Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
89 lines
3.5 KiB
C++
89 lines
3.5 KiB
C++
// Verifies the VibeVoice fine-tune overlay: the diffusion head + connectors fully replace their
|
|
// base tensors and the LM LoRA changes the decoder linears. Run with the base 7B model and the
|
|
// mp1 adapter (defaults below), or pass <model_dir> <adapter_dir>.
|
|
#include "engine/framework/assets/tensor_source.h"
|
|
#include "engine/framework/assets/torch_bin.h"
|
|
#include "engine/framework/debug/trace.h"
|
|
#include "engine/models/vibevoice/assets.h"
|
|
#include "engine/models/vibevoice/lora.h"
|
|
|
|
#include <cmath>
|
|
#include <cstdio>
|
|
#include <filesystem>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
namespace {
|
|
|
|
void print8(const char * label, const std::vector<float> & values) {
|
|
std::printf("%-26s", label);
|
|
for (int i = 0; i < 8 && i < static_cast<int>(values.size()); ++i) {
|
|
std::printf(" % .6f", values[i]);
|
|
}
|
|
std::printf("\n");
|
|
}
|
|
|
|
bool equal8(const std::vector<float> & a, const std::vector<float> & b) {
|
|
for (int i = 0; i < 8; ++i) {
|
|
if (std::fabs(a[i] - b[i]) > 1.0e-6F * (1.0F + std::fabs(b[i]))) {
|
|
return false;
|
|
}
|
|
}
|
|
return true;
|
|
}
|
|
|
|
bool differ8(const std::vector<float> & a, const std::vector<float> & b) {
|
|
for (int i = 0; i < 8; ++i) {
|
|
if (std::fabs(a[i] - b[i]) > 1.0e-7F) {
|
|
return true;
|
|
}
|
|
}
|
|
return false;
|
|
}
|
|
|
|
} // namespace
|
|
|
|
int main(int argc, char ** argv) {
|
|
engine::debug::configure_logging(engine::debug::LoggingConfig{true, std::nullopt});
|
|
|
|
const std::filesystem::path model = argc > 1 ? argv[1] : "models/VibeVoice-7B";
|
|
const std::filesystem::path adapter =
|
|
argc > 2 ? argv[2] : "E:/REPOS/VIBEVOICE-MAIN/VibeVoice/loras/mp1";
|
|
|
|
const auto assets = engine::models::vibevoice::load_vibevoice_assets(model);
|
|
const auto base = assets->model_weights;
|
|
const auto overlay = engine::models::vibevoice::make_vibevoice_finetune_source(base, adapter, -1.0F);
|
|
|
|
bool ok = true;
|
|
|
|
const auto acoustic = engine::assets::open_torch_bin_tensor_source(
|
|
adapter / "acoustic_connector" / "pytorch_model.bin");
|
|
const auto acoustic_ref = acoustic->require_f32("fc1.weight");
|
|
const auto acoustic_overlay = overlay->require_f32("model.acoustic_connector.fc1.weight");
|
|
print8("acoustic.fc1 adapter", acoustic_ref);
|
|
print8("acoustic.fc1 overlay", acoustic_overlay);
|
|
const bool a_ok = equal8(acoustic_ref, acoustic_overlay);
|
|
std::printf("(a) acoustic connector override: %s\n\n", a_ok ? "PASS" : "FAIL");
|
|
ok = ok && a_ok;
|
|
|
|
const auto head = engine::assets::open_tensor_source(adapter / "diffusion_head" / "model.safetensors");
|
|
const auto head_ref = head->require_f32("cond_proj.weight");
|
|
const auto head_overlay = overlay->require_f32("model.prediction_head.cond_proj.weight");
|
|
print8("dhead.cond_proj adapter", head_ref);
|
|
print8("dhead.cond_proj overlay", head_overlay);
|
|
const bool b_ok = equal8(head_ref, head_overlay);
|
|
std::printf("(b) diffusion head override: %s\n\n", b_ok ? "PASS" : "FAIL");
|
|
ok = ok && b_ok;
|
|
|
|
const std::string lora_weight = "model.language_model.layers.0.mlp.gate_proj.weight";
|
|
const auto lora_base = base->require_f32(lora_weight);
|
|
const auto lora_overlay = overlay->require_f32(lora_weight);
|
|
print8("gate_proj base", lora_base);
|
|
print8("gate_proj overlay", lora_overlay);
|
|
const bool c_ok = differ8(lora_base, lora_overlay);
|
|
std::printf("(c) LM LoRA merge changed weights: %s\n\n", c_ok ? "PASS" : "FAIL");
|
|
ok = ok && c_ok;
|
|
|
|
std::printf("=== OVERLAY CHECK %s ===\n", ok ? "PASS" : "FAIL");
|
|
return ok ? 0 : 1;
|
|
}
|