cui-llama.rn 1.6.1 → 1.7.0

package/cpp/ggml-backend.cpp

@@ -56,7 +56,7 @@ size_t lm_ggml_backend_buft_get_max_size(lm_ggml_backend_buffer_type_t buft) {
     return SIZE_MAX;
 }
 
-size_t lm_ggml_backend_buft_get_alloc_size(lm_ggml_backend_buffer_type_t buft, struct lm_ggml_tensor * tensor) {
+size_t lm_ggml_backend_buft_get_alloc_size(lm_ggml_backend_buffer_type_t buft, const struct lm_ggml_tensor * tensor) {
     // get_alloc_size is optional, defaults to lm_ggml_nbytes
     if (buft->iface.get_alloc_size) {
         size_t size = buft->iface.get_alloc_size(buft, tensor);
@@ -152,7 +152,7 @@ size_t lm_ggml_backend_buffer_get_max_size(lm_ggml_backend_buffer_t buffer) {
     return lm_ggml_backend_buft_get_max_size(lm_ggml_backend_buffer_get_type(buffer));
 }
 
-size_t lm_ggml_backend_buffer_get_alloc_size(lm_ggml_backend_buffer_t buffer, struct lm_ggml_tensor * tensor) {
+size_t lm_ggml_backend_buffer_get_alloc_size(lm_ggml_backend_buffer_t buffer, const struct lm_ggml_tensor * tensor) {
     return lm_ggml_backend_buft_get_alloc_size(lm_ggml_backend_buffer_get_type(buffer), tensor);
 }
 
@@ -674,6 +674,8 @@ struct lm_ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
+    bool op_offload;
+
     int debug;
 };
 
@@ -766,7 +768,7 @@ static int lm_ggml_backend_sched_backend_id_from_cur(lm_ggml_backend_sched_t sch
         if (tensor->op != LM_GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == LM_GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = lm_ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1 && lm_ggml_backend_buffer_is_host(src->buffer)) {
+            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && lm_ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (lm_ggml_backend_supports_op(sched->backends[b], tensor) && lm_ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
@@ -1109,7 +1111,7 @@ static void lm_ggml_backend_sched_split_graph(lm_ggml_backend_sched_t sched, str
 
             const int node_backend_id = tensor_backend_id(node);
 
-            assert(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
@@ -1452,7 +1454,8 @@ lm_ggml_backend_sched_t lm_ggml_backend_sched_new(
         lm_ggml_backend_buffer_type_t * bufts,
         int n_backends,
         size_t graph_size,
-        bool parallel) {
+        bool parallel,
+        bool op_offload) {
     LM_GGML_ASSERT(n_backends > 0);
     LM_GGML_ASSERT(n_backends <= LM_GGML_SCHED_MAX_BACKENDS);
     LM_GGML_ASSERT(lm_ggml_backend_dev_type(lm_ggml_backend_get_device(backends[n_backends - 1])) == LM_GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1497,6 +1500,7 @@ lm_ggml_backend_sched_t lm_ggml_backend_sched_new(
     }
 
     sched->galloc = lm_ggml_gallocr_new_n(sched->bufts, n_backends);
+    sched->op_offload = op_offload;
 
     lm_ggml_backend_sched_reset(sched);
 

package/cpp/ggml-backend.h

@@ -38,7 +38,7 @@ extern "C" {
     LM_GGML_API lm_ggml_backend_buffer_t lm_ggml_backend_buft_alloc_buffer  (lm_ggml_backend_buffer_type_t buft, size_t size);
     LM_GGML_API size_t                lm_ggml_backend_buft_get_alignment (lm_ggml_backend_buffer_type_t buft);
     LM_GGML_API size_t                lm_ggml_backend_buft_get_max_size  (lm_ggml_backend_buffer_type_t buft);
-    LM_GGML_API size_t                lm_ggml_backend_buft_get_alloc_size(lm_ggml_backend_buffer_type_t buft, struct lm_ggml_tensor * tensor);
+    LM_GGML_API size_t                lm_ggml_backend_buft_get_alloc_size(lm_ggml_backend_buffer_type_t buft, const struct lm_ggml_tensor * tensor);
     LM_GGML_API bool                  lm_ggml_backend_buft_is_host       (lm_ggml_backend_buffer_type_t buft);
     LM_GGML_API lm_ggml_backend_dev_t    lm_ggml_backend_buft_get_device    (lm_ggml_backend_buffer_type_t buft);
 
