Faster per-token fp8 group quant packed kernel for blackwell (#41326)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Yongye Zhu
2026-04-30 18:09:55 -07:00
committed by GitHub
co-authored by Roger Wang
parent 9c61864bf8
commit a3c83ff2fd
3 changed files with 366 additions and 89 deletions
@@ -236,17 +236,41 @@ void per_token_group_quant_8bit(const torch::stable::Tensor& input,
#undef LAUNCH_KERNEL
}
template <typename T, typename DST_DTYPE>
__global__ void per_token_group_quant_8bit_packed_kernel(
// Register-resident fast path for group_size==128.
//
// Each thread holds 16 source elements (32 B = uint4 x 2) in registers across
// the absmax reduce -> scale compute -> quantize pipeline. No shared memory.
// UE8M0 scale extracted via bit math (bit-exact with exp2f(ceilf(log2f))).
//
// Loads two contiguous uint4s (16 B + 16 B = 32 B) per thread; on Blackwell
// nvcc fuses these into a single 256-bit LDG.E.256.
//
// Constraints: GROUP_SIZE % (THREADS_PER_GROUP * VEC_SIZE) == 0; for
// THREADS_PER_GROUP=8 and bf16/fp16 (VEC_SIZE=16), this means GROUP_SIZE=128.
template <typename T, typename DST_DTYPE, int GROUP_SIZE>
__global__ void per_token_group_quant_8bit_packed_register_kernel(
const T* __restrict__ input, void* __restrict__ output_q,
unsigned int* __restrict__ output_s_packed, const int group_size,
const int num_groups_padded, const int groups_per_block,
const int padded_groups_per_row, const int groups_per_row, const int mn,
const int tma_aligned_mn, const int num_scale_elems, const float eps,
unsigned int* __restrict__ output_s_packed, const int64_t num_groups_padded,
const int groups_per_block, const int padded_groups_per_row,
const int groups_per_row, const int mn, const int output_q_mn_extent,
const int tma_aligned_mn, const int64_t num_scale_elems, const float eps,
const float min_8bit, const float max_8bit) {
const int threads_per_group = 16;
const int64_t local_group_id = threadIdx.x / threads_per_group;
const int lane_id = threadIdx.x % threads_per_group;
static_assert(GROUP_SIZE == 128, "fast path supports GROUP_SIZE==128");
constexpr int THREADS_PER_GROUP = 8;
constexpr int VEC_SIZE = 32 / sizeof(T); // 16 for bf16/fp16
static_assert(GROUP_SIZE == THREADS_PER_GROUP * VEC_SIZE,
"GROUP_SIZE must equal THREADS_PER_GROUP * VEC_SIZE");
// Each group's 8 threads must live in a single warp octet so the
// 0xffu << (threadIdx.x & 24u) shuffle mask selects exactly the lanes
// that share a group. Requires 32 % THREADS_PER_GROUP == 0 and the host
// to launch num_threads as a multiple of THREADS_PER_GROUP (which it does
// via num_threads = groups_per_block * THREADS_PER_GROUP).
static_assert(32 % THREADS_PER_GROUP == 0,
"THREADS_PER_GROUP must divide warp size for the shuffle "
"mask to be valid");
const int local_group_id = threadIdx.x / THREADS_PER_GROUP;
const int lane_id = threadIdx.x % THREADS_PER_GROUP;
const int64_t block_group_id = blockIdx.x * groups_per_block;
const int64_t global_group_id = block_group_id + local_group_id;
@@ -254,141 +278,207 @@ __global__ void per_token_group_quant_8bit_packed_kernel(
return;
}
// map flat group id to 2D indices (mn_idx, sf_k_idx)
const int sf_k_idx =
static_cast<int>(global_group_id % padded_groups_per_row);
const int mn_idx = static_cast<int>(global_group_id / padded_groups_per_row);
// whether it is a valid group (not padding)
const bool is_valid_group = (mn_idx < mn) && (sf_k_idx < groups_per_row);
// shared memory to cache each group's data to avoid double DRAM reads.
extern __shared__ __align__(16) char smem_raw[];
T* smem = reinterpret_cast<T*>(smem_raw);
T* smem_group = smem + local_group_id * group_size;
// compute scale for valid groups
float y_s = 0.f;
// Load 16 input elements (32 B) into registers as two adjacent uint4
// loads. nvcc keeps these as 2x LDG.E.128 on sm_100; the per-thread cost
// is dominated by HBM bandwidth at large MN, so a fused 256-bit load via
// inline PTX gave no measurable speedup.
