Sync from v0.13

This commit is contained in:
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions

View File

@@ -0,0 +1,263 @@
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include "../../dispatch_utils.h"
#include "layernorm_utils.cuh"
#include "quant_conversions.cuh"
namespace vllm {
template <typename scalar_t, typename scalar_out_t, bool has_residual = false>
__device__ void rms_norm_dynamic_per_token_quant_vec(
scalar_out_t* __restrict__ out, // [..., hidden_size]
float* __restrict__ scales, // [num_tokens]
scalar_t const* __restrict__ input, // [..., hidden_size]
scalar_t const* __restrict__ weight, // [hidden_size]
float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
scalar_t* __restrict__ residual = nullptr) {
float rms = 0.0f;
float token_scale = 0.0f;
// Compute rms
vllm::vectorized::compute_rms<scalar_t, has_residual>(
&rms, input, hidden_size, var_epsilon, residual);
// Compute scale
vllm::vectorized::compute_dynamic_per_token_scales<scalar_t, scalar_out_t,
has_residual>(
&token_scale, scales, input, weight, rms, scale_ub, hidden_size,
residual);
// RMS Norm + Quant
if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
token_scale = 1.0f / token_scale;
vllm::vectorized::norm_and_quant<scalar_t, scalar_out_t, true,
has_residual>(
out, input, weight, rms, &token_scale, hidden_size, residual);
} else {
// FP8 - Do not invert token_scale for exact match with FBGemm
vllm::vectorized::norm_and_quant<scalar_t, scalar_out_t, false,
has_residual>(
out, input, weight, rms, &token_scale, hidden_size, residual);
}
}
// RMS norm + quant kernel
template <typename scalar_t, typename scalar_out_t, bool has_residual = false>
__global__ void rms_norm_dynamic_per_token_quant_kernel(
scalar_out_t* __restrict__ out, // [..., hidden_size]
float* __restrict__ scales, // [num_tokens]
scalar_t const* __restrict__ input, // [..., hidden_size]
scalar_t const* __restrict__ weight, // [hidden_size]
float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
scalar_t* __restrict__ residual = nullptr) {
// For vectorization, token_input and token_output pointers need to be
// aligned at 8-byte and 4-byte addresses respectively.
bool const can_vectorize = hidden_size % 4 == 0;
if (can_vectorize) {
return rms_norm_dynamic_per_token_quant_vec<scalar_t, scalar_out_t,
has_residual>(
out, scales, input, weight, scale_ub, var_epsilon, hidden_size,
residual);
}
float rms = 0.0f;
float token_scale = 0.0f;
// Compute RMS
vllm::compute_rms<scalar_t, has_residual>(&rms, input, hidden_size,
var_epsilon, residual);
// Compute Scale
vllm::compute_dynamic_per_token_scales<scalar_t, scalar_out_t, has_residual>(
&token_scale, scales, input, weight, rms, scale_ub, hidden_size,
residual);
// RMS Norm + Quant
if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
token_scale = 1.0f / token_scale;
vllm::norm_and_quant<scalar_t, scalar_out_t, true, has_residual>(
out, input, weight, rms, &token_scale, hidden_size, residual);
} else {
// FP8 - Do not invert s_token_scale for exact match with FBGemm
vllm::norm_and_quant<scalar_t, scalar_out_t, false, has_residual>(
out, input, weight, rms, &token_scale, hidden_size, residual);
}
}
// RMS norm + quant kernel
template <typename scalar_t, typename scalar_out_t, bool has_residual = false,
bool is_scale_transposed = false, int32_t group_size = 0>
__global__ void rms_norm_per_block_quant_kernel(
scalar_out_t* __restrict__ out, // [..., hidden_size]
float* __restrict__ scales, // [num_tokens, hidden_size / group_size]
// or
// [hidden_size / group_size, num_tokens]
scalar_t const* __restrict__ input, // [..., hidden_size]
scalar_t const* __restrict__ weight, // [hidden_size]
float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
scalar_t* __restrict__ residual = nullptr) {
float rms;
// Compute RMS
// Always able to vectorize due to constraints on hidden_size
vllm::vectorized::compute_rms<scalar_t, has_residual>(
&rms, input, hidden_size, var_epsilon, residual);
// Compute Scale
// Always able to vectorize due to constraints on hidden_size and group_size
vllm::vectorized::compute_dynamic_per_token_scales<
scalar_t, scalar_out_t, has_residual, is_scale_transposed, group_size>(
nullptr, scales, input, weight, rms, scale_ub, hidden_size, residual);
// RMS Norm + Quant
// Always able to vectorize due to constraints on hidden_size
// For int8, don't invert token_scale here: do it inside the norm_and_quant
// kernel. We do it because particular elements of token_scale can be shared
// between multiple threads, so this way, we avoid extra synchronization
// overhead.
