Use CU_MEMCPY_SRC_ACCESS_ORDER_ANY for batch KV cache swaps (#39306)
Signed-off-by: Itay Etelis <itay.etelis@ibm.com> Signed-off-by: Itay Etelis <etelis2019@gmail.com> Signed-off-by: Itay Etelis <92247226+Etelis@users.noreply.github.com> Co-authored-by: Itay Etelis <itay.etelis@ibm.com> Co-authored-by: Or Ozeri <oro@il.ibm.com> Co-authored-by: Itay Etelis <etelis2019@gmail.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
Itay Etelis
Or Ozeri
Itay Etelis
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
0d382ecde8
commit
00b0618a03
+2
-1
@@ -12,7 +12,8 @@ void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
|
||||
|
||||
void swap_blocks_batch(const torch::Tensor& src_ptrs,
|
||||
const torch::Tensor& dst_ptrs,
|
||||
const torch::Tensor& sizes);
|
||||
const torch::Tensor& sizes,
|
||||
bool is_src_access_order_any);
|
||||
|
||||
void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
|
||||
torch::Tensor& key_cache, torch::Tensor& value_cache,
|
||||
|
||||
@@ -77,7 +77,8 @@ void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
|
||||
|
||||
void swap_blocks_batch(const torch::Tensor& src_ptrs,
|
||||
const torch::Tensor& dst_ptrs,
|
||||
const torch::Tensor& sizes) {
|
||||
const torch::Tensor& sizes,
|
||||
bool is_src_access_order_any) {
|
||||
TORCH_CHECK(src_ptrs.device().is_cpu(), "src_ptrs must be on CPU");
|
||||
TORCH_CHECK(dst_ptrs.device().is_cpu(), "dst_ptrs must be on CPU");
|
||||
TORCH_CHECK(sizes.device().is_cpu(), "sizes must be on CPU");
|
||||
@@ -124,7 +125,12 @@ void swap_blocks_batch(const torch::Tensor& src_ptrs,
|
||||
|
||||
if (batch_fn != nullptr) {
|
||||
CUmemcpyAttributes attr = {};
|
||||
attr.srcAccessOrder = CU_MEMCPY_SRC_ACCESS_ORDER_STREAM;
|
||||
// ANY lets the DMA engine prefetch source bytes out of stream order,
|
||||
// which is only safe when no GPU stream is concurrently writing the
|
||||
// source.
|
||||
attr.srcAccessOrder = is_src_access_order_any
|
||||
? CU_MEMCPY_SRC_ACCESS_ORDER_ANY
|
||||
: CU_MEMCPY_SRC_ACCESS_ORDER_STREAM;
|
||||
size_t attrs_idx = 0;
|
||||
size_t fail_idx = 0;
|
||||
CUresult result = batch_fn(reinterpret_cast<CUdeviceptr*>(dst_data),
|
||||
|
||||
@@ -553,7 +553,8 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
|
||||
// Batch swap: submit all block copies in a single driver call.
|
||||
cache_ops.def(
|
||||
"swap_blocks_batch(Tensor src_ptrs, Tensor dst_ptrs,"
|
||||
" Tensor sizes) -> ()");
|
||||
" Tensor sizes,"
|
||||
" bool is_src_access_order_any=False) -> ()");
|
||||
cache_ops.impl("swap_blocks_batch", torch::kCPU, &swap_blocks_batch);
|
||||
|
||||
// Reshape the key and value tensors and cache them.
|
||||
|
||||
Reference in New Issue
Block a user