EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[core] Support custom ascendc kernels in vllm-ascend (#233)

Browse Source 
This PR add custom ascendc kernel rotary_embedding support in
vllm-ascend, related CMakeLists and setuptools is also added in this PR.

Related: https://github.com/vllm-project/vllm-ascend/issues/156

---------

Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
This commit is contained in:
Pleaplusone
2025-04-03 14:52:34 +08:00
committed by GitHub
parent 14d9a64047
commit ce8259975e
15 changed files with 1378 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

367
csrc/kernels/pos_encoding_kernels.cpp Normal file
View File

@@ -0,0 +1,367 @@
/*
* Copyright (c) Huawei Technologies Co., Ltd. 2024. 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.
*/
#include "kernel_operator.h"
#include "kernel_tpipe_impl.h"
#include "kernel_tensor_impl.h"
#include "kernel_type.h"
#include "kernel_operator_intf.h"
#include "inner_interface/inner_kernel_operator_intf.h"
#include <stdio.h>
#include "types.h"
#include "utils.h"
using vllm_ascend::AccType;
using vllm_ascend::local_mem_copy;
template <typename scalar_t, bool isNeox> class RotaryEmbedding {
// NOTE(ganyi): we use 32K as load stride for pipe, need to find another way to
// retrive this size from runtime for more Soc support
static int constexpr loadSize = 1024 * 4;
using dst_t = scalar_t;
using acc_t = typename AccType<scalar_t>::type;
// only half tensor have cast instruct to int8, hardcode acc_dst_t as half
using local_scalar_t = AscendC::LocalTensor<scalar_t>;
using local_acc_t = AscendC::LocalTensor<acc_t>;
using local_dst_t = AscendC::LocalTensor<dst_t>;
public:
__aicore__ inline RotaryEmbedding()
{
}
// Allocate buffers for input and output queue and the temp buffer used during kernel compute process,
// this init process happens only in the kernel compute on a single vector core.
__aicore__ inline void init(__gm__ int64_t *positions, __gm__ void *queryDst, __gm__ void *keyDst,
__gm__ scalar_t *query, __gm__ scalar_t *key, __gm__ scalar_t *cosSinCache,
const int rotDim, const int64_t dstQueryStride,
const int64_t dstKeyStride, const int64_t queryStride, const int64_t keyStride,
const int numHeads, const int numKvHeads, const int headSize, AscendC::TPipe *pipe)
{
pipe_ = pipe;
rotDim_ = rotDim;
// query stride and key stride is used to handle the strided tensor which is not contiguous on num_tokens dim
queryStride_ = queryStride;
keyStride_ = keyStride;
dstQueryStride_ = dstQueryStride;
dstKeyStride_ = dstKeyStride;
numHeads_ = numHeads;
numKvHeads_ = numKvHeads;
headSize_ = headSize;
embedDim_ = rotDim / 2;
pipe_->InitBuffer(inQue_, 1 /* buffer_num */, loadSize /* buffer_size */);
pipe_->InitBuffer(inQueSinCos_, 1 /* buffer_num */, rotDim_ * sizeof(scalar_t) /* buffer_size */);
pipe_->InitBuffer(outQue_, 1 /* buffer_num */, loadSize /* buffer_size */);
// 2 temperary calculation buffer
calcTmpBufferOffset_ = 0;
// 1 upcast buffer for bf16 (headSize)
upcastInputBufferOffset_ = calcTmpBufferOffset_ + sizeof(acc_t) * embedDim_ * 2;
