Migrate header files to torch stable abi (#44013)
This commit is contained in:
@@ -0,0 +1,100 @@
|
||||
/*
|
||||
* Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Licensed under the Apache License, Version 2.0 (the "License");
|
||||
* you may not use this file except in compliance with the License.
|
||||
* You may obtain a copy of the License at
|
||||
*
|
||||
* http://www.apache.org/licenses/LICENSE-2.0
|
||||
*
|
||||
* Unless required by applicable law or agreed to in writing, software
|
||||
* distributed under the License is distributed on an "AS IS" BASIS,
|
||||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
* See the License for the specific language governing permissions and
|
||||
* limitations under the License.
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
namespace vllm {
|
||||
namespace cuda_async {
|
||||
|
||||
__device__ __forceinline__ void cp_async_shared_global_16_cg(
|
||||
void* smem_ptr, const void* glob_ptr) {
|
||||
#if defined(USE_ROCM)
|
||||
*reinterpret_cast<int4*>(smem_ptr) = *reinterpret_cast<const int4*>(glob_ptr);
|
||||
#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
|
||||
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
|
||||
asm volatile("cp.async.cg.shared.global [%0], [%1], 16;\n"
|
||||
:
|
||||
: "r"(smem), "l"(glob_ptr));
|
||||
#elif defined(__CUDA_ARCH__)
|
||||
*reinterpret_cast<int4*>(smem_ptr) = *reinterpret_cast<const int4*>(glob_ptr);
|
||||
#else
|
||||
(void)smem_ptr;
|
||||
(void)glob_ptr;
|
||||
#endif
|
||||
}
|
||||
|
||||
__device__ __forceinline__ void cp_async_shared_global_ca(void* smem_ptr,
|
||||
const void* glob_ptr,
|
||||
int size_bytes) {
|
||||
#if defined(USE_ROCM)
|
||||
if (size_bytes == 4) {
|
||||
*reinterpret_cast<uint32_t*>(smem_ptr) =
|
||||
*reinterpret_cast<const uint32_t*>(glob_ptr);
|
||||
} else if (size_bytes == 8) {
|
||||
*reinterpret_cast<uint64_t*>(smem_ptr) =
|
||||
*reinterpret_cast<const uint64_t*>(glob_ptr);
|
||||
} else {
|
||||
*reinterpret_cast<int4*>(smem_ptr) =
|
||||
*reinterpret_cast<const int4*>(glob_ptr);
|
||||
}
|
||||
#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
|
||||
uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
|
||||
if (size_bytes == 4) {
|
||||
asm volatile("cp.async.ca.shared.global [%0], [%1], 4;\n"
|
||||
:
|
||||
: "r"(smem), "l"(glob_ptr));
|
||||
} else if (size_bytes == 8) {
|
||||
asm volatile("cp.async.ca.shared.global [%0], [%1], 8;\n"
|
||||
:
|
||||
: "r"(smem), "l"(glob_ptr));
|
||||
} else {
|
||||
asm volatile("cp.async.ca.shared.global [%0], [%1], 16;\n"
|
||||
:
|
||||
: "r"(smem), "l"(glob_ptr));
|
||||
}
|
||||
#elif defined(__CUDA_ARCH__)
|
||||
if (size_bytes == 4) {
|
||||
*reinterpret_cast<uint32_t*>(smem_ptr) =
|
||||
*reinterpret_cast<const uint32_t*>(glob_ptr);
|
||||
} else if (size_bytes == 8) {
|
||||
*reinterpret_cast<uint64_t*>(smem_ptr) =
|
||||
*reinterpret_cast<const uint64_t*>(glob_ptr);
|
||||
} else {
|
||||
*reinterpret_cast<int4*>(smem_ptr) =
|
||||
*reinterpret_cast<const int4*>(glob_ptr);
|
||||
}
|
||||
#else
|
||||
(void)smem_ptr;
|
||||
(void)glob_ptr;
|
||||
(void)size_bytes;
|
||||
#endif
|
||||
}
|
||||
|
||||
__device__ __forceinline__ void cp_async_commit_group() {
|
||||
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800 && !defined(USE_ROCM)
|
||||
asm volatile("cp.async.commit_group;\n" ::);
|
||||
#endif
|
||||
}
|
||||
|
||||
template <int n>
|
||||
__device__ __forceinline__ void cp_async_wait_group() {
|
||||
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800 && !defined(USE_ROCM)
|
||||
asm volatile("cp.async.wait_group %0;\n" : : "n"(n));
|
||||
#endif
|
||||
}
|
||||
|
||||
} // namespace cuda_async
|
||||
} // namespace vllm
|
||||
@@ -0,0 +1,497 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights
|
||||
*reserved. SPDX-License-Identifier: BSD-3-Clause
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions are met:
|
||||
*
|
||||
* 1. Redistributions of source code must retain the above copyright notice,
|
||||
*this list of conditions and the following disclaimer.
|
||||
*
|
||||
* 2. Redistributions in binary form must reproduce the above copyright notice,
|
||||
* this list of conditions and the following disclaimer in the documentation
|
||||
* and/or other materials provided with the distribution.
|
||||
*
|
||||
* 3. Neither the name of the copyright holder nor the names of its
|
||||
* contributors may be used to endorse or promote products derived from
|
||||
* this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
||||
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
||||
*ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
|
||||
*LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
|
||||
*CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
|
||||
*SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
|
||||
*INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
|
||||
*CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
|
||||
*ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
|
||||
*POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
|
||||
//
|
||||
// This file is a modified excerpt of
|
||||
// include/cutlass/epilogue/fusion/visitor_load.hpp from
|
||||
// https://github.com/NVIDIA/cutlass v3.5.0
|
||||
// It has been modified to support either
|
||||
// row/column or scalar broadcasting where the tensor being loaded from is
|
||||
// always passed in via a device pointer. This lets one compiled kernel handle
|
||||
// all cases of per-tensor or per-channel/per-token quantization.
|
||||
//
|
||||
// This interface also allows the scales to be passed in as tensors that
|
||||
// consistently reside on the device, which avoids an issue with a previous
|
||||
// implementation where scalars needed to be on the CPU since they
|
||||
// were passed in via float values. This created a potential performance hazard
|
||||
// if scales were initially on the device, and caused torch.compile graph
|
||||
// breaks when moving scales to the CPU.
|
||||
//
|
||||
#pragma once
|
||||
|
||||
// Turn off clang-format for the entire file to keep it close to upstream
|
||||
// clang-format off
|
||||
|
||||
#include "cutlass/epilogue/threadblock/fusion/visitor_2x.hpp"
|
||||
#include "cutlass/epilogue/threadblock/fusion/visitors.hpp"
|
||||
#include "cute/tensor.hpp"
|
||||
|
||||
namespace cutlass::epilogue::threadblock {
|
||||
|
||||
using namespace cute;
|
||||
using namespace detail;
|
||||
|
||||
template<
|
||||
class ThreadMap,
|
||||
class Element,
|
||||
class StrideMNL
|
||||
>
|
||||
struct VisitorRowOrScalarBroadcast {
|
||||
|
||||
// This struct has been modified to have a bool indicating that ptr_row is a
|
||||
// scalar that must be broadcast.
