ggml : add AArch64 optimized GEMV and GEMM Q4 kernels (#5780)

* Arm AArch64: optimized GEMV and GEMM kernels for q4_0_q8_0, and q8_0_q8_0 quantization

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add copyright claim only to ggml-aarch64.cpp and ggml-aarch64.h files

* Arm AArch64: minor code refactoring for rebase

* Arm AArch64: minor code refactoring for resolving a build issue with cmake

* Arm AArch64: minor code refactoring to split the Q4_0_AARC64 type into three separate types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: minor code change for resolving a build issue with server-windows

* retrigger checks

* Arm AArch64: minor code changes for rebase

* Arm AArch64: minor changes to skip the pr#7433 vec_dot code for arm cpus with SVE VL not equal to 256 bits

* Arm AArch64: remove stale LLAMA_QKK_64 from CMakeLists.txt and delete build.zig

* Arm AArch64: add reference scalar gemm and gemv, and avoid dynamic memory allocations during quantization for Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: add multithreaded quantization support for the new types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: minor code refactoring

* Arm AArch64: simplify logic for calling gemm and gemv functions in ggml_compute_forward_mul_mat

* Arm AArch64: minimize changes in ggml_compute_forward_mul_mat

* Arm AArch64: minor code refactoring, and add reference scalar code to quantize routines for new quant types

* Arm AArch64: minor code refactoring

* Arm AArch64: minor code refactoring

* Arm AArch64: minor code refactoring

* rebase on the latest master commit 3fd62a6 and adapt to the new directory structure

* Arm AArch64: remove a redundant comment

* Arm AArch64: add pragma in ggml-aarch64.c to turn -Woverlength-strings warning off

* Arm AArch64: use __aarch64__ check to guard 64-bit neon kernels

* Arm AArch64: update docs/build.md README to include compile time flags for buiilding the Q4_0_4_4 quant type

Makefile

@@ -835,7 +835,8 @@ OBJ_GGML += \
 	ggml/src/ggml.o \
 	ggml/src/ggml-alloc.o \
 	ggml/src/ggml-backend.o \
-	ggml/src/ggml-quants.o
+	ggml/src/ggml-quants.o \
+	ggml/src/ggml-aarch64.o
 
 OBJ_LLAMA = \
 	src/llama.o \
@@ -969,6 +970,13 @@ ggml/src/ggml-quants.o: \
 	ggml/src/ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
+ggml/src/ggml-aarch64.o: \
+	ggml/src/ggml-aarch64.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
 ggml/src/ggml-blas.o: \
 	ggml/src/ggml-blas.cpp \
 	ggml/include/ggml-blas.h

Package.swift

@@ -10,6 +10,7 @@ var sources = [
     "ggml/src/ggml-alloc.c",
     "ggml/src/ggml-backend.c",
     "ggml/src/ggml-quants.c",
+    "ggml/src/ggml-aarch64.c",
 ]
 
 var resources: [Resource] = []

docs/build.md

@@ -28,6 +28,7 @@ In order to build llama.cpp you have four different options.
         ```
 
   - Notes:
+    - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
     - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
     - For faster repeated compilation, install [ccache](https://ccache.dev/).
     - For debug builds, run `make LLAMA_DEBUG=1`
@@ -41,6 +42,7 @@ In order to build llama.cpp you have four different options.
 
   **Notes**:
 
+    - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
     - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
     - For faster repeated compilation, install [ccache](https://ccache.dev/).
     - For debug builds, there are two cases:

examples/quantize/quantize.cpp

@@ -46,6 +46,9 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B",  },
     { "Q6_K",   LLAMA_FTYPE_MOSTLY_Q6_K,   " 6.14G, +0.0217 ppl @ Llama-3-8B",  },
     { "Q8_0",   LLAMA_FTYPE_MOSTLY_Q8_0,   " 7.96G, +0.0026 ppl @ Llama-3-8B",  },
+    { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
+    { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
+    { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
     { "F16",    LLAMA_FTYPE_MOSTLY_F16,    "14.00G, +0.0020 ppl @ Mistral-7B",  },
     { "BF16",   LLAMA_FTYPE_MOSTLY_BF16,   "14.00G, -0.0050 ppl @ Mistral-7B",  },
     { "F32",    LLAMA_FTYPE_ALL_F32,       "26.00G              @ 7B",          },

ggml/include/ggml.h

@@ -383,6 +383,9 @@ extern "C" {
         GGML_TYPE_F64     = 28,
         GGML_TYPE_IQ1_M   = 29,
         GGML_TYPE_BF16    = 30,
+        GGML_TYPE_Q4_0_4_4 = 31,
+        GGML_TYPE_Q4_0_4_8 = 32,
+        GGML_TYPE_Q4_0_8_8 = 33,
         GGML_TYPE_COUNT,
     };
 
