Skip to content

Introduce 6-bit quantization for Llama in torchchat #1007

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

Merged
merged 1 commit into from
Oct 3, 2024
Merged

Introduce 6-bit quantization for Llama in torchchat #1007

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
107 changes: 101 additions & 6 deletions torchao/experimental/kernels/cpu/aarch64/benchmarks/benchmark_bitpacking.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,7 @@
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint2.h>
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint3.h>
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint4.h>
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint6.h>
#include <torchao/experimental/kernels/cpu/aarch64/tests/test_utils.h>
#include <cassert>

Expand All @@ -36,7 +37,7 @@ TORCHAO_ALWAYS_INLINE void pack_uint_odd_bit_values(
int variant) {
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(packed_size % variant == 0);
assert(unpacked_size % variant == 0);

uint8x16_t unpacked0;
uint8x16_t unpacked1;
Expand Down Expand Up @@ -103,7 +104,7 @@ TORCHAO_ALWAYS_INLINE void unpack_uint_odd_bit_values(
int variant) {
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(packed_size % variant == 0);
assert(unpacked_size % variant == 0);

uint8x16_t unpacked0;
uint8x16_t unpacked1;
Expand Down Expand Up @@ -222,7 +223,7 @@ void pack_uint_values<2>(
constexpr int nbit = 2;
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(packed_size % variant == 0);
assert(unpacked_size % variant == 0);

uint8x8_t unpacked0_8x8;
uint8x8_t unpacked1_8x8;
Expand Down Expand Up @@ -287,7 +288,7 @@ void unpack_uint_values<2>(
constexpr int nbit = 2;
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(packed_size % variant == 0);
assert(unpacked_size % variant == 0);

uint8x8_t unpacked0_8x8;
uint8x8_t unpacked1_8x8;
Expand Down Expand Up @@ -394,7 +395,7 @@ void pack_uint_values<4>(
constexpr int nbit = 4;
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(packed_size % variant == 0);
assert(unpacked_size % variant == 0);

uint8x16_t unpacked0;
uint8x16_t unpacked1;
Expand Down Expand Up @@ -435,7 +436,7 @@ void unpack_uint_values<4>(
constexpr int nbit = 4;
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(packed_size % variant == 0);
assert(unpacked_size % variant == 0);

uint8x16_t unpacked0;
uint8x16_t unpacked1;
Expand Down Expand Up @@ -507,6 +508,98 @@ void unpack_uint_values<5>(
variant);
}

// Benchmark utility to compare variants of uint6 packing
template <>
void pack_uint_values<6>(
uint8_t* packed,
uint8_t* unpacked,
int packed_size,
int unpacked_size,
int variant) {
constexpr int nbit = 6;
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(unpacked_size % variant == 0);

uint8x16_t unpacked0;
uint8x16_t unpacked1;
uint8x16_t unpacked2;
uint8x16_t unpacked3;

switch (variant) {
case 4:
for (int i = 0; i < unpacked_size; i += 4) {
torchao::bitpacking::internal::pack_4_uint6_values(
packed + ((i * nbit) / bitsPerByte), unpacked + i);
}
break;
case 32:
for (int i = 0; i < unpacked_size; i += 32) {
unpacked0 = vld1q_u8(unpacked + i);
unpacked1 = vld1q_u8(unpacked + 16 + i);
torchao::bitpacking::internal::vec_pack_32_uint6_values(
packed + ((i * nbit) / bitsPerByte), unpacked0, unpacked1);
}
break;
case 64:
for (int i = 0; i < unpacked_size; i += 64) {
unpacked0 = vld1q_u8(unpacked + i);
unpacked1 = vld1q_u8(unpacked + 16 + i);
unpacked2 = vld1q_u8(unpacked + 32 + i);
unpacked3 = vld1q_u8(unpacked + 48 + i);
torchao::bitpacking::internal::vec_pack_64_uint6_values(
packed + ((i * nbit) / bitsPerByte), unpacked0, unpacked1, unpacked2, unpacked3);
}
break;
}
}