// alignas(16) is required so the uint4* reinterpret_cast below is
// well-defined for T == bf16/fp16 (default alignof is 2).
alignas(16) T regs[VEC_SIZE];
float local_absmax = eps;
if (is_valid_group) {
const T* group_input =
input + static_cast<int64_t>(mn_idx) * groups_per_row * group_size +
sf_k_idx * group_size;
y_s = ComputeGroupScale<T, true>(group_input, smem_group, group_size,
lane_id, threads_per_group, eps, max_8bit);
input + static_cast<int64_t>(mn_idx) * groups_per_row * GROUP_SIZE +
sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE;
uint4* dst = reinterpret_cast<uint4*>(&regs[0]);
const uint4* src = reinterpret_cast<const uint4*>(group_input);
dst[0] = src[0];
dst[1] = src[1];
#pragma unroll
for (int i = 0; i < VEC_SIZE; ++i) {
float v = fabsf(static_cast<float>(regs[i]));
local_absmax = fmaxf(local_absmax, v);
}
}
// pack 4 scales into a uint32 exponent
// 8-lane subgroup shuffle reduce (octet of the warp). The mask selects the
// 8 lanes within the warp that share a group.
unsigned mask = 0xffu << (threadIdx.x & 24u);
local_absmax = fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 4));
local_absmax = fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 2));
local_absmax = fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 1));
float y_s = local_absmax / max_8bit;
y_s = fmaxf(y_s, 1e-10f);
uint32_t bits = __float_as_uint(y_s);
uint32_t exp_bits = (bits >> 23) & 0xffu;
uint32_t mant_bits = bits & 0x7fffffu;
uint8_t exp_byte =
static_cast<uint8_t>(exp_bits + (mant_bits != 0u ? 1u : 0u));
// Lane 0 writes the packed scale byte.
if (lane_id == 0) {
// each uint32 in output_s_packed stores 4 packed scales
const int sf_k_pack_idx = sf_k_idx / 4;
const int pos = sf_k_idx % 4;
const int out_idx = sf_k_pack_idx * tma_aligned_mn + mn_idx;
if (is_valid_group) {
// reinterpret the UE8M0 scale y_s as IEEE bits, extract the 8-bit
// exponent, and place it into the correct byte of the 32-bit word.
const unsigned int bits = __float_as_uint(y_s);
const uint8_t exponent = static_cast<uint8_t>((bits >> 23u) & 0xffu);
reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = exponent;
reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = exp_byte;
} else if (out_idx < num_scale_elems) {
// write zero for padding groups if within bounds of output_s_packed
reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = 0;
}
}
__syncthreads();
if (is_valid_group) {
DST_DTYPE* group_output =
static_cast<DST_DTYPE*>(output_q) +
static_cast<int64_t>(mn_idx) * groups_per_row * group_size +
sf_k_idx * group_size;
QuantizeGroup<T, DST_DTYPE>(smem_group, group_output, group_size, lane_id,
threads_per_group, y_s, min_8bit, max_8bit);
// For padded mn rows that fall within output_q's allocated extent, write
// a uint4 of zeros to keep the buffer clean for downstream TMA loads.
// Skip writes for sf_k padding (those positions don't exist in output_q).
if (!is_valid_group) {
if (sf_k_idx < groups_per_row && mn_idx >= mn &&
mn_idx < output_q_mn_extent) {
DST_DTYPE* group_output =
static_cast<DST_DTYPE*>(output_q) +
static_cast<int64_t>(mn_idx) * groups_per_row * GROUP_SIZE +
sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE;
*reinterpret_cast<uint4*>(group_output) = make_uint4(0, 0, 0, 0);
}
return;
}
// Reconstruct y_s as a power-of-2 float and use its reciprocal.
float y_s_q = __uint_as_float(static_cast<uint32_t>(exp_byte) << 23);
float inv_y = 1.0f / y_s_q;
// Quantize and pack into 16 fp8/int8 bytes (= uint4). VEC_SIZE==16 so we
// fill four 32-bit words, four bytes each.