// 1 upcast temp buffer for bf16 (2 * embed_dim)
upcastTempBufferOffset_ = upcastInputBufferOffset_ + sizeof(acc_t) * headSize_;
// 2 sin cos upcast buffer for bf16
cosSinUpcastBufferOffset_ = upcastTempBufferOffset_ + sizeof(acc_t) * 2 * embedDim_;
// 2. bf16 path: needs 2 cos sin upcast buffer size
// 3. fp16 path: needs 2 temperary calculation buffer size
tempBufferSize_ = cosSinUpcastBufferOffset_ + 2 * embedDim_ * sizeof(acc_t);
// need to consider upcast the bf16 to fp32, so we might need 4 buffer just in case
// 2 temperary buffer, 2 input buffer, 1 cos buffer, 1 sin buffer, 2 scale buffer (headSize), 2 zp
// buffer(headSize int8), 1 dst_temp buffer(headSize, int32)
pipe_->InitBuffer(calcBuf_, tempBufferSize_ /* buffer_size */);
if constexpr (!std::is_same_v<scalar_t, acc_t>) {
pipe_->InitBuffer(copyBuf_, loadSize);
}
}
__aicore__ inline void update_mem_offset(__gm__ int64_t *positions, __gm__ void *queryDst, __gm__ void *keyDst,
__gm__ scalar_t *query, __gm__ scalar_t *key, __gm__ scalar_t *cosSinCache,
const int rotDim, const int64_t dstQueryStride, const int64_t dstKeyStride,
const int64_t queryStride, const int64_t keyStride, const int numHeads,
const int numKvHeads, const int headSize, const int64_t idx)
{
int64_t pos = positions[idx];
cosSin_.SetGlobalBuffer(cosSinCache + pos * rotDim_, rotDim_);
query_.SetGlobalBuffer(query + queryStride * idx, headSize * numHeads_);
key_.SetGlobalBuffer(key + keyStride * idx, headSize * numKvHeads_);
queryDst_.SetGlobalBuffer(reinterpret_cast<__gm__ dst_t *>(queryDst) + dstQueryStride * idx,
headSize * numHeads_);
keyDst_.SetGlobalBuffer(reinterpret_cast<__gm__ dst_t *>(keyDst) + dstKeyStride * idx, headSize * numKvHeads_);
}
// compute per head for neox on bf16
template <typename acc_t_, typename std::enable_if<!std::is_same_v<acc_t_, scalar_t>, void>::type * = nullptr>
__aicore__ inline void
neox_compute(local_scalar_t src, local_dst_t dst, AscendC::LocalTensor<acc_t_> sin, AscendC::LocalTensor<acc_t_> cos,
AscendC::LocalTensor<acc_t_> upcastInputBuffer, AscendC::LocalTensor<acc_t_> calcTmpBuffer)
{
// slice dst
local_dst_t dstX = dst;
local_dst_t dstY = dst[embedDim_];
// slice src
local_scalar_t srcX = src;
local_scalar_t srcY = src[embedDim_];
// slice temp buffer
local_acc_t calcTmpBufferX = calcTmpBuffer;
local_acc_t calcTmpBufferY = calcTmpBuffer[embedDim_];
// slice upcast input buffer
local_acc_t upcastBufferX = upcastInputBuffer;
local_acc_t upcastBufferY = upcastBufferX[embedDim_];
// dst x calc
Cast(upcastInputBuffer, src, AscendC::RoundMode::CAST_NONE, headSize_);
Mul(calcTmpBufferX, upcastBufferX, cos, embedDim_);
Mul(calcTmpBufferY, upcastBufferY, sin, embedDim_);
Sub(calcTmpBufferX, calcTmpBufferX, calcTmpBufferY, embedDim_);
Cast(dstX, calcTmpBufferX, AscendC::RoundMode::CAST_TRUNC, embedDim_);
// dst y calc
Mul(calcTmpBufferX, upcastBufferX, sin, embedDim_);
Mul(calcTmpBufferY, upcastBufferY, cos, embedDim_);
Add(calcTmpBufferX, calcTmpBufferX, calcTmpBufferY, embedDim_);