|
||||
struct Arguments {
|
||||
Element const* ptr_row = nullptr;
|
||||
bool row_broadcast = true;
|
||||
StrideMNL dRow = {};
|
||||
};
|
||||
|
||||
using Params = Arguments;
|
||||
|
||||
template <class ProblemShape>
|
||||
static constexpr Params
|
||||
to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
|
||||
return args;
|
||||
}
|
||||
|
||||
template <class ProblemShape>
|
||||
static size_t
|
||||
get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
struct SharedStorage {};
|
||||
|
||||
// Global load type
|
||||
static int constexpr vec_bits = ThreadMap::kElementsPerAccess * sizeof_bits<Element>::value;
|
||||
using VecType = uint_bit_t<cute::min(128, vec_bits)>;
|
||||
static int constexpr VecLength = sizeof(VecType) / sizeof(Element);
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
VisitorRowOrScalarBroadcast() { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
VisitorRowOrScalarBroadcast(Params const& params, SharedStorage const& shared_storage)
|
||||
: params_ptr(¶ms) { }
|
||||
|
||||
Params const* params_ptr;
|
||||
|
||||
template <class GTensor, class RTensor, class CTensor, class ProblemShape>
|
||||
struct Callbacks : EmptyCallbacks {
|
||||
CUTLASS_DEVICE
|
||||
Callbacks(
|
||||
GTensor&& tC_gRow,
|
||||
RTensor&& tC_rRow,
|
||||
CTensor&& tC_cRow,
|
||||
ProblemShape problem_shape,
|
||||
Params const* params_ptr
|
||||
):
|
||||
tC_gRow(cute::forward<GTensor>(tC_gRow)),
|
||||
tC_rRow(cute::forward<RTensor>(tC_rRow)),
|
||||
tC_cRow(cute::forward<CTensor>(tC_cRow)),
|
||||
n(get<1>(problem_shape)),
|
||||
params_ptr(params_ptr) { }
|
||||
|
||||
GTensor tC_gRow;
|
||||
RTensor tC_rRow;
|
||||
CTensor tC_cRow;
|
||||
Params const* params_ptr;
|
||||
int n;
|
||||
|
||||
// This function is modified from VisitorRowBroadcast
|
||||
CUTLASS_DEVICE void
|
||||
begin_epilogue() {
|
||||
clear(tC_rRow);
|
||||
auto src_v = filter(tC_gRow);
|
||||
auto coord_v = filter(tC_cRow);
|
||||
auto dst_v = filter(tC_rRow);
|
||||
|
||||
if (params_ptr->row_broadcast) {
|
||||
// In this case we are loading from a row vector and broadcasting
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < size(src_v); ++i) {
|
||||
bool guard = get<1>(coord_v(i)) < n;
|
||||
cutlass::arch::global_load<VecType, sizeof(VecType)>(
|
||||
dst_v(i), (void const*)&src_v(i), guard);
|
||||
}
|
||||
} else {
|
||||
// In this case we are loading from a scalar and broadcasting
|
||||
VecType filled_vec;
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < VecLength; i++) {
|
||||
reinterpret_cast<Element*>(&filled_vec)[i] = *(params_ptr->ptr_row);
|
||||
}
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < size(src_v); ++i) {
|
||||
if (get<1>(coord_v(i)) < n) {
|
||||
dst_v(i) = filled_vec;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <class ElementAccumulator, int FragmentSize>
|
||||
CUTLASS_DEVICE auto // returns an Array
|
||||
visit(int iter_idx, int row_idx, int column_idx, int frg_idx,
|
||||
Array<ElementAccumulator, FragmentSize> const& frg_acc) {
|
||||
Tensor rRow_frg = recast<Array<Element, FragmentSize>>(coalesce(tC_rRow));
|
||||
return rRow_frg(column_idx);
|
||||
}
|
||||
};
|
||||
|
||||
template <class ProblemShape>
|
||||
CUTLASS_DEVICE auto
|
||||
get_callbacks(
|
||||
gemm::GemmCoord threadblock_tile_offset,
|
||||
int thread_idx,
|
||||
ProblemShape problem_shape
|
||||
) {
|
||||
Tensor mRow = make_tensor(
|
||||
make_gmem_ptr(params_ptr->ptr_row),
|
||||
problem_shape,
|
||||
params_ptr->dRow);
|
||||
|
||||
// VECTOR, FRAGMENT_COLUMN
|
||||
Tensor tC_gRow = recast<VecType>(
|
||||
ThreadMap::partition(mRow, thread_idx, threadblock_tile_offset)
|
||||
)(_,_,_0{},_0{},_0{},_0{});
|
||||
Tensor tC_rRow = make_tensor_like(tC_gRow);
|
||||
|
||||
// Generate the pred tensor
|
||||
Tensor cRow = make_identity_tensor(mRow.shape());
|
||||
Tensor tC_cRow = outer_partition(
|
||||
ThreadMap::partition(cRow, thread_idx, threadblock_tile_offset)(_,_,_0{},_0{},_0{},_0{}),
|
||||
Shape<Int<VecLength>>{},
|
||||
(_0{})
|
||||
);
|
||||
|
||||
return Callbacks<
|
||||
decltype(tC_gRow), decltype(tC_rRow),
|
||||
decltype(tC_cRow), ProblemShape>(
|
||||
cute::move(tC_gRow),
|
||||
cute::move(tC_rRow),
|
||||
cute::move(tC_cRow),
|
||||
problem_shape,
|
||||
params_ptr
|
||||
);
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// This is a modified RowBroadcast that will broadcast 0 if ptr_row is null
|
||||
template<
|
||||
class ThreadMap,
|
||||
class Element,
|
||||
class StrideMNL
|
||||
>
|
||||
struct VisitorRowOrZeroBroadcast {
|
||||
|
||||
// This struct has been modified to remove null_default (because it's always 0)
|
||||
struct Arguments {
|
||||
Element const* ptr_row = nullptr;
|
||||
StrideMNL dRow = {};
|
||||
};
|
||||
|
||||
using Params = Arguments;
|
||||
|
||||
template <class ProblemShape>
|
||||
static constexpr Params
|
||||
to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
|
||||
return args;
|
||||
}
|
||||
|
||||
template <class ProblemShape>
|
||||
static size_t
|
||||
get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
struct SharedStorage {};
|
||||
|
||||
// Global load type
|
||||
static int constexpr vec_bits = ThreadMap::kElementsPerAccess * sizeof_bits<Element>::value;
|
||||
using VecType = uint_bit_t<cute::min(128, vec_bits)>;
|
||||
static int constexpr VecLength = sizeof(VecType) / sizeof(Element);
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
VisitorRowOrZeroBroadcast() { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
VisitorRowOrZeroBroadcast(Params const& params, SharedStorage const& shared_storage)
|
||||
: params_ptr(¶ms) { }
|
||||
|
||||
Params const* params_ptr;
|
||||
|
||||
template <class GTensor, class RTensor, class CTensor, class ProblemShape>
|
||||
struct Callbacks : EmptyCallbacks {
|
||||
CUTLASS_DEVICE
|
||||
Callbacks(
|
||||
GTensor&& tC_gRow,
|
||||
RTensor&& tC_rRow,
|
||||
CTensor&& tC_cRow,
|
||||
ProblemShape problem_shape,
|
||||
Params const* params_ptr
|
||||
):
|
||||
tC_gRow(cute::forward<GTensor>(tC_gRow)),
|
||||
tC_rRow(cute::forward<RTensor>(tC_rRow)),
|
||||
tC_cRow(cute::forward<CTensor>(tC_cRow)),
|
||||
n(get<1>(problem_shape)),
|
||||
params_ptr(params_ptr) { }
|
||||
|
||||
GTensor tC_gRow;
|
||||
RTensor tC_rRow;
|
||||
CTensor tC_cRow;
|
||||
Params const* params_ptr;
|
||||
int n;
|
||||
|
||||
// This function is modified from VisitorRowBroadcast
|
||||
CUTLASS_DEVICE void
|
||||
begin_epilogue() {
|
||||
clear(tC_rRow);
|
||||
auto src_v = filter(tC_gRow);
|
||||
auto coord_v = filter(tC_cRow);
|
||||
auto dst_v = filter(tC_rRow);
|
||||
|
||||
if (params_ptr->ptr_row != nullptr) {
|
||||
// In this case we are loading from a row vector and broadcasting
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < size(src_v); ++i) {
|
||||
bool guard = get<1>(coord_v(i)) < n;
|
||||
cutlass::arch::global_load<VecType, sizeof(VecType)>(
|
||||
dst_v(i), (void const*)&src_v(i), guard);
|
||||
}
|
||||
} else {
|
||||
// In this case we are broadcasting 0
|
||||
VecType filled_vec;
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < VecLength; i++) {
|
||||
reinterpret_cast<Element*>(&filled_vec)[i] = Element{0};
|
||||
}
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < size(src_v); ++i) {
|
||||
if (get<1>(coord_v(i)) < n) {
|
||||
dst_v(i) = filled_vec;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <class ElementAccumulator, int FragmentSize>
|
||||
CUTLASS_DEVICE auto // returns an Array
|
||||
visit(int iter_idx, int row_idx, int column_idx, int frg_idx,
|
||||
Array<ElementAccumulator, FragmentSize> const& frg_acc) {
|
||||
Tensor rRow_frg = recast<Array<Element, FragmentSize>>(coalesce(tC_rRow));
|
||||
return rRow_frg(column_idx);
|
||||
}
|
||||
};
|
||||
|
||||
template <class ProblemShape>
|
||||
CUTLASS_DEVICE auto
|
||||
get_callbacks(
|
||||
gemm::GemmCoord threadblock_tile_offset,
|
||||
int thread_idx,
|
||||
ProblemShape problem_shape
|
||||
) {
|
||||
Tensor mRow = make_tensor(
|
||||
make_gmem_ptr(params_ptr->ptr_row),
|
||||
problem_shape,
|
||||
params_ptr->dRow);
|
||||
|
||||
// VECTOR, FRAGMENT_COLUMN
|
||||
Tensor tC_gRow = recast<VecType>(
|
||||
ThreadMap::partition(mRow, thread_idx, threadblock_tile_offset)
|
||||
)(_,_,_0{},_0{},_0{},_0{});
|
||||
Tensor tC_rRow = make_tensor_like(tC_gRow);
|
||||
|
||||
// Generate the pred tensor
|
||||
Tensor cRow = make_identity_tensor(mRow.shape());
|
||||
Tensor tC_cRow = outer_partition(
|
||||
ThreadMap::partition(cRow, thread_idx, threadblock_tile_offset)(_,_,_0{},_0{},_0{},_0{}),
|
||||
Shape<Int<VecLength>>{},
|
||||
(_0{})
|
||||
);
|
||||
|
||||
return Callbacks<
|
||||
decltype(tC_gRow), decltype(tC_rRow),
|
||||
decltype(tC_cRow), ProblemShape>(
|
||||
cute::move(tC_gRow),
|
||||
cute::move(tC_rRow),
|
||||
cute::move(tC_cRow),
|
||||
problem_shape,
|
||||
params_ptr
|
||||
);
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
// Column vector broadcast
|
||||
template<
|
||||
class ThreadMap,
|
||||
class Element,
|
||||
class StrideMNL = Stride<_1,_0,_0>
|
||||
>
|
||||
struct VisitorColOrScalarBroadcast {
|
||||
|
||||
// This struct has been modified to have a bool indicating that ptr_col is a
|
||||
// scalar that must be broadcast.
