minor simplifications

ggml/src/ggml-cuda/allreduce.cu

@@ -520,7 +520,6 @@ bool ggml_cuda_ar_allreduce(
     }
 
     const size_t max_chunk_elems = p->buf_bytes / type_size;
-    GGML_ASSERT(max_chunk_elems > 0);
 
     // Insert chunked kernels into each GPU's existing compute stream via events:
     //   record(app, compute_stream)       — capture "upstream done"

ggml/src/ggml-cuda/allreduce.cuh

@@ -14,8 +14,9 @@ struct ggml_cuda_ar_pipeline;
 
 // Allocate a pipeline for n_devices GPUs.
 // devices[] holds the CUDA device IDs in rank order.
-// max_bytes is the staging buffer size per device; must be at least as large
-// as the largest tensor that will be reduced.
+// max_bytes is the staging buffer size per device, which also bounds the
+// per-launch chunk size; tensors larger than this are reduced via multiple
+// chunked launches.
 // Returns nullptr on allocation failure.
 ggml_cuda_ar_pipeline * ggml_cuda_ar_pipeline_init(
     const int * devices, size_t n_devices, size_t max_bytes);

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

@@ -1327,11 +1327,6 @@ static bool ggml_backend_cuda_comm_allreduce_nccl(
 }
 #endif // GGML_USE_NCCL
 
-static bool ggml_backend_cuda_comm_allreduce_internal(
-        ggml_backend_cuda_comm_context * comm_ctx, struct ggml_tensor ** tensors) {
-    return ggml_cuda_ar_allreduce(comm_ctx->ar_pipeline, comm_ctx->backends.data(), tensors);
-}
-
 enum ggml_cuda_comm_allreduce_result {
     GGML_CUDA_COMM_ALLREDUCE_SUCCESS,
     GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED,
@@ -1389,7 +1384,7 @@ static ggml_cuda_comm_allreduce_result ggml_backend_cuda_comm_try_allreduce_inte
         }
     }
 
-    return ggml_backend_cuda_comm_allreduce_internal(comm_ctx, tensors)
+    return ggml_cuda_ar_allreduce(comm_ctx->ar_pipeline, comm_ctx->backends.data(), tensors)
         ? GGML_CUDA_COMM_ALLREDUCE_SUCCESS
         : GGML_CUDA_COMM_ALLREDUCE_FAILED;
 }