vllm::vectorized::norm_and_quant<
scalar_t, scalar_out_t, std::is_same_v<scalar_out_t, int8_t>,
has_residual, is_scale_transposed, group_size>(
out, input, weight, rms, scales, hidden_size, residual);
}
} // namespace vllm
// Residual add + RMS norm + dynamic per token
template <typename scalar_in_t>
void rms_norm_dynamic_per_token_quant_dispatch(
torch::Tensor& out, // [..., hidden_size]
torch::Tensor const& input, // [..., hidden_size]
torch::Tensor const& weight, // [hidden_size]
torch::Tensor& scales, // [num_tokens]
double const var_epsilon, // Variance epsilon used in norm calculation
std::optional<at::Tensor> const& scale_ub,
std::optional<at::Tensor>& residual) {
int32_t hidden_size = input.size(-1);
auto num_tokens = input.numel() / hidden_size;
dim3 grid(num_tokens);
dim3 block(std::min(hidden_size, 1024));
const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
VLLM_DISPATCH_BOOL(residual.has_value(), has_residual, [&] {
VLLM_DISPATCH_QUANT_TYPES(
out.scalar_type(), "rms_norm_dynamic_per_token_quant_kernel", [&] {
vllm::rms_norm_dynamic_per_token_quant_kernel<scalar_in_t, scalar_t,
has_residual>
<<<grid, block, 0, stream>>>(
out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
input.data_ptr<scalar_in_t>(), weight.data_ptr<scalar_in_t>(),
scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
var_epsilon, hidden_size,
has_residual ? residual->data_ptr<scalar_in_t>() : nullptr);
});
});
}
void rms_norm_dynamic_per_token_quant(
torch::Tensor& out, // [..., hidden_size]
torch::Tensor const& input, // [..., hidden_size]
torch::Tensor const& weight, // [hidden_size]
torch::Tensor& scales, // [num_tokens]
double const var_epsilon, // Variance epsilon used in norm calculation
std::optional<at::Tensor> scale_ub, std::optional<at::Tensor> residual) {
static c10::ScalarType kFp8Type = is_fp8_ocp()
? c10::ScalarType::Float8_e4m3fn
: c10::ScalarType::Float8_e4m3fnuz;
TORCH_CHECK(out.dtype() == kFp8Type || out.dtype() == torch::kInt8);
TORCH_CHECK(out.is_contiguous() && input.is_contiguous());
if (scale_ub.has_value()) {
TORCH_CHECK(out.dtype() == kFp8Type);
}
TORCH_CHECK(weight.dtype() == input.dtype());
TORCH_CHECK(scales.dtype() == torch::kFloat32);
if (residual) {
TORCH_CHECK(residual->scalar_type() == input.scalar_type());
}
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "rms_norm_dynamic_per_token_quant_dispatch", [&] {
rms_norm_dynamic_per_token_quant_dispatch<scalar_t>(
out, input, weight, scales, var_epsilon, scale_ub, residual);
});
}
// Residual add + RMS norm + dynamic per token
