[Misc] Consolidate csrc shared headers and prune dead code (#60721)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Misha Goin
2026-10-12 09:46:22 -07:00
committed by GitHub
co-authored by Claude
parent e0317b28ce
commit d79b52b182
242 changed files with 2783 additions and 3638 deletions
@@ -5,7 +5,7 @@
#include <hip/hip_bf16.h>
#include <hip/hip_bfloat16.h>
#include "../../../../attention/attention_dtypes.h"
#include "core/vec_dtypes/vec_dtypes.h"
#include <torch/headeronly/core/ScalarType.h>
namespace vllm {
+506
View File
@@ -0,0 +1,506 @@
#include "common.cuh"
#include "core/cuda_compat.h"
#include "../../../dispatch_utils.h"
#include "../../../cub_helpers.h"
#include "../../vectorization_utils.cuh"
#include "../../../torch_utils.h"
#include <torch/csrc/stable/macros.h>
#include <algorithm>
namespace vllm {
namespace {
// 16 fp8 outputs form one 16-byte store; the matching 16-element input load
// is 32 B for 16-bit types and 64 B for fp32.
constexpr int kFp8QuantVecSize = 16;
constexpr int kFp8QuantBlockSize = 256;
template <typename T>
using fp8_quant_vec_t = vec_n_t<T, kFp8QuantVecSize>;
template <typename T>
bool ptr_vec_aligned(const T* ptr) {
return reinterpret_cast<uintptr_t>(ptr) % alignof(fp8_quant_vec_t<T>) == 0;
}
__device__ __forceinline__ float warp_reduce_max(float v) {
#pragma unroll
for (int offset = WARP_SIZE / 2; offset > 0; offset >>= 1) {
v = fmaxf(v, VLLM_SHFL_XOR_SYNC(v, offset));
}
return v;
}
// Block-wide max, returned to every thread. Warp partials go through one
// shared-memory exchange, so the whole reduction costs a single
// __syncthreads. blockDim.x must be a multiple of WARP_SIZE and at most
// kFp8QuantBlockSize.
__device__ __forceinline__ float block_reduce_max(float v, float* warp_max) {
const int lane = threadIdx.x % WARP_SIZE;
const int warp = threadIdx.x / WARP_SIZE;
const int num_warps = blockDim.x / WARP_SIZE;
v = warp_reduce_max(v);
if (lane == 0) {
warp_max[warp] = v;
}
__syncthreads();
v = warp_max[0];
for (int w = 1; w < num_warps; ++w) {
v = fmaxf(v, warp_max[w]);
}
return v;
}
} // namespace
// STRIDE_I_ZERO: true if scale_stride_i == 0 (per-tensor or per-channel)
// STRIDE_J_ZERO: true if scale_stride_j == 0 (per-tensor or per-token)
template <typename scalar_t, typename fp8_type, bool STRIDE_I_ZERO,
bool STRIDE_J_ZERO>
__global__ void scaled_fp8_quant_kernel_strided_group_shape(
fp8_type* __restrict__ out, const scalar_t* __restrict__ input,
const float* __restrict__ scale, int hidden_size, int64_t in_row_stride,
int64_t out_row_stride, int group_m, int group_n, int64_t scale_stride_i,
int64_t scale_stride_j) {
const int64_t token_idx = blockIdx.x;
const int tid = threadIdx.x;
const scalar_t* token_in = input + token_idx * in_row_stride;
fp8_type* token_out = out + token_idx * out_row_stride;
// Precompute row-level base offset for scale access (compile-time eliminated
// when STRIDE_I_ZERO)
const int64_t scale_row_base =
STRIDE_I_ZERO ? 0
: static_cast<int>(token_idx) / group_m * scale_stride_i;
auto get_inv_scale = [&](int gj) {
return 1.0f / scale[scale_row_base + gj * scale_stride_j];
};
int cached_gj = -1;
float cached_inv_scale = 0.0f;
auto get_inv_scale_cached = [&](int gj) {
if (gj != cached_gj) {
cached_inv_scale = 1.0f / scale[scale_row_base + gj * scale_stride_j];
cached_gj = gj;
}
return cached_inv_scale;
};
constexpr int VEC_SIZE = 16; // FP8 so vectorize to 128 bits
auto scaled_fp8_conversion_vectorized = [&](const scalar_t* in, fp8_type* out,
int size, float inv_scale) {
vectorize_with_alignment<VEC_SIZE>(
in, out, size, tid, blockDim.x,
[=] __device__(fp8_type & dst, const scalar_t& src) {
dst = scaled_fp8_conversion<true, fp8_type>(static_cast<float>(src),
inv_scale);
});
};
if (STRIDE_J_ZERO && hidden_size % VEC_SIZE == 0) {
// Per-tensor or per-token: single scale per row, vectorize full row
scaled_fp8_conversion_vectorized(token_in, token_out, hidden_size,
get_inv_scale(0));
} else if (group_n % VEC_SIZE == 0) {
// Multiple column groups with vectorization
const int num_groups_n = hidden_size / group_n;
for (int gj = 0; gj < num_groups_n; gj++) {
scaled_fp8_conversion_vectorized(token_in + gj * group_n,
token_out + gj * group_n, group_n,
get_inv_scale(gj));
}
} else {
// Scalar path for small column groups (group_n < VEC_SIZE)
for (int n = tid; n < hidden_size; n += blockDim.x) {
const int gj = n / group_n;
token_out[n] = scaled_fp8_conversion<true, fp8_type>(
static_cast<float>(token_in[n]), get_inv_scale_cached(gj));
}
}
}
// Per-tensor abs-max of a contiguous tensor viewed as a flat array of
// 16-element vectors: grid-stride loop, one atomic per block.
template <typename scalar_t, typename fp8_type>
__global__ void __launch_bounds__(kFp8QuantBlockSize)
absmax_reduction_flat(float* __restrict__ scale,
const scalar_t* __restrict__ input,
int64_t num_vecs) {
using in_vec_t = fp8_quant_vec_t<scalar_t>;
const auto* v_in = reinterpret_cast<const in_vec_t*>(input);
const int64_t stride = static_cast<int64_t>(gridDim.x) * blockDim.x;
float thread_max = 0.0f;
#pragma unroll 4
for (int64_t v = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
v < num_vecs; v += stride) {
const in_vec_t vec = v_in[v];
#pragma unroll
for (int i = 0; i < kFp8QuantVecSize; ++i) {
thread_max = fmaxf(thread_max, fabsf(static_cast<float>(vec.val[i])));
}
}
__shared__ float warp_max[kFp8QuantBlockSize / 32];
const float block_max = block_reduce_max(thread_max, warp_max);
if (threadIdx.x == 0) {
atomicMaxFloat(scale, block_max / quant_type_max_v<fp8_type>);
}
}
// Per-tensor abs-max for strided or unaligned rows: one block per token.