|
||||
struct Arguments {
|
||||
Element const* ptr_col = nullptr;
|
||||
bool col_broadcast = true;
|
||||
StrideMNL dCol = {};
|
||||
};
|
||||
|
||||
using Params = Arguments;
|
||||
|
||||
template <class ProblemShape>
|
||||
static constexpr Params
|
||||
to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
|
||||
return args;
|
||||
}
|
||||
|
||||
template <class ProblemShape>
|
||||
static size_t
|
||||
get_workspace_size(ProblemShape const& problem_shape, Arguments const& args) {
|
||||
return 0;
|
||||
}
|
||||
|
||||
struct SharedStorage { };
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
VisitorColOrScalarBroadcast() { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
VisitorColOrScalarBroadcast(Params const& params, SharedStorage const& shared_storage)
|
||||
: params_ptr(¶ms) { }
|
||||
|
||||
Params const* params_ptr;
|
||||
|
||||
template <class GTensor, class RTensor, class CTensor, class ProblemShape>
|
||||
struct Callbacks : EmptyCallbacks {
|
||||
CUTLASS_DEVICE
|
||||
Callbacks(
|
||||
GTensor&& tC_gCol,
|
||||
RTensor&& tC_rCol,
|
||||
CTensor&& tC_cCol,
|
||||
ProblemShape problem_shape,
|
||||
Params const* params_ptr
|
||||
):
|
||||
tC_gCol(cute::forward<GTensor>(tC_gCol)),
|
||||
tC_rCol(cute::forward<RTensor>(tC_rCol)),
|
||||
tC_cCol(cute::forward<CTensor>(tC_cCol)),
|
||||
m(get<0>(problem_shape)),
|
||||
params_ptr(params_ptr) { }
|
||||
|
||||
GTensor tC_gCol;
|
||||
RTensor tC_rCol;
|
||||
CTensor tC_cCol;
|
||||
Params const* params_ptr;
|
||||
int m;
|
||||
|
||||
// This function is modified from VisitorColBroadcast
|
||||
CUTLASS_DEVICE void
|
||||
begin_epilogue() {
|
||||
clear(tC_rCol);
|
||||
|
||||
Tensor pred = make_tensor<bool>(shape(tC_gCol));
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < size(pred); ++i) {
|
||||
pred(i) = get<0>(tC_cCol(i)) < m;
|
||||
}
|
||||
|
||||
if (params_ptr->col_broadcast) {
|
||||
// In this case we are loading from a column vector and broadcasting
|
||||
copy_if(pred, tC_gCol, tC_rCol);
|
||||
} else {
|
||||
// In this case we are loading from a scalar and broadcasting
|
||||
auto dst_v = filter(tC_rCol);
|
||||
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < size(dst_v); ++i) {
|
||||
if (pred(i)) {
|
||||
dst_v(i) = *(params_ptr->ptr_col);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <class ElementAccumulator, int FragmentSize>
|
||||
CUTLASS_DEVICE auto // returns an Array
|
||||
visit(int iter_idx, int row_idx, int column_idx, int frg_idx,
|
||||
Array<ElementAccumulator, FragmentSize> const& frg_acc) {
|
||||
Array<Element, FragmentSize> frg_col;
|
||||
frg_col.fill(tC_rCol(row_idx,iter_idx));
|
||||
return frg_col;
|
||||
}
|
||||
};
|
||||
|
||||
template <class ProblemShape>
|
||||
CUTLASS_DEVICE auto
|
||||
get_callbacks(
|
||||
gemm::GemmCoord threadblock_tile_offset,
|
||||
int thread_idx,
|
||||
ProblemShape problem_shape
|
||||
) {
|
||||
Tensor mCol = make_tensor(
|
||||
make_gmem_ptr(params_ptr->ptr_col),
|
||||
problem_shape,
|
||||
params_ptr->dCol);
|
||||
|
||||
// VECTOR, FRAGMENT_COLUMN, FRAGMENT_ROW, ITERATION_ROW, ITERATION_GROUP, ITERATION_CLUSTER
|
||||
Tensor tC_gCol = group_modes<1,4>(
|
||||
ThreadMap::partition(mCol, thread_idx, threadblock_tile_offset)(_0{},_0{},_,_,_,_));
|
||||
Tensor tC_rCol = make_tensor_like(tC_gCol);
|
||||
|
||||
// Generate the pred tensor
|
||||
Tensor cCol = make_identity_tensor(mCol.shape());
|
||||
Tensor tC_cCol = group_modes<1,4>(
|
||||
ThreadMap::partition(cCol, thread_idx, threadblock_tile_offset)(_0{},_0{},_,_,_,_));
|
||||
|
||||
return Callbacks<
|
||||
decltype(tC_gCol), decltype(tC_rCol),
|
||||
decltype(tC_cCol), ProblemShape>(
|
||||
cute::move(tC_gCol),
|
||||
cute::move(tC_rCol),
|
||||
cute::move(tC_cCol),
|
||||
problem_shape,
|
||||
params_ptr
|
||||
);
|
||||
}
|
||||
};
|
||||
|
||||
}
|
||||
@@ -2,7 +2,7 @@
|
||||
|
||||
#include <torch/csrc/stable/tensor.h>
|
||||
|
||||
#include "cutlass_extensions/epilogue/broadcast_load_epilogue_c2x.hpp"
|
||||
#include "broadcast_load_epilogue_c2x.hpp"
|
||||
|
||||
/*
|
||||
This file defines custom epilogues for fusing channel scales, token scales,
|
||||
|
||||
@@ -20,7 +20,7 @@
|
||||
|
||||
#include "torch_utils.h"
|
||||
|
||||
#include "../async_util.cuh"
|
||||
#include "async_util.cuh"
|
||||
#include "../cuda_compat.h"
|
||||
#include "../type_convert.cuh"
|
||||
#include "dispatch_utils.h"
|
||||
|
||||
@@ -0,0 +1,64 @@
|
||||
#pragma once
|
||||
|
||||
#include <cuda_runtime_api.h>
|
||||
#include <algorithm>
|
||||
|
||||
// maximum blocks per SM cap
|
||||
#ifndef VLLM_LAUNCH_BLOCKS_CAP
|
||||
#define VLLM_LAUNCH_BLOCKS_CAP 4
|
||||
#endif
|
||||
|
||||
// Compile-time estimate of max threads per SM for launch bounds.