void rms_norm_per_block_quant_dispatch(
torch::Tensor& out, // [..., hidden_size]
torch::Tensor const& input, // [..., hidden_size]
torch::Tensor const& weight, // [hidden_size]
torch::Tensor& scales, // [num_tokens, hidden_size / group_size] or
// [hidden_size / group_size, num_tokens]
int32_t group_size,
double const var_epsilon, // Variance epsilon used in norm calculation
std::optional<at::Tensor> const& scale_ub,
std::optional<at::Tensor>& residual, bool is_scale_transposed) {
int32_t hidden_size = input.size(-1);
auto num_tokens = input.numel() / hidden_size;
dim3 grid(num_tokens);
const int max_block_size = (num_tokens <= 256) ? 512 : 256;
dim3 block(std::min(hidden_size, max_block_size));
const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "rms_norm_per_block_quant_fp_dispatch", [&] {
using scalar_in_t = scalar_t;
VLLM_DISPATCH_GROUP_SIZE(group_size, gs, [&] {
VLLM_DISPATCH_BOOL(residual.has_value(), has_residual, [&] {
VLLM_DISPATCH_BOOL(is_scale_transposed, transpose_scale, [&] {
VLLM_DISPATCH_QUANT_TYPES(
out.scalar_type(), "rms_norm_per_block_quant_kernel", [&] {
vllm::rms_norm_per_block_quant_kernel<scalar_in_t, scalar_t,
has_residual,
transpose_scale, gs>
<<<grid, block, 0, stream>>>(
out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
input.data_ptr<scalar_in_t>(),
weight.data_ptr<scalar_in_t>(),
scale_ub.has_value() ? scale_ub->data_ptr<float>()
: nullptr,
var_epsilon, hidden_size,
has_residual ? residual->data_ptr<scalar_in_t>()
: nullptr);
});
});
});
});
});
}
void rms_norm_per_block_quant(torch::Tensor& out, torch::Tensor const& input,
torch::Tensor const& weight,
torch::Tensor& scales, double const var_epsilon,
std::optional<torch::Tensor> scale_ub,
std::optional<torch::Tensor> residual,
int64_t group_size, bool is_scale_transposed) {
static c10::ScalarType kFp8Type = is_fp8_ocp()
? c10::ScalarType::Float8_e4m3fn
: c10::ScalarType::Float8_e4m3fnuz;
TORCH_CHECK(out.dtype() == kFp8Type || out.dtype() == torch::kInt8);
TORCH_CHECK(out.is_contiguous() && input.is_contiguous());
if (scale_ub.has_value()) {
TORCH_CHECK(out.dtype() == kFp8Type);
}
TORCH_CHECK(weight.dtype() == input.dtype());
TORCH_CHECK(scales.dtype() == torch::kFloat32);
if (residual) {
TORCH_CHECK(residual->scalar_type() == input.scalar_type());
}
TORCH_CHECK(group_size == 128 || group_size == 64,
"Unsupported group size: ", group_size);
rms_norm_per_block_quant_dispatch(out, input, weight, scales, group_size,
var_epsilon, scale_ub, residual,
is_scale_transposed);
}

View File

@@ -0,0 +1,539 @@
#pragma once
/**
* __device__ layernorm utilities.
