From 8647bda84ab5d40570588c4a39845a32a0849ee4 Mon Sep 17 00:00:00 2001
From: kkkzbh <145212963+kkkzbh@users.noreply.github.com>
Date: Sat, 13 Jun 2026 19:30:33 +0800
Subject: [PATCH 1/2] fix(cpu): preserve raw fp16 bits in x86 quantize

---
 .../kernels/common/ggml/quantize/quantize.hpp | 25 ++++++++++++++++---
 1 file changed, 22 insertions(+), 3 deletions(-)

diff --git a/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp b/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
index e318451a1..ba315dba0 100644
--- a/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
+++ b/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
@@ -30,6 +30,7 @@
 #include <cassert>
 #include <cmath>
 #include <cstring>
+#include <type_traits>
 
 #include "mllm/core/DataTypes.hpp"
 
@@ -105,8 +106,26 @@ inline static float lookup_fp16_to_fp32(uint16_t f) {
 
 #else
 namespace mllm::cpu {
-#define MLLM_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define MLLM_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+template<typename T>
+inline static uint16_t mllm_fp16_bits(const T& f) {
+  if constexpr (std::is_integral_v<std::decay_t<T>>) {
+    return static_cast<uint16_t>(f);
+  } else {
+    static_assert(sizeof(T) == sizeof(uint16_t), "fp16 type must be 16 bits");
+    uint16_t s;
+    memcpy(&s, &f, sizeof(s));
+    return s;
+  }
+}
+
+inline static mllm_fp16_t mllm_fp16_from_bits(uint16_t bits) {
+  mllm_fp16_t f;
+  memcpy(&f, &bits, sizeof(bits));
+  return f;
+}
+
+#define MLLM_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(mllm_fp16_bits(x))
+#define MLLM_COMPUTE_FP32_TO_FP16(x) mllm_fp16_from_bits(_cvtss_sh(x, 0))
 
 static float table_f32_f16[1 << 16];
 static bool table_f32_f16_init = false;
@@ -127,7 +146,7 @@ inline static float lookup_fp16_to_fp32(uint16_t f) {
 }
 
 #ifndef MLLM_FP16_TO_FP32
-#define MLLM_FP16_TO_FP32(x) lookup_fp16_to_fp32(x)
+#define MLLM_FP16_TO_FP32(x) lookup_fp16_to_fp32(mllm_fp16_bits(x))
 #endif
 #ifndef MLLM_FP32_TO_FP16
 #define MLLM_FP32_TO_FP16(x) MLLM_COMPUTE_FP32_TO_FP16(x)

From 7011227bef4c38dbdd121c5e94863f5a7c0e34fc Mon Sep 17 00:00:00 2001
From: kkkzbh <145212963+kkkzbh@users.noreply.github.com>
Date: Sat, 13 Jun 2026 19:52:04 +0800
Subject: [PATCH 2/2] docs(cpu): explain x86 fp16 bit helpers

---
 mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp b/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
index ba315dba0..48e9540ec 100644
--- a/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
+++ b/mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
@@ -106,6 +106,8 @@ inline static float lookup_fp16_to_fp32(uint16_t f) {
 
 #else
 namespace mllm::cpu {
+// Extract a raw 16-bit fp16 bit pattern without numeric conversion. Integral inputs
+// are already bit patterns; non-integral fp16-like values are copied byte-for-byte.
 template<typename T>
 inline static uint16_t mllm_fp16_bits(const T& f) {
   if constexpr (std::is_integral_v<std::decay_t<T>>) {
@@ -118,6 +120,7 @@ inline static uint16_t mllm_fp16_bits(const T& f) {
   }
 }
 
+// Construct an mllm_fp16_t value from a raw 16-bit fp16 bit pattern.
 inline static mllm_fp16_t mllm_fp16_from_bits(uint16_t bits) {
   mllm_fp16_t f;
   memcpy(&f, &bits, sizeof(bits));