template <typename scalar_t, typename fp8_type>
__global__ void __launch_bounds__(kFp8QuantBlockSize)
segmented_max_reduction_strided(float* __restrict__ scale,
const scalar_t* __restrict__ input,
int hidden_size, int64_t in_row_stride,
int64_t num_tokens) {
const int64_t token_idx = blockIdx.x;
if (token_idx >= num_tokens) {
return;
}
const scalar_t* row_ptr = input + token_idx * in_row_stride;
float thread_max = 0.0f;
vectorize_read_with_alignment<kFp8QuantVecSize>(
row_ptr, hidden_size, threadIdx.x, blockDim.x,
[&] __device__(const scalar_t& v) {
thread_max = fmaxf(thread_max, fabsf(static_cast<float>(v)));
});
__shared__ float warp_max[kFp8QuantBlockSize / 32];
const float block_max = block_reduce_max(thread_max, warp_max);
if (threadIdx.x == 0) {
atomicMaxFloat(scale, block_max / quant_type_max_v<fp8_type>);
}
}
template <typename scalar_t, typename fp8_type>
__global__ void scaled_fp8_quant_kernel_strided_dynamic(
fp8_type* __restrict__ out, const scalar_t* __restrict__ input,
const float* __restrict__ scale, int hidden_size, int64_t in_row_stride,
int64_t out_row_stride) {
const int64_t token_idx = blockIdx.x;
const int tid = threadIdx.x;
const scalar_t* token_in = input + token_idx * in_row_stride;
fp8_type* token_out = out + token_idx * out_row_stride;
const float reciprocal_scale = 1.0f / (*scale);
vectorize_with_alignment<16>(
token_in, token_out, hidden_size, tid, blockDim.x,
[=] __device__(fp8_type & dst, const scalar_t& src) {
dst = scaled_fp8_conversion<true, fp8_type>(static_cast<float>(src),
reciprocal_scale);
});
}
template <typename scalar_t, typename fp8_type>
__global__ void dynamic_per_token_scaled_fp8_quant_kernel_strided(
fp8_type* __restrict__ out, float* __restrict__ scale,
const scalar_t* __restrict__ input, const float* __restrict__ scale_ub,
int hidden_size, int64_t in_row_stride, int64_t out_row_stride) {
const int64_t token_idx = blockIdx.x;
const int tid = threadIdx.x;
// Use int64 to avoid overflowing an int32 when calculating this offset
int64_t in_offset = static_cast<int64_t>(token_idx) * in_row_stride;
int64_t out_offset = static_cast<int64_t>(token_idx) * out_row_stride;
const scalar_t* token_in = input + in_offset;
fp8_type* token_out = out + out_offset;
// 1) per-token absmax
float absmax_val = 0.f;
vectorize_read_with_alignment<16>(
token_in, hidden_size, tid, blockDim.x, [&] __device__(scalar_t v) {
absmax_val = fmaxf(absmax_val, fabsf(static_cast<float>(v)));
});
using BlockReduce = cub::BlockReduce<float, 256>;
__shared__ typename BlockReduce::TempStorage tmp;
const float block_max =
BlockReduce(tmp).Reduce(absmax_val, CubMaxOp{}, blockDim.x);
__shared__ float token_scale;
if (tid == 0) {
token_scale = scale_ub ? fminf(block_max, *scale_ub) : block_max;
token_scale = fmaxf(token_scale / quant_type_max_v<fp8_type>,
min_scaling_factor<fp8_type>::val());
scale[token_idx] = token_scale;
}
__syncthreads();
// 2) quantize
vectorize_with_alignment<16>(
token_in, token_out, hidden_size, tid, blockDim.x,
[=] __device__(fp8_type & dst, const scalar_t& src) {
dst = scaled_fp8_conversion<false, fp8_type>(static_cast<float>(src),
token_scale);
});
}
} // namespace vllm
void static_scaled_fp8_quant(
torch::stable::Tensor& out, // [..., d]
torch::stable::Tensor const& input, // [..., d]
torch::stable::Tensor const& scale, // various shapes
std::optional<torch::headeronly::IntHeaderOnlyArrayRef>
opt_group_shape) // optional explicit [group_m, group_n]
{
STD_TORCH_CHECK(input.stride(-1) == 1,
"last dimension of input must be contiguous");
STD_TORCH_CHECK(out.stride(-1) == 1,
"last dimension of output must be contiguous");
const int hidden_size = input.size(-1); // N (columns)
const int num_tokens = input.numel() / hidden_size; // M (rows)
// Determine group_m, group_n, and scale strides from scale shape
// Scale indexing: scale[gi * scale_stride_j + gj * scale_stride_i]
// where gi = m / group_m, gj = n / group_n
int group_m, group_n;
int64_t scale_stride_i, scale_stride_j;
if (scale.dim() == 0 || scale.numel() == 1) {
// Per-tensor: one scale for the entire tensor
group_m = num_tokens;
group_n = hidden_size;
scale_stride_i = 0;
scale_stride_j = 0;
} else if (scale.dim() == 1) {
// 1D scale: require explicit group_shape to disambiguate per-channel vs
// per-token (avoids edge case where num_tokens == hidden_size)
STD_TORCH_CHECK(
opt_group_shape.has_value(),
"1D scale requires explicit group_shape to disambiguate "
"per-channel vs per-token quantization. "
"Use group_shape=(-1, 1) for per-channel or group_shape=(1, "
"-1) for per-token.");
STD_TORCH_CHECK(opt_group_shape->size() == 2,
"group_shape must have exactly 2 elements, got ",
opt_group_shape->size());
const auto opt_group_m = (*opt_group_shape)[0];
const auto opt_group_n = (*opt_group_shape)[1];
group_m = opt_group_m == -1 ? num_tokens : static_cast<int>(opt_group_m);
group_n = opt_group_n == -1 ? hidden_size : static_cast<int>(opt_group_n);
// Validate the explicit group shape matches the 1D scale
const int64_t scale_len = scale.numel();
const int64_t expected_scale_m = num_tokens / group_m;
const int64_t expected_scale_n = hidden_size / group_n;
const int64_t expected_scale_numel = expected_scale_m * expected_scale_n;
STD_TORCH_CHECK(scale_len == expected_scale_numel, "1D scale length (",
scale_len, ") does not match expected size (",
expected_scale_numel, ") for group_shape (", opt_group_m,
", ", opt_group_n, ") with input shape (", num_tokens, ", ",
hidden_size, ")");
// For 1D scale, determine strides based on which dim is trivial
// Scale indexing: scale[gi * scale_stride_i + gj * scale_stride_j]
// where gi = m / group_m (row group), gj = n / group_n (col group)
if (expected_scale_m == 1) {
// Per-channel style: one scale in M dim, scale varies along N
// gi = 0 always, gj varies, so stride_1 traverses the scale
scale_stride_i = 0;
scale_stride_j = scale.stride(0);
} else if (expected_scale_n == 1) {
// Per-token style: one scale in N dim, scale varies along M
// gj = 0 always, gi varies, so stride_0 traverses the scale
scale_stride_i = scale.stride(0);
scale_stride_j = 0;
} else {
STD_TORCH_CHECK(
false,
"1D scale can only be used when one of the scale dimensions is 1. "
"For 2D group scaling, use a 2D scale tensor.");
}
} else if (scale.dim() == 2) {
// 2D scale: infer group sizes from scale dimensions (or use explicit if
