtest/gtest.h"
-
-#include "core/mlas/lib/mlasi.h"
-
-#if defined(__linux__)
-#include <unistd.h>
-#include <signal.h>
-#include <dlfcn.h>
-#endif
-
-namespace {
-
-constexpr const char* kTag = "MLAS-DIAG: ";
-
-// Print a formatted message to both stdout and stderr. CI infrastructures
-// often capture only one of the two; doing both maximizes the chance the
-// diagnostic shows up.
-void PrintBoth(const char* fmt, ...) {
-  va_list ap;
-  va_start(ap, fmt);
-  va_list ap2;
-  va_copy(ap2, ap);
-  std::vfprintf(stderr, fmt, ap);
-  std::vfprintf(stdout, fmt, ap2);
-  va_end(ap);
-  va_end(ap2);
-  std::fflush(stderr);
-  std::fflush(stdout);
-}
-
-void PrintBuildMacros() {
-  PrintBoth("%sbuild macros:", kTag);
-#if defined(__AVX__)
-  PrintBoth(" __AVX__");
-#endif
-#if defined(__AVX2__)
-  PrintBoth(" __AVX2__");
-#endif
-#if defined(__FMA__)
-  PrintBoth(" __FMA__");
-#endif
-#if defined(__AVX512F__)
-  PrintBoth(" __AVX512F__");
-#endif
-#if defined(__AVX512BW__)
-  PrintBoth(" __AVX512BW__");
-#endif
-#if defined(__AVX512DQ__)
-  PrintBoth(" __AVX512DQ__");
-#endif
-#if defined(__AVX512VL__)
-  PrintBoth(" __AVX512VL__");
-#endif
-#if defined(__AVX512VNNI__)
-  PrintBoth(" __AVX512VNNI__");
-#endif
-#if defined(__GNUC__)
-  PrintBoth(" __GNUC__=%d", __GNUC__);
-#endif
-#if defined(_MSC_VER)
-  PrintBoth(" _MSC_VER=%d", _MSC_VER);
-#endif
-  PrintBoth("\n");
-}
-
-void PrintMlasPlatform() {
-  const auto& p = GetMlasPlatform();
-  PrintBoth("%sMlasPlatform: Avx2Supported_=%d Avx512Supported_=%d\n",
-            kTag,
-            static_cast<int>(p.Avx2Supported_),
-            static_cast<int>(p.Avx512Supported_));
-
-  const void* dispatch = static_cast<const void*>(p.QNBitGemmDispatch);
-  const char* name = "unknown/null";
-#if defined(MLAS_TARGET_AMD64) || defined(MLAS_TARGET_IX86)
-  if (dispatch == &MlasSQNBitGemmDispatchAvx2) {
-    name = "Avx2";
-  } else if (dispatch == &MlasSQNBitGemmDispatchAvx2vnni) {
-    name = "Avx2vnni";
-  } else if (dispatch == &MlasSQNBitGemmDispatchAvx512) {
-    name = "Avx512";
-  } else if (dispatch == &MlasSQNBitGemmDispatchAvx512vnni) {
-    name = "Avx512vnni";
-  } else if (dispatch == nullptr) {
-    name = "null";
-  }
-#endif
-  PrintBoth("%sQNBitGemmDispatch=%p (%s)\n",
-            kTag, dispatch, name);
-}
-
-#if defined(__linux__) && (defined(__x86_64__) || defined(__i386__))
-
-void PrintRawCpuid() {
-  auto cpuid = [](uint32_t leaf, uint32_t subleaf,
-                  uint32_t& eax, uint32_t& ebx, uint32_t& ecx, uint32_t& edx) {
-    __asm__ volatile(
-        "cpuid"
-        : "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx)
-        : "a"(leaf), "c"(subleaf));
-  };
-  uint32_t a, b, c, d;
-  cpuid(1, 0, a, b, c, d);
-  PrintBoth("%sCPUID(1,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
-            kTag, a, b, c, d);
-  cpuid(7, 0, a, b, c, d);