|
||||
// Families: 1024, 1536, 2048 threads/SM.
|
||||
#ifndef VLLM_MAX_THREADS_PER_SM
|
||||
#ifdef __CUDA_ARCH__
|
||||
|
||||
/* 1024 thr/SM: Turing (sm_75) */
|
||||
#if (__CUDA_ARCH__ == 750)
|
||||
#define VLLM_MAX_THREADS_PER_SM 1024
|
||||
|
||||
/* 1536 thr/SM: Ampere GA10x (sm_86/87), Ada (sm_89),
|
||||
GB20x consumer (sm_120/121), Thor (sm_101 or sm_110) */
|
||||
#elif (__CUDA_ARCH__ == 860) || (__CUDA_ARCH__ == 870) || \
|
||||
(__CUDA_ARCH__ == 890) || (__CUDA_ARCH__ == 1010) || \
|
||||
(__CUDA_ARCH__ == 1100) || (__CUDA_ARCH__ == 1200) || \
|
||||
(__CUDA_ARCH__ == 1210)
|
||||
#define VLLM_MAX_THREADS_PER_SM 1536
|
||||
|
||||
/* 2048 thr/SM: Volta (sm_70/72), Ampere GA100 (sm_80),
|
||||
Hopper (sm_90), Blackwell (sm_100/103) */
|
||||
#elif (__CUDA_ARCH__ == 700) || (__CUDA_ARCH__ == 720) || \
|
||||
(__CUDA_ARCH__ == 800) || (__CUDA_ARCH__ == 900) || \
|
||||
(__CUDA_ARCH__ == 1000) || (__CUDA_ARCH__ == 1030)
|
||||
#define VLLM_MAX_THREADS_PER_SM 2048
|
||||
|
||||
/* Fallback: use 2048 for unknown future CCs */
|
||||
#else
|
||||
#define VLLM_MAX_THREADS_PER_SM 2048
|
||||
#endif
|
||||
|
||||
#else
|
||||
/* Host pass (no __CUDA_ARCH__): neutral default */
|
||||
#define VLLM_MAX_THREADS_PER_SM 2048
|
||||
#endif
|
||||
#endif
|
||||
|
||||
// compute the number of blocks per SM to request in __launch_bounds__
|
||||
#define VLLM_BLOCKS_DIV(VAL) (VLLM_MAX_THREADS_PER_SM / (VAL))
|
||||
#define VLLM_CLAMP_BLOCKS_PER_SM(VAL) \
|
||||
(((VAL) <= 0) \
|
||||
? 1 \
|
||||
: (((VAL) < VLLM_LAUNCH_BLOCKS_CAP) ? (VAL) : VLLM_LAUNCH_BLOCKS_CAP))
|
||||
#define VLLM_BLOCKS_PER_SM(BLOCK_THREADS) \
|
||||
VLLM_CLAMP_BLOCKS_PER_SM(VLLM_BLOCKS_DIV(BLOCK_THREADS))
|
||||
|
||||
// runtime-time helper to compute blocks/SM
|
||||
static inline int vllm_runtime_blocks_per_sm(int block_threads) {
|
||||
int device = -1;
|
||||
cudaGetDevice(&device);
|
||||
int max_threads_per_sm = VLLM_MAX_THREADS_PER_SM;
|
||||
cudaDeviceGetAttribute(&max_threads_per_sm,
|
||||
cudaDevAttrMaxThreadsPerMultiProcessor, device);
|
||||
int blocks = (block_threads > 0) ? (max_threads_per_sm / block_threads) : 1;
|
||||
return VLLM_CLAMP_BLOCKS_PER_SM(blocks);
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -25,7 +25,7 @@
|
||||
#include <cuda_fp8.h>
|
||||
|
||||
#include "cuda_utils.h"
|
||||
#include "launch_bounds_utils.h"
|
||||
#include "libtorch_stable/launch_bounds_utils.h"
|
||||
|
||||
// Define before including nvfp4_utils.cuh so the header
|
||||
// can use this macro during compilation.
|
||||
|
||||
@@ -34,7 +34,7 @@
|
||||
static_assert(CVT_FP4_ELTS_PER_THREAD == 16,
|
||||
"MXFP4 experts quant requires PACK16 mode (CUDA >= 12.9)");
|
||||
|
||||
#include "launch_bounds_utils.h"
|
||||
#include "libtorch_stable/launch_bounds_utils.h"
|
||||
|
||||
namespace vllm {
|
||||
|
||||
|
||||
@@ -26,7 +26,7 @@
|
||||
|
||||
#include "cuda_utils.h"
|
||||
#include "nvfp4_utils.cuh"
|
||||
#include "launch_bounds_utils.h"
|
||||
#include "libtorch_stable/launch_bounds_utils.h"
|
||||
|
||||
namespace vllm {
|
||||
|
||||
|
||||
@@ -26,7 +26,7 @@
|
||||
#include "../../cuda_vec_utils.cuh"
|
||||
|
||||
#include "cuda_utils.h"
|
||||
#include "launch_bounds_utils.h"
|
||||
#include "libtorch_stable/launch_bounds_utils.h"
|
||||
|
||||
// Define before including nvfp4_utils.cuh so the header
|
||||
// can use this macro during compilation.