@@ -424,6 +427,9 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
     };
 
     // available tensor operations:
@@ -2406,6 +2412,12 @@ extern "C" {
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_vec_dot_t)   (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+    typedef void (*ggml_from_float_to_mat_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t nr,
+                                             int64_t k, int64_t bx);
+    typedef void (*ggml_gemv_t)      (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                      const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_gemm_t)      (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                      const void * GGML_RESTRICT y, int nr, int nc);
 
     typedef struct {
         const char      * type_name;
@@ -2418,6 +2430,11 @@ extern "C" {
         ggml_vec_dot_t    vec_dot;
         enum ggml_type    vec_dot_type;
         int64_t           nrows; // number of rows to process simultaneously;
+        int64_t           ncols; // number of columns to process simultaneously;
+        int64_t           interleave_blcksize; // interleave elements in blocks of interleave_blcksize;
+        ggml_from_float_to_mat_t from_float_to_mat;
+        ggml_gemv_t       gemv;
+        ggml_gemm_t       gemm;
     } ggml_type_traits_t;
 
     GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);

ggml/src/CMakeLists.txt

@@ -1153,6 +1153,7 @@ add_library(ggml
             ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
             ${GGML_SOURCES_BLAS}      ${GGML_HEADERS_BLAS}
             ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
+            ggml-aarch64.c            ggml-aarch64.h
             )
 
 if (EMSCRIPTEN)

ggml/src/ggml-aarch64.h

@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t interleave_blcksize);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+// GEMV
+void ggml_gemv_q4_0_4x4_q8_0  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// GEMM
+void ggml_gemm_q4_0_4x4_q8_0  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#ifdef __cplusplus
+}
+#endif
+

ggml/src/ggml-common.h

@@ -199,6 +199,30 @@ typedef struct {
 } block_q8_1;
 static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
 
+typedef struct {
+    ggml_half d[4];        // deltas for 4 q4_0 blocks
+    uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks
+} block_q4_0x4;
+static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding");
+
+typedef struct {
+    ggml_half d[8];        // deltas for 8 q4_0 blocks
+    uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks
+} block_q4_0x8;
+static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding");
+
+typedef struct {
+    ggml_half d[4];        // deltas for 4 q8_0 blocks
+    int8_t qs[QK8_0 * 4];  // quants for 4 q8_0 blocks
+} block_q8_0x4;
+static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding");
+
+typedef struct {
+    ggml_half d[8];        // deltas for 8 q8_0 blocks
+    int8_t qs[QK8_0 * 8];  // quants for 8 q8_0 blocks
+} block_q8_0x8;
+static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding");
+
 //
 // Super-block quantization structures
 //

ggml/src/ggml-impl.h

@@ -609,6 +609,10 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
 // precomputed f32 table for f16 (256 KB)
 // defined in ggml.c, initialized in ggml_init()
 extern float ggml_table_f32_f16[1 << 16];

ggml/src/ggml-quants.c

@@ -3814,6 +3814,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     }
 #endif
 #if defined(__ARM_FEATURE_SVE)
+    if (svcntb() == QK8_0) {
         const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
         const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
 
@@ -3850,7 +3851,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         }
 
         *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-#elif defined(__ARM_NEON)
+        return;
+    }
+#endif
+#if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
@@ -5422,6 +5426,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     }
 #endif
 #if defined(__ARM_FEATURE_SVE)
+    if (svcntb() == QK8_0) {
         svfloat32_t sumv0 = svdup_n_f32(0.0f);
         svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
@@ -5446,7 +5451,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
         }
 
         *s = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-#elif defined(__ARM_NEON)
+        return;
+    }
+#endif
+#if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
@@ -14760,6 +14768,16 @@ static bool validate_fp16(ggml_fp16_t f, size_t i) {
         } \
     }
 
+#define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        for (size_t j = 0; j < (nr); ++j) { \
+            if (!validate_fp16(q[i].d[j], i)) { \
+                return false; \
+            } \
+        } \
+    }
+
 bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes) {
     if (type < 0 || type >= GGML_TYPE_COUNT) {
         fprintf(stderr, "%s: invalid type %d\n", __func__, type);
@@ -14977,6 +14995,16 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
             } break;
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+            {
+                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4);
+            } break;
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8);
+            } break;
+
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:

ggml/src/ggml.c

@@ -4,6 +4,7 @@
 #include "ggml-impl.h"
 #include "ggml-quants.h"
 #include "ggml.h"
+#include "ggml-aarch64.h"
 
