Fix Windows/MinGW build and inference

demo/asr_infer.cpp

@@ -15,6 +15,8 @@
 #include "../utils/audio_io.h"
 #include "../utils/prompt_builder.h"
 
+#include "time_compat.h"
+
 #include <algorithm>
 #include <cmath>
 #include <cstdio>
@@ -220,6 +222,9 @@ int main(int argc, char ** argv) {
 
     llama_model_params lm_mparams = llama_model_default_params();
     lm_mparams.n_gpu_layers = 0;  // CPU only
+#ifdef _WIN32
+    lm_mparams.use_mmap = false;  // MinGW/Windows lacks PrefetchVirtualMemory; mmap load fails
+#endif
 
     llama_model * lm_model = llama_load_model_from_file(params.lm_model_path.c_str(), lm_mparams);
     if (!lm_model) {

include/time_compat.h

@@ -0,0 +1,30 @@
+// Portability shim: provide clock_gettime(CLOCK_MONOTONIC, ...) on platforms
+// (e.g. MinGW/Windows) where it is not available. No-op on POSIX systems that
+// already provide it.
+#ifndef VIBEASR_TIME_COMPAT_H
+#define VIBEASR_TIME_COMPAT_H
+
+#include <time.h>
+
+#if defined(_WIN32) && !defined(CLOCK_MONOTONIC)
+
+#include <chrono>
+
+#ifndef CLOCK_MONOTONIC
+#define CLOCK_MONOTONIC 1
+#endif
+
+static inline int vibeasr_clock_gettime(int /*clk_id*/, struct timespec *ts) {
+    const auto now = std::chrono::steady_clock::now().time_since_epoch();
+    const auto secs = std::chrono::duration_cast<std::chrono::seconds>(now);
+    const auto nsecs = std::chrono::duration_cast<std::chrono::nanoseconds>(now - secs);
+    ts->tv_sec = static_cast<time_t>(secs.count());
+    ts->tv_nsec = static_cast<long>(nsecs.count());
+    return 0;
+}
+
+#define clock_gettime(clk, ts) vibeasr_clock_gettime((clk), (ts))
+
+#endif // _WIN32 && !CLOCK_MONOTONIC
+
+#endif // VIBEASR_TIME_COMPAT_H

src/asr_server.cpp

@@ -17,6 +17,8 @@
 #include "../utils/audio_io.h"
 #include "../utils/prompt_builder.h"
 
+#include "time_compat.h"
+
 #include <algorithm>
 #include <cmath>
 #include <cstdio>
@@ -363,6 +365,9 @@ int main(int argc, char ** argv) {
 
     llama_model_params lm_mparams = llama_model_default_params();
     lm_mparams.n_gpu_layers = 0;
+#ifdef _WIN32
+    lm_mparams.use_mmap = false;  // MinGW/Windows lacks PrefetchVirtualMemory; mmap load fails
+#endif
 
     llama_model * lm_model = llama_load_model_from_file(params.lm_model_path.c_str(), lm_mparams);
     if (!lm_model) {

src/vae.cpp

@@ -5,6 +5,8 @@
 #include "ggml-backend.h"
 #include "ggml-vae-i8_s-mad.h"
 
+#include "time_compat.h"
+
 #include <algorithm>
 #include <cassert>
 #include <cmath>
@@ -646,8 +648,15 @@ static int32_t vae_encode_impl(
 
     // Create computation context with sufficient memory
     // F32 models need more memory than I8_S due to 4x larger intermediate tensors
+#ifdef _WIN32
+    // Windows has no memory overcommit: a 128 GB reservation fails outright.
+    // 6 GB is ample for the VAE intermediate tensors at typical audio lengths.
+    const size_t vae_ctx_mem_size = (size_t)6 * 1024 * 1024 * 1024;
+#else
+    const size_t vae_ctx_mem_size = (size_t)128 * 1024 * 1024 * 1024;
+#endif
     struct ggml_init_params ctx_params = {
-        /*.mem_size   =*/ (size_t)128 * 1024 * 1024 * 1024,
+        /*.mem_size   =*/ vae_ctx_mem_size,
         /*.mem_buffer =*/ nullptr,
         /*.no_alloc   =*/ false,  // Let ggml allocate tensors
     };