// provided)
const int64_t scale_size_0 = scale.size(0);
const int64_t scale_size_1 = scale.size(1);
STD_TORCH_CHECK(num_tokens % scale_size_0 == 0, "num_tokens (", num_tokens,
") must be divisible by scale.size(0) (", scale_size_0,
")");
STD_TORCH_CHECK(hidden_size % scale_size_1 == 0, "hidden_size (",
hidden_size, ") must be divisible by scale.size(1) (",
scale_size_1, ")");
// Infer from 2D scale shape
int inferred_group_m = num_tokens / scale_size_0;
int inferred_group_n = hidden_size / scale_size_1;
// Use explicit if provided, otherwise use inferred
if (opt_group_shape.has_value()) {
STD_TORCH_CHECK(opt_group_shape->size() == 2,
"group_shape must have exactly 2 elements, got ",
opt_group_shape->size());
const auto opt_group_m = (*opt_group_shape)[0];
const auto opt_group_n = (*opt_group_shape)[1];
group_m = opt_group_m == -1 ? num_tokens : static_cast<int>(opt_group_m);
group_n = opt_group_n == -1 ? hidden_size : static_cast<int>(opt_group_n);
// Validate explicit matches inferred
STD_TORCH_CHECK(
group_m == inferred_group_m && group_n == inferred_group_n,
"Explicit group_shape (", opt_group_m, ", ", opt_group_n,
") does not match inferred group shape (", inferred_group_m, ", ",
inferred_group_n, ") from 2D scale tensor shape (", scale_size_0,
", ", scale_size_1, ")");
} else {
group_m = inferred_group_m;
group_n = inferred_group_n;
}
scale_stride_i = scale.stride(0);
scale_stride_j = scale.stride(1);
} else {
STD_TORCH_CHECK(false, "scale must be 0D, 1D, or 2D tensor, but got ",
scale.dim(), "D");
}
const int block_size = 256;
dim3 grid(num_tokens);
dim3 block(block_size);
const int64_t in_row_stride = input.stride(-2);
const int64_t out_row_stride = out.stride(-2);
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
const cudaStream_t stream = get_current_cuda_stream();
// Dispatch to template-specialized kernel based on stride pattern
VLLM_STABLE_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "scaled_fp8_quant_kernel_scalar_type", [&] {
VLLM_STABLE_DISPATCH_FP8_TYPES(
out.scalar_type(), "scaled_fp8_quant_kernel_fp8_type", [&] {
VLLM_STABLE_DISPATCH_BOOL(scale_stride_i == 0, S0_ZERO, [&] {
VLLM_STABLE_DISPATCH_BOOL(scale_stride_j == 0, S1_ZERO, [&] {
vllm::scaled_fp8_quant_kernel_strided_group_shape<
scalar_t, fp8_t, S0_ZERO, S1_ZERO>
<<<grid, block, 0, stream>>>(
out.mutable_data_ptr<fp8_t>(),
input.const_data_ptr<scalar_t>(),
scale.const_data_ptr<float>(), hidden_size,
in_row_stride, out_row_stride, group_m, group_n,
scale_stride_i, scale_stride_j);
});
});
});
});
}
void dynamic_scaled_fp8_quant(torch::stable::Tensor& out, // [..., d]
torch::stable::Tensor const& input, // [..., d]
torch::stable::Tensor& scale) // [1]
{
STD_TORCH_CHECK(input.stride(-1) == 1,
"last dimension of input must be contiguous");
STD_TORCH_CHECK(out.stride(-1) == 1,
"last dimension of output must be contiguous");
const int hidden_size = input.size(-1);
const int num_tokens = input.numel() / hidden_size;
const int block_size = vllm::kFp8QuantBlockSize;
dim3 grid(num_tokens);
dim3 block(block_size);
const int64_t in_row_stride = input.stride(-2);
const int64_t out_row_stride = out.stride(-2);
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
const cudaStream_t stream = get_current_cuda_stream();
// scale tensor should be initialised to <=0 before reduction
STD_CUDA_CHECK(cudaMemsetAsync(scale.mutable_data_ptr<float>(), 0,
sizeof(float), stream));
VLLM_STABLE_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "scaled_fp8_quant_kernel_scalar_type", [&] {
VLLM_STABLE_DISPATCH_FP8_TYPES(
out.scalar_type(), "scaled_fp8_quant_kernel_fp8_type", [&] {
const scalar_t* in_ptr = input.const_data_ptr<scalar_t>();
const bool flat = in_row_stride == hidden_size &&
hidden_size % vllm::kFp8QuantVecSize == 0 &&
vllm::ptr_vec_aligned(in_ptr);
if (flat) {
const int64_t num_vecs = input.numel() / vllm::kFp8QuantVecSize;
const int64_t max_blocks =
static_cast<int64_t>(
get_device_prop()->multiProcessorCount) *
8;
const int64_t num_blocks = std::max<int64_t>(
1,
std::min<int64_t>(
max_blocks, (num_vecs + block_size - 1) / block_size));
vllm::absmax_reduction_flat<scalar_t, fp8_t>
<<<num_blocks, block, 0, stream>>>(
scale.mutable_data_ptr<float>(), in_ptr, num_vecs);
} else {
vllm::segmented_max_reduction_strided<scalar_t, fp8_t>
<<<grid, block, 0, stream>>>(
scale.mutable_data_ptr<float>(), in_ptr, hidden_size,
in_row_stride, static_cast<int64_t>(num_tokens));
}
vllm::scaled_fp8_quant_kernel_strided_dynamic<scalar_t, fp8_t>
<<<grid, block, 0, stream>>>(out.mutable_data_ptr<fp8_t>(),
input.const_data_ptr<scalar_t>(),
scale.const_data_ptr<float>(),
hidden_size, in_row_stride,
out_row_stride);
});
});
}
void dynamic_per_token_scaled_fp8_quant(
torch::stable::Tensor& out, // [..., d]
torch::stable::Tensor const& input, // [..., d]
torch::stable::Tensor& scales,
std::optional<torch::stable::Tensor> const& scale_ub) {
STD_TORCH_CHECK(input.stride(-1) == 1,
"last dimension of input must be contiguous");
STD_TORCH_CHECK(out.stride(-1) == 1,
"last dimension of output must be contiguous");
const int hidden_size = input.size(-1);
const int num_tokens = input.numel() / hidden_size;
const int block_size = 256;
dim3 grid(num_tokens);
dim3 block(std::min(hidden_size, block_size));
const int64_t in_row_stride = input.stride(-2);
const int64_t out_row_stride = out.stride(-2);
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
const cudaStream_t stream = get_current_cuda_stream();
VLLM_STABLE_DISPATCH_FLOATING_TYPES(
input.scalar_type(),
"dynamic_per_token_scaled_fp8_quant_kernel_scalar_type", [&] {
VLLM_STABLE_DISPATCH_FP8_TYPES(
out.scalar_type(),
"dynamic_per_token_scaled_fp8_quant_kernel_fp8_type", [&] {
vllm::dynamic_per_token_scaled_fp8_quant_kernel_strided<scalar_t,
fp8_t>
<<<grid, block, 0, stream>>>(
out.mutable_data_ptr<fp8_t>(),
scales.mutable_data_ptr<float>(),
input.const_data_ptr<scalar_t>(),
scale_ub.has_value() ? scale_ub->const_data_ptr<float>()
: nullptr,
hidden_size, in_row_stride, out_row_stride);
});
});
}
+8 -8
View File
@@ -1,17 +1,17 @@
#pragma once
#include "libtorch_stable/quantization/vectorization.cuh"
#include "../../utils.cuh"
#include "quantization/vectorization.cuh"
#include "quantization/utils.cuh"
#include <cmath>
// This header is shared between _C and _C_stable_libtorch targets.