Cast(dstY, calcTmpBufferX, AscendC::RoundMode::CAST_TRUNC, embedDim_);
}
// compute per head output for neox
template <typename acc_t_, typename std::enable_if<std::is_same_v<acc_t_, scalar_t>, void>::type * = nullptr>
__aicore__ inline void
neox_compute(local_scalar_t src, local_dst_t dst, AscendC::LocalTensor<acc_t_> sin, AscendC::LocalTensor<acc_t_> cos,
AscendC::LocalTensor<acc_t_> upcastInputBuffer, AscendC::LocalTensor<acc_t_> calcTmpBuffer)
{
// slice dst buffer
local_dst_t dstX = dst;
local_dst_t dstY = dst[embedDim_];
// slice src buffer
local_scalar_t srcX = src;
local_scalar_t srcY = src[embedDim_];
// slice temp buffer
local_acc_t calcTmpBufferX = calcTmpBuffer;
local_acc_t calcTmpBufferY = calcTmpBuffer[embedDim_];
// dst x calc
Mul(calcTmpBufferX, srcX, cos, embedDim_);
Mul(calcTmpBufferY, srcY, sin, embedDim_);
Sub(dstX, calcTmpBufferX, calcTmpBufferY, embedDim_);
// dst y calc
Mul(calcTmpBufferX, srcX, sin, embedDim_);
Mul(calcTmpBufferY, srcY, cos, embedDim_);
Add(dstY, calcTmpBufferX, calcTmpBufferY, embedDim_);
}
__aicore__ inline void compute_qk(AscendC::GlobalTensor<scalar_t> srcG, AscendC::GlobalTensor<dst_t> dstG,
local_acc_t localCos, local_acc_t localSin, local_acc_t upcastInputBuffer,
local_acc_t calcTmpBuffer, int loopCnt, int tailHeads, int loadStride,
int headNumPerLoad)
{
for (int loopNum = 0; loopNum < loopCnt; ++loopNum) {
local_scalar_t src = inQue_.AllocTensor<scalar_t>();
local_dst_t dst = outQue_.AllocTensor<dst_t>();
AscendC::DataCopy(src, srcG[loopNum * loadStride], loadStride);
inQue_.EnQue(src);
local_scalar_t srcDeque = inQue_.DeQue<scalar_t>();
if constexpr (!std::is_same_v<scalar_t, acc_t>) {
int elem_num = loadStride / sizeof(scalar_t);
AscendC::LocalTensor<acc_t> upBuffer = copyBuf_.GetWithOffset<acc_t>(elem_num, 0);
Cast(upBuffer, srcDeque, AscendC::RoundMode::CAST_TRUNC, elem_num);
Cast(dst, upBuffer, AscendC::RoundMode::CAST_TRUNC, elem_num);
} else {
local_mem_copy(dst, srcDeque, loadStride);
}
for (int i = 0; i < headNumPerLoad; ++i) {
neox_compute(srcDeque[i * headSize_], dst[i * headSize_], localSin, localCos, upcastInputBuffer,
calcTmpBuffer);
}
outQue_.EnQue(dst);
local_dst_t dstDeque = outQue_.DeQue<dst_t>();
AscendC::DataCopy(dstG[loopNum * loadStride], dstDeque, loadStride);
outQue_.FreeTensor(dstDeque);
inQue_.FreeTensor(srcDeque);
}
// process tail
{
local_scalar_t src = inQue_.AllocTensor<scalar_t>();
local_dst_t dst = outQue_.AllocTensor<dst_t>();
AscendC::DataCopy(src, srcG[loopCnt * loadStride], tailHeads * headSize_);
inQue_.EnQue(src);
local_scalar_t srcDeque = inQue_.DeQue<scalar_t>();
if constexpr (!std::is_same_v<scalar_t, acc_t>) {
int elem_num = tailHeads * headSize_ / sizeof(scalar_t);
AscendC::LocalTensor<acc_t> upBuffer = copyBuf_.GetWithOffset<acc_t>(elem_num, 0);
Cast(upBuffer, srcDeque, AscendC::RoundMode::CAST_TRUNC, elem_num);
Cast(dst, upBuffer, AscendC::RoundMode::CAST_TRUNC, elem_num);
} else {
local_mem_copy(dst, srcDeque, tailHeads * headSize_);
}
for (int i = 0; i < tailHeads; ++i) {