*/
#include "quantization/vectorization.cuh"
#include "quantization/utils.cuh"
#include "quant_conversions.cuh"
#include "../../cub_helpers.h"
#include "../../cuda_compat.h"
namespace vllm {
// has_residual must be true, if residual is not a nullptr
template <typename scalar_t, bool has_residual = false>
__device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
int32_t const hidden_size, float const epsilon,
scalar_t const* __restrict__ residual = nullptr) {
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
// sum of squares
float ss = 0.0f;
for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
float x = static_cast<float>(input[token_offset + i]);
if constexpr (has_residual) {
x += static_cast<float>(residual[token_offset + i]);
}
ss += x * x;
}
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
ss = BlockReduce(reduceStore).Reduce(ss, CubAddOp{}, blockDim.x);
__shared__ float s_rms;
if (threadIdx.x == 0) {
s_rms = rsqrtf(ss / hidden_size + epsilon);
}
__syncthreads();
*rms = s_rms;
}
__device__ float warpReduceMaxSpecialized(volatile float* val, int64_t tid,
int64_t thread_in_warp,
int64_t reduced_elems) {
static_assert(WARP_SIZE == 32 || WARP_SIZE == 64);
if constexpr (WARP_SIZE == 64) {
if (thread_in_warp + 64 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 64]);
}
if (thread_in_warp + 32 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 32]);
if (thread_in_warp + 16 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 16]);
if (thread_in_warp + 8 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 8]);
if (thread_in_warp + 4 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 4]);
if (thread_in_warp + 2 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 2]);
if (thread_in_warp + 1 < reduced_elems)
val[tid] = fmaxf(val[tid], val[tid + 1]);
return val[tid];
}
template <typename scalar_t, typename scalar_out_t, bool has_residual = false,
bool is_scale_transposed = false>
__device__ void compute_dynamic_per_token_scales(
float* __restrict__ token_scale, float* __restrict__ all_token_scales,
scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
float const rms, float const* __restrict__ scale_ub,
int32_t const hidden_size, scalar_t const* __restrict__ residual = nullptr,
int32_t const group_size = 0) {
float block_absmax_val_maybe = 0.0f;
constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
__syncthreads();
if (group_size > 0) {
__shared__ float s_max_vals[1024];
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
int64_t num_groups = hidden_size / group_size;
int64_t const threads_per_group = blockDim.x / num_groups;
int64_t const thread_in_group = threadIdx.x % threads_per_group;
int64_t const group_offset = threadIdx.x / threads_per_group * group_size;
int64_t const thread_offset = group_offset + thread_in_group;
int64_t const thread_end =
min(group_offset + group_size, static_cast<int64_t>(hidden_size));
for (auto i = thread_offset; i < thread_end; i += threads_per_group) {
float x = static_cast<float>(input[token_offset + i]);
if constexpr (has_residual) {
x += static_cast<float>(residual[token_offset + i]);
}
x = static_cast<float>(static_cast<scalar_t>(x * rms) * weight[i]);
block_absmax_val_maybe = fmaxf(block_absmax_val_maybe, fabsf(x));
}
s_max_vals[threadIdx.x] = block_absmax_val_maybe;
__syncthreads();
int64_t const warp_size = WARP_SIZE;
int64_t const num_warps = blockDim.x / warp_size;
int64_t const warp_id = threadIdx.x / warp_size;
int64_t const thread_in_warp = threadIdx.x % warp_size;
int64_t const groups_per_warp = (num_groups + num_warps - 1) / num_warps;
for (auto i = 0; i < groups_per_warp; ++i) {
int64_t const group_id = i * num_warps + warp_id;
if (group_id < num_groups) {
int64_t warp_start = group_id * threads_per_group;
int64_t const start = warp_start + thread_in_warp;
int64_t const warp_end = min(warp_start + threads_per_group,
static_cast<int64_t>(hidden_size));
for (auto j = start; j + warp_size < warp_end; j += warp_size) {
s_max_vals[start] =
fmaxf(s_max_vals[start], s_max_vals[j + warp_size]);
}
warpReduceMaxSpecialized(s_max_vals, start, thread_in_warp,
min(warp_end - warp_start, warp_size));
}
}
__syncthreads();
if (thread_in_group == 0 && thread_offset < thread_end) {
block_absmax_val_maybe = s_max_vals[threadIdx.x];
float scale = 0.0f;
if (scale_ub) {
scale = min(block_absmax_val_maybe, *scale_ub);
} else {
scale = block_absmax_val_maybe;
}
// token scale computation
scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
// Global output store
if constexpr (is_scale_transposed) {
all_token_scales[(threadIdx.x / threads_per_group) * gridDim.x +
