ViperEkura
3439e3104e
- Thread a cudaStream_t through attn dispatchers onto torch's current stream - Scope the device guard to the entry function so kernels run on tensor device - DISPATCH_HEAD_DIM now forwards varargs so stream reaches each dispatch - Parallelize CPU reference kernels with OpenMP (paged test 31s -> 7s) - Merge decode/prefill standalone tests into attn_test.cu with correctness tables - Drop bench error column (CPU ref too slow at large sizes) - Update cuda_kernels.md for the merged test layout
47 lines
1.5 KiB
Plaintext
47 lines
1.5 KiB
Plaintext
#include "attn_dispatchers.cuh"
|
|
#include "attn_entry_utils.cuh"
|
|
|
|
torch::Tensor attn_paged_decode(
|
|
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,
|
|
int64_t max_seq_len,
|
|
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();
|
|
|
|
PagedAttentionParams<bf16> p;
|
|
attn_pack_paged_decode_params(q, k_cache, v_cache,
|
|
req_to_token, req_pool_indices, kv_indptr,
|
|
max_seq_len, mask, causal_offset, scale, p);
|
|
|
|
auto O = torch::empty({q.size(0), q.size(1), q.size(2)}, q.options());
|
|
p.o = (bf16*)O.data_ptr();
|
|
|
|
alloc_split_partials(p);
|
|
DISPATCH_HEAD_DIM(p.head_dim, dispatch_paged_decode, p, stream);
|
|
C10_CUDA_CHECK(cudaGetLastError());
|
|
return O;
|
|
}
|
|
|
|
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
|
|
m.def("attn_paged_decode", &attn_paged_decode,
|
|
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("max_seq_len"),
|
|
py::arg("mask") = py::none(),
|
|
py::arg("causal_offset") = -1,
|
|
py::arg("scale") = 0.0,
|
|
"SGLang-style paged decode: flat KV pool + req_to_token + kv_indptr.");
|
|
}
|