Skip to content

fix: validate vision projector output dim against LLM hidden size #1918

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
leejet merged 6 commits into from
Sep 13, 2026
+20 −8
Merged

fix: validate vision projector output dim against LLM hidden size #1918

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
66e2352
fix: validate vision projector output dim against LLM hidden size
fszontagh Aug 27, 2026
916c514
Merge remote-tracking branch 'upstream/master' into fix/llm-vision-di…
fszontagh Aug 27, 2026
db2310b
Merge remote-tracking branch 'upstream/master' into fix/llm-vision-di…
fszontagh Sep 2, 2026
6b732fc
Merge remote-tracking branch 'upstream/master' into fix/llm-vision-di…
fszontagh Sep 7, 2026
fb1c859
refactor: move LLM vision validation into config detection
leejet Sep 13, 2026
71893f3
Merge branch 'master' into fix/llm-vision-dim-mismatch
leejet Sep 13, 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
28 changes: 20 additions & 8 deletions src/model/te/llm.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,7 +139,8 @@ namespace LLM {

static LLMConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
const std::string& prefix,
LLMArch arch) {
LLMArch arch,
bool& enable_vision) {
LLMConfig config;
config.arch = arch;
if (arch == LLMArch::MISTRAL_SMALL_3_2 || arch == LLMArch::MINISTRAL_3_3B) {
Expand Down Expand Up @@ -230,8 +231,9 @@ namespace LLM {
config.num_experts_per_tok = 4;
}

config.num_layers = 0;
int detected_vision_layers = 0;
config.num_layers = 0;
int detected_vision_layers = 0;
bool out_hidden_size_detected = false;
for (const auto& [name, tensor_storage] : tensor_storage_map) {
if (!starts_with(name, prefix)) {
continue;
Expand Down Expand Up @@ -277,6 +279,7 @@ namespace LLM {
if (ends_with(name, "visual.merger.linear_fc2.weight") ||
ends_with(name, "visual.merger.mlp.2.weight")) {
config.vision.out_hidden_size = tensor_storage.ne[1];
out_hidden_size_detected = true;
}
continue;
}
Expand Down Expand Up @@ -330,6 +333,19 @@ namespace LLM {
config.vocab_size,
config.hidden_size,
config.intermediate_size);
if (enable_vision && !config.have_vision_weight) {
LOG_WARN("no vision weights detected, vision disabled");
enable_vision = false;
}
// The default would reject valid models, so only compare a detected dim.
if (enable_vision && out_hidden_size_detected &&
config.vision.out_hidden_size != config.hidden_size) {
LOG_ERROR("vision projector output size (%" PRId64 ") does not match LLM hidden size (%" PRId64 "), "
"the vision weights (mmproj) likely belong to a different LLM variant, vision disabled",
config.vision.out_hidden_size,
config.hidden_size);
enable_vision = false;
}
return config;
}
};
Expand Down Expand Up @@ -1886,12 +1902,8 @@ namespace LLM {
bool enable_vision_ = false,
std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
: GGMLRunner(backend, weight_manager),
config(LLMConfig::detect_from_weights(tensor_storage_map, prefix, arch)),
config(LLMConfig::detect_from_weights(tensor_storage_map, prefix, arch, enable_vision_)),
enable_vision(enable_vision_) {
if (enable_vision && !config.have_vision_weight) {
LOG_WARN("no vision weights detected, vision disabled");
enable_vision = false;
}
if (enable_vision) {
LOG_VERBOSE("enable llm vision");
if (config.llama_cpp_style) {
Expand Down
Loading