// This header is shared between _rocm_C and _C_stable_libtorch targets.
// torch_utils.h provides get_device_prop(). We need to pass USE_CUDA
// to the .so to expose some of the shims used by torch_utils.h. For now
// this is only done for _C_stable_libtorch and not for _C, so we use the
// non stable at::cuda::getCurrentDeviceProperties for _C for now.
// this is only done for _C_stable_libtorch and not for _rocm_C, so we use
// the non stable at::cuda::getCurrentDeviceProperties for _rocm_C for now.
#ifdef TORCH_TARGET_VERSION
#include "../../../libtorch_stable/torch_utils.h"
#include "torch_utils.h"
#else
#ifdef USE_ROCM
#include <ATen/hip/HIPContext.h>
@@ -19,9 +19,9 @@
#endif
#ifndef USE_ROCM
#include "nvidia/quant_utils.cuh"
#include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
#else
#include "amd/quant_utils.cuh"
#include "quantization/w8a8/fp8/amd/quant_utils.cuh"
#endif
// Determines the preferred FP8 type for the current platform.
@@ -1,6 +1,6 @@
#pragma once
#include "../../../../attention/attention_dtypes.h"
#include "core/vec_dtypes/vec_dtypes.h"
#include <torch/headeronly/core/ScalarType.h>
#include <assert.h>
#include <float.h>
@@ -243,7 +243,7 @@ __inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, __nv_bfloat16>(
}
// torch stable-ABI (headeronly) scalar types delegate to the CUDA-native
// conversions, so libtorch_stable kernels dispatched on c10::BFloat16 /
// conversions, so stable-ABI kernels dispatched on c10::BFloat16 /
// c10::Half quantize correctly without manual casts.
template <>
__inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, c10::BFloat16>(
@@ -0,0 +1,881 @@
#include <torch/csrc/stable/tensor.h>
#include <torch/csrc/stable/ops.h>
#include <torch/headeronly/util/Exception.h>
#include <torch/headeronly/core/ScalarType.h>
#include <cmath>
#ifdef USE_ROCM
#include <hip/hip_fp8.h>
#else
#include <cuda_fp8.h>
#endif
#include "quantization/vectorization.cuh"
#include "quantization/vectorization_utils.cuh"
#include "dispatch_utils.h"
#include "torch_utils.h"
__device__ __forceinline__ float GroupReduceMax(float val) {
#ifdef USE_ROCM
// 16-thread logical groups may pack up to four per 64-lane wavefront; use a
// 64-bit mask and explicit width so shuffles stay within each group.
const int lane_in_wave = threadIdx.x % warpSize;
const unsigned long long mask = 0xFFFFull << ((lane_in_wave / 16) * 16);
val = fmaxf(val, __shfl_xor_sync(mask, val, 8, 16));
val = fmaxf(val, __shfl_xor_sync(mask, val, 4, 16));
val = fmaxf(val, __shfl_xor_sync(mask, val, 2, 16));
val = fmaxf(val, __shfl_xor_sync(mask, val, 1, 16));
#else
unsigned mask = threadIdx.x % 32 >= 16 ? 0xffff0000 : 0x0000ffff;
val = fmaxf(val, __shfl_xor_sync(mask, val, 8));
val = fmaxf(val, __shfl_xor_sync(mask, val, 4));
val = fmaxf(val, __shfl_xor_sync(mask, val, 2));
val = fmaxf(val, __shfl_xor_sync(mask, val, 1));
#endif
return val;
}
template <typename T, bool SCALE_UE8M0>
__device__ __forceinline__ float ComputeGroupScale(
const T* __restrict__ group_input, T* __restrict__ smem_group,
const int group_size, const int lane_id, const int threads_per_group,
const float eps, const float max_8bit) {
float local_absmax = eps;
constexpr int vec_size = 16 / sizeof(T);
// copy global -> shared & compute absmax
auto scalar_op_cache = [&] __device__(T & dst, const T& src) {
float abs_v = fabsf(static_cast<float>(src));
local_absmax = fmaxf(local_absmax, abs_v);
dst = src;
};
vllm::vectorize_with_alignment<vec_size>(
group_input, // in
smem_group, // out (shared)
group_size, // elements per group
lane_id, // thread id
threads_per_group, // stride in group
scalar_op_cache); // scalar handler
local_absmax = GroupReduceMax(local_absmax);
float y_s = local_absmax / max_8bit;
if constexpr (SCALE_UE8M0) {
y_s = exp2f(ceilf(log2f(fmaxf(fabsf(y_s), 1e-10f))));
}
return y_s;
}
template <typename T, typename DST_DTYPE>
__device__ __forceinline__ void QuantizeGroup(
const T* __restrict__ smem_group, DST_DTYPE* __restrict__ group_output,
const int group_size, const int lane_id, const int threads_per_group,
const float y_s, const float min_8bit, const float max_8bit) {
constexpr int vec_size = 16 / sizeof(T);
// quantize shared -> global 8-bit
auto scalar_op_quant = [&] __device__(DST_DTYPE & dst, const T& src) {
float q = fminf(fmaxf(static_cast<float>(src) / y_s, min_8bit), max_8bit);
dst = DST_DTYPE(q);
};
vllm::vectorize_with_alignment<vec_size>(
smem_group, // in (shared)
group_output, // out (global quant tensor)
group_size, // elements
lane_id, // tid
threads_per_group, // stride
scalar_op_quant); // scalar handler
}
template <typename T, typename DST_DTYPE, bool IS_COLUMN_MAJOR = false,
bool SCALE_UE8M0 = false, typename scale_packed_t = float,