neox_compute(srcDeque[i * headSize_], dst[i * headSize_], localSin, localCos, upcastInputBuffer,
calcTmpBuffer);
}
outQue_.EnQue(dst);
local_dst_t dstDeque = outQue_.DeQue<dst_t>();
AscendC::DataCopy(dstG[loopCnt * loadStride], dstDeque, tailHeads * headSize_);
outQue_.FreeTensor(dstDeque);
inQue_.FreeTensor(srcDeque);
}
}
__aicore__ inline void compute_function()
{
local_scalar_t cosSinLocal = inQueSinCos_.AllocTensor<scalar_t>();
AscendC::DataCopy(cosSinLocal, cosSin_, embedDim_ * 2);
inQueSinCos_.EnQue(cosSinLocal);
local_scalar_t localSinCosDeque = inQueSinCos_.DeQue<scalar_t>();
local_scalar_t localCos = localSinCosDeque;
local_scalar_t localSin = localSinCosDeque[embedDim_];
local_acc_t calcTmpBuffer;
local_acc_t upcastInputBuffer;
local_acc_t upcastTempBuffer;
local_acc_t cosSinUpcastBuffer;
local_acc_t scaleBuffer;
local_acc_t offsetBuffer;
calcTmpBuffer = calcBuf_.GetWithOffset<acc_t>(embedDim_ * 2, calcTmpBufferOffset_);
upcastInputBuffer = calcBuf_.GetWithOffset<acc_t>(headSize_, upcastInputBufferOffset_);
upcastTempBuffer = calcBuf_.GetWithOffset<acc_t>(embedDim_ * 2, upcastTempBufferOffset_);
cosSinUpcastBuffer = calcBuf_.GetWithOffset<acc_t>(embedDim_ * 2, cosSinUpcastBufferOffset_);
local_acc_t cosAccBuffer;
local_acc_t sinAccBuffer;
if constexpr (!std::is_same_v<scalar_t, acc_t>) {
Cast(cosSinUpcastBuffer, localSinCosDeque, AscendC::RoundMode::CAST_NONE, 2 * embedDim_);
cosAccBuffer = cosSinUpcastBuffer;
sinAccBuffer = cosSinUpcastBuffer[embedDim_];
} else {
cosAccBuffer = localCos;
sinAccBuffer = localSin;
}
constexpr const int loadSizeByElem = loadSize / sizeof(scalar_t);
int64_t headNumPerLoad = loadSizeByElem / headSize_;
int64_t loopCnt = numHeads_ / headNumPerLoad;
int64_t tailHeads = numHeads_ - loopCnt * headNumPerLoad;
int64_t loadStride = headNumPerLoad * headSize_;
int64_t loopCntKv = numKvHeads_ / headNumPerLoad;
int64_t tailHeadsKv = numKvHeads_ - loopCntKv * headNumPerLoad;
compute_qk(query_, queryDst_, cosAccBuffer, sinAccBuffer, upcastInputBuffer,
calcTmpBuffer, loopCnt, tailHeads, loadStride, headNumPerLoad);
compute_qk(key_, keyDst_, cosAccBuffer, sinAccBuffer, upcastInputBuffer, calcTmpBuffer,
loopCntKv, tailHeadsKv, loadStride, headNumPerLoad);
inQueSinCos_.FreeTensor(localSinCosDeque);
}
private:
AscendC::TPipe *pipe_;
AscendC::TQue<AscendC::QuePosition::VECIN, 1> inQue_, inQueSinCos_;
AscendC::TQue<AscendC::QuePosition::VECOUT, 1> outQue_;
AscendC::TBuf<AscendC::TPosition::VECCALC> calcBuf_;
AscendC::TBuf<AscendC::TPosition::VECCALC> copyBuf_;
AscendC::GlobalTensor<dst_t> queryDst_;
AscendC::GlobalTensor<dst_t> keyDst_;
AscendC::GlobalTensor<scalar_t> query_;
AscendC::GlobalTensor<scalar_t> key_;
AscendC::GlobalTensor<scalar_t> cosSin_;
int rotDim_;
int embedDim_;
int64_t queryStride_;
int64_t keyStride_;
int64_t dstQueryStride_;
int64_t dstKeyStride_;
int numHeads_;
int numKvHeads_;
int headSize_;
int calcTmpBufferOffset_;
int upcastInputBufferOffset_;
int upcastTempBufferOffset_;
int cosSinUpcastBufferOffset_;
int tempBufferSize_;
};