// Benchmark utility to compare variants of uint6 unpacking
template <>
void unpack_uint_values<6>(
uint8_t* unpacked,
uint8_t* packed,
int unpacked_size,
int packed_size,
int variant) {
constexpr int nbit = 6;
constexpr int bitsPerByte = 8;
assert(unpacked_size * nbit / bitsPerByte == packed_size);
assert(unpacked_size % variant == 0);

uint8x16_t unpacked0;
uint8x16_t unpacked1;
uint8x16_t unpacked2;
uint8x16_t unpacked3;

switch (variant) {
case 4:
for (int i = 0; i < unpacked_size; i += 4) {
torchao::bitpacking::internal::unpack_4_uint6_values(
unpacked + i, packed + ((i * nbit) / bitsPerByte));
}
break;
case 32:
for (int i = 0; i < unpacked_size; i += 32) {
torchao::bitpacking::internal::vec_unpack_32_uint6_values(
unpacked0, unpacked1, packed + ((i * nbit) / bitsPerByte));
vst1q_u8(unpacked + i, unpacked0);
vst1q_u8(unpacked + 16 + i, unpacked1);
}
break;
case 64:
for (int i = 0; i < unpacked_size; i += 64) {
torchao::bitpacking::internal::vec_unpack_64_uint6_values(
unpacked0, unpacked1, unpacked2, unpacked3, packed + ((i * nbit) / bitsPerByte));
vst1q_u8(unpacked + i, unpacked0);
vst1q_u8(unpacked + 16 + i, unpacked1);
vst1q_u8(unpacked + 32 + i, unpacked2);
vst1q_u8(unpacked + 48 + i, unpacked3);
}
break;
}
}

} // namespace

template <int nbit>
Expand Down Expand Up @@ -557,6 +650,8 @@ BENCHMARK(benchmark_pack_uint_values<4>)->ArgsProduct({{128}, {2, 16, 32}});
BENCHMARK(benchmark_unpack_uint_values<4>)->ArgsProduct({{128}, {2, 16, 32}});
BENCHMARK(benchmark_pack_uint_values<5>)->ArgsProduct({{128}, {8, 64, 128}});
BENCHMARK(benchmark_unpack_uint_values<5>)->ArgsProduct({{128}, {8, 64, 128}});
BENCHMARK(benchmark_pack_uint_values<6>)->ArgsProduct({{128}, {8, 64, 128}});
BENCHMARK(benchmark_unpack_uint_values<6>)->ArgsProduct({{128}, {4, 32, 64}});

// Run the benchmark
BENCHMARK_MAIN();
Expand Down
6 changes: 6 additions & 0 deletions torchao/experimental/kernels/cpu/aarch64/benchmarks/benchmark_linear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -238,6 +238,8 @@ BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x1x32_F32_NEONDOT
4);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x1x32_F32_NEONDOT(
5);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x1x32_F32_NEONDOT(
6);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x4x16_F32_NEONDOT(
1);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x4x16_F32_NEONDOT(
Expand All @@ -248,6 +250,8 @@ BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x4x16_F32_NEONDOT
4);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x4x16_F32_NEONDOT(
5);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x4x16_F32_NEONDOT(
6);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x4x16_F32_NEONDOT(
1);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x8x16_F32_NEONDOT(
Expand All @@ -258,6 +262,8 @@ BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x8x16_F32_NEONDOT
4);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x8x16_F32_NEONDOT(
5);
BENCHMARK_CHANNELWISE_8BIT_ACTIVATION_GROUPWISE_LOWBIT_WEIGHT_1x8x16_F32_NEONDOT(
6);

// Run the benchmark
BENCHMARK_MAIN();
41 changes: 35 additions & 6 deletions torchao/experimental/kernels/cpu/aarch64/bitpacking/bitpack.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint3.h>
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint4.h>
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint5.h>
#include <torchao/experimental/kernels/cpu/aarch64/bitpacking/uint6.h>
#include <cassert>

namespace torchao {
Expand Down Expand Up @@ -80,7 +81,7 @@ TORCHAO_ALWAYS_INLINE inline void vec_pack_32_lowbit_values(

// Currently supported values
static_assert(nbit >= 1);
static_assert(nbit <= 5);
static_assert(nbit <= 6);

// Shift unpacked values to nonnegative range
int8x16_t shift = vdupq_n_s8(1 << (nbit - 1));
Expand Down Expand Up @@ -138,6 +139,10 @@ TORCHAO_ALWAYS_INLINE inline void vec_pack_32_lowbit_values(
torchao::bitpacking::internal::pack_8_uint5_values(
packed + 15, buffer5 + 24);
break;
case 6:
torchao::bitpacking::internal::vec_pack_32_uint6_values(
packed, shifted0, shifted1);
break;
default:
assert(false);
}
Expand All @@ -153,7 +158,7 @@ TORCHAO_ALWAYS_INLINE inline void vec_unpack_32_lowbit_values(

