vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix Vulkan validation errors on embedding models with no offloaded layers

Browse source
This commit is contained in:
0cc4m 2024-05-19 09:55:36 +02:00
parent 8dbde1f033
commit 6db8ec3a71
1 changed files with 7 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml-vulkan.cpp
View file

@ -2739,7 +2739,7 @@ static void ggml_vk_matmul(
uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
uint32_t expert_stride_b, uint32_t expert_stride_d, uint32_t idx, uint32_t nbi1, uint32_t n_as) { uint32_t expert_stride_b, uint32_t expert_stride_d, uint32_t idx, uint32_t nbi1, uint32_t n_as) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << split_k_buffer.buffer->buffer << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl; std::cerr << "ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl;
#endif #endif
ggml_vk_sync_buffers(subctx); ggml_vk_sync_buffers(subctx);
if (split_k == 1) { if (split_k == 1) {
@ -6158,6 +6158,10 @@ GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
ctx->transfer_ctx = nullptr; ctx->transfer_ctx = nullptr;
} }
static bool ggml_vk_is_empty(ggml_tensor * node) {
return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
}
GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
#ifdef GGML_VULKAN_DEBUG #ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)" << std::endl; std::cerr << "ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)" << std::endl;
@ -6172,7 +6176,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
int last_node = cgraph->n_nodes - 1; int last_node = cgraph->n_nodes - 1;
// If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
while (last_node > 0 && ggml_is_empty(cgraph->nodes[last_node])) { while (last_node > 0 && ggml_vk_is_empty(cgraph->nodes[last_node])) {
last_node -= 1; last_node -= 1;
} }
@ -6186,7 +6190,7 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
for (int i = 0; i < cgraph->n_nodes; i++) { for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i]; ggml_tensor * node = cgraph->nodes[i];
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) { if (ggml_vk_is_empty(node)) {
continue; continue;
} }