bool CHECK_BOUNDS = false>
__global__ void per_token_group_quant_8bit_kernel(
const T* __restrict__ input, void* __restrict__ output_q,
scale_packed_t* __restrict__ output_s, const int group_size,
const int num_groups, const int groups_per_block, const float eps,
const float min_8bit, const float max_8bit, const int scale_num_rows = 0,
const int scale_stride = 0) {
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;
const int64_t block_group_id = blockIdx.x * groups_per_block;
const int64_t global_group_id = block_group_id + local_group_id;
const int64_t block_group_offset = global_group_id * group_size;
using scale_element_t = float;
static_assert(sizeof(scale_packed_t) % sizeof(scale_element_t) == 0);
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaGridDependencySynchronize();
#endif
// 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;
const bool valid_group = !CHECK_BOUNDS || global_group_id < num_groups;
float y_s = 1.0f;
if (valid_group) {
const T* group_input = input + block_group_offset;
scale_element_t* scale_output;
if constexpr (IS_COLUMN_MAJOR) {
const int num_elems_per_pack =
static_cast<int>(sizeof(scale_packed_t) / sizeof(scale_element_t));
const int scale_num_rows_element = scale_num_rows * num_elems_per_pack;
const int row_idx = global_group_id / scale_num_rows_element;
const int col_idx_raw = global_group_id % scale_num_rows_element;
const int col_idx = col_idx_raw / num_elems_per_pack;
const int pack_idx = col_idx_raw % num_elems_per_pack;
scale_output = reinterpret_cast<scale_element_t*>(output_s) +
(col_idx * scale_stride * num_elems_per_pack +
row_idx * num_elems_per_pack + pack_idx);
} else {
scale_output = output_s + global_group_id;
}
y_s = ComputeGroupScale<T, SCALE_UE8M0>(group_input, smem_group, group_size,
lane_id, threads_per_group, eps,
max_8bit);
scale_element_t y_s_quant = y_s;
if (lane_id == 0) {
*scale_output = y_s_quant;
}
}
__syncthreads();
if (valid_group) {
DST_DTYPE* group_output =
static_cast<DST_DTYPE*>(output_q) + block_group_offset;
QuantizeGroup<T, DST_DTYPE>(smem_group, group_output, group_size, lane_id,
threads_per_group, y_s, min_8bit, max_8bit);
}
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaTriggerProgrammaticLaunchCompletion();
#endif
}
// Register-resident variant for the unpacked float32 scale layout, mirroring
// per_token_group_quant_8bit_packed_register_kernel below, which float32-scale
// callers cannot reach because it is hardwired to UE8M0 exponents packed
// 4-per-int32 with TMA-aligned strides.
//
// A group_size-128 group of 16-bit input is 256 B, so it fits in registers and
// the smem staging plus the __syncthreads() that the kernel above needs between
// its absmax and quantize passes are both avoidable.
//
// Requires GROUP_SIZE == THREADS_PER_GROUP * VEC_SIZE, i.e. group_size 128 with
// 16-bit input; CanUseRegisterPath enforces that.
template <typename T, typename DST_DTYPE, bool IS_COLUMN_MAJOR = false,
bool SCALE_UE8M0 = false, int GROUP_SIZE = 128>
__global__ void per_token_group_quant_8bit_register_kernel(
const T* __restrict__ input, void* __restrict__ output_q,
float* __restrict__ output_s, const int64_t num_groups, const float eps,
const float min_8bit, const float max_8bit, const int scale_num_rows,
const int scale_stride) {
constexpr int THREADS_PER_GROUP = 8;
constexpr int VEC_SIZE = 32 / sizeof(T);
static_assert(32 % THREADS_PER_GROUP == 0,
"THREADS_PER_GROUP must divide the warp size so the shuffle "
"mask stays inside one group");
static_assert(sizeof(DST_DTYPE) == 1,
"the uint4 store packs one byte per element");
// CanUseRegisterPath only routes 16-bit inputs here, but
// VLLM_STABLE_DISPATCH_FLOATING_TYPES instantiates every floating type, so
// gate the body on the shape this schedule can actually serve instead of a
// static_assert.
if constexpr (GROUP_SIZE != THREADS_PER_GROUP * VEC_SIZE) {
return;
} else {
const int64_t group_id =
(static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x) /
THREADS_PER_GROUP;
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaGridDependencySynchronize();
#endif
if (group_id >= num_groups) {
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaTriggerProgrammaticLaunchCompletion();
#endif
return;
}
const int lane_id = threadIdx.x % THREADS_PER_GROUP;
// alignas(16) is required for the uint4 reinterpret_cast: the default
// alignof for 16-bit T is 2.
alignas(16) T regs[VEC_SIZE];
{
const T* group_input = input + group_id * 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];
}
float local_absmax = eps;
#pragma unroll
for (int i = 0; i < VEC_SIZE; ++i) {
local_absmax = fmaxf(local_absmax, fabsf(static_cast<float>(regs[i])));
}
// Reduce within the 8-lane subgroup that shares this group.