-  PrintBoth("%sCPUID(7,0): eax=%08x ebx=%08x ecx=%08x edx=%08x\n",
-            kTag, a, b, c, d);
-}
-
-// Async-signal-safe writer: integer -> hex into a small buffer.
-size_t HexWrite(uintptr_t v, char* out, size_t out_len) {
-  static const char kHex[] = "0123456789abcdef";
-  // Reserve room for "0x" + up to 16 hex digits + null.
-  char tmp[20];
-  size_t n = 0;
-  if (v == 0) {
-    tmp[n++] = '0';
-  } else {
-    while (v != 0 && n < sizeof(tmp)) {
-      tmp[n++] = kHex[v & 0xF];
-      v >>= 4;
-    }
-  }
-  // Now reverse into out, prefixed with "0x".
-  size_t w = 0;
-  if (out_len > 0) out[w++] = '0';
-  if (out_len > 1) out[w++] = 'x';
-  while (n > 0 && w < out_len) {
-    out[w++] = tmp[--n];
-  }
-  return w;
-}
-
-void SafeWriteStr(const char* s) {
-  if (s == nullptr) return;
-  size_t n = 0;
-  while (s[n] != '\0') ++n;
-  (void)write(STDERR_FILENO, s, n);
-  (void)write(STDOUT_FILENO, s, n);
-}
-
-void SafeWriteHex(uintptr_t v) {
-  char buf[20];
-  size_t n = HexWrite(v, buf, sizeof(buf));
-  (void)write(STDERR_FILENO, buf, n);
-  (void)write(STDOUT_FILENO, buf, n);
-}
-
-void SigillHandler(int sig, siginfo_t* info, void* /*ucontext*/) {
-  SafeWriteStr(kTag);
-  SafeWriteStr("SIGILL caught. si_addr=");
-  uintptr_t addr = reinterpret_cast<uintptr_t>(info ? info->si_addr : nullptr);
-  SafeWriteHex(addr);
-  SafeWriteStr("\n");
-
-  // Try to resolve module + nearest symbol. dladdr is not strictly
-  // async-signal-safe, but in practice on glibc it is safe enough for a
-  // best-effort crash dump; if it deadlocks we still get the si_addr above.
-  Dl_info dlinfo{};
-  if (info && dladdr(info->si_addr, &dlinfo) != 0) {
-    SafeWriteStr(kTag);
-    SafeWriteStr("dladdr: fname=");
-    SafeWriteStr(dlinfo.dli_fname ? dlinfo.dli_fname : "(null)");
-    SafeWriteStr(" fbase=");
-    SafeWriteHex(reinterpret_cast<uintptr_t>(dlinfo.dli_fbase));
-    SafeWriteStr(" sname=");
-    SafeWriteStr(dlinfo.dli_sname ? dlinfo.dli_sname : "(null)");
-    SafeWriteStr(" saddr=");
-    SafeWriteHex(reinterpret_cast<uintptr_t>(dlinfo.dli_saddr));
-    SafeWriteStr("\n");
-  }
-
-  // Dump first 16 bytes at si_addr so we can decode the instruction offline.
-  if (info && info->si_addr != nullptr) {
-    SafeWriteStr(kTag);
-    SafeWriteStr("bytes@si_addr:");
-    const unsigned char* p = static_cast<const unsigned char*>(info->si_addr);
-    for (int i = 0; i < 16; ++i) {
-      char two[4];
-      static const char kHex[] = "0123456789abcdef";
-      two[0] = ' ';
-      two[1] = kHex[(p[i] >> 4) & 0xF];
-      two[2] = kHex[p[i] & 0xF];
-      two[3] = '\0';
-      (void)write(STDERR_FILENO, two, 3);
-      (void)write(STDOUT_FILENO, two, 3);
-    }
-    SafeWriteStr("\n");
-  }
-
-  // Re-raise with default disposition so gtest still records the crash.
-  struct sigaction dfl{};
-  dfl.sa_handler = SIG_DFL;
