chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

support smem in per_token_quant_fp8 kernel (#16725)

Browse Source 
Co-authored-by: zhangxin81 <969206500@qq.com>
This commit is contained in:
zhangxin81
2026-02-02 17:18:50 +08:00
committed by GitHub
parent a0757c9624
commit e3021b65fe
1 changed files with 97 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

123
sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
View File

@@ -6,13 +6,15 @@
#include "utils.h"
static constexpr int kWarpSize = 32;
static constexpr int DEFAULT_SHARED_MEM_THRESHOLD_KB = 48; // Default shared memory quota in KB
// ---------------------------------------------------------------------------
// 1. Warp‑local, no shared memory
// 1. Warp‑local with configurable shared memory
// • One warp handles one token.
// • Eight tokens per 256‑thread CTA.
// • Shared memory usage is configurable via template parameter.
// ---------------------------------------------------------------------------
template <typename T, typename DST_DTYPE, int kTokensPerCTA = 8, int kVecSize = 16>
template <typename T, typename DST_DTYPE, int kTokensPerCTA = 8, int kVecSize = 16, bool USE_SMEM = true>
__global__ void per_token_quant_fp8_kernel(
const T* __restrict__ input,
DST_DTYPE* __restrict__ output_q,
@@ -29,8 +31,14 @@ __global__ void per_token_quant_fp8_kernel(
DST_DTYPE* token_output = output_q + token_id * hidden_dim;
float* token_scale = output_s + token_id;
extern __shared__ char smem_buffer[];
const int smem_padding = 32; // Pad to bank boundary (32 banks * 4 bytes = 128 bytes)
const int warp_smem_stride = (hidden_dim * sizeof(T) + smem_padding - 1) / smem_padding * smem_padding;
const int warp_smem_offset = warp_id * warp_smem_stride;
T* shared_input = reinterpret_cast<T*>(smem_buffer + warp_smem_offset);
//
// Pass-1: Perform a warp reduce to find the max_value of a token's hidden_dim
// Pass-1: Load data and compute max_value
//
float max_value = 0.f;
using vec_t = flashinfer::vec_t<T, kVecSize>;
@@ -40,12 +48,26 @@ __global__ void per_token_quant_fp8_kernel(
vec_t input_vec;
input_vec.cast_load(token_input + i * kVecSize);
// Store to shared memory if USE_SMEM=true
if constexpr (USE_SMEM) {
#pragma unroll
for (uint32_t j = 0; j < kVecSize; ++j) {
shared_input[i * kVecSize + j] = input_vec[j];
}
}
// Compute max value in parallel
#pragma unroll
for (uint32_t j = 0; j < kVecSize; ++j) {
max_value = fmaxf(max_value, fabsf(static_cast<float>(input_vec[j])));
}
}
// Ensure all threads in the warp have finished writing to shared memory
if constexpr (USE_SMEM) {
__syncwarp();
}
float warp_max = warpReduceMax(max_value);
// NOTE: one CTA has multiple warps (each warp handles one token), so `scale`
@@ -58,11 +80,22 @@ __global__ void per_token_quant_fp8_kernel(
const float scale_inv = (scale == 0.f) ? 0.f : 1.0f / scale;
//
// Pass-2: quantize and write back
// Pass-2: Quantize and write back
//
for (int i = lane_id; i < num_vec_elems; i += kWarpSize) {
vec_t input_vec;
input_vec.cast_load(token_input + i * kVecSize);
if constexpr (USE_SMEM) {
// Load from shared memory
#pragma unroll
for (uint32_t j = 0; j < kVecSize; ++j) {
input_vec[j] = shared_input[i * kVecSize + j];
}
} else {
// Reload from global memory
input_vec.cast_load(token_input + i * kVecSize);
}
DST_DTYPE output_arr[kVecSize];
#pragma unroll
for (uint32_t j = 0; j < kVecSize; ++j) {
@@ -164,6 +197,48 @@ __global__ void per_token_quant_fp8_small_batch_kernel(
}
}
template <bool USE_SMEM, typename scalar_t, int TOKENS_PER_CTA>
static inline void launch_per_token_quant_fp8_warp_kernel(
const dim3& grid,
const dim3& block,
size_t dynamicSmemSz,
cudaStream_t stream,
bool use_vec16,
bool use_vec8,