blockIdx.x] = scale;
} else {
all_token_scales[blockIdx.x * num_groups +
threadIdx.x / threads_per_group] = scale;
}
}
__syncthreads();
} else {
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
float x = static_cast<float>(input[token_offset + i]);
if constexpr (has_residual) {
x += static_cast<float>(residual[token_offset + i]);
}
x = static_cast<float>(static_cast<scalar_t>(x * rms) * weight[i]);
block_absmax_val_maybe = fmaxf(block_absmax_val_maybe, fabsf(x));
}
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
block_absmax_val_maybe =
BlockReduce(reduceStore)
.Reduce(block_absmax_val_maybe, CubMaxOp{}, blockDim.x);
__shared__ float s_token_scale;
if (threadIdx.x == 0) {
float scale = 0.0f;
if (scale_ub) {
scale = min(block_absmax_val_maybe, *scale_ub);
} else {
scale = block_absmax_val_maybe;
}
// token scale computation
scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
s_token_scale = scale; // Shared memory store
all_token_scales[blockIdx.x] = scale; // Global output store
}
__syncthreads();
*token_scale = s_token_scale;
}
}
template <typename scalar_t, typename scalar_out_t, bool is_scale_inverted,
bool has_residual = false, bool is_scale_transposed = false>
__device__ void norm_and_quant(scalar_out_t* __restrict__ output,
scalar_t const* __restrict__ input,
scalar_t const* __restrict__ weight,
float const rms, float* const scale,
int32_t const hidden_size,
scalar_t* __restrict__ residual = nullptr,
int32_t const group_size = 0) {
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
float x = static_cast<float>(input[token_offset + i]);
if constexpr (has_residual) {
x += static_cast<float>(residual[token_offset + i]);
residual[token_offset + i] = static_cast<scalar_t>(x);
}
// Norm
x = static_cast<float>(static_cast<scalar_t>(x * rms) * weight[i]);
// Quant
// If groupwise is_scale_inverted is true, so we invert the scale here.
int64_t scale_idx = 0;
if (group_size > 0) {
if constexpr (is_scale_transposed) {
scale_idx = (i / group_size) * gridDim.x + blockIdx.x;
} else {
scale_idx = blockIdx.x * (hidden_size / group_size) + i / group_size;
}
}
auto scale_val =
(group_size > 0
? (is_scale_inverted ? 1.0f / scale[scale_idx] : scale[scale_idx])
: *scale);
output[token_offset + i] =
ScaledQuant<scalar_out_t, is_scale_inverted>::quant_fn(x, scale_val);
}
}
namespace vectorized {
// Compute 1.0/rms(input)
// hidden_size must be a multiple of 4
template <typename scalar_t, bool has_residual = false>
__device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
int32_t const hidden_size, float const epsilon,
scalar_t const* __restrict__ residual = nullptr) {
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
// Vectorized input/output to better utilize memory bandwidth.
vec4_t<scalar_t> const* vec_input =
reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
vec4_t<scalar_t> const* vec_residual = nullptr;
if constexpr (has_residual) {
vec_residual =
reinterpret_cast<vec4_t<scalar_t> const*>(&residual[token_offset]);
}
// sum of squares
float ss = 0.0f;
const int VEC_SIZE = 4;
int32_t const num_vec_elems = hidden_size >> 2;
#pragma unroll 4
for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
vec4_t<scalar_t> in = vec_input[i];
vec4_t<float> x;
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] = static_cast<float>(in.val[j]);
}
if constexpr (has_residual) {
vec4_t<scalar_t> r = vec_residual[i];
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] += static_cast<float>(r.val[j]);
}
}
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
ss += x.val[j] * x.val[j];
}
}
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
ss = BlockReduce(reduceStore).Reduce(ss, CubAddOp{}, blockDim.x);
__shared__ float s_rms;
if (threadIdx.x == 0) {
s_rms = rsqrtf(ss / hidden_size + epsilon);
}
__syncthreads();
*rms = s_rms;
}
// Vectorized version of vllm::compute_dynamic_per_token_scales
// hidden_size must be a multiple of 4
template <typename scalar_t, typename scalar_out_t, bool has_residual = false,
bool is_scale_transposed = false, int32_t group_size = 0>
__device__ void compute_dynamic_per_token_scales(
float* __restrict__ token_scale, float* __restrict__ all_token_scales,
scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
float const rms, float const* __restrict__ scale_ub,
int32_t const hidden_size,
scalar_t const* __restrict__ residual = nullptr) {
constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
const int VEC_SIZE = 4;
float block_absmax_val_maybe = 0.0f;
// Vectorized input/weight/residual to better utilize memory bandwidth.