#ifdef USE_ROCM
const int lane_in_wave = threadIdx.x % warpSize;
const unsigned long long mask = 0xFFull << (lane_in_wave & ~7);
local_absmax =
fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 4, 8));
local_absmax =
fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 2, 8));
local_absmax =
fmaxf(local_absmax, __shfl_xor_sync(mask, local_absmax, 1, 8));
#else
const 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));
#endif
float y_s = local_absmax / max_8bit;
if constexpr (SCALE_UE8M0) {
// Bit-exact with exp2f(ceilf(log2f(fmaxf(fabsf(y_s), 1e-10f)))): bump the
// exponent whenever the mantissa is non-zero, then clear the mantissa.
y_s = fmaxf(fabsf(y_s), 1e-10f);
const uint32_t bits = __float_as_uint(y_s);
const uint32_t exp_bits = (bits >> 23) & 0xffu;
const uint32_t mant_bits = bits & 0x7fffffu;
y_s = __uint_as_float((exp_bits + (mant_bits != 0u ? 1u : 0u)) << 23);
}
if (lane_id == 0) {
if constexpr (IS_COLUMN_MAJOR) {
const int64_t row_idx = group_id / scale_num_rows;
const int64_t col_idx = group_id % scale_num_rows;
output_s[col_idx * scale_stride + row_idx] = y_s;
} else {
output_s[group_id] = y_s;
}
}
// A UE8M0 scale is an exact power of two, so multiplying by the reciprocal
// matches dividing. Otherwise the reciprocal costs an ulp and can move
// values that land exactly on an 8-bit boundary, so divide instead to stay
// bit-exact with the smem kernel.
const float inv_y = 1.0f / y_s;
uint32_t packed[VEC_SIZE / 4] = {};
#pragma unroll
for (int i = 0; i < VEC_SIZE; ++i) {
const float v = static_cast<float>(regs[i]);
const float scaled = SCALE_UE8M0 ? (v * inv_y) : (v / y_s);
const float q = fminf(fmaxf(scaled, min_8bit), max_8bit);
DST_DTYPE qb = DST_DTYPE(q);
const uint8_t byte = *reinterpret_cast<uint8_t*>(&qb);
packed[i >> 2] |= static_cast<uint32_t>(byte) << ((i & 3) * 8);
}
DST_DTYPE* group_output = static_cast<DST_DTYPE*>(output_q) +
group_id * GROUP_SIZE + lane_id * VEC_SIZE;
*reinterpret_cast<uint4*>(group_output) =
make_uint4(packed[0], packed[1], packed[2], packed[3]);
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaTriggerProgrammaticLaunchCompletion();
#endif
}
}
// The register path needs group_size 128 with 16-bit input so that one group is
// exactly THREADS_PER_GROUP(8) x VEC_SIZE(16) elements, and it loads/stores
// through uint4, so both buffers must be 16 B aligned.
inline bool CanUseRegisterPath(const torch::stable::Tensor& input,
const torch::stable::Tensor& output_q,
int64_t group_size) {
const auto dtype = input.scalar_type();
if (group_size != 128) return false;
if (dtype != torch::headeronly::ScalarType::Half &&
dtype != torch::headeronly::ScalarType::BFloat16) {
return false;
}
if (!input.is_contiguous() || !output_q.is_contiguous()) return false;
const auto in_addr = reinterpret_cast<uintptr_t>(input.data_ptr());
const auto out_addr = reinterpret_cast<uintptr_t>(output_q.data_ptr());
return (in_addr % 16 == 0) && (out_addr % 16 == 0);
}
inline int GetGroupsPerBlock(int64_t num_groups) {
if (num_groups % 16 == 0) {
return 16;
}
if (num_groups % 8 == 0) {
return 8;
}
if (num_groups % 4 == 0) {
return 4;
}
if (num_groups % 2 == 0) {
return 2;
}
return 1;
}
// Largest divisor of padded_groups_per_row that is <= 16. ry = 16 / kx.
inline int GetGroupsPerBlockX(int64_t padded_groups_per_row) {
if (padded_groups_per_row % 16 == 0) {
return 16;
}
if (padded_groups_per_row % 8 == 0) {
return 8;
}
return 4;
}
// 8-bit per-token-group quantization helper used by both FP8 and INT8
static void per_token_group_quant_8bit(const torch::stable::Tensor& input,
torch::stable::Tensor& output_q,
torch::stable::Tensor& output_s,
int64_t group_size, double eps,
double min_8bit, double max_8bit,
bool scale_ue8m0 = false) {
STD_TORCH_CHECK(input.is_contiguous());
STD_TORCH_CHECK(output_q.is_contiguous());
const int num_groups = input.numel() / group_size;
STD_TORCH_CHECK(input.numel() % group_size == 0);
STD_TORCH_CHECK(output_s.dim() == 2);
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
cudaStream_t stream = get_current_cuda_stream();
constexpr int THREADS_PER_GROUP = 16;
if (num_groups == 0) {
return;
}
STD_TORCH_CHECK(input.size(-1) % group_size == 0);
const int64_t groups_per_row = input.size(-1) / group_size;
const int64_t num_rows = num_groups / groups_per_row;
// Odd row counts inherit the next even row count's launch geometry.
const int64_t policy_num_groups =
(num_rows + (num_rows & 1)) * groups_per_row;
const int groups_per_block = GetGroupsPerBlock(policy_num_groups);
const bool needs_bounds_check = num_groups % groups_per_block != 0;
auto dst_type = output_q.scalar_type();
const int num_blocks = (num_groups + groups_per_block - 1) / groups_per_block;
const int num_threads = groups_per_block * THREADS_PER_GROUP;
const bool is_column_major = output_s.stride(0) < output_s.stride(1);
const int scale_num_rows = output_s.size(1);
const int scale_stride = output_s.stride(1);
const bool use_register_path =
CanUseRegisterPath(input, output_q, group_size);
constexpr int REG_THREADS_PER_GROUP = 8;
constexpr int REG_BLOCK_THREADS = 256;
const int64_t reg_groups_per_block =
REG_BLOCK_THREADS / REG_THREADS_PER_GROUP;
const int reg_num_blocks = static_cast<int>(
(num_groups + reg_groups_per_block - 1) / reg_groups_per_block);
#ifndef USE_ROCM
#define LAUNCH_KERNEL_IMPL(T, DST_DTYPE, COL_MAJOR, UE8M0, CHECK_BOUNDS, \
SMEM_BYTES) \
