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

This commit is contained in:
fszontagh
2026-09-13 23:42:46 +08:00
committed by GitHub
parent 0bd72f075a
commit ca37fad89a
+20 -8
View File
@@ -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) {
@@ -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;
@@ -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;
}
@@ -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;
}
};
@@ -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) {