Skip to content

[QNN-EP] Add additional guards for file mapping #27871

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
adrianlizarraga merged 12 commits into from
Apr 10, 2026
Merged

[QNN-EP] Add additional guards for file mapping #27871

Changes from 2 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
30a066f
[QNN-EP] Add additional guards for file mapping
Mar 26, 2026
47c67ec
Fix typo in log
Mar 26, 2026
5f0d87d
address pr comments
Mar 27, 2026
aa7b319
Use unique_ptrs w/ deleters to properly clean up QnnSystemContext_Han…
Mar 27, 2026
d035d34
lint
Mar 27, 2026
1fa741e
Address copilot comment
Mar 31, 2026
c4a5776
Address pr comments
Apr 1, 2026
bf8f5bd
Resolve logic errors
Apr 1, 2026
2998f5e
Cast unsigned char* to void* when retrieving bin version in
Apr 1, 2026
7ae34b5
comments
Apr 2, 2026
6fe0e9a
Move checks for function pointers for sys context handle
Apr 4, 2026
bbeac4d
pr comments
Apr 8, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
41 changes: 36 additions & 5 deletions onnxruntime/core/providers/qnn/builder/qnn_backend_manager.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1351,8 +1351,9 @@ Status QnnBackendManager::LoadCachedQnnContextFromBuffer(char* buffer, uint64_t

void* bin_buffer = nullptr;
#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
bool use_file_mapping = file_mapped_weights_enabled_;
// A nonzero buffer length implies an embedded context
if (file_mapped_weights_enabled_ && buffer_length == 0) {
if (use_file_mapping && buffer_length == 0) {
ORT_RETURN_IF(!file_mapper_, "Attemping to use File Mapping feature but file_mapper_ is uninitialized");
Comment thread
quic-calvnguy marked this conversation as resolved.
Outdated
Comment thread
quic-calvnguy marked this conversation as resolved.
Outdated

ORT_RETURN_IF_ERROR(GetFileSizeIfValid(context_bin_filepath, buffer_length));
Expand All @@ -1361,6 +1362,11 @@ Status QnnBackendManager::LoadCachedQnnContextFromBuffer(char* buffer, uint64_t
ORT_RETURN_IF_ERROR(file_mapper_->GetContextBinMappedMemoryPtr(context_bin_filepath, &bin_buffer));

Comment thread
quic-calvnguy marked this conversation as resolved.
} else {
if (use_file_mapping) {
use_file_mapping = false;
LOGS(*logger_, WARNING) << "Node " << node_name << " is using an embedded cache."
<< "Disabling file mapping for this node.";
Comment thread
quic-calvnguy marked this conversation as resolved.
Outdated
}
ORT_RETURN_IF(buffer == nullptr, "Attempting to load QNN context from buffer but buffer is null");
bin_buffer = static_cast<void*>(buffer);
}
Expand All @@ -1385,28 +1391,52 @@ Status QnnBackendManager::LoadCachedQnnContextFromBuffer(char* buffer, uint64_t
// Binary info to graph info
// retrieve Qnn graph info from binary info
ORT_RETURN_IF(nullptr == binary_info, "Qnn cached binary info is nullptr.");

#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
Qnn_Version_t blob_version;
Comment thread
quic-calvnguy marked this conversation as resolved.
Outdated
#endif

uint32_t graph_count = 0;
QnnSystemContext_GraphInfo_t* graphs_info = nullptr;
if (binary_info->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_1) {
graph_count = binary_info->contextBinaryInfoV1.numGraphs;
graphs_info = binary_info->contextBinaryInfoV1.graphs;
#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
blob_version = binary_info->contextBinaryInfoV1.contextBlobVersion;
#endif
}
#if QNN_API_VERSION_MAJOR == 2 && (QNN_API_VERSION_MINOR >= 15) // starts from 2.22
else if (binary_info->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_2) {
graph_count = binary_info->contextBinaryInfoV2.numGraphs;
graphs_info = binary_info->contextBinaryInfoV2.graphs;
#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
blob_version = binary_info->contextBinaryInfoV2.contextBlobVersion;
#endif
}
#endif
#if QNN_API_VERSION_MAJOR == 2 && (QNN_API_VERSION_MINOR >= 21) // starts from 2.28
else if (binary_info->version == QNN_SYSTEM_CONTEXT_BINARY_INFO_VERSION_3) {
graph_count = binary_info->contextBinaryInfoV3.numGraphs;
graphs_info = binary_info->contextBinaryInfoV3.graphs;
#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
blob_version = binary_info->contextBinaryInfoV3.contextBlobVersion;
#endif
}
#endif
else {
return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, "Unsupported context binary info version.");
}

#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
// Cannot use contextCreateFromBinaryWithCallback() unless context bin version is >= 3.3.3
if (use_file_mapping && (blob_version.major < 3 || blob_version.minor < 3 || blob_version.patch < 3)) {
std::string version_str(std::to_string(blob_version.major) + "." + std::to_string(blob_version.minor) + "." + std::to_string(blob_version.patch));
LOGS(*logger_, WARNING) << "Context binary of " << node_name << " is " + version_str
<< "File mapping is only supported for versions >= 3.3.3. Disabling file mapping for this node.";
Comment thread
quic-calvnguy marked this conversation as resolved.
Outdated
use_file_mapping = false;
}
Comment thread
quic-calvnguy marked this conversation as resolved.
Outdated
#endif

ORT_RETURN_IF(graph_count < 1 || graphs_info == nullptr, "Failed to get graph info from Qnn cached context.");
LOGS(*logger_, VERBOSE) << "Graph count from QNN context: " << graph_count;

Expand Down Expand Up @@ -1452,7 +1482,7 @@ Status QnnBackendManager::LoadCachedQnnContextFromBuffer(char* buffer, uint64_t

#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
Qnn_ContextBinaryCallback_t callbacks;
if (file_mapped_weights_enabled_ && file_mapper_) {
if (use_file_mapping && file_mapper_) {
ORT_RETURN_IF(nullptr == qnn_interface_.contextCreateFromBinaryWithCallback,
"Invalid function pointer for contextCreateFromBinaryWithCallback.");

Expand All @@ -1479,7 +1509,7 @@ Status QnnBackendManager::LoadCachedQnnContextFromBuffer(char* buffer, uint64_t

#ifdef QNN_FILE_MAPPED_WEIGHTS_AVAILABLE
std::vector<char> backup_buffer;
if (file_mapped_weights_enabled_ && file_mapper_) {
if (use_file_mapping && file_mapper_) {
rt = qnn_interface_.contextCreateFromBinaryWithCallback(backend_handle_,
device_handle_,
context_configs,
Expand All @@ -1501,9 +1531,10 @@ Status QnnBackendManager::LoadCachedQnnContextFromBuffer(char* buffer, uint64_t
bin_buffer = static_cast<void*>(backup_buffer.data());
}
}
#endif // QNN_FILE_MAPPED_WEIGHTS_AVAILABLE

if (!file_mapped_weights_enabled_ || rt != QNN_SUCCESS) {
if (!use_file_mapping || rt != QNN_SUCCESS)
#endif
Comment thread
quic-calvnguy marked this conversation as resolved.
{
rt = qnn_interface_.contextCreateFromBinary(backend_handle_,
device_handle_,
context_configs,
Expand Down
Loading