fix: add TurboQuant for Windows

ggml/src/ggml-cpu/ops.cpp

@@ -4894,6 +4894,8 @@ void ggml_compute_forward_get_rows(
     //}
 }
 
+extern "C" GGML_API int turbo3_cpu_wht_group_size;
+
 template<typename idx_t>
 static void ggml_compute_forward_set_rows_f32(
         const ggml_compute_params * params,
@@ -4928,7 +4930,6 @@ static void ggml_compute_forward_set_rows_f32(
 
     // For turbo types: communicate WHT group size to the quantize function via global
     if (dst->type == GGML_TYPE_TURBO3_0 || dst->type == GGML_TYPE_TURBO4_0 || dst->type == GGML_TYPE_TURBO2_0) {
-        extern int turbo3_cpu_wht_group_size;
         int gs = 0;
         memcpy(&gs, dst->op_params, sizeof(int));
         turbo3_cpu_wht_group_size = (gs == 64 || gs == 128) ? gs : 0;

ggml/src/ggml-cuda/turbo-innerq.cuh

@@ -1,5 +1,7 @@
 #pragma once
 
+#include "ggml.h"
+
 // TurboQuant InnerQ per-channel equalization — cross-TU shared state
 // The host-side state lives in turbo-innerq.cu; device-side state is per-TU
 // in turbo-quant.cuh (only set-rows.cu needs device access).
@@ -8,14 +10,14 @@
 
 // Host-side shared state (defined in turbo-innerq.cu)
 extern bool  g_innerq_finalized;
-extern float g_innerq_scale_inv_host[INNERQ_MAX_CHANNELS];
+GGML_API float g_innerq_scale_inv_host[INNERQ_MAX_CHANNELS];
 
 // Called from set-rows.cu after InnerQ finalization to publish scale_inv
 void turbo_innerq_publish(const float * scale_inv, int group_size);
 
 // Called from llama-kv-cache.cpp (or equivalent) to check if tensor needs update
 // Returns true if there are new scale_inv values to upload
-bool turbo_innerq_needs_tensor_update(void);
+GGML_API bool turbo_innerq_needs_tensor_update(void);
 
 // Called after tensor update to clear the flag
-void turbo_innerq_mark_tensor_updated(void);
+GGML_API void turbo_innerq_mark_tensor_updated(void);

ggml/src/ggml-turbo-quant.c

@@ -10,13 +10,17 @@
 #include "ggml-common.h"
 #include "ggml-impl.h"
 
+#if defined(_WIN32) && !defined(_USE_MATH_DEFINES)
+#define _USE_MATH_DEFINES
+#endif 
+
 #include <math.h>
 #include <string.h>
 #include <assert.h>
 #include <stdlib.h>
 
 /* Global: WHT group size for CPU quantize path (set by CPU SET_ROWS handler) */
-int turbo3_cpu_wht_group_size = 0;
+GGML_API int turbo3_cpu_wht_group_size = 0;
 
 /* ---------- constants ---------- */
 
@@ -246,7 +250,6 @@ void quantize_row_turbo3_0_ref(const float * GGML_RESTRICT x, block_turbo3_0 * G
 
     // Read WHT group size from global (set by CPU SET_ROWS handler before each call).
     // Fallback: 128 if row is 128-aligned, else 64.
-    extern int turbo3_cpu_wht_group_size;
     int group_size = turbo3_cpu_wht_group_size;
     if (group_size != 64 && group_size != 128) {
         group_size = (k % 128 == 0) ? 128 : 64;
@@ -341,7 +344,6 @@ size_t quantize_turbo3_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT d
 void quantize_row_turbo2_0_ref(const float * GGML_RESTRICT x, block_turbo2_0 * GGML_RESTRICT y, int64_t k) {
     assert(k % QK_TURBO2 == 0);
 
-    extern int turbo3_cpu_wht_group_size;
     int group_size = turbo3_cpu_wht_group_size;
     if (group_size != 64 && group_size != 128) {
         group_size = (k % 128 == 0) ? 128 : 64;

src/llama-kv-cache.cpp

@@ -21,10 +21,10 @@
 #endif
 
 #ifdef GGML_USE_CUDA
-extern bool  g_innerq_finalized;
-extern float g_innerq_scale_inv_host[INNERQ_MAX_CHANNELS];
-extern bool turbo_innerq_needs_tensor_update(void);
-extern void turbo_innerq_mark_tensor_updated(void);
+GGML_API bool  g_innerq_finalized;
+GGML_API float g_innerq_scale_inv_host[INNERQ_MAX_CHANNELS];
+GGML_API bool turbo_innerq_needs_tensor_update(void);
+GGML_API void turbo_innerq_mark_tensor_updated(void);
 #else
 static bool  g_innerq_finalized = false;
 static float g_innerq_scale_inv_host[INNERQ_MAX_CHANNELS] = {};