happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[bugfix]make sure single node op will have the same type (#29) 

* debug

* disable reshape

* make sure single node op have same type

* fix warning at the logger

* Revert "disable reshape"

This reverts commit 5aeca4ba9bec6db3f047f9da803df20f9f6612b3.
This commit is contained in:
nullname 2025-02-28 19:18:16 +08:00 committed by GitHub
parent c867641222
commit f289752664
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 32 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
ggml/src/ggml-qnn/backend-ops.cpp
View File

@ -369,6 +369,31 @@ bool ggnl_qnn_supports_op_tensor(ggml_backend_qnn_device_context * ctx, const gg
return true;
}
bool ggml_qnn_have_same_tensor_types(ggml_backend_qnn_device_context * ctx, const ggml_tensor * op) {
auto * src0 = op->src[0];
auto * src1 = op->src[1];
if (src1) {
if (src0->type != op->type || src1->type != op->type) {
QNN_LOG_DEBUG("[%s][%s]type src0(%s), src1(%s) and op(%s) are not equal\n",
qnn::get_backend_name(ctx->device), ggml_op_name(op->op), ggml_type_name(src0->type),
ggml_type_name(src1->type), ggml_type_name(op->type));
return false;
}
} else {
if (src0->type != op->type) {
QNN_LOG_DEBUG("[%s][%s]type src0(%s) and op(%s) are not equal\n", qnn::get_backend_name(ctx->device),
ggml_op_name(op->op), ggml_type_name(src0->type), ggml_type_name(op->type));
return false;
}
}
#ifdef NDEBUG
GGML_UNUSED(ctx);
#endif
return true;
}
bool ggml_qnn_supports_matmul_op(ggml_backend_qnn_device_context * ctx, const ggml_tensor * op) {
constexpr const size_t kMaxNpuTensorSize = 8192L * 2048 + 8192 * 512 + 2048 * 512;
constexpr const auto get_tensor_size = [](const ggml_tensor * tensor) -> size_t {
@ -393,10 +418,8 @@ bool ggml_qnn_supports_matmul_op(ggml_backend_qnn_device_context * ctx, const gg
// fall through, from test here, the convert op is super slow on NPU:
// https://github.com/usefulsensors/qc_npu_benchmark
case QNN_BACKEND_GPU:
if (src0->type != src1->type || src0->type != op->type) {
if (ggml_qnn_have_same_tensor_types(ctx, op)) {
// there's no convert op for GPU.
QNN_LOG_DEBUG("[qnn-gpu][MUL_MAT]type src0(%s), src1(%s) and op(%s) are not equal\n",
ggml_type_name(src0->type), ggml_type_name(src1->type), ggml_type_name(op->type));
return false;
}
break;
@ -472,7 +495,7 @@ bool device_supports_op(ggml_backend_qnn_device_context * ctx, const ggml_tensor
break;
default:
// default to supported
is_op_supported = ggml_qnn_have_same_tensor_types(ctx, op);
break;
}
}

2
ggml/src/ggml-qnn/graph.cpp
View File

@ -246,7 +246,7 @@ qnn_graph::qnn_graph(const std::string & graph_name, QNNBackend device, std::sha
return;
}
QNN_LOG_INFO("[%s][%s]create succeed\n", get_backend_name(device), graph_name.c_str());
QNN_LOG_DEBUG("[%s][%s]create succeed\n", get_backend_name(device), graph_name.c_str());
_graph_handle = graph_handle;
_qnn_interface = qnn_interface;
}

5
ggml/src/ggml-qnn/logger.cpp
View File

@ -13,7 +13,7 @@ void qnn::sdk_logcallback(const char * fmt, QnnLog_Level_t level, uint64_t /*tim
static std::mutex log_mutex;
static char s_ggml_qnn_logbuf[4096];
char log_level_desc = 'U';
char log_level_desc;
switch (level) {
case QNN_LOG_LEVEL_ERROR:
log_level_desc = 'E';
@ -30,6 +30,9 @@ void qnn::sdk_logcallback(const char * fmt, QnnLog_Level_t level, uint64_t /*tim
case QNN_LOG_LEVEL_VERBOSE:
log_level_desc = 'V';
break;
default:
log_level_desc = 'U';
break;
}
{