ViperEkura
16a55bb474
- move attention kernels to csrc/kernels/attention/ and rotary to rotary/ - add shared common/mma.cuh (mma_sync, ldmatrix) and device.cuh (sm checks) - split fp8_mm into three-layer fp8/common.h, gemm.cuh, mm.cu - fix fused FP8 GEMM ldmatrix lane indexing to fix OOB shared reads - update extension ops, loader, and kernel tests
52 lines
1.7 KiB
Plaintext
52 lines
1.7 KiB
Plaintext
#include "dispatchers.cuh"
|
|
#include "entry_utils.cuh"
|
|
|
|
torch::Tensor attn_paged_prefill(
|
|
torch::Tensor q,
|
|
torch::Tensor k_cache,
|
|
torch::Tensor v_cache,
|
|
torch::Tensor req_to_token,
|
|
torch::Tensor req_pool_indices,
|
|
torch::Tensor kv_indptr,
|
|
torch::Tensor qo_indptr,
|
|
torch::Tensor q_tile_to_batch,
|
|
torch::Tensor q_tile_to_index,
|
|
c10::optional<torch::Tensor> mask,
|
|
int64_t causal_offset,
|
|
double scale
|
|
) {
|
|
const at::cuda::OptionalCUDAGuard device_guard(device_of(q));
|
|
auto stream = at::cuda::getCurrentCUDAStream();
|
|
|
|
AttentionParams<bf16> p;
|
|
attn_pack_paged_prefill_params(q, k_cache, v_cache,
|
|
req_to_token, req_pool_indices,
|
|
kv_indptr, qo_indptr,
|
|
q_tile_to_batch, q_tile_to_index, mask,
|
|
causal_offset, scale, p);
|
|
|
|
auto O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
|
|
p.o_ptr = (bf16*)O.data_ptr();
|
|
|
|
DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_prefill, p, stream);
|
|
C10_CUDA_CHECK(cudaGetLastError());
|
|
return O;
|
|
}
|
|
|
|
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
|
|
m.def("attn_paged_prefill", &attn_paged_prefill,
|
|
py::arg("q"),
|
|
py::arg("k_cache"),
|
|
py::arg("v_cache"),
|
|
py::arg("req_to_token"),
|
|
py::arg("req_pool_indices"),
|
|
py::arg("kv_indptr"),
|
|
py::arg("qo_indptr"),
|
|
py::arg("q_tile_to_batch"),
|
|
py::arg("q_tile_to_index"),
|
|
py::arg("mask") = py::none(),
|
|
py::arg("causal_offset") = -1,
|
|
py::arg("scale") = 0.0,
|
|
"SGLang-style paged prefill: flat KV pool + ragged batch.");
|
|
}
|