Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[POC] [Do not merge] BatchPrefill without custom mask support non-cont kv-cache #508

Closed
wants to merge 2 commits into from
Closed

[POC] [Do not merge] BatchPrefill without custom mask support non-cont kv-cache #508

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6e2a33a
BatchPrefill without custom mask support non-cont kv-cache
reyoung Sep 25, 2024
709eb82
Merge branch 'flashinfer-ai:main' into feature/support_non_contiguous…
reyoung Sep 26, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 26 additions & 0 deletions include/flashinfer/page.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,32 @@ struct paged_kv_t {
last_page_len(nullptr),
rope_pos_offset(nullptr) {}

__host__ __forceinline__ paged_kv_t(uint32_t num_heads, uint32_t page_size, uint32_t head_dim,
uint32_t batch_size, QKVLayout layout, DType* kv_data,
DType* k_data, DType* v_data, const int64_t * kv_strides,
IdType* indices, IdType* indptr,
IdType* last_page_len, IdType* rope_pos_offset = nullptr)
: num_heads(num_heads),
page_size(page_size),
head_dim(head_dim),
batch_size(batch_size),
indices(indices),
indptr(indptr),
last_page_len(last_page_len),
rope_pos_offset(rope_pos_offset) {
bool kv_defined = kv_data != nullptr;
if (kv_defined) {
this->k_data = kv_data;
this->v_data = kv_data + kv_strides[1];
} else {
this->k_data = k_data;
this->v_data = v_data;
}
stride_page = kv_strides[0];
stride_n = layout == QKVLayout::kHND ? kv_strides[2 + kv_defined] : kv_strides[1 + kv_defined];
stride_h = layout == QKVLayout::kHND ? kv_strides[1 + kv_defined] : kv_strides[2 + kv_defined];
}

/*!
* \brief Construct a paged key-value cache
* \param num_heads The number of heads
Expand Down
22 changes: 19 additions & 3 deletions python/csrc/batch_prefill.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,10 +76,13 @@ std::vector<torch::Tensor> BatchPrefillWithPagedKVCachePyTorchWrapper::Run(
CHECK_INPUT(q);
CHECK_INPUT(qo_indptr);
if (paged_kv_defined) {
CHECK_INPUT(paged_kv_cache.value());
CHECK_CUDA(paged_kv_cache.value());
CHECK_LAST_DIM_CONTIGUOUS(paged_kv_cache.value());
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe we need a new macro here, like CHECK_INPUT_LAST_DIM_CONTIGUOUS()?

@yzh119

} else {
CHECK_INPUT(paged_k_cache.value());
CHECK_INPUT(paged_v_cache.value());
CHECK_CUDA(paged_k_cache.value());
CHECK_LAST_DIM_CONTIGUOUS(paged_k_cache.value());
CHECK_CUDA(paged_v_cache.value());
CHECK_LAST_DIM_CONTIGUOUS(paged_v_cache.value());
}
CHECK_INPUT(paged_kv_indptr);
CHECK_INPUT(paged_kv_indices);
Expand Down Expand Up @@ -163,6 +166,17 @@ std::vector<torch::Tensor> BatchPrefillWithPagedKVCachePyTorchWrapper::Run(
auto kv_scalar_type =
paged_kv_defined ? paged_kv_cache->scalar_type() : paged_k_cache->scalar_type();


const int64_t* kv_cache_strides = nullptr;
if (paged_kv_cache.has_value()) {
kv_cache_strides = paged_kv_cache->strides().data();
} else {
auto k_strides = paged_k_cache->strides();
auto v_strides = paged_v_cache->strides();
TORCH_CHECK(k_strides == v_strides, "k/v cache strides not match");
kv_cache_strides = k_strides.data();
}

if (q_scalar_type == kv_scalar_type) {
DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(q_scalar_type, c_type, [&] {
return DISPATCH_logits_post_hook(logits_post_hook, LOGITS_POST_HOOK, [&] {
Expand All @@ -171,6 +185,7 @@ std::vector<torch::Tensor> BatchPrefillWithPagedKVCachePyTorchWrapper::Run(
static_cast<c_type*>(paged_kv_cache.has_value() ? paged_kv_cache->data_ptr() : nullptr),
static_cast<c_type*>(paged_k_cache.has_value() ? paged_k_cache->data_ptr() : nullptr),
static_cast<c_type*>(paged_v_cache.has_value() ? paged_v_cache->data_ptr() : nullptr),
kv_cache_strides,
static_cast<int32_t*>(paged_kv_indices.data_ptr()),
static_cast<int32_t*>(paged_kv_indptr.data_ptr()),
static_cast<int32_t*>(paged_kv_last_page_len.data_ptr()));
Expand Down Expand Up @@ -214,6 +229,7 @@ std::vector<torch::Tensor> BatchPrefillWithPagedKVCachePyTorchWrapper::Run(
: nullptr),
static_cast<kv_type*>(paged_v_cache.has_value() ? paged_v_cache->data_ptr()
: nullptr),
kv_cache_strides,
static_cast<int32_t*>(paged_kv_indices.data_ptr()),
static_cast<int32_t*>(paged_kv_indptr.data_ptr()),
static_cast<int32_t*>(paged_kv_last_page_len.data_ptr()));
Expand Down
3 changes: 3 additions & 0 deletions python/csrc/pytorch_extension_utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,6 +235,9 @@ inline constexpr uint32_t pack_u16(uint16_t a, uint16_t b) {

#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")

#define CHECK_LAST_DIM_CONTIGUOUS(x) \
TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "'s last dim must be contiguous")

#define CHECK_INPUT(x) \
CHECK_CUDA(x); \
CHECK_CONTIGUOUS(x)
Expand Down