Fix segfault in cudaMemcpyBatchAsync on CUDA 13.0 (#23136)

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>

sgl-kernel/csrc/kvcacheio/transfer.cu

@@ -806,17 +806,29 @@ inline void transfer_kv_page_first_direct_impl(
   }
 
   // Symbol gate: runtime may not expose cudaMemcpyBatchAsync in some environments.
-  using CudaMemcpyBatchAsyncFn =
+  static void* cuda_memcpy_batch_async_sym = dlsym(RTLD_DEFAULT, "cudaMemcpyBatchAsync");
-      cudaError_t (*)(void**, void**, size_t*, size_t, cudaMemcpyAttributes*, size_t*, size_t, size_t*, cudaStream_t);
+  if (cuda_memcpy_batch_async_sym == nullptr) {
-  static CudaMemcpyBatchAsyncFn cuda_memcpy_batch_async = []() {
-    void* symbol = dlsym(RTLD_DEFAULT, "cudaMemcpyBatchAsync");
-    return reinterpret_cast<CudaMemcpyBatchAsyncFn>(symbol);
-  }();
-  if (cuda_memcpy_batch_async == nullptr) {
     fallback_to_page_copy();
     return;
   }
 
+  // CUDA 13.0 removed the failIdx parameter from cudaMemcpyBatchAsync. The ABI
+  // of the dlsym'd symbol is determined by the libcudart loaded in this process,
+  // not the host driver — a cu12 runtime on a cu13 driver host (common in
+  // containers) still exposes the 9-param v12 signature. Dispatching on the
+  // driver version here would segfault in that case (verified empirically).
+  // Use cudaRuntimeGetVersion so the signature follows the runtime. The
+  // runtime version is process-constant, so cache the query (static init is
+  // thread-safe in C++11+) to keep the KV-transfer hot path free of a redundant
+  // runtime API call per invocation.
+  static int runtime_version = 0;
+  static cudaError_t runtime_version_err = cudaRuntimeGetVersion(&runtime_version);
+  if (runtime_version_err != cudaSuccess) {
+    fallback_to_page_copy();
+    return;
+  }
+  static const bool use_v13_signature = runtime_version >= 13000;
+
   size_t num_copies = 0;
   std::vector<void*> batch_srcs;
   std::vector<void*> batch_dsts;
@@ -916,9 +928,27 @@ inline void transfer_kv_page_first_direct_impl(
 
   TORCH_CHECK(batch_srcs.size() == num_copies, "Batch memcpy count mismatch");
   if (num_copies > 0) {
+    cudaError_t err;
     size_t fail_idx = std::numeric_limits<size_t>::max();
-    cudaError_t err = cuda_memcpy_batch_async(
+    if (use_v13_signature) {
-        batch_dsts.data(),
+      using FnV13 = cudaError_t (*)(
+          void* const*,
+          const void* const*,
+          const size_t*,
+          size_t,
+          cudaMemcpyAttributes*,
+          size_t*,
+          size_t,
+          cudaStream_t);
+      auto fn = reinterpret_cast<FnV13>(cuda_memcpy_batch_async_sym);
+      err = fn(
+          batch_dsts.data(), batch_srcs.data(), batch_sizes.data(), num_copies, &attrs, attrs_idxs.data(), 1, stream);
+    } else {
+      using FnV12 = cudaError_t (*)(
+          void**, void**, size_t*, size_t, cudaMemcpyAttributes*, size_t*, size_t, size_t*, cudaStream_t);
+      auto fn = reinterpret_cast<FnV12>(cuda_memcpy_batch_async_sym);
+      err =
+          fn(batch_dsts.data(),
              batch_srcs.data(),
              batch_sizes.data(),
              num_copies,
@@ -927,6 +957,7 @@ inline void transfer_kv_page_first_direct_impl(
              1,
              &fail_idx,
              stream);
+    }
     if (err == cudaErrorNotSupported || err == cudaErrorCallRequiresNewerDriver) {
       fallback_to_page_copy();
       return;