ggml: add Q1_0 1-bit quantization support (CPU) (#21273)

* ggml: add Q1_0 and Q1_0_g128 1-bit quantization support (CPU)

* add generic fallback for x86

* remove Q1_0 (group size 32)

* rename Q1_0_g128 => Q1_0

* fix Q1_0 LlamaFileType Enum

* Fix trailing spaces; add generic fallback for othre backends

* Apply suggestions from code review

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* fix /r/n spacing + arch-fallback

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

ggml/src/ggml-cpu/arch-fallback.h

@@ -16,6 +16,7 @@
 #define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -82,6 +83,7 @@
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
@@ -112,6 +114,7 @@
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
@@ -160,6 +163,7 @@
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -200,6 +204,7 @@
 #elif defined(__riscv)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
@@ -240,6 +245,7 @@
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -303,6 +309,7 @@
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -137,6 +137,109 @@ void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 
 //===================================== Dot products =================================
 
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK1_0;  // 128
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q1_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    float sumf = 0.0f;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv = vdupq_n_f32(0.0f);
+
+    for (int i = 0; i < nb; i++) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        // Process 4 Q8_0 blocks (each has 32 elements)
+        for (int k = 0; k < 4; k++) {
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 4 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
+
+            // Get the 4 bytes of bits for this Q8_0 block (32 bits = 4 bytes)
+            // Bits are at offset k*4 bytes in x[i].qs
+            const uint8_t * bits = &x[i].qs[k * 4];
+
+            // Load 32 int8 values from y
+            const int8x16_t y0 = vld1q_s8(yb->qs);
+            const int8x16_t y1 = vld1q_s8(yb->qs + 16);
+
+            // Byte 0-1: bits for y0[0..15]
+            const uint64_t expand0 = table_b2b_0[bits[0]];
+            const uint64_t expand1 = table_b2b_0[bits[1]];
+            // Byte 2-3: bits for y1[0..15]
+            const uint64_t expand2 = table_b2b_0[bits[2]];
+            const uint64_t expand3 = table_b2b_0[bits[3]];
+
+            // Build the sign vectors by reinterpreting the table values
+            uint8x8_t e0 = vcreate_u8(expand0);
+            uint8x8_t e1 = vcreate_u8(expand1);
+            uint8x8_t e2 = vcreate_u8(expand2);
+            uint8x8_t e3 = vcreate_u8(expand3);
+
+            // Shift right by 4 to get 0 or 1
+            int8x8_t s0 = vreinterpret_s8_u8(vshr_n_u8(e0, 4));
+            int8x8_t s1 = vreinterpret_s8_u8(vshr_n_u8(e1, 4));
+            int8x8_t s2 = vreinterpret_s8_u8(vshr_n_u8(e2, 4));
+            int8x8_t s3 = vreinterpret_s8_u8(vshr_n_u8(e3, 4));
+
+            // Convert 0/1 to -1/+1: sign = 2*val - 1
+            int8x8_t one = vdup_n_s8(1);
+            s0 = vsub_s8(vadd_s8(s0, s0), one);  // 2*s0 - 1
+            s1 = vsub_s8(vadd_s8(s1, s1), one);
+            s2 = vsub_s8(vadd_s8(s2, s2), one);
+            s3 = vsub_s8(vadd_s8(s3, s3), one);
+
+            // Combine into 16-element vectors
+            int8x16_t signs0 = vcombine_s8(s0, s1);
+            int8x16_t signs1 = vcombine_s8(s2, s3);
+
+            // Multiply signs with y values and accumulate
+            // dot(signs, y) where signs are +1/-1
+            int32x4_t p0 = ggml_vdotq_s32(vdupq_n_s32(0), signs0, y0);
+            int32x4_t p1 = ggml_vdotq_s32(p0, signs1, y1);
+
+            // Scale by d1 and accumulate
+            sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(p1), d0 * d1);
+        }
+    }
+
+    sumf = vaddvq_f32(sumv);
+#else
+    // Scalar fallback
+    for (int i = 0; i < nb; i++) {
+        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+
+        // Process 4 Q8_0 blocks
+        for (int k = 0; k < 4; k++) {
+            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
+
+            int sumi = 0;
+            for (int j = 0; j < QK8_0; j++) {
+                const int bit_index = k * QK8_0 + j;
+                const int byte_index = bit_index / 8;
+                const int bit_offset = bit_index % 8;
+
+                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
+                sumi += xi * y[i*4 + k].qs[j];
+            }
+            sumf += d0 * d1 * sumi;
+        }
+    }
+#endif
+
+    *s = sumf;
+}
+
+
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/arch/loongarch/quants.c

@@ -2156,4 +2156,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-

ggml/src/ggml-cpu/arch/powerpc/quants.c

@@ -2302,4 +2302,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -1463,4 +1463,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-

ggml/src/ggml-cpu/arch/wasm/quants.c

@@ -1218,4 +1218,3 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-

ggml/src/ggml-cpu/ggml-cpu.c

@@ -217,6 +217,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_F16,
         .nrows                    = 1,
     },
+    [GGML_TYPE_Q1_0] = {
+        .from_float               = quantize_row_q1_0,
+        .vec_dot                  = ggml_vec_dot_q1_0_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q4_0] = {
         .from_float               = quantize_row_q4_0,
         .vec_dot                  = ggml_vec_dot_q4_0_q8_0,

ggml/src/ggml-cpu/ops.cpp

@@ -4829,6 +4829,7 @@ void ggml_compute_forward_get_rows(
     const ggml_tensor * src0 = dst->src[0];
 
     switch (src0->type) {
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -5554,6 +5555,7 @@ void ggml_compute_forward_clamp(
                 ggml_compute_forward_clamp_f16(params, dst);
             } break;
         case GGML_TYPE_BF16:
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:

ggml/src/ggml-cpu/quants.c

@@ -22,6 +22,10 @@
 
 #define UNUSED GGML_UNUSED
 
+void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q1_0_ref(x, y, k);
+}
+
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
     quantize_row_q4_0_ref(x, y, k);
 }
@@ -116,6 +120,51 @@ void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRI
 
 //===================================== Dot products =================================
 
+void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK1_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q1_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    float sumf = 0.0;
+
+    for (int i = 0; i < nb; i++) {
+        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+
+        float sumi = 0.0f;
+
+        for (int k = 0; k < 4; k++) {
+            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
+
+            int sumi_block = 0;
+
+            for (int j = 0; j < QK8_0; j++) {
+                const int bit_index = k * QK8_0 + j;
+                const int byte_index = bit_index / 8;
+                const int bit_offset = bit_index % 8;
+
+                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
+                sumi_block += xi * y[i*4 + k].qs[j];
+            }
+
+            sumi += d1 * sumi_block;
+        }
+
+        sumf += d0 * sumi;
+    }
+
+    *s = sumf;
+}
+
+
 void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/quants.h

@@ -12,6 +12,7 @@ extern "C" {
 #endif
 
 // Quantization
+void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -36,6 +37,7 @@ void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y,
 void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 // Dot product
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -68,6 +70,7 @@ void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const
 void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
 void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);