torch::Tensor input,
torch::Tensor output_q,
torch::Tensor output_s,
const int64_t hidden_dim,
const int64_t num_tokens) {
const size_t smem_size = USE_SMEM ? dynamicSmemSz : 0;
if (use_vec16) {
per_token_quant_fp8_kernel<scalar_t, __nv_fp8_e4m3, TOKENS_PER_CTA, 16, USE_SMEM>
<<<grid, block, smem_size, stream>>>(
static_cast<const scalar_t*>(input.data_ptr()),
static_cast<__nv_fp8_e4m3*>(output_q.data_ptr()),
static_cast<float*>(output_s.data_ptr()),
hidden_dim,
num_tokens);
} else if (use_vec8) {
per_token_quant_fp8_kernel<scalar_t, __nv_fp8_e4m3, TOKENS_PER_CTA, 8, USE_SMEM>
<<<grid, block, smem_size, stream>>>(
static_cast<const scalar_t*>(input.data_ptr()),
static_cast<__nv_fp8_e4m3*>(output_q.data_ptr()),
static_cast<float*>(output_s.data_ptr()),
hidden_dim,
num_tokens);
} else {
per_token_quant_fp8_kernel<scalar_t, __nv_fp8_e4m3, TOKENS_PER_CTA, 4, USE_SMEM>
<<<grid, block, smem_size, stream>>>(
static_cast<const scalar_t*>(input.data_ptr()),
static_cast<__nv_fp8_e4m3*>(output_q.data_ptr()),
static_cast<float*>(output_s.data_ptr()),
hidden_dim,
num_tokens);
}
}
void sgl_per_token_quant_fp8(torch::Tensor input, torch::Tensor output_q, torch::Tensor output_s) {
CHECK_INPUT(input);
CHECK_INPUT(output_q);
@@ -180,34 +255,30 @@ void sgl_per_token_quant_fp8(torch::Tensor input, torch::Tensor output_q, torch:
const bool use_vec16 = (hidden_dim % 16 == 0);
const bool use_vec8 = (hidden_dim % 8 == 0);
const int sizeof_T = input.scalar_type() == torch::kFloat16 ? 2 : (input.scalar_type() == torch::kBFloat16 ? 2 : 4);
const int smem_padding = 32; // Pad to bank boundary to avoid conflicts
const int warp_smem_stride = (hidden_dim * sizeof_T + smem_padding - 1) / smem_padding * smem_padding;
const size_t dynamicSmemSz = warp_smem_stride * TOKENS_PER_CTA;
bool use_smem = (hidden_dim < 2048);
if (dynamicSmemSz >= DEFAULT_SHARED_MEM_THRESHOLD_KB) {
use_smem = false; // Disable shared memory if >= 48KB to avoid allocation failures
}
DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FLOAT_FP16(input.scalar_type(), scalar_t, [&] {
if (use_warp_kernel) {
// -------- warp‑local ---------------------------------------------------
constexpr int THREADS = TOKENS_PER_CTA * kWarpSize; // 256
constexpr int THREADS = TOKENS_PER_CTA * kWarpSize;
dim3 grid((num_tokens + TOKENS_PER_CTA - 1) / TOKENS_PER_CTA);
dim3 block(THREADS);
if (use_vec16) {
per_token_quant_fp8_kernel<scalar_t, __nv_fp8_e4m3, TOKENS_PER_CTA, 16><<<grid, block, 0, stream>>>(
static_cast<const scalar_t*>(input.data_ptr()),
static_cast<__nv_fp8_e4m3*>(output_q.data_ptr()),
static_cast<float*>(output_s.data_ptr()),
hidden_dim,
num_tokens);
} else if (use_vec8) {
per_token_quant_fp8_kernel<scalar_t, __nv_fp8_e4m3, TOKENS_PER_CTA, 8><<<grid, block, 0, stream>>>(
static_cast<const scalar_t*>(input.data_ptr()),
static_cast<__nv_fp8_e4m3*>(output_q.data_ptr()),
static_cast<float*>(output_s.data_ptr()),
hidden_dim,
num_tokens);
if (use_smem) {
launch_per_token_quant_fp8_warp_kernel</*USE_SMEM=*/true, scalar_t, TOKENS_PER_CTA>(
grid, block, dynamicSmemSz, stream, use_vec16, use_vec8, input, output_q, output_s, hidden_dim, num_tokens);
} else {
per_token_quant_fp8_kernel<scalar_t, __nv_fp8_e4m3, TOKENS_PER_CTA, 4><<<grid, block, 0, stream>>>(
static_cast<const scalar_t*>(input.data_ptr()),
static_cast<__nv_fp8_e4m3*>(output_q.data_ptr()),
static_cast<float*>(output_s.data_ptr()),
hidden_dim,
num_tokens);
launch_per_token_quant_fp8_warp_kernel</*USE_SMEM=*/false, scalar_t, TOKENS_PER_CTA>(
grid, block, dynamicSmemSz, stream, use_vec16, use_vec8, input, output_q, output_s, hidden_dim, num_tokens);
}
} else {
// -------- baseline -----------------------------------------------------