|
||||
|
||||
@@ -0,0 +1,571 @@
|
||||
// copied and adapted from https://github.com/ggerganov/llama.cpp/blob/b2899/ggml-cuda/convert.cu
|
||||
// Dequant functions
|
||||
static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
||||
const block_q4_0 * x = (const block_q4_0 *) vx;
|
||||
|
||||
const dfloat d = x[ib].d;
|
||||
|
||||
const int vui = x[ib].qs[iqs];
|
||||
|
||||
v.x = __int2half_rn(vui & 0xF);
|
||||
v.y = __int2half_rn(vui >> 4);
|
||||
|
||||
v = __hsub2(v, __floats2half2_rn(8.0f, 8.0f));
|
||||
v = __hmul2(v, {d, d});
|
||||
}
|
||||
|
||||
static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
||||
const block_q4_1 * x = (const block_q4_1 *) vx;
|
||||
|
||||
const dfloat d = __low2half(x[ib].dm);
|
||||
const dfloat m = __high2half(x[ib].dm);
|
||||
|
||||
const int vui = x[ib].qs[iqs];
|
||||
|
||||
v.x = __int2half_rn(vui & 0xF);
|
||||
v.y = __int2half_rn(vui >> 4);
|
||||
|
||||
v = __hmul2(v, {d, d});
|
||||
v = __hadd2(v, {m, m});
|
||||
}
|
||||
|
||||
static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
||||
const block_q5_0 * x = (const block_q5_0 *) vx;
|
||||
|
||||
const dfloat d = x[ib].d;
|
||||
|
||||
uint32_t qh;
|
||||
memcpy(&qh, x[ib].qh, sizeof(qh));
|
||||
|
||||
const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
||||
const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
||||
|
||||
v.x = __int2half_rn((x[ib].qs[iqs] & 0xf) | xh_0);
|
||||
v.y = __int2half_rn((x[ib].qs[iqs] >> 4) | xh_1);
|
||||
|
||||
v = __hsub2(v, __floats2half2_rn(16.0f, 16.0f));
|
||||
v = __hmul2(v, {d, d});
|
||||
}
|
||||
|
||||
static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
||||
const block_q5_1 * x = (const block_q5_1 *) vx;
|
||||
|
||||
const dfloat d = __low2half(x[ib].dm);
|
||||
const dfloat m = __high2half(x[ib].dm);
|
||||
|
||||
uint32_t qh;
|
||||
memcpy(&qh, x[ib].qh, sizeof(qh));
|
||||
|
||||
const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
|
||||
const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
|
||||
|
||||
v.x = __int2half_rn((x[ib].qs[iqs] & 0xf) | xh_0);
|
||||
v.y = __int2half_rn((x[ib].qs[iqs] >> 4) | xh_1);
|
||||
|
||||
v = __hmul2(v, {d, d});
|
||||
v = __hadd2(v, {m, m});
|
||||
}
|
||||
|
||||
static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
|
||||
const block_q8_0 * x = (const block_q8_0 *) vx;
|
||||
|
||||
const dfloat d = x[ib].d;
|
||||
|
||||
v.x = __int2half_rn(x[ib].qs[iqs + 0]);
|
||||
v.y = __int2half_rn(x[ib].qs[iqs + 1]);
|
||||
|
||||
v = __hmul2(v, {d, d});
|
||||
}
|
||||
|
||||
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
|
||||
static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
|
||||
const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x);
|
||||
|
||||
if (i >= k) {
|
||||
return;
|
||||
}
|
||||
|
||||
const int ib = i/qk; // block index
|
||||
const int iqs = (i%qk)/qr; // quant index
|
||||
const int iybs = i - i%qk; // y block start index
|
||||
const int y_offset = qr == 1 ? 1 : qk/2;
|
||||
|
||||
// dequantize
|
||||
dfloat2 v;
|
||||
dequantize_kernel(vx, ib, iqs, v);
|
||||
|
||||
y[iybs + iqs + 0] = convert_from_half<dst_t>(v.x);
|
||||
y[iybs + iqs + y_offset] = convert_from_half<dst_t>(v.y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_q2_K * x = (const block_q2_K *) vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int n = tid/32;
|
||||
const int l = tid - 32*n;
|
||||
const int is = 8*n + l/16;
|
||||
|
||||
const uint8_t q = x[i].qs[32*n + l];
|
||||
dst_t * y = yy + i*QK_K + 128*n;
|
||||
|
||||
half dall = __low2half(x[i].dm);
|
||||
half dmin = __high2half(x[i].dm);
|
||||
y[l+ 0] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+0] >> 4))));
|
||||
y[l+32] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+2] >> 4))));
|
||||
y[l+64] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+4] >> 4))));
|
||||
y[l+96] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin, __int2half_rn(x[i].scales[is+6] >> 4))));
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_q3_K * x = (const block_q3_K *) vx;
|
||||
|
||||
const auto r = threadIdx.x/4;
|
||||
const int tid = r/2;
|
||||
const int is0 = r%2;
|
||||
const int l0 = 16*is0 + 4*(threadIdx.x%4);
|
||||
const int n = tid / 4;
|
||||
const int j = tid - 4*n;
|
||||
|
||||
uint8_t m = 1 << (4*n + j);
|
||||
int is = 8*n + 2*j + is0;
|
||||
int shift = 2*j;
|
||||
|
||||
int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
|
||||
is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
|
||||
is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
|
||||
(x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
|
||||
half d_all = x[i].d;
|
||||
half dl = __hmul(d_all, __int2half_rn(us - 32));
|
||||
|
||||
dst_t * y = yy + i*QK_K + 128*n + 32*j;
|
||||
const uint8_t * q = x[i].qs + 32*n;
|
||||
const uint8_t * hm = x[i].hmask;
|
||||
|
||||
for (int l = l0; l < l0+4; ++l) {
|
||||
y[l] = convert_from_half<dst_t>(__hmul(dl, __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4))));
|
||||
}
|
||||
}
|
||||
|
||||
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
|
||||
if (j < 4) {
|
||||
d = q[j] & 63; m = q[j + 4] & 63;
|
||||
} else {
|
||||
d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
|
||||
m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
const block_q4_K * x = (const block_q4_K *) vx;
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
|
||||
// assume 32 threads
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8;
|
||||
const int ir = tid%8;
|
||||
const int is = 2*il;
|
||||
const int n = 4;
|
||||
|
||||
dst_t * y = yy + i*QK_K + 64*il + n*ir;
|
||||
|
||||
const half dall = __low2half(x[i].dm);
|
||||
const half dmin = __high2half(x[i].dm);
|
||||
|
||||
const uint8_t * q = x[i].qs + 32*il + n*ir;
|
||||
|
||||
uint8_t sc, m;
|
||||
get_scale_min_k4(is + 0, x[i].scales, sc, m);
|
||||
const half d1 = __hmul(dall, __int2half_rn(sc));
|
||||
const half m1 = __hmul(dmin, __int2half_rn(m));
|
||||
get_scale_min_k4(is + 1, x[i].scales, sc, m);
|
||||
const half d2 = __hmul(dall, __int2half_rn(sc));
|
||||
const half m2 = __hmul(dmin, __int2half_rn(m));
|
||||
for (int l = 0; l < n; ++l) {
|
||||
y[l + 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1));
|
||||
y[l +32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn(q[l] >> 4)), m2));
|
||||
}
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
const block_q5_K * x = (const block_q5_K *) vx;
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
|
||||
// assume 64 threads - this is very slightly better than the one below
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/16; // il is in 0...3
|
||||
const int ir = tid%16; // ir is in 0...15
|
||||
const int is = 2*il; // is is in 0...6
|
||||
|
||||
dst_t * y = yy + i*QK_K + 64*il + 2*ir;
|
||||
|
||||
const half dall = __low2half(x[i].dm);
|
||||
const half dmin = __high2half(x[i].dm);
|
||||
|
||||
const uint8_t * ql = x[i].qs + 32*il + 2*ir;
|
||||
const uint8_t * qh = x[i].qh + 2*ir;
|
||||
|
||||
uint8_t sc, m;
|
||||
get_scale_min_k4(is + 0, x[i].scales, sc, m);
|
||||
const half d1 = __hmul(dall, __int2half_rn(sc)); const half m1 = __hmul(dmin, __int2half_rn(m));
|
||||
get_scale_min_k4(is + 1, x[i].scales, sc, m);
|
||||
const half d2 = __hmul(dall, __int2half_rn(sc)); const half m2 = __hmul(dmin, __int2half_rn(m));
|
||||
|
||||
uint8_t hm = 1 << (2*il);
|
||||
y[ 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1));
|
||||
y[ 1] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1));
|
||||
hm <<= 1;
|
||||
y[32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[0] >> 4) + (qh[0] & hm ? 16 : 0))), m2));
|
||||
y[33] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[1] >> 4) + (qh[1] & hm ? 16 : 0))), m2));
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
const block_q6_K * x = (const block_q6_K *) vx;
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
|
||||
// assume 64 threads - this is very slightly better than the one below
|
||||
const auto tid = threadIdx.x;
|
||||
const int ip = tid/32; // ip is 0 or 1
|
||||
const int il = tid - 32*ip; // 0...32
|
||||
const int is = 8*ip + il/16;
|
||||
|
||||
dst_t * y = yy + i*QK_K + 128*ip + il;
|
||||
|
||||
const half d = x[i].d;
|
||||
|
||||
const uint8_t * ql = x[i].ql + 64*ip + il;
|
||||
const uint8_t qh = x[i].qh[32*ip + il];
|
||||
const int8_t * sc = x[i].scales + is;
|
||||
|
||||
y[ 0] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32))));