vec4_t<scalar_t> const* vec_input = nullptr;
vec4_t<scalar_t> const* vec_weight = nullptr;
vec4_t<scalar_t> const* vec_residual = nullptr;
if constexpr (group_size > 0) {
__shared__ float s_max_vals[1024];
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
int64_t const num_groups = hidden_size / group_size;
int64_t const threads_per_group = blockDim.x / num_groups;
int64_t const thread_in_group = threadIdx.x % threads_per_group;
int64_t const group_offset =
threadIdx.x / threads_per_group * (group_size >> 2);
int64_t const thread_offset = group_offset + thread_in_group;
int64_t const thread_end = min(group_offset + (group_size >> 2),
static_cast<int64_t>(hidden_size >> 2));
vec_input = reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
vec_weight = reinterpret_cast<vec4_t<scalar_t> const*>(weight);
if constexpr (has_residual) {
vec_residual =
reinterpret_cast<vec4_t<scalar_t> const*>(&residual[token_offset]);
}
int32_t const num_vec_elems = thread_end;
#pragma unroll 4
for (auto i = thread_offset; i < num_vec_elems; i += threads_per_group) {
vec4_t<scalar_t> in = vec_input[i];
vec4_t<scalar_t> const w = vec_weight[i];
vec4_t<float> x;
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] = static_cast<float>(in.val[j]);
}
if constexpr (has_residual) {
vec4_t<scalar_t> r = vec_residual[i];
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] += static_cast<float>(r.val[j]);
}
}
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
block_absmax_val_maybe =
fmaxf(block_absmax_val_maybe,
fabs(static_cast<scalar_t>(x.val[j] * rms) * w.val[j]));
}
}
s_max_vals[threadIdx.x] = block_absmax_val_maybe;
__syncthreads();
int64_t const warp_size = WARP_SIZE;
int64_t const num_warps = blockDim.x / warp_size;
int64_t const warp_id = threadIdx.x / warp_size;
int64_t const thread_in_warp = threadIdx.x % warp_size;
int64_t const groups_per_warp = (num_groups + num_warps - 1) / num_warps;
for (auto i = 0; i < groups_per_warp; ++i) {
int64_t const group_id = i * num_warps + warp_id;
if (group_id < num_groups) {
int64_t warp_start = group_id * threads_per_group;
int64_t const start = warp_start + thread_in_warp;
int64_t const warp_end = min(warp_start + threads_per_group,
static_cast<int64_t>(hidden_size));
for (auto j = start; j + warp_size < warp_end; j += warp_size) {
s_max_vals[start] =
fmaxf(s_max_vals[start], s_max_vals[j + warp_size]);
}
warpReduceMaxSpecialized(s_max_vals, start, thread_in_warp,
min(warp_end - warp_start, warp_size));
}
}
__syncthreads();
if (thread_in_group == 0 && thread_offset < thread_end) {
block_absmax_val_maybe = s_max_vals[threadIdx.x];
float scale = 0.0f;
if (scale_ub) {
scale = min(block_absmax_val_maybe, *scale_ub);
} else {
scale = block_absmax_val_maybe;
}
// token scale computation
scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
// Global output store
if constexpr (is_scale_transposed) {
all_token_scales[(threadIdx.x / threads_per_group) * gridDim.x +
blockIdx.x] = scale;
} else {
all_token_scales[blockIdx.x * num_groups +
threadIdx.x / threads_per_group] = scale;
}
}
__syncthreads();
} else {
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
vec_input = reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
vec_weight = reinterpret_cast<vec4_t<scalar_t> const*>(weight);
if constexpr (has_residual) {
vec_residual =
