WIP: add TurboQuant KV cache types (turbo3, turbo4)

New types: GGML_TYPE_TURBO3_0 (3-bit) and GGML_TYPE_TURBO4_0 (4-bit)
Implements PolarQuant + QJL compression per the ICLR 2026 paper.

Block size = 128 (matching head_dim for optimal rotation Gaussianization)
turbo3: 52 bytes per 128 values = 3.25 bits/value (4.9× vs fp16)
turbo4: 68 bytes per 128 values = 4.25 bits/value (3.8× vs fp16)

Status:
- ✅ Type definitions in ggml.h
- ✅ Block structures in ggml-common.h
- ✅ Quantize/dequantize C implementation in ggml-turbo-quant.c
- ✅ Registered in ggml.c type traits
- ✅ Added to kv_cache_types in arg.cpp
- ✅ Builds successfully
- ✅ Shows in --help output
- ❌ Metal SET_ROWS kernel not implemented (blocks GPU inference)
- ❌ Needs Metal dequantize kernels for attention computation

Co-Authored-By: tturney@psyguard.ai
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: TheTom

common/arg.cpp

@@ -390,6 +390,8 @@ const std::vector<ggml_type> kv_cache_types = {
     GGML_TYPE_IQ4_NL,
     GGML_TYPE_Q5_0,
     GGML_TYPE_Q5_1,
+    GGML_TYPE_TURBO3_0,
+    GGML_TYPE_TURBO4_0,
 };
 
 static ggml_type kv_cache_type_from_str(const std::string & s) {

ggml/include/ggml.h

@@ -429,7 +429,9 @@ 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_COUNT   = 44,
     };
 
     // precision

ggml/src/CMakeLists.txt

@@ -206,6 +206,7 @@ add_library(ggml-base
             ggml-threading.h
             ggml-quants.c
             ggml-quants.h
+            ggml-turbo-quant.c
             gguf.cpp)
 
 set_target_properties(ggml-base PROPERTIES

ggml/src/ggml-common.h

@@ -277,6 +277,31 @@ typedef struct {
 } block_tq2_0;
 static_assert(sizeof(block_tq2_0) == sizeof(ggml_half) + QK_K / 4, "wrong tq2_0 block size/padding");
 
+// TurboQuant 3-bit: 2-bit PolarQuant indices + 1-bit QJL signs
+// Block size = 128 (matches typical head_dim for optimal rotation Gaussianization)
+// Per block: norm(fp16) + residual_norm(fp16) + 2-bit indices (32 bytes) + 1-bit signs (16 bytes)
+// = 52 bytes per 128 values = 3.25 bits/value → 4.9× compression vs fp16
+#define QK_TURBO3 128
+typedef struct {
+    ggml_half  norm;                    //  2 bytes: vector L2 norm (for rescaling)
+    ggml_half  rnorm;                   //  2 bytes: QJL residual L2 norm
+    uint8_t    qs[QK_TURBO3 / 4];      // 32 bytes: 2-bit PolarQuant indices (4 per byte)
+    uint8_t    signs[QK_TURBO3 / 8];   // 16 bytes: 1-bit QJL signs (8 per byte)
+} block_turbo3_0;                       // 52 bytes total
+static_assert(sizeof(block_turbo3_0) == 2*sizeof(ggml_half) + QK_TURBO3/4 + QK_TURBO3/8, "wrong turbo3_0 block size/padding");
+
+// TurboQuant 4-bit: 3-bit PolarQuant indices + 1-bit QJL signs
+// Per block: norm(fp16) + residual_norm(fp16) + 3-bit indices (48 bytes) + 1-bit signs (16 bytes)
+// = 68 bytes per 128 values = 4.25 bits/value → 3.8× compression vs fp16
+#define QK_TURBO4 128
+typedef struct {
+    ggml_half  norm;                    //  2 bytes
+    ggml_half  rnorm;                   //  2 bytes
+    uint8_t    qs[QK_TURBO4 * 3 / 8];  // 48 bytes: 3-bit PolarQuant indices
+    uint8_t    signs[QK_TURBO4 / 8];   // 16 bytes: 1-bit QJL signs
+} block_turbo4_0;                       // 68 bytes total
+static_assert(sizeof(block_turbo4_0) == 2*sizeof(ggml_half) + QK_TURBO4*3/8 + QK_TURBO4/8, "wrong turbo4_0 block size/padding");
+
 //
 // Super-block quantization structures
 //

ggml/src/ggml-quants.h

@@ -102,6 +102,14 @@ GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 GGML_API size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
+// TurboQuant KV cache compression (arXiv 2504.19874)
+GGML_API void quantize_row_turbo3_0_ref(const float * GGML_RESTRICT x, block_turbo3_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_turbo4_0_ref(const float * GGML_RESTRICT x, block_turbo4_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_turbo3_0(const block_turbo3_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_turbo4_0(const block_turbo4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API size_t quantize_turbo3_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_turbo4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
 GGML_API void iq2xs_init_impl(enum ggml_type type);
 GGML_API void iq2xs_free_impl(enum ggml_type type);
 GGML_API void iq3xs_init_impl(int grid_size);