|
||||
y[32] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32))));
|
||||
y[64] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32))));
|
||||
y[96] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32))));
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq2_xxs * x = (const block_iq2_xxs *) vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const uint16_t * q2 = x[i].qs + 4*ib;
|
||||
const uint8_t * aux8 = (const uint8_t *)q2;
|
||||
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
|
||||
const uint32_t aux32 = q2[2] | (q2[3] << 16);
|
||||
const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.25f;
|
||||
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
|
||||
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq2_xs * x = (const block_iq2_xs *) vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const uint16_t * q2 = x[i].qs + 4*ib;
|
||||
const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
|
||||
const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
|
||||
const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
|
||||
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
|
||||
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq2_s * x = (const block_iq2_s *) vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
|
||||
const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
|
||||
const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
|
||||
for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq3_xxs * x = (const block_iq3_xxs *) vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const uint8_t * q3 = x[i].qs + 8*ib;
|
||||
const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
|
||||
const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
|
||||
const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
|
||||
const uint32_t aux32 = gas[0] | (gas[1] << 16);
|
||||
const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.5f;
|
||||
const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
|
||||
for (int j = 0; j < 4; ++j) {
|
||||
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
|
||||
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq3_s * x = (const block_iq3_s *) vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const uint8_t * qs = x[i].qs + 8*ib;
|
||||
const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
|
||||
const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
|
||||
const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f;
|
||||
const uint8_t signs = x[i].signs[4*ib + il];
|
||||
for (int j = 0; j < 4; ++j) {
|
||||
y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
|
||||
y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const int64_t i = blockIdx.x;
|
||||
const block_iq1_s * x = (const block_iq1_s *) vx;
|
||||
|
||||
const int64_t tid = threadIdx.x;
|
||||
const int64_t il = tid/8; // 0...3
|
||||
const int64_t ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
|
||||
const float d = __half2float(x[i].d) * (2*((x[i].qh[ib] >> 12) & 7) + 1);
|
||||
uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
|
||||
grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
|
||||
grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
|
||||
grid32[0] &= 0x0f0f0f0f;
|
||||
for (int j = 0; j < 8; ++j) {
|
||||
y[j] = d * (q[j] + delta);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const int64_t i = blockIdx.x;
|
||||
const block_iq1_m * x = (const block_iq1_m *) vx;
|
||||
|
||||
const int64_t tid = threadIdx.x;
|
||||
const int64_t il = tid/8; // 0...3
|
||||
const int64_t ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
|
||||
const uint16_t * sc = (const uint16_t *)x[i].scales;
|
||||
iq1m_scale_t scale;
|
||||
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
|
||||
const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
|
||||
const float d = __half2float(scale.f16) * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
|
||||
const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
|
||||
uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
|
||||
grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
|
||||
grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
|
||||
grid32[0] &= 0x0f0f0f0f;
|
||||
for (int j = 0; j < 8; ++j) {
|
||||
y[j] = d * (q[j] + delta);
|
||||
}
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
|
||||
const uint8_t * q4 = x[ib].qs + 4*il;
|
||||
const float d = __half2float(x[ib].d);
|
||||
for (int j = 0; j < 4; ++j) {
|
||||
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
|
||||
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
|
||||
const auto i = blockIdx.x;
|
||||
const block_iq4_xs * x = (const block_iq4_xs *)vx;
|
||||
|
||||
const auto tid = threadIdx.x;
|
||||
const int il = tid/8; // 0...3
|
||||
const int ib = tid%8; // 0...7
|
||||
dst_t * y = yy + i*QK_K + 32*ib + 4*il;
|
||||
const uint8_t * q4 = x[i].qs + 16*ib + 4*il;
|
||||
const float d = __half2float(x[i].d) * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
|
||||
for (int j = 0; j < 4; ++j) {
|
||||
y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
|
||||
y[j+16] = d * kvalues_iq4nl[q4[j] >> 4];
|
||||
}
|
||||
}
|
||||
|
||||
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
|
||||
static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
|
||||
const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
|
||||
dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = k / QK_K;
|
||||
dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = (k + QK_K - 1) / QK_K;
|
||||
dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
|
||||
const int nb = (k + QK_K - 1) / QK_K;
|
||||
dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
template<typename dst_t>
|
||||
static to_cuda_ggml_t<dst_t> ggml_get_to_cuda(int64_t type) {
|
||||
switch (type) {
|
||||
case 2:
|
||||
return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
|
||||
case 3:
|
||||
return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
|
||||
case 6:
|
||||
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
|
||||
case 7:
|
||||
return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
|
||||
case 8:
|
||||
return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
|
||||
case 10:
|
||||
return dequantize_row_q2_K_cuda;
|
||||
case 11:
|
||||
return dequantize_row_q3_K_cuda;
|
||||
case 12:
|
||||
return dequantize_row_q4_K_cuda;
|
||||
case 13:
|
||||
return dequantize_row_q5_K_cuda;
|
||||
case 14:
|
||||
return dequantize_row_q6_K_cuda;
|
||||
case 16:
|
||||
return dequantize_row_iq2_xxs_cuda;
|
||||
case 17:
|
||||
return dequantize_row_iq2_xs_cuda;
|
||||
case 18:
|
||||
return dequantize_row_iq3_xxs_cuda;
|
||||
case 19:
|
||||
return dequantize_row_iq1_s_cuda;
|
||||
case 20:
|
||||
return dequantize_row_iq4_nl_cuda;
|
||||
case 21:
|
||||
return dequantize_row_iq3_s_cuda;
|
||||
case 22:
|
||||
return dequantize_row_iq2_s_cuda;
|
||||
case 23:
|
||||
return dequantize_row_iq4_xs_cuda;
|
||||
case 29:
|
||||
return dequantize_row_iq1_m_cuda;
|
||||
default:
|
||||
return nullptr;
|
||||
}
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -7,14 +7,11 @@
|
||||
|
||||
#include <torch/csrc/stable/ops.h>
|
||||
|
||||
// NOTE: These headers are intentionally kept in csrc/quantization/gguf/ (not
|
||||
// moved to libtorch_stable) to avoid unnecessary reformatting that would break
|
||||
// git rename detection and pollute blame history.
|
||||
#include "../../../quantization/gguf/ggml-common.h"
|
||||
#include "../../../quantization/gguf/vecdotq.cuh"
|
||||
#include "../../../quantization/gguf/dequantize.cuh"
|
||||
#include "../../../quantization/gguf/mmvq.cuh"
|
||||
#include "../../../quantization/gguf/mmq.cuh"
|
||||
#include "ggml-common.h"
|
||||
#include "vecdotq.cuh"
|
||||
#include "dequantize.cuh"
|
||||
#include "mmvq.cuh"
|
||||
#include "mmq.cuh"
|
||||
#include "moe.cuh"
|
||||
#include "moe_vec.cuh"
|
||||
|
||||
|
||||
@@ -0,0 +1,610 @@
|
||||
// copied from https://github.com/ggerganov/llama.cpp/blob/b2899/ggml-cuda/mmq.cu
|
||||
template <typename scalar_t, int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
|
||||
allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
|
||||
static __device__ __forceinline__ void mul_mat_q(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
|
||||
const block_q_t * x = (const block_q_t *) vx;
|
||||
const block_q8_1 * y = (const block_q8_1 *) vy;
|
||||
|
||||