reinterpret_cast<vec4_t<scalar_t> const*>(&residual[token_offset]);
}
int32_t const num_vec_elems = (hidden_size >> 2);
#pragma unroll 4
for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
vec4_t<scalar_t> in = vec_input[i];
vec4_t<scalar_t> const w = vec_weight[i];
vec4_t<float> x;
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] = static_cast<float>(in.val[j]);
}
if constexpr (has_residual) {
vec4_t<scalar_t> r = vec_residual[i];
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] += static_cast<float>(r.val[j]);
}
}
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
block_absmax_val_maybe =
fmaxf(block_absmax_val_maybe,
fabs(static_cast<scalar_t>(x.val[j] * rms) * w.val[j]));
}
}
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
block_absmax_val_maybe =
BlockReduce(reduceStore)
.Reduce(block_absmax_val_maybe, CubMaxOp{}, blockDim.x);
__shared__ float s_token_scale;
if (threadIdx.x == 0) {
float scale = 0.0f;
if (scale_ub) {
scale = min(block_absmax_val_maybe, *scale_ub);
} else {
scale = block_absmax_val_maybe;
}
// token scale computation
scale = max(scale / qmax, min_scaling_factor<scalar_out_t>::val());
s_token_scale = scale; // shared memory store
all_token_scales[blockIdx.x] = scale; // global output store
}
__syncthreads();
*token_scale = s_token_scale;
}
}
// hidden_size must be a multiple of 4
template <typename scalar_t, typename scalar_out_t, bool is_scale_inverted,
bool has_residual = false, bool is_scale_transposed = false,
int32_t group_size = 0>
__device__ void norm_and_quant(scalar_out_t* __restrict__ output,
scalar_t const* __restrict__ input,
scalar_t const* __restrict__ weight,
float const rms, float* const scale,
int32_t const hidden_size,
scalar_t* __restrict__ residual = nullptr) {
int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
// Vectorized input/output/weight/residual to better utilize memory bandwidth.
vec4_t<scalar_t> const* vec_input =
reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
vec4_t<scalar_t> const* vec_weight =
reinterpret_cast<vec4_t<scalar_t> const*>(weight);
q8x4_t<scalar_out_t>* vec_output =
reinterpret_cast<q8x4_t<scalar_out_t>*>(&output[token_offset]);
vec4_t<scalar_t>* vec_residual = nullptr;
if constexpr (has_residual) {
vec_residual = reinterpret_cast<vec4_t<scalar_t>*>(&residual[token_offset]);
}
const int VEC_SIZE = 4;
int32_t const num_vec_elems = hidden_size >> 2;
// TODO(luka/varun) extract into type-agnostic vectorized quant function to
// replace scaled_fp8_conversion_vec
#pragma unroll 4
for (auto i = threadIdx.x; i < num_vec_elems; i += blockDim.x) {
vec4_t<scalar_t> const in = vec_input[i];
vec4_t<scalar_t> const w = vec_weight[i];
vec4_t<float> x;
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] = static_cast<float>(in.val[j]);
}
if constexpr (has_residual) {
vec4_t<scalar_t> r = vec_residual[i];
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
x.val[j] += static_cast<float>(r.val[j]);
}
// Update residual
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
r.val[j] = static_cast<scalar_t>(x.val[j]);
}
vec_residual[i] = r;
}
q8x4_t<scalar_out_t> out;
float scale_val;
if constexpr (group_size > 0) {
int64_t const num_groups = hidden_size / group_size;
int64_t scale_idx = 0;
if constexpr (is_scale_transposed) {
scale_idx = (i * VEC_SIZE / group_size) * gridDim.x + blockIdx.x;
