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 <redacted>

diff --git a/ggml/src/ggml-backend-meta.cpp b/ggml/src/ggml-backend-meta.cpp
index 0a36f099000f910cfa3af33d29e354987563b68c..7bd32916480e1561059498ef02189963ed7ee8bf 100644 (file)
--- a/ggml/src/ggml-backend-meta.cpp
+++ b/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);