vulkan: skip zero size tensors in backend copies (llama/20233)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 0bf7d2e24730accca957e4055400a5d20fbc1995..70e992ce233cbf41c3dd5c4fd551868be2b55ae9 100644 (file)
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -13253,6 +13253,10 @@ static void ggml_backend_vk_buffer_memset_tensor(ggml_backend_buffer_t buffer, g
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
     vk_buffer buf = buf_ctx->dev_buffer;
 
+    if (size == 0) {
+        return;
+    }
+
     uint32_t val32 = (uint32_t)value * 0x01010101;
     ggml_vk_buffer_memset(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, val32, size);
 }
@@ -13262,6 +13266,10 @@ static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
     vk_buffer buf = buf_ctx->dev_buffer;
 
+    if (size == 0) {
+        return;
+    }
+
     ggml_vk_buffer_write(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
 }
 
@@ -13269,12 +13277,20 @@ static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, cons
     VK_LOG_DEBUG("ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
 
+    if (size == 0) {
+        return;
+    }
+
     vk_buffer buf = buf_ctx->dev_buffer;
 
     ggml_vk_buffer_read(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
 }
 
 static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+    if (ggml_nbytes(src) == 0) {
+        return true;
+    }
+
     if (ggml_backend_buffer_is_vk(src->buffer)) {
         ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context;
         ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
@@ -13464,6 +13480,10 @@ static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
 
+    if (size == 0) {
+        return;
+    }
+
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
 
     vk_context cpy_ctx;
@@ -13507,6 +13527,10 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
 
+    if (size == 0) {
+        return;
+    }
+
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
 
     vk_context compute_ctx = ggml_vk_get_compute_ctx(ctx);
@@ -13533,9 +13557,14 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 }
 
 static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    VK_LOG_DEBUG("ggml_backend_vk_cpy_tensor_async()");
+    VK_LOG_DEBUG("ggml_backend_vk_cpy_tensor_async(" << src << " -> " << dst << ", size=" << ggml_nbytes(src) << ")");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend_dst->context;
 
+    // Skip zero-size tensors
+    if (ggml_nbytes(src) == 0) {
+        return true;
+    }
+
     if (dst->buffer->buft != ggml_backend_vk_get_default_buffer_type(backend_dst)) {
         return false;
     }