@@ -59,7 +59,7 @@ extern "C" {
     LM_GGML_API enum lm_ggml_status               lm_ggml_backend_buffer_init_tensor   (lm_ggml_backend_buffer_t buffer, struct lm_ggml_tensor * tensor);
     LM_GGML_API size_t                         lm_ggml_backend_buffer_get_alignment (lm_ggml_backend_buffer_t buffer);
     LM_GGML_API size_t                         lm_ggml_backend_buffer_get_max_size  (lm_ggml_backend_buffer_t buffer);
-    LM_GGML_API size_t                         lm_ggml_backend_buffer_get_alloc_size(lm_ggml_backend_buffer_t buffer, struct lm_ggml_tensor * tensor);
+    LM_GGML_API size_t                         lm_ggml_backend_buffer_get_alloc_size(lm_ggml_backend_buffer_t buffer, const struct lm_ggml_tensor * tensor);
     LM_GGML_API void                           lm_ggml_backend_buffer_clear         (lm_ggml_backend_buffer_t buffer, uint8_t value);
     LM_GGML_API bool                           lm_ggml_backend_buffer_is_host       (lm_ggml_backend_buffer_t buffer);
     LM_GGML_API void                           lm_ggml_backend_buffer_set_usage     (lm_ggml_backend_buffer_t buffer, enum lm_ggml_backend_buffer_usage usage);
@@ -248,7 +248,7 @@ extern "C" {
         // preferrably to run on the same backend as the buffer
         lm_ggml_backend_buffer_set_usage(buf_weights, LM_GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        sched = lm_ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, LM_GGML_DEFAULT_GRAPH_SIZE, false);
+        sched = lm_ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, LM_GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         // initialize buffers from a max size graph (optional)
         reserve_graph = build_graph(sched, max_batch_size);
@@ -289,7 +289,7 @@ extern "C" {
     typedef bool (*lm_ggml_backend_sched_eval_callback)(struct lm_ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
-    LM_GGML_API lm_ggml_backend_sched_t lm_ggml_backend_sched_new(lm_ggml_backend_t * backends, lm_ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    LM_GGML_API lm_ggml_backend_sched_t lm_ggml_backend_sched_new(lm_ggml_backend_t * backends, lm_ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
     LM_GGML_API void                 lm_ggml_backend_sched_free(lm_ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph

package/cpp/ggml-cpu/ggml-cpu-aarch64.cpp

@@ -72,8 +72,6 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(lm_ggml_half) + QK4_NL * 2, "
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
-#elif defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
 #define UNUSED LM_GGML_UNUSED

package/cpp/ggml-cpu/ggml-cpu-quants.c

@@ -20,12 +20,6 @@
 #define GROUP_MAX_EPS_IQ1_M 1e-7f
 #define GROUP_MAX_EPS_IQ1_S 1e-12f
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid warnings for hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-#endif
-
 #define UNUSED LM_GGML_UNUSED
 
 // some compilers don't provide _mm256_set_m128i, e.g. gcc 7
@@ -6596,7 +6590,118 @@ void lm_ggml_vec_dot_q3_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, con
     }
 
     *s = hsum_float_8(acc);
+#elif defined(__VXE__) || defined(__VXE2__)
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    const int32x4_t v_z = vec_splat_s32(0);
+    const uint8x16_t v_3m = vec_splat_u8(0x03);
+
+    const uint8x16_t v_0c = vec_splat_u8(1);
+    const uint8x16_t v_1c = vec_sl(v_0c, 1);
+    const uint8x16_t v_2c = vec_sl(v_0c, 2);
+    const uint8x16_t v_3c = vec_sl(v_0c, 3);
+
+    uint8x16_t q3h[4];
+    uint8x16_t q3b[2];
+    int8x16_t q3bytes[4];
+    int8x16_t q8bytes[4];
+    uint8x16_t qhbits[2];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * LM_GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict x0l = x[i].qs;
+        const uint8_t * restrict x0h = x[i].hmask;
+        const int8_t  * restrict y0  = y[i].qs;
+
+        qhbits[0] = vec_xl(0 , x0h);
+        qhbits[1] = vec_xl(16, x0h);
+
+        int32_t isum = 0;
+
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            int32x4_t isum0, isum1, isum2, isum3;
+
+            q3b[0] = vec_xl(0 , x0l);
+            q3b[1] = vec_xl(16, x0l);
+            x0l += 32;
+
+            q8bytes[0] = vec_xl(0  , y0);
+            q8bytes[1] = vec_xl(16 , y0);
+            q8bytes[2] = vec_xl(32 , y0);
+            q8bytes[3] = vec_xl(48 , y0);
+            q8bytes[4] = vec_xl(64 , y0);
+            q8bytes[5] = vec_xl(80 , y0);
+            q8bytes[6] = vec_xl(96 , y0);
+            q8bytes[7] = vec_xl(112, y0);
+            y0 += 128;
+
+            q3h[0] = vec_sl(vec_andc(v_0c, qhbits[0]), 2);
+            q3h[1] = vec_sl(vec_andc(v_0c, qhbits[1]), 2);
+            q3h[2] = vec_sl(vec_andc(v_1c, qhbits[0]), 1);
+            q3h[3] = vec_sl(vec_andc(v_1c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(q3b[0], v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(q3b[1], v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 2), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 2), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = lm_ggml_vec_dot(v_z, q3bytes[0], q8bytes[0]);
+            isum1 = lm_ggml_vec_dot(v_z, q3bytes[1], q8bytes[1]);
+            isum2 = lm_ggml_vec_dot(v_z, q3bytes[2], q8bytes[2]);
+            isum3 = lm_ggml_vec_dot(v_z, q3bytes[3], q8bytes[3]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            q3h[0] = vec_andc(v_2c, qhbits[0]);
+            q3h[1] = vec_andc(v_2c, qhbits[1]);
+            q3h[2] = vec_sr(vec_andc(v_3c, qhbits[0]), 1);
+            q3h[3] = vec_sr(vec_andc(v_3c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 4), v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 4), v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 6), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 6), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = lm_ggml_vec_dot(v_z, q3bytes[0], q8bytes[4]);
+            isum1 = lm_ggml_vec_dot(v_z, q3bytes[1], q8bytes[5]);
+            isum2 = lm_ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
+            isum3 = lm_ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits[0] = vec_sr(qhbits[0], 4);
+                qhbits[1] = vec_sr(qhbits[1], 4);
+            }
+        }
+
+        sum += d * isum;
+    }
 