// Note: Need to use macro to instaniate all the target functions here, for the current build system dose not support template call in cpp
// We use C style symbol here for kernel compilation, cpp style kernel entry may lead to compilation failure
#define ROPE_CUSTOM_KERNEL_TYPE_DECLARE(TYPE, NEOX) \
extern "C" __global__ __aicore__ void rope_custom_##NEOX##_##TYPE( \
__gm__ int64_t* positions, __gm__ void* queryDst, __gm__ void* keyDst, __gm__ TYPE* query, __gm__ TYPE* key, \
__gm__ TYPE* cosSinCache, const int rotDim, const int64_t queryStride, const int64_t keyStride, \
const int64_t dstQueryStride, const int64_t dstKeyStride, const int numHeads, const int numKvHeads, \
const int headSize, const int64_t numTokens, const int loopNum, const int coreNum) \
{ \
AscendC::TPipe pipe; \
RotaryEmbedding<TYPE, NEOX> op{}; \
op.init(positions, queryDst, keyDst, query, key, cosSinCache, rotDim, dstQueryStride, dstKeyStride, \
queryStride, keyStride, numHeads, numKvHeads, headSize, &pipe); \
for (int64_t i = AscendC::GetBlockIdx(); i < numTokens; i += coreNum) { \
op.update_mem_offset(positions, queryDst, keyDst, query, key, cosSinCache, rotDim, dstQueryStride, dstKeyStride, \
queryStride, keyStride, numHeads, numKvHeads, headSize, i); \
op.compute_function(); \
} \
}
#define ROPE_CUSTOM_KERNEL_DECLARE(TYPE) \
ROPE_CUSTOM_KERNEL_TYPE_DECLARE(TYPE, true); \
ROPE_CUSTOM_KERNEL_TYPE_DECLARE(TYPE, false);
// Declare all the kernel entry here
ROPE_CUSTOM_KERNEL_DECLARE(half)
ROPE_CUSTOM_KERNEL_DECLARE(bfloat16_t)
namespace vllm_ascend {
#define ROTARY_EMBEDDING_KERNEL_CALL(TYPE) \
if (isNeox) \
rope_custom_true_##TYPE<<<blockDim, nullptr, stream>>>( \
positions, queryDst, keyDst, reinterpret_cast<TYPE *>(query), reinterpret_cast<TYPE *>(key), \
reinterpret_cast<TYPE *>(cosSinCache), rotDim, queryStride, keyStride, dstQueryStride, dstKeyStride, \
numHeads, numKvHeads, headSize, numTokens, loopCnt, blockDim); \
else \
rope_custom_false_##TYPE<<<blockDim, nullptr, stream>>>( \
positions, queryDst, keyDst, reinterpret_cast<TYPE *>(query), reinterpret_cast<TYPE *>(key), \
reinterpret_cast<TYPE *>(cosSinCache), rotDim, queryStride, keyStride, dstQueryStride, dstKeyStride, \
numHeads, numKvHeads, headSize, numTokens, loopCnt, blockDim);
// maximum number for runtime to launch a ascendc kernel.
// we use this to constrain the maximum number of block size
static const int64_t maxParallelSize = 65535;
extern void rotary_embedding_impl(AscendType type, bool isNeox, void *stream, int64_t *positions, void *queryDst,
void *keyDst, void *query, void *key, void *cosSinCache, const int rotDim,
const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride,
const int64_t dstKeyStride, const int numHeads, const int numKvHeads,
const int headSize, const int64_t numTokens, const uint32_t loopCnt,
uint32_t aivNum)
{
int blockDim = maxParallelSize > numTokens ? numTokens : maxParallelSize;
if (type == AscendType::FP16) {
ROTARY_EMBEDDING_KERNEL_CALL(half);
} else if (type == AscendType::BF16) {
ROTARY_EMBEDDING_KERNEL_CALL(bfloat16_t);
} else {
return;
}
}
} // namespace vllm_ascend