uint32_t packed_lo = 0;
uint32_t packed_lo_hi = 0;
uint32_t packed_hi_lo = 0;
uint32_t packed_hi = 0;
#pragma unroll
for (int i = 0; i < VEC_SIZE; ++i) {
float q =
fminf(fmaxf(static_cast<float>(regs[i]) * inv_y, min_8bit), max_8bit);
DST_DTYPE qb = DST_DTYPE(q);
uint8_t byte = *reinterpret_cast<uint8_t*>(&qb);
const int shift = (i & 3) * 8;
if (i < 4) {
packed_lo |= static_cast<uint32_t>(byte) << shift;
} else if (i < 8) {
packed_lo_hi |= static_cast<uint32_t>(byte) << shift;
} else if (i < 12) {
packed_hi_lo |= static_cast<uint32_t>(byte) << shift;
} else {
packed_hi |= static_cast<uint32_t>(byte) << shift;
}
}
uint4 packed_out =
make_uint4(packed_lo, packed_lo_hi, packed_hi_lo, packed_hi);
DST_DTYPE* group_output =
static_cast<DST_DTYPE*>(output_q) +
static_cast<int64_t>(mn_idx) * groups_per_row * GROUP_SIZE +
sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE;
*reinterpret_cast<uint4*>(group_output) = packed_out;
}
// Public entry point: register-resident packed quant kernel.
// Constraints: group_size == 128 and bf16/fp16 input.
void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
torch::stable::Tensor& output_q,
torch::stable::Tensor& output_s_packed,
int64_t group_size, double eps,
double min_8bit, double max_8bit) {
STD_TORCH_CHECK(group_size == 128,
"per_token_group_quant_8bit_packed only supports "
"group_size==128, got ",
group_size, ".");
const auto in_dtype = input.scalar_type();
STD_TORCH_CHECK(
in_dtype == torch::headeronly::ScalarType::Half ||
in_dtype == torch::headeronly::ScalarType::BFloat16,
"per_token_group_quant_8bit_packed only supports bf16/fp16 input.");
STD_TORCH_CHECK(input.is_contiguous());
STD_TORCH_CHECK(output_q.is_contiguous());
const int64_t k = input.size(-1);
STD_TORCH_CHECK(k % group_size == 0, "Last dimension (", k,
") must be divisible by group_size (", group_size, ").");
STD_TORCH_CHECK(k % group_size == 0, "input last dim k=", k,
" is not divisible by group_size=", group_size, ".");
const int64_t mn = input.numel() / k;
const int64_t groups_per_row = k / group_size;
STD_TORCH_CHECK(output_s_packed.dim() == 2,
"output_s_packed must be 2D, got dim=", output_s_packed.dim(),
".");
const int64_t k_num_packed_sfk = (groups_per_row + 3) / 4;
const int64_t tma_aligned_mn = ((mn + 3) / 4) * 4;
// output_q may be allocated with extra padded mn rows (e.g.,
// (tma_aligned_mn, k)) so the kernel can zero-fill them in-line and the
// caller can use torch.empty instead of torch.zeros. The grid only covers
// up to tma_aligned_mn, so we cap the extent there.
const int64_t output_q_mn_actual = output_q.numel() / k;
STD_TORCH_CHECK(output_q_mn_actual >= mn,
"output_q must have at least mn rows; got ",
output_q_mn_actual, " rows for mn=", mn, ".");
const int64_t output_q_mn_extent =
output_q_mn_actual < tma_aligned_mn ? output_q_mn_actual : tma_aligned_mn;
STD_TORCH_CHECK(
output_s_packed.scalar_type() == torch::headeronly::ScalarType::Int,
"output_s_packed must have dtype int32 for UE8M0-packed scales.");
// DeepGEMM expects SFA scales in MN-major form with shape
// [mn, ceil_div(K, 128 * 4)] and TMA-aligned stride on the last
// dimension.
"output_s_packed must be int32 for UE8M0-packed scales.");
STD_TORCH_CHECK(output_s_packed.size(0) == mn &&
output_s_packed.size(1) == k_num_packed_sfk,
"output_s_packed shape must be [", mn, ", ", k_num_packed_sfk,
"], but got [", output_s_packed.size(0), ", ",
"]; got [", output_s_packed.size(0), ", ",
output_s_packed.size(1), "].");
// Verify column-major TMA-aligned layout
STD_TORCH_CHECK(output_s_packed.stride(0) == 1 &&
output_s_packed.stride(1) == tma_aligned_mn,
"output_s_packed must have strides [1, ", tma_aligned_mn,
"], but got [", output_s_packed.stride(0), ", ",
"output_s_packed strides must be [1, ", tma_aligned_mn,
"]; got [", output_s_packed.stride(0), ", ",
output_s_packed.stride(1), "].");
cudaStream_t stream = get_current_cuda_stream();
constexpr int THREADS_PER_GROUP = 16;
// Expand the grid to cover MN and K padding so every byte in
// output_s_packed is written (padding bytes get zeroed by the kernel).
