merge: TurboQuant KV cache types onto upstream master (MTP built-in)

ggml/include/ggml.h

@@ -429,7 +429,10 @@ extern "C" {
         GGML_TYPE_MXFP4   = 39, // MXFP4 (1 block)
         GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
         GGML_TYPE_Q1_0    = 41,
-        GGML_TYPE_COUNT   = 42,
+        GGML_TYPE_TURBO3_0 = 42, // TurboQuant 3-bit KV cache: 2-bit PolarQuant + 1-bit QJL
+        GGML_TYPE_TURBO4_0 = 43, // TurboQuant 4-bit KV cache: 3-bit PolarQuant + 1-bit QJL
+        GGML_TYPE_TURBO2_0 = 44, // TurboQuant 2-bit KV cache: 2-bit PolarQuant (no QJL)
+        GGML_TYPE_COUNT   = 45,
     };
 
     // precision
@@ -567,6 +570,7 @@ extern "C" {
         GGML_OP_RWKV_WKV7,
         GGML_OP_SOLVE_TRI,
         GGML_OP_GATED_DELTA_NET,
+        GGML_OP_TURBO_WHT,
 
         GGML_OP_UNARY,
 
@@ -2555,6 +2559,16 @@ extern "C" {
             struct ggml_tensor  * beta,
             struct ggml_tensor  * state);
 
+    // TurboQuant Walsh-Hadamard Transform (O(d log d) rotation for KV cache compression)
+    // Applies WHT rotation to 128-element groups along ne[0]: sign1 → butterfly → sign2 → normalize
+    // direction: 0 = forward (signs1 → WHT → signs2), 1 = inverse (signs2 → WHT → signs1)
+    GGML_API struct ggml_tensor * ggml_turbo_wht(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   direction,
+            int                   group_size,    // 0 = auto (64 or 128 from ne[0])
+            struct ggml_tensor  * scale);        // NULL = no InnerQ scaling
+
     // custom operators
 
     typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);