Properly track aborts and add a test for that

dfyz · May 21, 2024 · fee4a6b · fee4a6b
1 parent eadad54
commit fee4a6b
Show file tree

Hide file tree

Showing 2 changed files with 30 additions and 4 deletions.
diff --git a/src/device/mixed_precision_reduce_scatter.h b/src/device/mixed_precision_reduce_scatter.h
@@ -204,6 +204,7 @@ class MixedPrecisionReduceScatterPrims {
                        RoleWaitSend = 0x08,
                        RolePostSend = 0x10,
                        RolePostRecv = 0x20,
+                       Aborted = 0x40,
                        OffsFifoEnabled = 0x80,
                        SizesFifoEnabled = 0x100,
                        ThreadsSynced = 0x800;
@@ -235,7 +236,8 @@ class MixedPrecisionReduceScatterPrims {
         waitPeer<Send, Recv>(sliceSize);
         subBarrier();
 
-        // For simplicity, don't track aborts here.
+        // `prims_simple.h` tries to avoid doing unnecessary reduceCopy() if we are already aborted,
+        // but we don't really mind doing some extra work.
         auto& group = ncclShmem.groups[0];
         reduceCopyMixedPrecision<
           ncclCollUnroll(), RedOp, T, TInput,
@@ -340,9 +342,21 @@ class MixedPrecisionReduceScatterPrims {
   __device__ __forceinline__ void waitPeer(int nelts) {
     const bool isSendNotRecv = (Send && Recv) ? (flags & RoleWaitSend) : Send;
     if (flags & (Recv * RoleWaitRecv | Send * RoleWaitSend)) {
+      int spins = 0;
       while (connStepCache + (isSendNotRecv ? NCCL_STEPS : 0) < step + StepPerSlice) {
-        // For simplicity, don't track aborts here.
         connStepCache = ld_volatile_global(connStepPtr);
+        // Check for kernel abort.
+        spins++;
+        if (!(flags & Aborted) && spins == NCCL_SPINS_BEFORE_CHECK_ABORT) {
+          if (*ncclShmem.comm.abortFlag) {
+            flags |= Aborted;
+            ncclShmem.aborted = 1;
+          }
+          spins = 0;
+        }
+        if (flags & Aborted) {
+          break;
+        }
       }
 
       if (isSendNotRecv && (flags & SizesFifoEnabled)) {

diff --git a/tests/test_rs.cu b/tests/test_rs.cu
@@ -1,5 +1,6 @@
 // nvcc -O2 -I /usr/local/mpi/include -I ~/nccl/build/include -L /usr/local/mpi/lib -L ~/nccl/build/lib -gencode=arch=compute_80,code=sm_80 -o test_rs test_rs.cu -lnccl -lmpi
 #include <iostream>
+#include <optional>
 #include <random>
 #include <string>
 #include <vector>
@@ -47,7 +48,7 @@ constexpr size_t WARMUP_ITERS = 10;
 constexpr size_t ITERS = 100;
 
 std::vector<__nv_bfloat16> DoReduceScatter(__nv_bfloat16* sendbuff, size_t elemCount, size_t recvCount, ncclComm_t comm, cudaStream_t stream,
-                                           ReduceScatterMode mode, int rank, float* elapsedMs) {
+                                           ReduceScatterMode mode, int rank, float* elapsedMs, std::optional<ReduceScatterMode> abortMode) {
     __nv_bfloat16* recvbuff = sendbuff + rank * recvCount;
     float* tmpbuff = {};
 
@@ -93,6 +94,12 @@ std::vector<__nv_bfloat16> DoReduceScatter(__nv_bfloat16* sendbuff, size_t elemC
         CudaOrDie(cudaGetLastError(), "fp32 -> bf16 conversion");
     }
 
+    if (abortMode.has_value() && mode == *abortMode) {
+        NcclOrDie(ncclCommAbort(comm), "abort NCCL communicator");
+        std::cerr << "Aborted NCCL communicator on rank " << rank << std::endl;
+        exit(42);
+    }
+
     CudaOrDie(cudaStreamSynchronize(stream), "synchronize GPU");
 
     std::vector<__nv_bfloat16> result(recvCount);
@@ -161,6 +168,11 @@ int main(int argc, char** argv) {
     CudaOrDie(cudaStreamCreate(&stream), "create GPU stream");
     CudaOrDie(cudaMalloc(&sendbuff, elemCount * sizeof(__nv_bfloat16)), "allocate GPU buffer");
 
+    std::optional<ReduceScatterMode> abortMode;
+    if (const auto* abortModeEnv = getenv("ABORT_MODE")) {
+        abortMode = static_cast<ReduceScatterMode>(std::stoi(abortModeEnv));
+    }
+
     for (size_t rawMode = 0; rawMode < static_cast<size_t>(ReduceScatterMode::ModeCount); ++rawMode) {
         const auto mode = static_cast<ReduceScatterMode>(rawMode);
 
@@ -173,7 +185,7 @@ int main(int argc, char** argv) {
                 cudaMemcpy(sendbuff, gradients.data(), elemCount * sizeof(__nv_bfloat16), cudaMemcpyHostToDevice),
                 "copy to GPU"
             );
-            const auto result = DoReduceScatter(sendbuff, elemCount, recvCount, comm, stream, mode, rank, &elapsedMs);
+            const auto result = DoReduceScatter(sendbuff, elemCount, recvCount, comm, stream, mode, rank, &elapsedMs, abortMode);
 
             if (iter < WARMUP_ITERS) {
                 continue;