do { \
cudaLaunchConfig_t config = {}; \
config.gridDim = dim3(num_blocks); \
config.blockDim = dim3(num_threads); \
config.dynamicSmemBytes = (SMEM_BYTES); \
config.stream = stream; \
cudaLaunchAttribute attrs[1]; \
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization; \
attrs[0].val.programmaticStreamSerializationAllowed = 1; \
config.numAttrs = 1; \
config.attrs = attrs; \
cudaLaunchKernelEx( \
&config, \
per_token_group_quant_8bit_kernel<T, DST_DTYPE, COL_MAJOR, UE8M0, \
float, CHECK_BOUNDS>, \
static_cast<T*>(input.data_ptr()), output_q.data_ptr(), \
static_cast<float*>(output_s.data_ptr()), group_size, num_groups, \
groups_per_block, (float)eps, (float)min_8bit, (float)max_8bit, \
scale_num_rows, scale_stride); \
} while (0)
#else
#define LAUNCH_KERNEL_IMPL(T, DST_DTYPE, COL_MAJOR, UE8M0, CHECK_BOUNDS, \
SMEM_BYTES) \
do { \
per_token_group_quant_8bit_kernel<T, DST_DTYPE, COL_MAJOR, UE8M0, float, \
CHECK_BOUNDS> \
<<<dim3(num_blocks), dim3(num_threads), (SMEM_BYTES), stream>>>( \
static_cast<T*>(input.data_ptr()), output_q.data_ptr(), \
static_cast<float*>(output_s.data_ptr()), group_size, \
num_groups, groups_per_block, (float)eps, (float)min_8bit, \
(float)max_8bit, scale_num_rows, scale_stride); \
} while (0)
#endif
#define LAUNCH_KERNEL_INST(T, DST_DTYPE, COL_MAJOR, UE8M0, SMEM_BYTES) \
do { \
if (needs_bounds_check) { \
LAUNCH_KERNEL_IMPL(T, DST_DTYPE, COL_MAJOR, UE8M0, true, SMEM_BYTES); \
} else { \
LAUNCH_KERNEL_IMPL(T, DST_DTYPE, COL_MAJOR, UE8M0, false, SMEM_BYTES); \
} \
} while (0)
#ifndef USE_ROCM
#define LAUNCH_REGISTER_INST(T, DST_DTYPE, COL_MAJOR, UE8M0) \
do { \
cudaLaunchConfig_t config = {}; \
config.gridDim = dim3(reg_num_blocks); \
config.blockDim = dim3(REG_BLOCK_THREADS); \
config.dynamicSmemBytes = 0; \
config.stream = stream; \
cudaLaunchAttribute attrs[1]; \
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization; \
attrs[0].val.programmaticStreamSerializationAllowed = 1; \
config.numAttrs = 1; \
config.attrs = attrs; \
cudaLaunchKernelEx( \
&config, \
per_token_group_quant_8bit_register_kernel<T, DST_DTYPE, COL_MAJOR, \
UE8M0>, \
static_cast<T*>(input.data_ptr()), output_q.data_ptr(), \
static_cast<float*>(output_s.data_ptr()), \
static_cast<int64_t>(num_groups), (float)eps, (float)min_8bit, \
(float)max_8bit, scale_num_rows, scale_stride); \
} while (0)
#else
#define LAUNCH_REGISTER_INST(T, DST_DTYPE, COL_MAJOR, UE8M0) \
do { \
per_token_group_quant_8bit_register_kernel<T, DST_DTYPE, COL_MAJOR, \
UE8M0> \
<<<dim3(reg_num_blocks), dim3(REG_BLOCK_THREADS), 0, stream>>>( \
static_cast<T*>(input.data_ptr()), output_q.data_ptr(), \
static_cast<float*>(output_s.data_ptr()), \
static_cast<int64_t>(num_groups), (float)eps, (float)min_8bit, \
(float)max_8bit, scale_num_rows, scale_stride); \
} while (0)
#endif
#define LAUNCH_KERNEL(T, DST_DTYPE) \
do { \
size_t smem_bytes = \
static_cast<size_t>(groups_per_block) * group_size * sizeof(T); \
if (is_column_major) { \
if (scale_ue8m0) { \
if (use_register_path) { \
LAUNCH_REGISTER_INST(T, DST_DTYPE, true, true); \
} else { \
LAUNCH_KERNEL_INST(T, DST_DTYPE, true, true, smem_bytes); \
} \
} else { \
if (use_register_path) { \
LAUNCH_REGISTER_INST(T, DST_DTYPE, true, false); \
} else { \
LAUNCH_KERNEL_INST(T, DST_DTYPE, true, false, smem_bytes); \
} \
} \
} else { \
if (scale_ue8m0) { \
if (use_register_path) { \
LAUNCH_REGISTER_INST(T, DST_DTYPE, false, true); \
} else { \
LAUNCH_KERNEL_INST(T, DST_DTYPE, false, true, smem_bytes); \
} \
} else { \
if (use_register_path) { \
LAUNCH_REGISTER_INST(T, DST_DTYPE, false, false); \
} else { \
LAUNCH_KERNEL_INST(T, DST_DTYPE, false, false, smem_bytes); \
} \
} \
} \
} while (0)
VLLM_STABLE_DISPATCH_FLOATING_TYPES(
input.scalar_type(), "per_token_group_quant_8bit", ([&] {
if (dst_type == torch::headeronly::ScalarType::Char) {
LAUNCH_KERNEL(scalar_t, int8_t);
} else {
VLLM_STABLE_DISPATCH_FP8_TYPES(
dst_type, "per_token_group_quant_8bit_fp8",
([&] { LAUNCH_KERNEL(scalar_t, fp8_t); }));
}
}));
#undef LAUNCH_KERNEL
#undef LAUNCH_KERNEL_INST
#undef LAUNCH_KERNEL_IMPL
#undef LAUNCH_REGISTER_INST
}
// 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.
//
// Each thread quantizes 16 bf16/fp16 values, using 2 or 8 threads per group.
template <typename T, typename DST_DTYPE, int GROUP_SIZE, int kGroupsPerBlockX,
int kRowsPerBlock>
__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 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) {
static_assert(GROUP_SIZE == 32 || GROUP_SIZE == 128);
constexpr int VEC_SIZE = 32 / sizeof(T); // 16 for bf16/fp16
constexpr int THREADS_PER_GROUP = GROUP_SIZE / VEC_SIZE;
static_assert(GROUP_SIZE == THREADS_PER_GROUP * VEC_SIZE,
"GROUP_SIZE must equal THREADS_PER_GROUP * VEC_SIZE");
static_assert(32 % THREADS_PER_GROUP == 0,
"THREADS_PER_GROUP must divide warp size for the shuffle "
"mask to be valid");
static_assert(
kGroupsPerBlockX > 0 && (kGroupsPerBlockX & (kGroupsPerBlockX - 1)) == 0,
"kGroupsPerBlockX must be a positive power of 2");
static_assert(kRowsPerBlock > 0, "kRowsPerBlock must be positive");
const int local_group_id = threadIdx.x / THREADS_PER_GROUP;
const int lane_id = threadIdx.x % THREADS_PER_GROUP;
const int sf_k_local = local_group_id % kGroupsPerBlockX;
const int row_local = local_group_id / kGroupsPerBlockX;
// Rows on grid.x: mn scales with tokens and can exceed the 65535 grid.y cap.
