biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build] Enforce style for C++ and CUDA code with clang-format (#4722)

Browse Source
This commit is contained in:
Michael Goin
2024-05-22 03:18:41 -04:00
committed by GitHub
parent 9b9a10d6cb
commit 5f6d10c14c
64 changed files with 6398 additions and 6790 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
csrc/quantization/fp8/common.cu
View File

@@ -10,17 +10,20 @@
namespace vllm {
__device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
float old;
old = (value >= 0) ? __int_as_float(atomicMax((int*)addr, __float_as_int(value))) :
__uint_as_float(atomicMin((unsigned int*)addr, __float_as_uint(value)));
float old;
old = (value >= 0)
? __int_as_float(atomicMax((int*)addr, __float_as_int(value)))
: __uint_as_float(
atomicMin((unsigned int*)addr, __float_as_uint(value)));
return old;
return old;
}
#define FP8_E4M3_MAX std::numeric_limits<c10::Float8_e4m3fn>::max()
template<typename scalar_t>
__device__ __forceinline__ c10::Float8_e4m3fn scaled_fp8_conversion(const scalar_t val, const float scale) {
template <typename scalar_t>
__device__ __forceinline__ c10::Float8_e4m3fn scaled_fp8_conversion(
const scalar_t val, const float scale) {
float x = static_cast<float>(val) / scale;
float r = fmax(-FP8_E4M3_MAX, fmin(x, FP8_E4M3_MAX));
return static_cast<c10::Float8_e4m3fn>(r);
@@ -32,11 +35,10 @@ __device__ __forceinline__ c10::Float8_e4m3fn scaled_fp8_conversion(const scalar
// So to get the right answer, *scale needs to be initialized to
// a value <= 0.0 and we need to wait for all thread blocks to
// finish before consuming *scale.
template<typename scalar_t>
__global__ void segmented_max_reduction(
float* __restrict__ scale,
const scalar_t* __restrict__ input,
int64_t num_elems) {
template <typename scalar_t>
__global__ void segmented_max_reduction(float* __restrict__ scale,
const scalar_t* __restrict__ input,
int64_t num_elems) {
__shared__ float cache[1024];
int i = blockDim.x * blockIdx.x + threadIdx.x;
@@ -56,7 +58,7 @@ __global__ void segmented_max_reduction(
int ib = blockDim.x / 2;
while (ib != 0) {
if (threadIdx.x < ib && cache[threadIdx.x + ib] > cache[threadIdx.x]) {
cache[threadIdx.x] = cache[threadIdx.x + ib];
cache[threadIdx.x] = cache[threadIdx.x + ib];
}
__syncthreads();
ib /= 2;
@@ -64,16 +66,16 @@ __global__ void segmented_max_reduction(
// Finally, since cache[0] contains the maximum for this thread block,
// atomically write the max to the target location
if (threadIdx.x == 0) {
atomicMaxFloat(scale, cache[0] / std::numeric_limits<c10::Float8_e4m3fn>::max());
atomicMaxFloat(scale,
cache[0] / std::numeric_limits<c10::Float8_e4m3fn>::max());
}
}
template<typename scalar_t>
__global__ void scaled_fp8_quant_kernel(
c10::Float8_e4m3fn* __restrict__ out,
const scalar_t* __restrict__ input,
const float* __restrict__ scale,
int64_t num_elems) {
template <typename scalar_t>
__global__ void scaled_fp8_quant_kernel(c10::Float8_e4m3fn* __restrict__ out,
const scalar_t* __restrict__ input,
const float* __restrict__ scale,
int64_t num_elems) {
int i = blockDim.x * blockIdx.x + threadIdx.x;
while (i < num_elems) {
out[i] = scaled_fp8_conversion(input[i], *scale);
@@ -81,12 +83,11 @@ __global__ void scaled_fp8_quant_kernel(
}
}
} // namespace vllm
} // namespace vllm
void static_scaled_fp8_quant(
torch::Tensor& out, // [..., d]
torch::Tensor& input, // [..., d]
torch::Tensor& scale) // [1]
void static_scaled_fp8_quant(torch::Tensor& out, // [..., d]
torch::Tensor& input, // [..., d]
torch::Tensor& scale) // [1]
{
int64_t num_tokens = input.numel() / input.size(-1);
int64_t num_elems = input.numel();
@@ -95,21 +96,16 @@ void static_scaled_fp8_quant(
const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(),
"scaled_fp8_quant_kernel",
[&] {
vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
out.data_ptr<c10::Float8_e4m3fn>(),
input.data_ptr<scalar_t>(),
scale.data_ptr<float>(),
num_elems);
input.scalar_type(), "scaled_fp8_quant_kernel", [&] {
vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
out.data_ptr<c10::Float8_e4m3fn>(), input.data_ptr<scalar_t>(),
scale.data_ptr<float>(), num_elems);
});
}
void dynamic_scaled_fp8_quant(
torch::Tensor& out, // [..., d]
torch::Tensor& input, // [..., d]
torch::Tensor& scale) // [1]
void dynamic_scaled_fp8_quant(torch::Tensor& out, // [..., d]
torch::Tensor& input, // [..., d]
torch::Tensor& scale) // [1]
{
int64_t num_tokens = input.numel() / input.size(-1);
int64_t num_elems = input.numel();
@@ -118,18 +114,11 @@ void dynamic_scaled_fp8_quant(
const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
VLLM_DISPATCH_FLOATING_TYPES(
input.scalar_type(),
"scaled_fp8_quant_kernel",
[&] {
vllm::segmented_max_reduction<scalar_t><<<grid, block, 0, stream>>>(
scale.data_ptr<float>(),
input.data_ptr<scalar_t>(),
num_elems);
vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
out.data_ptr<c10::Float8_e4m3fn>(),
input.data_ptr<scalar_t>(),
scale.data_ptr<float>(),
num_elems);
input.scalar_type(), "scaled_fp8_quant_kernel", [&] {
vllm::segmented_max_reduction<scalar_t><<<grid, block, 0, stream>>>(
scale.data_ptr<float>(), input.data_ptr<scalar_t>(), num_elems);
vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
out.data_ptr<c10::Float8_e4m3fn>(), input.data_ptr<scalar_t>(),
scale.data_ptr<float>(), num_elems);
});
}