const int blocks_per_row_x = ncols_x / qk;
|
||||
const int blocks_per_col_y = nrows_y / QK8_1;
|
||||
const int blocks_per_warp = WARP_SIZE_GGUF / qi;
|
||||
|
||||
const int & ncols_dst = ncols_y;
|
||||
|
||||
const auto row_dst_0 = blockIdx.x*mmq_y;
|
||||
const int & row_x_0 = row_dst_0;
|
||||
|
||||
const auto col_dst_0 = blockIdx.y*mmq_x;
|
||||
const int & col_y_0 = col_dst_0;
|
||||
|
||||
int * tile_x_ql = nullptr;
|
||||
half2 * tile_x_dm = nullptr;
|
||||
int * tile_x_qh = nullptr;
|
||||
int * tile_x_sc = nullptr;
|
||||
|
||||
allocate_tiles(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc);
|
||||
|
||||
__shared__ int tile_y_qs[mmq_x * WARP_SIZE_GGUF];
|
||||
__shared__ half2 tile_y_ds[mmq_x * WARP_SIZE_GGUF/QI8_1];
|
||||
|
||||
float sum[mmq_y/WARP_SIZE_GGUF][mmq_x/nwarps] = {{0.0f}};
|
||||
|
||||
for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) {
|
||||
|
||||
load_tiles(x + row_x_0*blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc,
|
||||
threadIdx.y, nrows_x-row_x_0-1, threadIdx.x, blocks_per_row_x);
|
||||
|
||||
#pragma unroll
|
||||
for (int ir = 0; ir < qr && ib0 + ir * blocks_per_warp/qr < blocks_per_row_x; ++ir) {
|
||||
const auto kqs = ir*WARP_SIZE_GGUF + threadIdx.x;
|
||||
const int kbxd = kqs / QI8_1;
|
||||
|
||||
#pragma unroll
|
||||
for (int i = 0; i < mmq_x; i += nwarps) {
|
||||
const int col_y_eff = min(col_y_0 + threadIdx.y + i, ncols_y-1); // to prevent out-of-bounds memory accesses
|
||||
const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd];
|
||||
const int index_y = (threadIdx.y + i) * WARP_SIZE_GGUF + kqs % WARP_SIZE_GGUF;
|
||||
tile_y_qs[index_y] = get_int_from_int8_aligned(by0->qs, threadIdx.x % QI8_1);
|
||||
}
|
||||
|
||||
#pragma unroll
|
||||
for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) {
|
||||
const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE_GGUF/QI8_1)) % mmq_x;
|
||||
const auto kby = threadIdx.x % (WARP_SIZE_GGUF/QI8_1);
|
||||
const int col_y_eff = min(col_y_0 + ids, ncols_y-1);
|
||||
|
||||
// if the sum is not needed it's faster to transform the scale to f32 ahead of time
|
||||
const half2 * dsi_src = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + ir*(WARP_SIZE_GGUF/QI8_1) + kby].ds;
|
||||
half2 * dsi_dst = &tile_y_ds[ids * (WARP_SIZE_GGUF/QI8_1) + kby];
|
||||
if (need_sum) {
|
||||
*dsi_dst = *dsi_src;
|
||||
} else {
|
||||
float * dfi_dst = (float *) dsi_dst;
|
||||
*dfi_dst = __low2float(*dsi_src);
|
||||
}
|
||||
}
|
||||
|
||||
__syncthreads();
|
||||
|
||||
// #pragma unroll // unrolling this loop causes too much register pressure
|
||||
for (int k = ir*WARP_SIZE_GGUF/qr; k < (ir+1)*WARP_SIZE_GGUF/qr; k += vdr) {
|
||||
#pragma unroll
|
||||
for (int j = 0; j < mmq_x; j += nwarps) {
|
||||
#pragma unroll
|
||||
for (int i = 0; i < mmq_y; i += WARP_SIZE_GGUF) {
|
||||
sum[i/WARP_SIZE_GGUF][j/nwarps] += vec_dot(
|
||||
tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, tile_y_qs, tile_y_ds,
|
||||
threadIdx.x + i, threadIdx.y + j, k);
|
||||
}
|
||||
}
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
#pragma unroll
|
||||
for (int j = 0; j < mmq_x; j += nwarps) {
|
||||
const auto col_dst = col_dst_0 + j + threadIdx.y;
|
||||
if (col_dst >= ncols_dst) {
|
||||
return;
|
||||
}
|
||||
|
||||
#pragma unroll
|
||||
for (int i = 0; i < mmq_y; i += WARP_SIZE_GGUF) {
|
||||
const auto row_dst = row_dst_0 + threadIdx.x + i;
|
||||
if (row_dst >= nrows_dst) {
|
||||
continue;
|
||||
}
|
||||
dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE_GGUF][j/nwarps];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q4_0 64
|
||||
#define MMQ_Y_Q4_0 128
|
||||
#define NWARPS_Q4_0 8
|
||||
#else
|
||||
#define MMQ_X_Q4_0 4
|
||||
#define MMQ_Y_Q4_0 32
|
||||
#define NWARPS_Q4_0 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q4_0, 2)
|
||||
#endif
|
||||
mul_mat_q4_0(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q4_0;
|
||||
const int mmq_y = MMQ_Y_Q4_0;
|
||||
const int nwarps = NWARPS_Q4_0;
|
||||
|
||||
mul_mat_q<scalar_t, QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
|
||||
load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q4_0_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
int mmq_x = MMQ_X_Q4_0;
|
||||
int mmq_y = MMQ_Y_Q4_0;
|
||||
int nwarps = NWARPS_Q4_0;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q4_0<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q4_0<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q4_1 64
|
||||
#define MMQ_Y_Q4_1 128
|
||||
#define NWARPS_Q4_1 8
|
||||
#else
|
||||
#define MMQ_X_Q4_1 4
|
||||
#define MMQ_Y_Q4_1 32
|
||||
#define NWARPS_Q4_1 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q4_1, 2)
|
||||
#endif
|
||||
mul_mat_q4_1(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q4_1;
|
||||
const int mmq_y = MMQ_Y_Q4_1;
|
||||
const int nwarps = NWARPS_Q4_1;
|
||||
|
||||
mul_mat_q<scalar_t, QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
|
||||
load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q4_1_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
int mmq_x = MMQ_X_Q4_1;
|
||||
int mmq_y = MMQ_Y_Q4_1;
|
||||
int nwarps = NWARPS_Q4_1;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q4_1<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q4_1<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q5_0 64
|
||||
#define MMQ_Y_Q5_0 128
|
||||
#define NWARPS_Q5_0 8
|
||||
#else
|
||||
#define MMQ_X_Q5_0 4
|
||||
#define MMQ_Y_Q5_0 32
|
||||
#define NWARPS_Q5_0 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q5_0, 2)
|
||||
#endif
|
||||
mul_mat_q5_0(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q5_0;
|
||||
const int mmq_y = MMQ_Y_Q5_0;
|
||||
const int nwarps = NWARPS_Q5_0;
|
||||
|
||||
mul_mat_q<scalar_t, QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
|
||||
load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q5_0_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
const int mmq_x = MMQ_X_Q5_0;
|
||||
const int mmq_y = MMQ_Y_Q5_0;
|
||||
const int nwarps = NWARPS_Q5_0;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q5_0<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q5_0<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q5_1 64
|
||||
#define MMQ_Y_Q5_1 128
|
||||
#define NWARPS_Q5_1 8
|
||||
#else
|
||||
#define MMQ_X_Q5_1 4
|
||||
#define MMQ_Y_Q5_1 32
|
||||
#define NWARPS_Q5_1 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q5_1, 2)
|
||||
#endif
|
||||
mul_mat_q5_1(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q5_1;
|
||||
const int mmq_y = MMQ_Y_Q5_1;
|
||||
const int nwarps = NWARPS_Q5_1;
|
||||
|
||||
mul_mat_q<scalar_t, QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
|
||||
load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q5_1_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const int mmq_x = MMQ_X_Q5_1;
|
||||
const int mmq_y = MMQ_Y_Q5_1;
|
||||
const int nwarps = NWARPS_Q5_1;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q5_1<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q5_1<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q8_0 64
|
||||
#define MMQ_Y_Q8_0 128
|
||||
#define NWARPS_Q8_0 8
|
||||
#else
|
||||
#define MMQ_X_Q8_0 4
|
||||
#define MMQ_Y_Q8_0 32
|
||||
#define NWARPS_Q8_0 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q8_0, 2)
|
||||
#endif
|
||||
mul_mat_q8_0(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q8_0;
|
||||
const int mmq_y = MMQ_Y_Q8_0;
|
||||
const int nwarps = NWARPS_Q8_0;
|
||||
|
||||
mul_mat_q<scalar_t, QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
|
||||
load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q8_0_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const int mmq_x = MMQ_X_Q8_0;
|
||||
const int mmq_y = MMQ_Y_Q8_0;
|
||||
const int nwarps = NWARPS_Q8_0;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q8_0<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q8_0<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q2_K 64
|
||||
#define MMQ_Y_Q2_K 128
|
||||
#define NWARPS_Q2_K 8
|
||||
#else
|
||||
#define MMQ_X_Q2_K 4
|
||||
#define MMQ_Y_Q2_K 32
|
||||
#define NWARPS_Q2_K 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q2_K, 2)
|
||||
#endif
|
||||
mul_mat_q2_K(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q2_K;
|
||||
const int mmq_y = MMQ_Y_Q2_K;
|
||||
const int nwarps = NWARPS_Q2_K;
|
||||
|
||||
mul_mat_q<scalar_t, QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
|
||||