const int sf_k_idx = blockIdx.y * kGroupsPerBlockX + sf_k_local;
const int mn_idx = blockIdx.x * kRowsPerBlock + row_local;
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaGridDependencySynchronize();
#endif
if (mn_idx >= tma_aligned_mn) {
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaTriggerProgrammaticLaunchCompletion();
#endif
return;
}
const bool is_valid_group = (mn_idx < mn) && (sf_k_idx < groups_per_row);
// 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 + 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);
}
}
#ifdef USE_ROCM
const int lane_in_wave = threadIdx.x % warpSize;
const unsigned long long mask = ((1ull << THREADS_PER_GROUP) - 1)
<< (lane_in_wave & ~(THREADS_PER_GROUP - 1));
#else
const unsigned mask = ((1u << THREADS_PER_GROUP) - 1)
<< ((threadIdx.x % 32) & ~(THREADS_PER_GROUP - 1));
#endif
#pragma unroll
for (int offset = THREADS_PER_GROUP / 2; offset > 0; offset /= 2) {
local_absmax =
fmaxf(local_absmax,
__shfl_xor_sync(mask, local_absmax, offset, THREADS_PER_GROUP));
}
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) {
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_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = exp_byte;
} else if (out_idx < num_scale_elems) {
reinterpret_cast<uint8_t*>(output_s_packed)[out_idx * 4 + pos] = 0;
}
}
// 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;
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
cudaTriggerProgrammaticLaunchCompletion();
#endif
}
// Public entry point: register-resident packed quant kernel.
// Constraints: group_size in {32, 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 == 32 || group_size == 128,
"per_token_group_quant_8bit_packed only supports "
"group_size in {32, 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, "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;
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 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,
"]; got [", output_s_packed.size(0), ", ",
output_s_packed.size(1), "].");
STD_TORCH_CHECK(output_s_packed.stride(0) == 1 &&
output_s_packed.stride(1) == tma_aligned_mn,
"output_s_packed strides must be [1, ", tma_aligned_mn,
"]; got [", output_s_packed.stride(0), ", ",
output_s_packed.stride(1), "].");
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
cudaStream_t stream = get_current_cuda_stream();
const int threads_per_group = group_size / 16;
const int64_t padded_groups_per_row = k_num_packed_sfk * 4;
const int64_t num_scale_elems = mn + (k_num_packed_sfk - 1) * tma_aligned_mn;
STD_TORCH_CHECK(padded_groups_per_row % 4 == 0,
"padded_groups_per_row=", padded_groups_per_row,
" is not a multiple of 4.");
const int kx = GetGroupsPerBlockX(padded_groups_per_row);
const int ry = 16 / kx;
const int64_t row_blocks = (tma_aligned_mn + ry - 1) / ry;
const int64_t sf_k_blocks = padded_groups_per_row / kx;
const int num_threads = (kx * ry) * threads_per_group;
// CUDA caps grid.x at 2^31 - 1 and grid.y at 2^16 - 1 (65535).
constexpr int64_t kMaxGridDimYZ = 65535;
STD_TORCH_CHECK(row_blocks <= static_cast<int64_t>(INT32_MAX) &&
sf_k_blocks <= kMaxGridDimYZ,
"per_token_group_quant_8bit_packed grid too large: (",
row_blocks, ", ", sf_k_blocks, ").");
auto dst_type = output_q.scalar_type();
// PDL (Programmatic Dependent Launch) is NVIDIA-only; ROCm/HIP has no
// equivalent launch attribute, so fall back to a classic launch there.
#ifndef USE_ROCM
#define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, GROUP_SIZE, KX, RY) \
do { \
cudaLaunchConfig_t config = {}; \
config.gridDim = dim3(static_cast<unsigned int>(row_blocks), \
static_cast<unsigned int>(sf_k_blocks)); \
config.blockDim = dim3(num_threads); \
config.dynamicSmemBytes = 0; \
config.stream = stream; \
cudaLaunchAttribute attrs[1]; \
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization; \
attrs[0].val.programmaticStreamSerializationAllowed = 1; \
config.numAttrs = 1; \
config.attrs = attrs; \
cudaLaunchKernelEx( \
&config, \
per_token_group_quant_8bit_packed_register_kernel< \
T, DST_DTYPE, GROUP_SIZE, KX, RY>, \
static_cast<const T*>(input.data_ptr()), output_q.data_ptr(), \
reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()), \
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)
#else
#define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, GROUP_SIZE, KX, RY) \
do { \
dim3 grid(static_cast<unsigned int>(row_blocks), \
static_cast<unsigned int>(sf_k_blocks)); \
dim3 block(num_threads); \
per_token_group_quant_8bit_packed_register_kernel<T, DST_DTYPE, \
GROUP_SIZE, KX, RY> \
<<<grid, block, 0, stream>>>( \
static_cast<const T*>(input.data_ptr()), output_q.data_ptr(), \
reinterpret_cast<unsigned int*>(output_s_packed.data_ptr()), \
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)
#endif
#define LAUNCH_REG_KERNEL_GROUP_SIZE(T, DST_DTYPE, GROUP_SIZE) \
do { \
if (kx == 16) { \
LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, GROUP_SIZE, 16, 1); \
} else if (kx == 8) { \
LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, GROUP_SIZE, 8, 2); \
} else if (kx == 4) { \
LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, GROUP_SIZE, 4, 4); \
} else { \
STD_TORCH_CHECK(false, "Unsupported kx value ", kx); \
} \
} while (0)
#define LAUNCH_REG_KERNEL(T, DST_DTYPE) \
do { \
if (group_size == 32) { \
LAUNCH_REG_KERNEL_GROUP_SIZE(T, DST_DTYPE, 32); \
} else { \
LAUNCH_REG_KERNEL_GROUP_SIZE(T, DST_DTYPE, 128); \
} \
} while (0)
VLLM_STABLE_DISPATCH_HALF_TYPES(
input.scalar_type(), "per_token_group_quant_8bit_packed_register", ([&] {
if (dst_type == torch::headeronly::ScalarType::Char) {
LAUNCH_REG_KERNEL(scalar_t, int8_t);
} else {
VLLM_STABLE_DISPATCH_FP8_TYPES(
dst_type, "per_token_group_quant_8bit_packed_fp8",
([&] { LAUNCH_REG_KERNEL(scalar_t, fp8_t); }));
}
}));
#undef LAUNCH_REG_KERNEL
#undef LAUNCH_REG_KERNEL_GROUP_SIZE
#undef LAUNCH_REG_KERNEL_INST
}
void per_token_group_quant_fp8(const torch::stable::Tensor& input,
torch::stable::Tensor& output_q,
torch::stable::Tensor& output_s,
int64_t group_size, double eps, double fp8_min,
double fp8_max, bool scale_ue8m0,
bool dummy_is_scale_transposed = false,
bool dummy_is_tma_aligned = false) {
per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
fp8_min, fp8_max, scale_ue8m0);
}
void per_token_group_quant_int8(const torch::stable::Tensor& input,
torch::stable::Tensor& output_q,
torch::stable::Tensor& output_s,
int64_t group_size, double eps, double int8_min,
double int8_max) {
per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
int8_min, int8_max);
}