} else {
scale_idx = blockIdx.x * num_groups + i * VEC_SIZE / group_size;
}
scale_val =
is_scale_inverted ? 1.0f / scale[scale_idx] : scale[scale_idx];
} else {
scale_val = *scale;
}
#pragma unroll
for (int j = 0; j < VEC_SIZE; ++j) {
out.val[j] = ScaledQuant<scalar_out_t, is_scale_inverted>::quant_fn(
static_cast<scalar_t>(x.val[j] * rms) * w.val[j], scale_val);
}
vec_output[i] = out;
}
}
} // namespace vectorized
} // namespace vllm

View File

@@ -0,0 +1,90 @@
#pragma once
/**
* __device__ helper functions to deal with float -> quant datatype conversion
*/
#include "quantization/vectorization.cuh"
// TODO(luka/varun):refactor common.cuh to use this file instead
#include "quantization/w8a8/fp8/common.cuh"
namespace vllm {
// TODO(luka/varun): combine into common utilities for int8
// (with int8_quant_kernels.cu)
static __device__ __forceinline__ int8_t float_to_int8_rn(float const x) {
#ifdef USE_ROCM
static const float i8_min =
static_cast<float>(std::numeric_limits<int8_t>::min());
static const float i8_max =
static_cast<float>(std::numeric_limits<int8_t>::max());
// round
float dst = std::nearbyint(x);
// saturate
// See https://github.com/pytorch/pytorch/issues/127666
// See https://github.com/llvm/llvm-project/issues/95183
// hip-clang std::clamp __glibcxx_assert_fail host function when building on
// Arch/gcc14. The following replaces std::clamp usage with similar logic
// dst = std::clamp(dst, i8_min, i8_max);
dst = (dst < i8_min) ? i8_min : (dst > i8_max) ? i8_max : dst;
return static_cast<int8_t>(dst);
#else
// CUDA path
uint32_t dst;
asm volatile("cvt.rni.sat.s8.f32 %0, %1;" : "=r"(dst) : "f"(x));
return reinterpret_cast<const int8_t&>(dst);
#endif
}
template <typename fp8_type>
static __device__ __forceinline__ fp8_type float_to_fp8(float const x) {
float const r =
fmax(-quant_type_max_v<fp8_type>, fmin(x, quant_type_max_v<fp8_type>));
return static_cast<fp8_type>(r);
}
template <typename quant_type_t, bool is_scale_inverted, typename enable = void>
struct ScaledQuant;
template <typename quant_type_t, bool is_scale_inverted>
struct ScaledQuant<
quant_type_t, is_scale_inverted,
typename std::enable_if_t<std::is_same_v<quant_type_t, int8_t>>> {
static __device__ __forceinline__ quant_type_t quant_fn(float const x,
float const scale) {
if constexpr (is_scale_inverted) {
return float_to_int8_rn(x * scale);
} else {
return float_to_int8_rn(x / scale);
}
}
};
template <typename quant_type_t, bool is_scale_inverted>
struct ScaledQuant<quant_type_t, is_scale_inverted,
typename std::enable_if_t<
std::is_same_v<quant_type_t, c10::Float8_e4m3fn> ||
std::is_same_v<quant_type_t, c10::Float8_e4m3fnuz>>> {
static __device__ __forceinline__ quant_type_t quant_fn(float const x,
float const scale) {
if constexpr (is_scale_inverted) {
return float_to_fp8<quant_type_t>(x * scale);
} else {
return float_to_fp8<quant_type_t>(x / scale);
}
}
};
template <typename scalar_t, typename quant_type_t, bool is_scale_inverted>
__device__ void scaled_quant_conversion(quant_type_t* __restrict__ output,
scalar_t const* __restrict__ input,
float const scale, int const tid,
int const num_elements,
int const step) {
for (int i = tid; i < num_elements; i += step) {
output[i] = ScaledQuant<quant_type_t, is_scale_inverted>(input[i], scale);
}
}
} // namespace vllm