[6/n] Migrate activation kernels, gptq, gguf, non cutlass w8a8 to libtorch stable ABI (continued) (#42663)

Signed-off-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Signed-off-by: Chris Leonard <chleonar@redhat.com>
Co-authored-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
Chris Leonard
2026-05-20 00:18:12 -07:00
committed by GitHub
co-authored by Mikayla Gawarecki Shengqi Chen
parent 40651c0207
commit 07aeaf9d4d
28 changed files with 810 additions and 639 deletions
+16 -14
View File
@@ -7,23 +7,23 @@
*/
#include <cmath>
#include <torch/types.h>
#include <torch/headeronly/macros/Macros.h>
#ifndef USE_ROCM
#include <c10/util/Float8_e4m3fn.h>
#include <torch/headeronly/util/Float8_e4m3fn.h>
#define MAYBE_HOST_DEVICE C10_HOST_DEVICE
#else
#include <ATen/hip/HIPContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <c10/util/Float8_e4m3fnuz.h>
#include <torch/headeronly/util/Float8_e4m3fn.h>
#include <torch/headeronly/util/Float8_e4m3fnuz.h>
// ROCm doesn't seem to need C10_HOST_DEVICE for static constexpr
#define MAYBE_HOST_DEVICE
#endif
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, c10::Float8_e4m3fn> ||
std::is_same_v<T, c10::Float8_e4m3fnuz> ||
std::is_same_v<T, int8_t>>>
typename = std::enable_if_t<
std::is_same_v<T, torch::headeronly::Float8_e4m3fn> ||
std::is_same_v<T, torch::headeronly::Float8_e4m3fnuz> ||
std::is_same_v<T, int8_t>>>
struct quant_type_max {
static constexpr T val() { return std::numeric_limits<T>::max(); }
};
@@ -31,9 +31,10 @@ struct quant_type_max {
// Using the default max value from pytorch (240.0 0x7F) will cause accuracy
// issues when running dynamic quantization. Here use 224.0 0x7E for rocm.
template <>
struct quant_type_max<c10::Float8_e4m3fnuz> {
static constexpr c10::Float8_e4m3fnuz val() {
return c10::Float8_e4m3fnuz(0x7E, c10::Float8_e4m3fnuz::from_bits());
struct quant_type_max<torch::headeronly::Float8_e4m3fnuz> {
static constexpr torch::headeronly::Float8_e4m3fnuz val() {
return torch::headeronly::Float8_e4m3fnuz(
0x7E, torch::headeronly::Float8_e4m3fnuz::from_bits());
}
};
@@ -42,9 +43,10 @@ MAYBE_HOST_DEVICE static constexpr T quant_type_max_v =
quant_type_max<T>::val();
template <typename T,
typename = std::enable_if_t<std::is_same_v<T, c10::Float8_e4m3fn> ||
std::is_same_v<T, c10::Float8_e4m3fnuz> ||
std::is_same_v<T, int8_t>>>
typename = std::enable_if_t<
std::is_same_v<T, torch::headeronly::Float8_e4m3fn> ||
std::is_same_v<T, torch::headeronly::Float8_e4m3fnuz> ||
std::is_same_v<T, int8_t>>>
struct min_scaling_factor {
C10_DEVICE C10_ALWAYS_INLINE static float val() {
return 1.0f / (quant_type_max_v<T> * 512.0f);