constexpr int THREADS_PER_GROUP = 8;
const int64_t padded_groups_per_row = k_num_packed_sfk * 4;
const int64_t num_groups_padded = tma_aligned_mn * padded_groups_per_row;
// Number of elements in output_s_packed.
const int64_t num_scale_elems = mn + (k_num_packed_sfk - 1) * tma_aligned_mn;
const int groups_per_block = GetGroupsPerBlock(num_groups_padded);
auto dst_type = output_q.scalar_type();
const int num_blocks = num_groups_padded / groups_per_block;
const int64_t num_blocks = num_groups_padded / groups_per_block;
const int num_threads = groups_per_block * THREADS_PER_GROUP;
// CUDA caps grid.x at 2^31 - 1; this fits any realistic shape but guard
// against pathological inputs.
STD_TORCH_CHECK(num_blocks <= static_cast<int64_t>(INT32_MAX),
"per_token_group_quant_8bit_packed grid too large: ",
num_blocks, " blocks (max ", INT32_MAX, ").");
#define LAUNCH_PACKED_KERNEL(T, DST_DTYPE) \
do { \
dim3 grid(num_blocks); \
dim3 block(num_threads); \
size_t smem_bytes = \
static_cast<size_t>(groups_per_block) * group_size * sizeof(T); \
per_token_group_quant_8bit_packed_kernel<T, DST_DTYPE> \
<<<grid, block, smem_bytes, stream>>>( \
static_cast<const T*>(input.data_ptr()), output_q.data_ptr(), \
reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()), \
static_cast<int>(group_size), static_cast<int>(num_groups_padded), \
groups_per_block, static_cast<int>(padded_groups_per_row), \
static_cast<int>(groups_per_row), static_cast<int>(mn), \
static_cast<int>(tma_aligned_mn), \
static_cast<int>(num_scale_elems), static_cast<float>(eps), \
static_cast<float>(min_8bit), static_cast<float>(max_8bit)); \
#define LAUNCH_REG_KERNEL(T, DST_DTYPE) \
do { \
dim3 grid(static_cast<unsigned int>(num_blocks)); \
dim3 block(num_threads); \
per_token_group_quant_8bit_packed_register_kernel<T, DST_DTYPE, 128> \
<<<grid, block, 0, stream>>>( \
static_cast<const T*>(input.data_ptr()), output_q.data_ptr(), \
reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()), \
num_groups_padded, groups_per_block, \
static_cast<int>(padded_groups_per_row), \
static_cast<int>(groups_per_row), static_cast<int>(mn), \
static_cast<int>(output_q_mn_extent), \
static_cast<int>(tma_aligned_mn), num_scale_elems, \
static_cast<float>(eps), static_cast<float>(min_8bit), \
static_cast<float>(max_8bit)); \
} while (0)
VLLM_STABLE_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "per_token_group_quant_8bit_packed", ([&] {
VLLM_STABLE_DISPATCH_HALF_TYPES(
input.scalar_type(), "per_token_group_quant_8bit_packed_register", ([&] {
if (dst_type == torch::headeronly::ScalarType::Float8_e4m3fn) {
LAUNCH_PACKED_KERNEL(scalar_t, __nv_fp8_e4m3);
LAUNCH_REG_KERNEL(scalar_t, __nv_fp8_e4m3);
} else if (dst_type == torch::headeronly::ScalarType::Char) {
LAUNCH_PACKED_KERNEL(scalar_t, int8_t);
LAUNCH_REG_KERNEL(scalar_t, int8_t);
} else {
STD_TORCH_CHECK(
false,
@@ -397,7 +487,7 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
}
}));
#undef LAUNCH_PACKED_KERNEL
#undef LAUNCH_REG_KERNEL
}
void per_token_group_quant_fp8(const torch::stable::Tensor& input,
@@ -8,3 +8,13 @@ void per_token_group_quant_8bit(const torch::stable::Tensor& input,
torch::stable::Tensor& output_s,
int64_t group_size, double eps, double min_8bit,
double max_8bit, bool scale_ue8m0 = false);
// Public op: register-resident packed quant for the DeepGEMM Blackwell path.
// Restricted to group_size == 128 and bf16/fp16 input; other configurations
// raise STD_TORCH_CHECK. The legacy shared-memory fallback was removed because
// no production caller (deep_gemm_moe / input_quant_fp8) uses other shapes.
void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
torch::stable::Tensor& output_q,
torch::stable::Tensor& output_s_packed,
int64_t group_size, double eps,
double min_8bit, double max_8bit);