// Currently supported values
static_assert(nbit >= 1);
static_assert(nbit <= 5);
static_assert(nbit <= 6);

uint8x16_t shifted0;
uint8x16_t shifted1;
Expand Down Expand Up @@ -208,6 +213,10 @@ TORCHAO_ALWAYS_INLINE inline void vec_unpack_32_lowbit_values(
shifted0 = vld1q_u8(buffer5);
shifted1 = vld1q_u8(buffer5 + 16);
break;
case 6:
torchao::bitpacking::internal::vec_unpack_32_uint6_values(
shifted0, shifted1, packed);
break;
default:
assert(false);
}
Expand All @@ -230,7 +239,7 @@ TORCHAO_ALWAYS_INLINE inline void vec_pack_64_lowbit_values(

// Currently supported values
static_assert(nbit >= 1);
static_assert(nbit <= 5);
static_assert(nbit <= 6);

// Shift unpacked values to nonnegative range
int8x16_t shift = vdupq_n_s8(1 << (nbit - 1));
Expand Down Expand Up @@ -262,6 +271,10 @@ TORCHAO_ALWAYS_INLINE inline void vec_pack_64_lowbit_values(
torchao::bitpacking::internal::vec_pack_64_uint5_values(
packed, shifted0, shifted1, shifted2, shifted3);
break;
case 6:
torchao::bitpacking::internal::vec_pack_64_uint6_values(
packed, shifted0, shifted1, shifted2, shifted3);
break;
default:
assert(false);
}
Expand All @@ -279,7 +292,7 @@ TORCHAO_ALWAYS_INLINE inline void vec_unpack_64_lowbit_values(

// Currently supported values
static_assert(nbit >= 1);
static_assert(nbit <= 5);
static_assert(nbit <= 6);

uint8x16_t shifted0;
uint8x16_t shifted1;
Expand Down Expand Up @@ -309,6 +322,10 @@ TORCHAO_ALWAYS_INLINE inline void vec_unpack_64_lowbit_values(
torchao::bitpacking::internal::vec_unpack_64_uint5_values(
shifted0, shifted1, shifted2, shifted3, packed);
break;
case 6:
torchao::bitpacking::internal::vec_unpack_64_uint6_values(
shifted0, shifted1, shifted2, shifted3, packed);
break;
default:
assert(false);
}
Expand Down Expand Up @@ -337,7 +354,7 @@ TORCHAO_ALWAYS_INLINE inline void vec_pack_128_lowbit_values(

// Currently supported values
static_assert(nbit >= 1);
static_assert(nbit <= 5);
static_assert(nbit <= 6);

// Shift unpacked values to nonnegative range
int8x16_t shift = vdupq_n_s8(1 << (nbit - 1));
Expand Down Expand Up @@ -403,6 +420,12 @@ TORCHAO_ALWAYS_INLINE inline void vec_pack_128_lowbit_values(
shifted6,
shifted7);
break;
case 6:
torchao::bitpacking::internal::vec_pack_64_uint6_values(
packed, shifted0, shifted1, shifted2, shifted3);
torchao::bitpacking::internal::vec_pack_64_uint6_values(
packed + 48, shifted4, shifted5, shifted6, shifted7);
break;
default:
assert(false);
}
Expand All @@ -424,7 +447,7 @@ TORCHAO_ALWAYS_INLINE inline void vec_unpack_128_lowbit_values(

// Currently supported values
static_assert(nbit >= 1);
static_assert(nbit <= 5);
static_assert(nbit <= 6);

uint8x16_t shifted0;
uint8x16_t shifted1;
Expand Down Expand Up @@ -488,6 +511,12 @@ TORCHAO_ALWAYS_INLINE inline void vec_unpack_128_lowbit_values(
shifted7,
packed);
break;
case 6:
torchao::bitpacking::internal::vec_unpack_64_uint6_values(
shifted0, shifted1, shifted2, shifted3, packed);
torchao::bitpacking::internal::vec_unpack_64_uint6_values(
shifted4, shifted5, shifted6, shifted7, packed + 48);
break;
default:
assert(false);
}
Expand Down
Loading