+    *s = sum;
 #else
     // scalar version
     // This function is written like this so the compiler can manage to vectorize most of it
@@ -8414,7 +8519,11 @@ void lm_ggml_vec_dot_q5_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, con
 
 void lm_ggml_vec_dot_q6_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, const void * LM_GGML_RESTRICT vx, size_t bx, const void * LM_GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+#ifdef __ARM_FEATURE_MATMUL_INT8
+    assert((nrc == 2) || (nrc == 1));
+#else
     assert(nrc == 1);
+#endif
     UNUSED(nrc);
     UNUSED(bx);
     UNUSED(by);
@@ -8425,6 +8534,197 @@ void lm_ggml_vec_dot_q6_K_q8_K(int n, float * LM_GGML_RESTRICT s, size_t bs, con
 
     const int nb = n / QK_K;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q6_K * LM_GGML_RESTRICT x0 = x;
+        const block_q6_K * LM_GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
+        const block_q8_K * LM_GGML_RESTRICT y0 = y;
+        const block_q8_K * LM_GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
+
+        float32x4_t vfsum = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
+            const uint8_t * LM_GGML_RESTRICT ql0 = x0->ql;
+            const uint8_t * LM_GGML_RESTRICT ql1 = x1->ql;
+            const uint8_t * LM_GGML_RESTRICT qh0 = x0->qh;
+            const uint8_t * LM_GGML_RESTRICT qh1 = x1->qh;
+            const  int8_t * LM_GGML_RESTRICT qy0 = y0->qs;
+            const  int8_t * LM_GGML_RESTRICT qy1 = y1->qs;
+
+            const uint8x16_t mone = vdupq_n_u8(0x30);
+            const uint8x16_t  m4b = vdupq_n_u8(0x0f);
+
+            int32x4_t visum = vdupq_n_s32(0);
+
+            // process 8 blocks per iteration, totally 16 blocks
+            for (int j = 0; j < 2; ++j, qh0 += 32, ql0 += 64, qh1 += 32, ql1 += 64) {
+                int8x16_t vx0[8], vx1[8];
+
+                // de-quantize vx0[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh0);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql0);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx0[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx0[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx0[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx0[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx0[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx0[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // de-quantize vx1[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh1);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql1);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx1[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx1[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx1[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx1[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx1[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx1[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // process 16 elements (one block with same scale) per iteration
+                // - vx = concat(ql, qh) - 32
+                // - r1,r2,r3,r4 = smmla(vx, vy)
+                for (int k = 0; k < 8; ++k) {
+                    const int blk = j * 8 + k;
+
+                    const int8x16_t vy0 = vld1q_s8(qy0);
+                    const int8x16_t vy1 = vld1q_s8(qy1);
+                    qy0 += 16;
+                    qy1 += 16;
+
+                    const int32x4_t block_scale = {
+                        x0->scales[blk],
+                        x0->scales[blk],
+                        x1->scales[blk],
+                        x1->scales[blk],
+                    };
+
+                    // calculate four results at once with outer product
+                    const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    int32x4_t vr = vdupq_n_s32(0);
+                    vr = vmmlaq_s32(vr, vx_l, vy_l);
+                    vr = vmmlaq_s32(vr, vx_h, vy_h);
+
+                    // apply block scale, will NOT overflow
+                    // block_scale * sum_256(int6*int8) <= 2^(8+8+6+8) = 30 bits
+                    visum = vmlaq_s32(visum, vr, block_scale);
+                }
+            }
+
+            // adjust bias, apply superblock scale
+            {
+                int32_t bias[4];
+#ifdef __ARM_FEATURE_SVE
+                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
+                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
+                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
+                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
+                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
+                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
+                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
+                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
+                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
+                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
+                const svint64_t zero = svdup_n_s64(0);
+                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
+                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
+                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
+                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
+#else
+                // NEON doesn't support int16 dot product, fallback to separated mul and add
+                const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
+                const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
+
+                int8x16_t scales_s8 = vld1q_s8(x0->scales);
+                const int16x8x2_t q6scales0 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+                scales_s8 = vld1q_s8(x1->scales);
+                const int16x8x2_t q6scales1 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+
+                int32x4_t prod;
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[0] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[1] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[2] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[3] = vaddvq_s32(prod);
+
+#endif
+                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
+
+                const float32x4_t superblock_scale = {
+                    LM_GGML_FP16_TO_FP32(x0->d) * y0->d,
+                    LM_GGML_FP16_TO_FP32(x0->d) * y1->d,
+                    LM_GGML_FP16_TO_FP32(x1->d) * y0->d,
+                    LM_GGML_FP16_TO_FP32(x1->d) * y1->d,
+                };
+
+                visum = vsubq_s32(visum, vibias);
+                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
+            }
+        }
+
+        // vfsum = ABCD -> ACBD
+        // AC -> s, BD -> (s+bs)
+        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
+        vst1_f32(s,      vget_low_f32 (vfsum));
+        vst1_f32(s + bs, vget_high_f32(vfsum));
+
+        return;
+    }
+#endif
+
 #ifdef __ARM_FEATURE_SVE
     const int vector_length = lm_ggml_cpu_get_sve_cnt()*8;
     float sum = 0;

