13 lines
392 B
Django/Jinja
13 lines
392 B
Django/Jinja
#include <flashinfer/attention/hopper/prefill_sm90.cuh>
|
|
#include "single_prefill_sm90_config.inc"
|
|
|
|
using namespace flashinfer;
|
|
|
|
namespace flashinfer {
|
|
|
|
template cudaError_t SinglePrefillWithKVCacheDispatched
|
|
<{{ head_dim_qk }}, {{ head_dim_vo }}, {{ mask_mode }}, /*USE_SLIDING_WINDOW=*/{{ use_sliding_window }}, {{ variant_name }}, Params>(
|
|
Params& params, cudaStream_t stream);
|
|
|
|
};
|