-  sigemptyset(&dfl.sa_mask);
-  sigaction(sig, &dfl, nullptr);
-  raise(sig);
-}
-
-void InstallSigillHandler() {
-  static bool installed = false;
-  if (installed) return;
-  installed = true;
-  struct sigaction sa{};
-  sa.sa_sigaction = &SigillHandler;
-  sa.sa_flags = SA_SIGINFO | SA_RESETHAND;
-  sigemptyset(&sa.sa_mask);
-  if (sigaction(SIGILL, &sa, nullptr) != 0) {
-    std::fprintf(stderr, "%sfailed to install SIGILL handler\n", kTag);
-    std::fprintf(stdout, "%sfailed to install SIGILL handler\n", kTag);
-  } else {
-    std::fprintf(stderr, "%sSIGILL handler installed\n", kTag);
-    std::fprintf(stdout, "%sSIGILL handler installed\n", kTag);
-  }
-  std::fflush(stderr);
-  std::fflush(stdout);
-}
-
-#else  // !linux x86
-
-void PrintRawCpuid() {}
-void InstallSigillHandler() {}
-
-#endif
-
-class MlasDiagEnvironment : public ::testing::Environment {
- public:
-  void SetUp() override {
-    PrintBoth("%s---- MLAS diagnostic dump ----\n", kTag);
-    PrintBuildMacros();
-    PrintRawCpuid();
-    PrintMlasPlatform();
-    InstallSigillHandler();
-    PrintBoth("%s---- end MLAS diagnostic dump ----\n", kTag);
-    std::fflush(stderr);
-    std::fflush(stdout);
-  }
-};
-
-}  // namespace
-
-// Externally-visible entry point so test_main.cpp can force this TU to link
-// and register the diagnostic GTest environment. Without an externally
-// referenced symbol, the linker may discard the anonymous-namespace static
-// initializer that registers the environment.
-void RegisterMlasDiagEnvironment() {
-  // Install the SIGILL handler as soon as we are called -- before GTest
-  // even sets up the environment. This way a fault during static init or
-  // very early test setup is still captured.
-  InstallSigillHandler();
-
-  // Also print the basic platform dump immediately so CI logs have it
-  // even if a crash kills us before the Environment::SetUp() runs.
-  PrintBoth("%s---- MLAS early diagnostic dump ----\n", kTag);
-  PrintBuildMacros();
-  PrintRawCpuid();
-  PrintMlasPlatform();
-  PrintBoth("%s---- end MLAS early diagnostic dump ----\n", kTag);
-  std::fflush(stderr);
-  std::fflush(stdout);
-
-  // ::testing::AddGlobalTestEnvironment takes ownership of the pointer.
-  static bool registered = false;
-  if (registered) {
-    return;
-  }
-  registered = true;
-  ::testing::AddGlobalTestEnvironment(new MlasDiagEnvironment());
-}
diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm_v2.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
similarity index 95%
rename from onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm_v2.cpp
rename to onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 10e751367b152..049693eca6671 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm_v2.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -4,10 +4,6 @@ Copyright (c) Microsoft Corporation. All rights reserved.
 