load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q2_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const int mmq_x = MMQ_X_Q2_K;
|
||||
const int mmq_y = MMQ_Y_Q2_K;
|
||||
const int nwarps = NWARPS_Q2_K;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q2_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q2_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q3_K 64
|
||||
#define MMQ_Y_Q3_K 128
|
||||
#define NWARPS_Q3_K 8
|
||||
#else
|
||||
#define MMQ_X_Q3_K 4
|
||||
#define MMQ_Y_Q3_K 32
|
||||
#define NWARPS_Q3_K 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q3_K, 2)
|
||||
#endif
|
||||
mul_mat_q3_K(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
|
||||
const int mmq_x = MMQ_X_Q3_K;
|
||||
const int mmq_y = MMQ_Y_Q3_K;
|
||||
const int nwarps = NWARPS_Q3_K;
|
||||
|
||||
mul_mat_q<scalar_t, QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
|
||||
load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q3_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
const int mmq_x = MMQ_X_Q3_K;
|
||||
const int mmq_y = MMQ_Y_Q3_K;
|
||||
const int nwarps = NWARPS_Q3_K;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q3_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q3_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q4_K 64
|
||||
#define MMQ_Y_Q4_K 128
|
||||
#define NWARPS_Q4_K 8
|
||||
#else
|
||||
#define MMQ_X_Q4_K 4
|
||||
#define MMQ_Y_Q4_K 32
|
||||
#define NWARPS_Q4_K 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q4_K, 2)
|
||||
#endif
|
||||
mul_mat_q4_K(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q4_K;
|
||||
const int mmq_y = MMQ_Y_Q4_K;
|
||||
const int nwarps = NWARPS_Q4_K;
|
||||
|
||||
mul_mat_q<scalar_t, QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
|
||||
load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q4_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const int mmq_x = MMQ_X_Q4_K;
|
||||
const int mmq_y = MMQ_Y_Q4_K;
|
||||
const int nwarps = NWARPS_Q4_K;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q4_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q4_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q5_K 64
|
||||
#define MMQ_Y_Q5_K 128
|
||||
#define NWARPS_Q5_K 8
|
||||
#else
|
||||
#define MMQ_X_Q5_K 4
|
||||
#define MMQ_Y_Q5_K 32
|
||||
#define NWARPS_Q5_K 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q5_K, 2)
|
||||
#endif
|
||||
mul_mat_q5_K(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q5_K;
|
||||
const int mmq_y = MMQ_Y_Q5_K;
|
||||
const int nwarps = NWARPS_Q5_K;
|
||||
|
||||
mul_mat_q<scalar_t, QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
|
||||
load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q5_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
const int mmq_x = MMQ_X_Q5_K;
|
||||
const int mmq_y = MMQ_Y_Q5_K;
|
||||
const int nwarps = NWARPS_Q5_K;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q5_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q5_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
#define MMQ_X_Q6_K 64
|
||||
#define MMQ_Y_Q6_K 128
|
||||
#define NWARPS_Q6_K 8
|
||||
#else
|
||||
#define MMQ_X_Q6_K 4
|
||||
#define MMQ_Y_Q6_K 32
|
||||
#define NWARPS_Q6_K 4
|
||||
#endif
|
||||
|
||||
template<typename scalar_t, bool need_check> static __global__ void
|
||||
#if defined(USE_ROCM)
|
||||
__launch_bounds__(WARP_SIZE_GGUF*NWARPS_Q6_K, 2)
|
||||
#endif
|
||||
mul_mat_q6_K(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
|
||||
const int mmq_x = MMQ_X_Q6_K;
|
||||
const int mmq_y = MMQ_Y_Q6_K;
|
||||
const int nwarps = NWARPS_Q6_K;
|
||||
|
||||
mul_mat_q<scalar_t, QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
|
||||
load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void ggml_mul_mat_q6_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, scalar_t * dst, const int ncols_x, const int nrows_x,
|
||||
const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const int mmq_x = MMQ_X_Q6_K;
|
||||
const int mmq_y = MMQ_Y_Q6_K;
|
||||
const int nwarps = NWARPS_Q6_K;
|
||||
|
||||
const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
|
||||
const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
|
||||
const dim3 block_nums(block_num_x, block_num_y, 1);
|
||||
const dim3 block_dims(WARP_SIZE_GGUF, nwarps, 1);
|
||||
|
||||
if (nrows_x % mmq_y == 0) {
|
||||
const bool need_check = false;
|
||||
mul_mat_q6_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
} else {
|
||||
const bool need_check = true;
|
||||
mul_mat_q6_K<scalar_t, need_check><<<block_nums, block_dims, 0, stream>>>
|
||||
(vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,212 @@
|
||||
// copied and adapted from https://github.com/ggerganov/llama.cpp/blob/b2899/ggml-cuda/mmvq.cu
|
||||
template <typename scalar_t, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
|
||||
static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, scalar_t * __restrict__ dst, const int ncols, const int nrows, const int nvecs) {
|
||||
const auto row = blockIdx.x*blockDim.y + threadIdx.y;
|
||||
const auto vec = blockIdx.y;
|
||||
|
||||
if (row >= nrows || vec >= nvecs) {
|
||||
return;
|
||||
}
|
||||
|
||||
const int blocks_per_row = ncols / qk;
|
||||
const int blocks_per_warp = vdr * WARP_SIZE / qi;
|
||||
const int nrows_y = (ncols + 512 - 1) / 512 * 512;
|
||||
|
||||
|
||||
// partial sum for each thread
|
||||
float tmp = 0.0f;
|
||||
|
||||
const block_q_t * x = (const block_q_t *) vx;
|
||||
const block_q8_1 * y = (const block_q8_1 *) vy;
|
||||
|
||||
for (auto i = threadIdx.x / (qi/vdr); i < blocks_per_row; i += blocks_per_warp) {
|
||||
const int ibx = row*blocks_per_row + i; // x block index
|
||||
|
||||
const int iby = vec*(nrows_y/QK8_1) + i * (qk/QK8_1); // y block index that aligns with ibx
|
||||
|
||||
const int iqs = vdr * (threadIdx.x % (qi/vdr)); // x block quant index when casting the quants to int
|
||||
|
||||
tmp += vec_dot_q_cuda(&x[ibx], &y[iby], iqs);
|
||||
}
|
||||
|
||||
// sum up partial sums and write back result
|
||||
#pragma unroll
|
||||
for (int mask = WARP_SIZE/2; mask > 0; mask >>= 1) {
|
||||
tmp += VLLM_SHFL_XOR_SYNC(tmp, mask);
|
||||
}
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
dst[vec*nrows + row] = tmp;
|
||||
}
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q2_K_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q3_K_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_q6_K_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq2_xxs_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI2_XXS, block_iq2_xxs, 1, vec_dot_iq2_xxs_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq2_xs_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI2_XS, block_iq2_xs, 1, vec_dot_iq2_xs_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq2_s_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI2_S, block_iq2_s, 1, vec_dot_iq2_s_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq3_xxs_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq1_s_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI1_S, block_iq1_s, 1, vec_dot_iq1_s_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq1_m_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI1_M, block_iq1_m, 1, vec_dot_iq1_m_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq4_nl_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK4_NL, QI4_NL, block_iq4_nl, VDR_Q4_0_Q8_1_MMVQ, vec_dot_iq4_nl_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq4_xs_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI4_XS, block_iq4_xs, 1, vec_dot_iq4_xs_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
|
||||
template<typename scalar_t>
|
||||
static void mul_mat_vec_iq3_s_q8_1_cuda(const void * vx, const void * vy, scalar_t * dst, const int ncols, const int nrows, const int nvecs, cudaStream_t stream) {
|
||||
const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
|
||||
const dim3 block_nums(block_num_y, nvecs, 1);
|
||||
const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
|
||||
mul_mat_vec_q<scalar_t, QK_K, QI3_XS, block_iq3_s, 1, vec_dot_iq3_s_q8_1>
|
||||
<<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows, nvecs);
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -7,7 +7,7 @@
|
||||
#include "torch_utils.h"
|
||||
|
||||
#ifndef USE_ROCM
|
||||
#include "../persistent_topk.cuh"
|
||||
#include "persistent_topk.cuh"
|
||||
#endif
|
||||
|
||||
namespace {
|
||||
|
||||
Reference in New Issue
Block a user