split ablation on main

Author: jayhshah

hopper/flash_api.cpp

@@ -625,6 +625,9 @@ mha_fwd_get_scheduler_metadata(
     params.pack_gqa |= params.num_splits > 1;
     // printf("Num splits (metadata) = %d.\n", params.num_splits);
 
+    // ABLATION: set num splits to 1. Should cripple perf
+    params.num_splits = 1;
+
     bool is_varlen = true;
 
     // Otherwise the kernel will be launched from cuda:0 device
@@ -986,6 +989,8 @@ mha_fwd(at::Tensor &q,   // (b, s_q, h, d) or (total_q, h, d) if there is cu_seq
     params.pack_gqa = pack_gqa_.has_value() ? pack_gqa_.value() : get_pack_gqa(params);
     // Always enable PackGQA for Split
     params.pack_gqa |= (params.num_splits > 1);
+    // ABLATION: set num splits to 1. Should cripple perf
+    params.num_splits = 1;
 
     // This needs to be set after get_num_splits
     at::Tensor tile_count_semaphore;  // Contains the semaphore and optionally num_splits_dynamic