 Licensed under the MIT License.
 
-// CI-CACHE-BUSTER 2026-06-19T22:50: pair with sqnbitgemm_kernel_avx512_2bit.h
-// to force ccache/sccache invalidation after wrapping W2 scalar reference
-// tests in #if 0.
-
 Module Name:
 
     test_sqnbitgemm_2bit.cpp
@@ -306,45 +302,26 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 
 }  // namespace
 
-// =============================================================================
-// W2 scalar reference tests -- temporarily disabled via #if 0.
 //
-// These tests call internal scalar W2 symbols directly, bypassing
-// MlasIsQNBitGemmAvailable. The scalar reference TU
+// W2 scalar reference tests. These call internal scalar W2 symbols directly,
+// bypassing MlasIsQNBitGemmAvailable. The scalar reference TU
 // (sqnbitgemm_kernel_avx512_2bit.cpp) is compiled with AVX-512-capable
-// flags and the build uses LTO, so the compiler may autovectorize these
-// "scalar" functions into AVX-512 instructions. On some virtualized Linux
-// hosts (e.g. Azure Standard_D32ads_v6 / Zen 4 CI image) certain AVX-512
-// VNNI encodings still SIGILL even when CPUID reports support, killing the
-// whole onnxruntime_mlas_test binary and blocking all downstream test
-// suites including the production-path MatMul2Bits op tests in
-// onnxruntime_test_all.
-//
-// Note: GTest's DISABLED_ prefix did not skip these in CI in practice
-// (binary still showed [ RUN ] for the original name even after rename),
-// so we wrap them out of the binary completely via #if 0.
-//
-// Production paths are unaffected:
-//   * SQ2BitGemmKernel_BlkSum_CompInt8_Scalar is not installed in any
-//     dispatch table -- it is only used as an oracle in these tests.
-//   * SQ2BitGemmPackQuantBDataAndBlkSum_Scalar runs at model load time
-//     under the AVX-512 dispatch path, which is only selected on hosts
-//     where it has been observed to be safe in practice.
-//
-// TODO(W2-scalar-disabled): re-enable once the W2 scalar reference TU
-// codegen is pinned away from AVX-512 autovectorization, e.g. by adding
-// __attribute__((target("..."))) on the scalar entry points or by
-// excluding the TU from LTO.
-// =============================================================================
-
-#if 0
+// flags so its SIMD siblings can use intrinsics; the compiler may
+// autovectorize the scalar functions into AVX-512 instructions. To avoid
+// SIGILL on hosts that do not expose AVX-512 these tests are gated on
+// GetMlasPlatform().Avx512Supported_, matching the SIMD test guards in
+// this file.
 //
 // Scalar block-group test, no zero-points. Covers the same small synthetic
 // shapes + representative prefill sizes used by the production W2 tests. All
 // shapes have K as a multiple of (kBlkLen * kBlockGroupBlks) = 256. K=384 is
 // NOT a multiple of 256 so it's excluded; that shape will need a tail
 // handler in a follow-up.
+//
 TEST(MlasSq2BitTest, Scalar_BlkLen64) {
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
+  }
   struct Shape {
     size_t M, N, K;
   };
@@ -385,8 +362,10 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64) {
 //
 // Same coverage with per-block non-default zero points.
 //
-// See DISABLED_Scalar_BlkLen64 comment.
 TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
+  }
   struct Shape {
     size_t M, N, K;
   };
@@ -422,7 +401,6 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
     }
   }
 }
-#endif  // W2 scalar reference tests (BlkLen64 pair)
 
 //
 // SIMD block-group shape coverage. Phase-3+K-tail kernel requires:
@@ -842,9 +820,10 @@ constexpr struct {
 
 }  // namespace
 
-#if 0  // See top-of-file W2 scalar reference tests note.
-// See top-of-file note for why these are disabled.
 TEST(MlasSq2BitTest, Scalar_BlkLen128) {
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -857,8 +836,10 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128) {
   }
 }
 
-// See top-of-file note for why these are disabled.
 TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
       for (bool bias : {false, true}) {
@@ -870,7 +851,6 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
     }
   }
 }
-#endif  // W2 scalar reference tests (BlkLen128 pair)
 
 TEST(MlasSq2BitTest, BlkLen128_Avx512) {
   if (!GetMlasPlatform().Avx512Supported_) {
@@ -1200,9 +1180,10 @@ constexpr struct {
 