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
@@ -37,7 +38,7 @@
 #include <unistd.h>
 #endif
 
-#ifdef __ARM_FEATURE_MATMUL_INT8
+#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
 #undef GGML_USE_LLAMAFILE
 #endif
 
@@ -692,6 +693,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
 #else
         .nrows                    = 1,
 #endif
+        .from_float_to_mat        = quantize_mat_q8_0,
     },
     [GGML_TYPE_Q8_1] = {
         .type_name                = "q8_1",
@@ -889,6 +891,54 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
+    },
+    [GGML_TYPE_Q4_0_4_4] = {
+        .type_name                = "q4_0_4x4",
+        .blck_size                = QK4_0,
+        .type_size                = sizeof(block_q4_0),
+        .is_quantized             = true,
+        .to_float                 = NULL,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
+        .vec_dot                  = NULL,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+        .ncols                    = 4,
+        .interleave_blcksize      = 4,
+        .gemv                     = ggml_gemv_q4_0_4x4_q8_0,
+        .gemm                     = ggml_gemm_q4_0_4x4_q8_0,
+    },
+    [GGML_TYPE_Q4_0_4_8] = {
+        .type_name                = "q4_0_4x8",
+        .blck_size                = QK4_0,
+        .type_size                = sizeof(block_q4_0),
+        .is_quantized             = true,
+        .to_float                 = NULL,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
+        .vec_dot                  = NULL,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+        .ncols                    = 4,
+        .interleave_blcksize      = 8,
+        .gemv                     = ggml_gemv_q4_0_4x8_q8_0,
+        .gemm                     = ggml_gemm_q4_0_4x8_q8_0,
+    },
+    [GGML_TYPE_Q4_0_8_8] = {
+        .type_name                = "q4_0_8x8",
+        .blck_size                = QK4_0,
+        .type_size                = sizeof(block_q4_0),
+        .is_quantized             = true,
+        .to_float                 = NULL,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
+        .vec_dot                  = NULL,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+        .ncols                    = 8,
+        .interleave_blcksize      = 8,
+        .gemv                     = ggml_gemv_q4_0_8x8_q8_0,
+        .gemm                     = ggml_gemm_q4_0_8x8_q8_0,
     }
 };
 
@@ -3188,6 +3238,9 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
         case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
+        case GGML_FTYPE_MOSTLY_Q4_0_4_4:      wtype = GGML_TYPE_Q4_0_4_4; break;
+        case GGML_FTYPE_MOSTLY_Q4_0_4_8:      wtype = GGML_TYPE_Q4_0_4_8; break;
+        case GGML_FTYPE_MOSTLY_Q4_0_8_8:      wtype = GGML_TYPE_Q4_0_8_8; break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -9432,6 +9485,9 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_add_q_f32(params, dst);
             } break;
@@ -9807,6 +9863,9 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_add1_q_f32(params, dst);
             } break;
@@ -9932,6 +9991,9 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
         default:
             {
                 GGML_ASSERT(false);
@@ -12134,6 +12196,12 @@ static void ggml_compute_forward_mul_mat(
     enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
     ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
     int64_t           const vec_dot_num_rows      = type_traits[type].nrows;
+    int64_t           const matmul_num_cols       = type_traits[type].ncols;
+    int64_t           const interleave_blcksize   = type_traits[type].interleave_blcksize;
+    ggml_from_float_to_mat_t const from_float_to_mat
+                                                  = type_traits[vec_dot_type].from_float_to_mat;
+    ggml_gemv_t       const gemv                  = type_traits[type].gemv;
+    ggml_gemm_t       const gemm                  = type_traits[type].gemm;
 
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -12192,7 +12260,16 @@ UseGgmlGemm1:;
 
         for (int64_t i13 = 0; i13 < ne13; ++i13) {
             for (int64_t i12 = 0; i12 < ne12; ++i12) {
-                for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
+                int64_t i11_processed = 0;
+                if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) {
+                    for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+                        from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
+                                          (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
+                                          4, ne10, interleave_blcksize);
+                    }
+                    i11_processed = ne11 - ne11 % 4;
+                }
+                for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
                     from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
                                           (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
                                            ne10);
@@ -12273,6 +12350,28 @@ UseGgmlGemm2:;
     const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
     const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
 