25
csrc/kernels/types.h Normal file
View File

@@ -0,0 +1,25 @@
/*
* Copyright (c) Huawei Technologies Co., Ltd. 2024. 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_ascend {
enum struct AscendType {
FP16 = 0,
BF16 = 1,
FP32 = 2,
};
}

49
csrc/kernels/utils.h Normal file
View File

@@ -0,0 +1,49 @@
/*
* Copyright (c) Huawei Technologies Co., Ltd. 2024. 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
#include "kernel_type.h"
namespace vllm_ascend {
template <typename scalar_t> struct AccType;
template <> struct AccType<bfloat16_t> {
using type = float;
};
template <> struct AccType<half> {
using type = half;
};
template <> struct AccType<float> {
using type = float;
};
template <> struct AccType<int8_t> {
using type = int;
};
template <typename scalar_t>
__aicore__ inline void local_mem_copy(AscendC::LocalTensor<scalar_t> dst, AscendC::LocalTensor<scalar_t> src, int size)
{
constexpr int loadSize = 256 / sizeof(scalar_t);
int loopCnt = size / loadSize;
int tailSize = size % loadSize;
if (loopCnt)
AscendC::Copy(dst, src, loadSize, loopCnt, {1, 1, 8, 8});
AscendC::Copy(dst[loopCnt * loadSize], src[loopCnt * loadSize], tailSize, 1, {1, 1, 8, 8});
}
} // namespace vllm_ascend

32
csrc/ops.h Normal file
View File

@@ -0,0 +1,32 @@
/*
* Copyright (c) Huawei Technologies Co., Ltd. 2024. 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
#include <optional>
#include <torch/library.h>
#include <vector>
#include "kernels/types.h"
namespace vllm_ascend {
extern void rotary_embedding_impl(AscendType type, bool isNeox, void *stream, int64_t *positions, void *queryDst,
void *keyDst, void *query, void *key, void *cosSinCache, const int rotDim,
const int64_t queryStride, const int64_t keyStride, const int64_t dstQueryStride,
const int64_t dstKeyStride, const int numHeads, const int numKvHeads,
const int headSize, const int64_t numTokens, const uint32_t loopCnt,
uint32_t aivNum);
}

