Skip to content

fix(cpu): preserve raw fp16 bits in x86 quantize conversion #685

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kkkzbh wants to merge 3 commits into
base: main
Choose a base branch
from
Open

fix(cpu): preserve raw fp16 bits in x86 quantize conversion #685

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8647bda
fix(cpu): preserve raw fp16 bits in x86 quantize
kkkzbh Jun 13, 2026
7011227
docs(cpu): explain x86 fp16 bit helpers
kkkzbh Jun 13, 2026
bb82cb7
Merge branch 'main' into codex/fix-x86-fp16-quantize
chenghuaWang Jun 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 22 additions & 3 deletions mllm/backends/cpu/kernels/common/ggml/quantize/quantize.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
#include <cassert>
#include <cmath>
#include <cstring>
#include <type_traits>

#include "mllm/core/DataTypes.hpp"

Expand Down Expand Up @@ -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;
}
Comment thread
coderabbitai[bot] marked this conversation as resolved.

#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;
Expand All @@ -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)
Expand Down