+    if ((ggml_n_dims(src0) == 2) && gemv) {
+        const void * src1_wdata      = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+        const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11;
+        int64_t src0_start = (ith * ne01) / nth;
+        int64_t src0_end   = ((ith + 1) * ne01) / nth;
+        src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start;
+        src0_end   = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end;
+        if (src0_start >= src0_end) return;
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (gemm && (ne11 > 3)) {
+            gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01,
+                 (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) {
+            gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                 (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                 src0_end - src0_start);
+        }
+        return;
+    }
+
     // The first chunk comes from our thread_id, the rest will get auto-assigned.
     int current_chunk = ith;
 
@@ -12318,6 +12417,8 @@ static void ggml_compute_forward_mul_mat_id(
     ggml_vec_dot_t    const vec_dot               = type_traits[type].vec_dot;
     enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
     ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
+    int64_t           const matmul_num_cols       = type_traits[type].ncols;
+    ggml_gemv_t       const gemv                  = type_traits[type].gemv;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -12403,6 +12504,34 @@ static void ggml_compute_forward_mul_mat_id(
         const int64_t nr0 = ne01; // src0 rows
         const int64_t nr1 = cne1; // src1 rows
 
+        if (((ggml_n_dims(src0) - 1) == 2) && gemv) {
+            int64_t src0_cur_start = (ith * ne01) / nth;
+            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
+            src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start;
+            src0_cur_end   = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end;
+            if (src0_cur_start >= src0_cur_end) return;
+
+            for (int ir1 = 0; ir1 < nr1; ir1++) {
+                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
+                const int id       = row_mapping.i1; // selected expert index
+
+                const int64_t  i11 = id % ne11;
+                const int64_t  i12 = row_mapping.i2; // row index in src1
+
+                const int64_t  i1 = id;  // selected expert index
+                const int64_t  i2 = i12; // row
+
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                    ? (i11        + i12 * ne11) * row_size
+                    : (i11 * nb11 + i12 * nb12));
+
+                gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
+                     (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
+            }
+            continue;
+        }
+
         // distribute the thread work across the inner or outer loop based on which one is larger
 
         const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
@@ -12704,6 +12833,9 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_out_prod_q_f32(params, dst);
             } break;
@@ -12889,6 +13021,9 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
         default:
             {
                 GGML_ASSERT(false);
@@ -13148,6 +13283,9 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_get_rows_q(params, dst);
             } break;
@@ -13734,6 +13872,9 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_Q8_K:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:
@@ -20457,6 +20598,9 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_F16:
             {
                 size_t elemsize = sizeof(ggml_fp16_t);
@@ -21759,8 +21903,6 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_FEATURE_SVE)
-    // TODO: Currently, SVE 256 bit is only supported.
-    GGML_ASSERT(svcntb() == QK8_0);
     return 1;
 #else
     return 0;

include/llama.h

@@ -162,6 +162,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };

src/llama.cpp

@@ -3788,6 +3788,9 @@ struct llama_model_loader {
                 case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
                 case GGML_TYPE_IQ4_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS;  break;
                 case GGML_TYPE_IQ3_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ3_S;   break;
+                case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
+                case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
+                case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -4481,6 +4484,9 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_S:  return "IQ3_S - 3.4375 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_M:  return "IQ3_S mix - 3.66 bpw";
+        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
+        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
+        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
 
         default: return "unknown, may not work";
     }
@@ -17768,6 +17774,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
                 new_type = GGML_TYPE_IQ3_S;
             }
+            else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
+                     new_type == GGML_TYPE_Q4_0_8_8) {
+                new_type = GGML_TYPE_Q4_0;
+            }
         }
     } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
                ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M    || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
@@ -18080,6 +18090,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_IQ4_XS:  default_type = GGML_TYPE_IQ4_XS;  break;
         case LLAMA_FTYPE_MOSTLY_IQ3_S:   default_type = GGML_TYPE_IQ3_S;   break;
         case LLAMA_FTYPE_MOSTLY_IQ3_M:   default_type = GGML_TYPE_IQ3_S;   break;
+        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -18390,6 +18403,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 f32_data = (float *) f32_conv_buf.data();
             }
 
+            int chunk_size_multiplier = 1;
+            if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
+                if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
+                else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
+                if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
+                else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
+            }
+
             LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
             fflush(stdout);
 
@@ -18402,7 +18423,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             const int64_t nrows = tensor->ne[1];
 
             static const int64_t min_chunk_size = 32 * 512;
-            const int64_t chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);
+            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
+                                       chunk_size_multiplier;
 
             const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
             const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;