[Fix] Add missing stubs from cpu fp8 attention changes (#41387)

Signed-off-by: Li, Tianmu <tianmu.li@intel.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
Tianmu Li
2026-05-06 12:16:27 +08:00
committed by GitHub
co-authored by Li, Jiang <jiang1.li@intel.com>
parent aee190ac37
commit e47c98ef7a
4 changed files with 25 additions and 0 deletions
+4
View File
@@ -486,6 +486,10 @@ struct FP32Vec16 : public VectorizedRegWrapper<FP32Vec16, 4, float> {
explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
// FP8 stub: dead code on ARM (fp8 KV cache is x86-only), needed for
// load_b_pair_vec template to compile on all platforms.
explicit FP32Vec16(const BF16Vec32&, int) : Base() {}
explicit FP32Vec16(const FP16Vec16& v) {
reg.val[0] = Vectorized<float>(vcvt_f32_f16(vget_low_f16(v.reg.val[0])));
reg.val[1] = Vectorized<float>(vcvt_f32_f16(vget_high_f16(v.reg.val[0])));
+10
View File
@@ -6,6 +6,9 @@
namespace vec_op {
struct fp8_e4m3_tag {};
struct fp8_e5m2_tag {};
#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...) \
AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
@@ -145,6 +148,9 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
}
void save(void* ptr) const { *reinterpret_cast<f16x32_t*>(ptr) = reg; }
explicit BF16Vec32(const uint8_t*, fp8_e4m3_tag) : reg{} {}
explicit BF16Vec32(const uint8_t*, fp8_e5m2_tag) : reg{} {}
};
struct FP32Vec4 : public Vec<FP32Vec4> {
@@ -302,6 +308,10 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
// FP8 stub: dead code on scalar path (fp8 KV cache is x86-only), needed for
// load_b_pair_vec template to compile on all platforms.
explicit FP32Vec16(const BF16Vec32&, int) : reg{} {}
FP32Vec16 operator*(const FP32Vec16& b) const {
f32x16_t ret;
unroll_loop<int, VEC_ELEM_NUM>(
+7
View File
@@ -146,6 +146,9 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
: reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {}
void save(void* ptr) const { *reinterpret_cast<ss16x8x4_t*>(ptr) = reg; }
explicit BF16Vec32(const uint8_t*, fp8_e4m3_tag) : reg{} {}
explicit BF16Vec32(const uint8_t*, fp8_e5m2_tag) : reg{} {}
};
struct FP32Vec4 : public Vec<FP32Vec4> {
@@ -408,6 +411,10 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
// FP8 stub: dead code on PowerPC (fp8 KV cache is x86-only), needed for
// load_b_pair_vec template to compile on all platforms.
explicit FP32Vec16(const BF16Vec32&, int) : reg{} {}
explicit FP32Vec16(const INT32Vec16& v) {
reg.val[0] = vec_ctf(v.reg.val[0], 0);
reg.val[1] = vec_ctf(v.reg.val[1], 0);
+4
View File
@@ -688,6 +688,10 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
// FP8 stub: dead code on s390x (fp8 KV cache is x86-only), needed for
// load_b_pair_vec template to compile on all platforms.
explicit FP32Vec16(const BF16Vec32&, int) : reg{} {}
FP32Vec16 operator*(const FP32Vec16& b) const {
return FP32Vec16(f32x4x4_t({vec_mul(reg.val[0], b.reg.val[0]),
vec_mul(reg.val[1], b.reg.val[1]),