ggml : fix tensor-parallel + -ncmoe crash on MoE models (#25028)

Tensor parallelism (-sm tensor) combined with -ncmoe (CPU-offloaded MoE
experts) aborts during warm-up on MoE models with
GGML_ASSERT(ggml_is_contiguous(tensor)) in ggml-backend-meta.cpp.

The failing tensor is the MoE router output (ffn_moe_topk): it is mirrored
(GGML_BACKEND_SPLIT_AXIS_MIRRORED, replicated across backends since routing
must be identical) and happens to be a non-contiguous view.
ggml_backend_meta_buffer_{get,set}_tensor asserted contiguity before
consulting the split state, so a mirrored non-contiguous tensor tripped the
assert even though the GGML_BACKEND_SPLIT_AXIS_MIRRORED case right below
already handles it.

Move the split-state lookup above the assert and allow the mirrored case in
both get_tensor and set_tensor.

Diagnosis credit to the reporter (@nathanmp).

Fixes #24886

Signed-off-by: liminfei-amd <91481003+liminfei-amd@users.noreply.github.com>

ggml/src/ggml-backend-meta.cpp

@@ -1245,9 +1245,8 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
 
 static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
-    GGML_ASSERT(ggml_is_contiguous(tensor));
-
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
+    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
 
     if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
@@ -1360,9 +1359,8 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
 
 static void ggml_backend_meta_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
-    GGML_ASSERT(ggml_is_contiguous(tensor));
-
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
+    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
 
     if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -543,12 +543,42 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
             // the memory allocation handle is no longer needed after mapping
             CU_CHECK(cuMemRelease(handle));
 
-            // set access
-            CUmemAccessDesc access = {};
-            access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-            access.location.id = device;
-            access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
-            CU_CHECK(cuMemSetAccess((CUdeviceptr)((char *)(pool_addr) + pool_size), reserve_size, &access, 1));
+            // VMM Bug fix for P2P access if GGML_CUDA_P2P is set, or if NCCL build
+            bool use_peer_access = getenv("GGML_CUDA_P2P") != nullptr;
+#if defined(GGML_USE_NCCL)
+            use_peer_access = true;
+#endif // defined(GGML_USE_NCCL)
+
+            if (use_peer_access) {
+                // NCCL implicitly enables peer access (cudaDeviceEnablePeerAccess), and
+                // GGML_CUDA_P2P enables it explicitly. Unlike cudaMalloc buffers, VMM
+                // allocations do not become peer-accessible from that alone, so access
+                // must be granted explicitly here.
+                std::vector<CUmemAccessDesc> access_descs;
+                const int device_count = ggml_cuda_info().device_count;
+                for (int id = 0; id < device_count; ++id) {
+                    if (id != device) {
+                        int can_access_peer = 0;
+                        CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, device));
+                        if (!can_access_peer) {
+                            continue;
+                        }
+                    }
+                    CUmemAccessDesc access = {};
+                    access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+                    access.location.id = id;
+                    access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+                    access_descs.push_back(access);
+                }
+                CU_CHECK(cuMemSetAccess(start_ptr, reserve_size, access_descs.data(), access_descs.size()));
+            } else {
+                // set access for non P2P
+                CUmemAccessDesc access = {};
+                access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+                access.location.id = device;
+                access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+                CU_CHECK(cuMemSetAccess(start_ptr, reserve_size, &access, 1));
+            }
 
             // add to the pool
             pool_size += reserve_size;