 }  // namespace
 
-#if 0  // See top-of-file W2 scalar reference tests note.
-// See top-of-file note for why these are disabled.
 TEST(MlasSq2BitTest, Scalar_BlkLen32) {
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1215,8 +1196,10 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32) {
   }
 }
 
-// See top-of-file note for why these are disabled.
 TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
+  }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
       for (bool bias : {false, true}) {
@@ -1228,7 +1211,6 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
     }
   }
 }
-#endif  // W2 scalar reference tests (BlkLen32 pair)
 
 TEST(MlasSq2BitTest, BlkLen32_Avx512) {
   if (!GetMlasPlatform().Avx512Supported_) {

From ce168d26aeedbafa9010ab3fc6b1fa1d069c49ac Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Sat, 20 Jun 2026 02:53:25 -0700
Subject: [PATCH 12/15] Debug 8

---
 cmake/onnxruntime_mlas.cmake                  | 19 +------------------
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    |  9 ---------
 2 files changed, 1 insertion(+), 27 deletions(-)

diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
index acb5988956611..4058ee8cf8f0f 100644
--- a/cmake/onnxruntime_mlas.cmake
+++ b/cmake/onnxruntime_mlas.cmake
@@ -842,30 +842,14 @@ endif()
           ${MLAS_SRC_DIR}/x86_64/QgemmU8X8KernelAvx512Core.S
           ${MLAS_SRC_DIR}/x86_64/ConvSymKernelAvx512Core.S
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512.cpp
-          # NOTE: sqnbitgemm_kernel_avx512_2bit_blklen{32,64,128}.h are also AVX-512
-          # only, but they are header files included by sqnbitgemm_kernel_avx512.cpp
-          # and sqnbitgemm_kernel_avx512vnni.cpp, so they pick up those TUs'
-          # AVX-512 compile flags automatically. Listing them here just adds
-          # them to IDE project files for navigation.
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.h
+          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen64.h
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen128.h
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit_blklen32.h
         )
         set_source_files_properties(${mlas_platform_srcs_avx512core} PROPERTIES COMPILE_FLAGS "-mfma -mavx512vnni -mavx512bw -mavx512dq -mavx512vl")
 
-        # sqnbitgemm_kernel_avx512_2bit.cpp is named "avx512" but is intentionally
-        # pure scalar C++ -- it implements the W2 pack helper and a scalar
-        # reference oracle (see file header comment). It must NOT be compiled
-        # with AVX-512 flags, otherwise the compiler may autovectorize it to
-        # AVX-512 (e.g. VNNI vpdpbusd) instructions that some virtualized
-        # hosts do not accept, producing SIGILL when tests call those symbols
-        # directly.
-        set(mlas_platform_srcs_w2_scalar
-          ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp
-        )
-        set_source_files_properties(${mlas_platform_srcs_w2_scalar} PROPERTIES COMPILE_FLAGS "-mavx2 -mfma -mf16c")
-
         set(mlas_platform_srcs_avx512vnni
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512vnni.cpp
           ${MLAS_SRC_DIR}/qkv_quant_kernel_avx512vnni.cpp
@@ -883,7 +867,6 @@ endif()
           ${mlas_platform_srcs_avx2}
           ${mlas_platform_srcs_avx512f}
           ${mlas_platform_srcs_avx512core}
-          ${mlas_platform_srcs_w2_scalar}
           ${mlas_platform_srcs_avx512vnni}
         )
 
diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 049693eca6671..9e5a83d4fad99 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -302,15 +302,6 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 
 }  // namespace
 
