Merge remote-tracking branch 'github/main' into open-source-h

csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_epilogue_tma_warpspecialized.hpp

@@ -118,12 +118,15 @@ struct Sm100FmhaFwdEpilogueTmaWarpspecialized {
       auto cumulative_length_q = get<0>(problem_shape).cumulative_length;
       if (cumulative_length_q != nullptr) {
           int max_length_q = get<0>(problem_shape).max_length;
+          get<0>(problem_shape_O).max_length = max(1, max_length_q);
           // for variable sequence lenght, the batch is in units of row_stride
           get<2,1>(dO) = get<0>(dO);
-          get<2,1>(problem_shape_O) = max_length_q * (1 + get<2,1>(problem_shape_O));
+          get<2,1>(problem_shape_O) = max(1, max_length_q * (1 + get<2,1>(problem_shape_O)));
           // offset ptr by the amount we add back in later
           ptr_O -= max_length_q * get<0>(dO);
       }
+    } else {
+      get<0>(problem_shape_O) = max(1, get<0>(problem_shape_O));
     }
 
     auto tma_store_o = make_tma_copy(

csrc/sm100/prefill/dense/collective/sm100_fmha_fwd_mainloop_tma_warpspecialized.hpp

@@ -1155,10 +1155,6 @@ struct Sm100FmhaFwdMainloopTmaWarpspecialized {
     float lse = -INFINITY;
     int thread_idx = threadIdx.x % (4 * NumThreadsPerWarp);
 
-#define DSHOW(x) print(#x ": "); print(x); print("\n")
-    if (threadIdx.x % 128 == 0 && block0()) {
-      DSHOW(sO);
-    }
 #if 1
 
     using ElementOut = typename CollectiveEpilogue::ElementOut;

csrc/sm100/prefill/dense/collective/sm100_fmha_load_tma_warpspecialized.hpp

@@ -112,6 +112,9 @@ struct Sm100FmhaLoadTmaWarpspecialized {
       problem_shape_qk = problem_shape;
     }
 
+    get<0>(problem_shape_qk) = max(1, get<0>(problem_shape_qk));
+    get<1>(problem_shape_qk) = max(1, get<1>(problem_shape_qk));
+
     auto params_qk = CollectiveMmaQK::to_underlying_arguments(
         problem_shape_qk,
         typename CollectiveMmaQK::Arguments {

csrc/sm100/prefill/dense/collective/sm100_fmha_mla_fwd_mainloop_tma_warpspecialized.hpp

@@ -1162,10 +1162,6 @@ struct Sm100MlaFwdMainloopTmaWarpspecialized {
     float lse = -INFINITY;
     int thread_idx = threadIdx.x % (4 * NumThreadsPerWarp);
 
-#define DSHOW(x) print(#x ": "); print(x); print("\n")
-    if (threadIdx.x % 128 == 0 && block0()) {
-      DSHOW(sO);
-    }
 #if 1
 
     using ElementOut = typename CollectiveEpilogue::ElementOut;

csrc/sm100/prefill/dense/collective/sm100_fmha_mla_load_tma_warpspecialized.hpp

@@ -119,6 +119,9 @@ struct Sm100MlaFwdLoadTmaWarpspecialized {
       problem_shape_qk = replace<2>(problem_shape, get<2, 0>(problem_shape) + get<2, 1>(problem_shape));;
     }
 
+    get<0>(problem_shape_qk) = max(1, get<0>(problem_shape_qk));
+    get<1>(problem_shape_qk) = max(1, get<1>(problem_shape_qk));
+
     auto problem_shape_pv = replace<1>(select<0,2,1,3>(problem_shape_qk), get<2, 0>(problem_shape));
 
     auto params_qk = CollectiveMmaQK::to_underlying_arguments(

csrc/sm100/prefill/dense/device/fmha.hpp

@@ -208,6 +208,11 @@ public:
     dim3 const block = Kernel::get_block_shape();
     dim3 const grid = get_grid_shape(params);
 
+    // No need to launch the kernel
+    if(grid.x == 0 || grid.y == 0 || grid.z == 0) {
+      return Status::kSuccess; 
+    }
+
     // configure smem size and carveout
     int smem_size = Kernel::SharedStorageSize;
 

csrc/sm100/prefill/dense/kernel/fmha_causal_tile_scheduler.hpp

@@ -160,7 +160,7 @@ struct CausalPersistentTileScheduler {
 
     return Params {
       num_blocks,
-      { size<3,0>(problem_size) }, { num_m_blocks}, { size<3,1>(problem_size) },
+      { size<3,0>(problem_size) }, { max(1, num_m_blocks) }, { size<3,1>(problem_size) },
       hw_info
     };
   }

csrc/sm100/prefill/dense/kernel/fmha_tile_scheduler.hpp

@@ -123,7 +123,7 @@ struct PersistentTileScheduler {
 
     return Params {
       num_blocks,
-      { num_m_blocks}, { size<3,0>(problem_size) }, { size<3,1>(problem_size) },
+      { max(1, num_m_blocks)}, { size<3,0>(problem_size) }, { size<3,1>(problem_size) },
       hw_info
     };
   }