Use larger MMA shape (#227)
This commit is contained in:
@@ -87,7 +87,7 @@ static void smxx_fp8_mqa_logits(const torch::Tensor& q,
|
||||
constexpr int block_qh = 128;
|
||||
constexpr int block_kv = 256;
|
||||
constexpr int num_specialized_threads = 128;
|
||||
constexpr int num_math_threads = 512;
|
||||
constexpr int num_math_threads = 256;
|
||||
constexpr int num_q_stages = 3, num_kv_stages = 3;
|
||||
const int block_q = block_qh / num_heads;
|
||||
DG_HOST_ASSERT(block_qh % num_heads == 0);
|
||||
|
||||
Reference in New Issue
Block a user