108
csrc/torch_binding.cpp Normal file
View File

@@ -0,0 +1,108 @@
/*
* Copyright (c) Huawei Technologies Co., Ltd. 2024. 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.
*/
#include <torch/extension.h>
#include <torch/library.h>
#include <torch/version.h>
#include <torch_npu/csrc/core/npu/NPUStream.h>
#include <torch_npu/csrc/framework/OpCommand.h>
#include <torch_npu/csrc/npu/Module.h>
#include <pybind11/pybind11.h>
#include "acl/acl.h"
#include "tiling/platform/platform_ascendc.h"
#include "aclnn/opdev/platform.h"
#include "ops.h"
#include "utils.h"
namespace vllm_ascend {
void rotary_embedding(at::Tensor &positions, at::Tensor &query, at::Tensor &key,
int64_t head_size, at::Tensor &cos_sin_cache, bool is_neox)
{
int32_t deviceId = 0;
int64_t num_tokens = positions.numel();
int positions_ndim = positions.dim();
TORCH_CHECK(
positions_ndim == 1 || positions_ndim == 2,
"positions must have shape [num_tokens] or [batch_size, seq_len]");
if (positions_ndim == 1) {
TORCH_CHECK(
query.size(0) == positions.size(0) && key.size(0) == positions.size(0),
"query, key and positions must have the same number of tokens");
}
if (positions_ndim == 2) {
TORCH_CHECK(
query.size(0) == positions.size(0) &&
key.size(0) == positions.size(0) &&
query.size(1) == positions.size(1) &&
key.size(1) == positions.size(1),
"query, key and positions must have the same batch_size and seq_len");
}
int query_hidden_size = query.numel() / num_tokens;
int key_hidden_size = key.numel() / num_tokens;
TORCH_CHECK(query_hidden_size % head_size == 0);
TORCH_CHECK(key_hidden_size % head_size == 0);
// Make sure query and key have consistent number of heads
int num_heads = query_hidden_size / head_size;
int num_kv_heads = key_hidden_size / head_size;
TORCH_CHECK(num_heads % num_kv_heads == 0);
int rot_dim = cos_sin_cache.size(1);
int64_t *position_ids_ptr = positions.data_ptr<int64_t>();
void *query_ptr = query.data_ptr();
void *key_ptr = key.data_ptr();
void *cos_sin_cache_ptr = cos_sin_cache.data_ptr();
int64_t query_stride = query.stride(-2);
int64_t key_stride = key.stride(-2);
at::ScalarType scalar_type = query.scalar_type();
aclrtStream stream = c10_npu::getCurrentNPUStream().stream();
at_npu::native::OpCommand cmd;
cmd.Name("rotary_embedding");
cmd.SetCustomHandler([scalar_type, is_neox, num_tokens, stream, position_ids_ptr,
query_ptr, key_ptr, cos_sin_cache_ptr, rot_dim, query_stride, key_stride,
num_heads, num_kv_heads, head_size]() -> int {
auto dtype_num = get_dtype_from_torch(scalar_type);
fe::PlatFormInfos platform_infos;
int device_id = 0;
fe::PlatformInfoManager::GeInstance().GetRuntimePlatformInfosByDevice(device_id, platform_infos);
uint32_t aivNum = platform_infos.GetCoreNumByType("aiv");
uint32_t loop_cnt = (num_tokens + aivNum - 1) / aivNum;
rotary_embedding_impl(dtype_num, is_neox, stream, position_ids_ptr, query_ptr, key_ptr, query_ptr,
key_ptr, cos_sin_cache_ptr, rot_dim, query_stride, key_stride, query_stride,
key_stride, num_heads, num_kv_heads, head_size, num_tokens, loop_cnt, aivNum);
return 0;
});
cmd.Run();
return ;
}
} // namespace vllm_ascend
TORCH_LIBRARY_EXPAND(_C, ops)
{
// vLLM-Ascend custom ops
// Rotary embedding
// Apply GPT-NeoX style rotary embedding to query and key.
ops.def(
"rotary_embedding(Tensor positions, Tensor! query,"
" Tensor! key, int head_size,"
" Tensor cos_sin_cache, bool is_neox) -> ()");
ops.impl("rotary_embedding", torch::kPrivateUse1, &vllm_ascend::rotary_embedding);
}
REGISTER_EXTENSION(_C)

43
csrc/utils.h Normal file
View File

@@ -0,0 +1,43 @@
#pragma once
#include "kernels/types.h"
#include <c10/core/ScalarType.h>
#include <Python.h>
#define _CONCAT(A, B) A##B
#define CONCAT(A, B) _CONCAT(A, B)
#define _STRINGIFY(A) #A
#define STRINGIFY(A) _STRINGIFY(A)
// A version of the TORCH_LIBRARY macro that expands the NAME, i.e. so NAME
// could be a macro instead of a literal token.
#define TORCH_LIBRARY_EXPAND(NAME, MODULE) TORCH_LIBRARY(NAME, MODULE)
// A version of the TORCH_LIBRARY_IMPL macro that expands the NAME, i.e. so NAME
// could be a macro instead of a literal token.
#define TORCH_LIBRARY_IMPL_EXPAND(NAME, DEVICE, MODULE) \
TORCH_LIBRARY_IMPL(NAME, DEVICE, MODULE)
// REGISTER_EXTENSION allows the shared library to be loaded and initialized
// via python's import statement.
#define REGISTER_EXTENSION(NAME) \
PyMODINIT_FUNC CONCAT(PyInit_, NAME)() { \
static struct PyModuleDef module = {PyModuleDef_HEAD_INIT, \
STRINGIFY(NAME), nullptr, 0, nullptr}; \
return PyModule_Create(&module); \
}
namespace vllm_ascend {
AscendType get_dtype_from_torch(at::ScalarType scalarType)
{
if (scalarType == at::ScalarType::Float) {
return AscendType::FP32;
} else if (scalarType == at::ScalarType::BFloat16) {
return AscendType::BF16;
} else {
return AscendType::FP16;
}
}
} // namespace vllm_ascend