-//
-// W2 scalar reference tests. These call internal scalar W2 symbols directly,
-// bypassing MlasIsQNBitGemmAvailable. The scalar reference TU
-// (sqnbitgemm_kernel_avx512_2bit.cpp) is compiled with AVX-512-capable
-// flags so its SIMD siblings can use intrinsics; the compiler may
-// autovectorize the scalar functions into AVX-512 instructions. To avoid
-// SIGILL on hosts that do not expose AVX-512 these tests are gated on
-// GetMlasPlatform().Avx512Supported_, matching the SIMD test guards in
-// this file.
 //
 // Scalar block-group test, no zero-points. Covers the same small synthetic
 // shapes + representative prefill sizes used by the production W2 tests. All

From 492a5f71e4620ae41976f57c814ab2b86681b6c5 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Sun, 21 Jun 2026 00:04:05 -0700
Subject: [PATCH 13/15] PR feedback

---
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    | 34 ++++++++++++-------
 1 file changed, 22 insertions(+), 12 deletions(-)

diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 9e5a83d4fad99..0da5790c417c2 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -310,8 +310,13 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 // handler in a follow-up.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64) {
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "AVX-512 not available on this host";
+  // The scalar oracle TU is compiled with -mavx512vnni so the compiler may
+  // autovectorize the int8 dot-product loop to vpdpbusd. Require the
+  // AVX-512-VNNI dispatch to be active so we only run on hosts that accept
+  // those encodings.
+  if (!GetMlasPlatform().Avx512Supported_ ||
+      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
+    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
   }
   struct Shape {
     size_t M, N, K;
@@ -354,8 +359,9 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64) {
 // Same coverage with per-block non-default zero points.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "AVX-512 not available on this host";
+  if (!GetMlasPlatform().Avx512Supported_ ||
+      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
+    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
   }
   struct Shape {
     size_t M, N, K;
@@ -812,8 +818,9 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128) {
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "AVX-512 not available on this host";
+  if (!GetMlasPlatform().Avx512Supported_ ||
+      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
+    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
@@ -828,8 +835,9 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "AVX-512 not available on this host";
+  if (!GetMlasPlatform().Avx512Supported_ ||
+      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
+    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
@@ -1172,8 +1180,9 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32) {
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "AVX-512 not available on this host";
+  if (!GetMlasPlatform().Avx512Supported_ ||
+      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
+    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
@@ -1188,8 +1197,9 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
-  if (!GetMlasPlatform().Avx512Supported_) {
-    GTEST_SKIP() << "AVX-512 not available on this host";
+  if (!GetMlasPlatform().Avx512Supported_ ||
+      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
+    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {

From 5bf8ca76965f80e4258b6063c736b31fe6546d02 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Sun, 21 Jun 2026 00:57:28 -0700
Subject: [PATCH 14/15] Re-work PR feedback

---
 cmake/onnxruntime_mlas.cmake                  |  8 +++++
 .../unittest/test_sqnbitgemm_2bit_gemm.cpp    | 34 +++++++------------
 2 files changed, 20 insertions(+), 22 deletions(-)

diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
index 4058ee8cf8f0f..58f39f072895e 100644
--- a/cmake/onnxruntime_mlas.cmake
+++ b/cmake/onnxruntime_mlas.cmake
@@ -850,6 +850,14 @@ endif()
         )
         set_source_files_properties(${mlas_platform_srcs_avx512core} PROPERTIES COMPILE_FLAGS "-mfma -mavx512vnni -mavx512bw -mavx512dq -mavx512vl")
 
