[CPU] add faster KV-cache writes (#25874)

This commit is contained in:
Ma Mingfei
2026-05-25 10:28:52 +08:00
committed by GitHub
parent e1463bb2c2
commit 821d5f4a5b
6 changed files with 249 additions and 1 deletions
+9
View File
@@ -1,6 +1,7 @@
#pragma once
#include <ATen/ATen.h>
#include <ATen/Dispatch.h>
#include <ATen/Parallel.h>
#if defined(_OPENMP)
@@ -44,6 +45,14 @@ namespace {
} \
}()
// Half + BFloat16, plus one extra scalar type
#define AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES_AND(SCALARTYPE, ...) \
AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES(__VA_ARGS__) \
AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
#define AT_DISPATCH_REDUCED_FLOATING_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \
AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES_AND(SCALARTYPE, __VA_ARGS__))
// dispatch: bfloat16, float16, int8_t, fp8_e4m3, uint8_t(mxfp4/int4)
#define CPU_DISPATCH_PACKED_TYPES(TYPE, ...) \
[&] { \