+        # Strip -mavx512vnni from the W2 scalar oracle / pack-helper TU so the
+        # compiler cannot autovectorize its int8 dot-product loops to vpdpbusd.
+        # Needed because this helper runs at model load on AVX-512-only
+        # (non-VNNI) hosts via the AVX-512 W2 dispatch. TU is pure C++ -- no
+        # AVX-512 intrinsics inside.
+        set_source_files_properties(${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp PROPERTIES
+          COMPILE_FLAGS "-mfma -mavx512bw -mavx512dq -mavx512vl -mavx512f")
+
         set(mlas_platform_srcs_avx512vnni
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512vnni.cpp
           ${MLAS_SRC_DIR}/qkv_quant_kernel_avx512vnni.cpp
diff --git a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
index 0da5790c417c2..9e5a83d4fad99 100644
--- a/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
+++ b/onnxruntime/test/mlas/unittest/test_sqnbitgemm_2bit_gemm.cpp
@@ -310,13 +310,8 @@ void RunW2Case(size_t M, size_t N, size_t K, bool WithBias, uint32_t seed,
 // handler in a follow-up.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64) {
-  // The scalar oracle TU is compiled with -mavx512vnni so the compiler may
-  // autovectorize the int8 dot-product loop to vpdpbusd. Require the
-  // AVX-512-VNNI dispatch to be active so we only run on hosts that accept
-  // those encodings.
-  if (!GetMlasPlatform().Avx512Supported_ ||
-      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
-    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
   }
   struct Shape {
     size_t M, N, K;
@@ -359,9 +354,8 @@ TEST(MlasSq2BitTest, Scalar_BlkLen64) {
 // Same coverage with per-block non-default zero points.
 //
 TEST(MlasSq2BitTest, Scalar_BlkLen64_WithZeroPoints) {
-  if (!GetMlasPlatform().Avx512Supported_ ||
-      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
-    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
   }
   struct Shape {
     size_t M, N, K;
@@ -818,9 +812,8 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128) {
-  if (!GetMlasPlatform().Avx512Supported_ ||
-      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
-    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
@@ -835,9 +828,8 @@ TEST(MlasSq2BitTest, Scalar_BlkLen128) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen128_WithZeroPoints) {
-  if (!GetMlasPlatform().Avx512Supported_ ||
-      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
-    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen128) {
@@ -1180,9 +1172,8 @@ constexpr struct {
 }  // namespace
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32) {
-  if (!GetMlasPlatform().Avx512Supported_ ||
-      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
-    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {
@@ -1197,9 +1188,8 @@ TEST(MlasSq2BitTest, Scalar_BlkLen32) {
 }
 
 TEST(MlasSq2BitTest, Scalar_BlkLen32_WithZeroPoints) {
-  if (!GetMlasPlatform().Avx512Supported_ ||
-      GetMlasPlatform().QNBitGemmDispatch != &MlasSQNBitGemmDispatchAvx512vnni) {
-    GTEST_SKIP() << "AVX-512-VNNI dispatch not active on this host";
+  if (!GetMlasPlatform().Avx512Supported_) {
+    GTEST_SKIP() << "AVX-512 not available on this host";
   }
   for (uint32_t seed : {0xC0FFEEu, 0xBADC0DEu}) {
     for (const auto& s : kSimdShapes_BlkLen32) {

From 17d323c375a4b1adee9412a480c46f4667ea1a07 Mon Sep 17 00:00:00 2001
From: Hari Seshadri <shariharan91@gmail.com>
Date: Mon, 22 Jun 2026 10:24:30 -0700
Subject: [PATCH 15/15] PR feedback 2

---
 cmake/onnxruntime_mlas.cmake | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
index 58f39f072895e..b8ab7142b6b35 100644
--- a/cmake/onnxruntime_mlas.cmake
+++ b/cmake/onnxruntime_mlas.cmake
@@ -856,7 +856,7 @@ endif()
         # (non-VNNI) hosts via the AVX-512 W2 dispatch. TU is pure C++ -- no
         # AVX-512 intrinsics inside.
         set_source_files_properties(${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512_2bit.cpp PROPERTIES
-          COMPILE_FLAGS "-mfma -mavx512bw -mavx512dq -mavx512vl -mavx512f")
+          COMPILE_FLAGS "-mfma -mavx512bw -mavx512dq -mavx512vl")
 
         set(mlas_platform_srcs_avx512vnni
           ${MLAS_SRC_DIR}/sqnbitgemm_kernel_avx512vnni.cpp