package/cpp/ggml-cpu/ggml-cpu.c

@@ -50,19 +50,6 @@
 #include "llamafile/sgemm.h"
 #endif
 
-#if defined(_MSC_VER)
-// disable "possible loss of data" to avoid hundreds of casts
-// we should just be careful :)
-#pragma warning(disable: 4244 4267)
-
-// disable POSIX deprecation warnings
-// these functions are never going away, anyway
-#pragma warning(disable: 4996)
-
-// unreachable code because of multiple instances of code after LM_GGML_ABORT
-#pragma warning(disable: 4702)
-#endif
-
 // Note: once we move threading into a separate C++ file
 // will use std::hardware_destructive_interference_size instead of hardcoding it here
 // and we'll use C++ attribute syntax.
@@ -295,7 +282,11 @@ static const struct lm_ggml_type_traits_cpu type_traits_cpu[LM_GGML_TYPE_COUNT]
         .from_float               = quantize_row_q6_K,
         .vec_dot                  = lm_ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = LM_GGML_TYPE_Q8_K,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
         .nrows                    = 1,
+#endif
     },
     [LM_GGML_TYPE_IQ2_XXS] = {
         .from_float               = NULL,
@@ -2211,6 +2202,7 @@ static int lm_ggml_get_n_tasks(struct lm_ggml_tensor * node, int n_threads) {
                     } break;
 
                 case LM_GGML_UNARY_OP_GELU:
+                case LM_GGML_UNARY_OP_GELU_ERF:
                 case LM_GGML_UNARY_OP_GELU_QUICK:
                 case LM_GGML_UNARY_OP_SILU:
                     {

package/cpp/ggml-cpu/ggml-cpu.cpp

@@ -11,24 +11,26 @@
 #include <vector>
 
 #ifdef LM_GGML_USE_CPU_HBM
-#include "ggml-cpu-hbm.h"
+#    include "ggml-cpu-hbm.h"
 #endif
 
 #ifdef LM_GGML_USE_CPU_KLEIDIAI
-#include "kleidiai/kleidiai.h"
-#endif
-
-#if defined(__APPLE__)
-#include <sys/types.h>
-#include <sys/sysctl.h>
+#    include "kleidiai/kleidiai.h"
 #endif
 
 #if defined(_WIN32)
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-    #define NOMINMAX
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#else
+#    include <unistd.h>
 #endif
-#include <windows.h>
+
+#if defined(__APPLE__)
+#    include <sys/sysctl.h>
+#    include <sys/types.h>
 #endif
 
 // ggml-backend interface
@@ -70,8 +72,10 @@ static lm_ggml_backend_buffer_type_t * lm_ggml_backend_cpu_device_get_extra_buff
 }
 
 static bool lm_ggml_backend_cpu_is_extra_buffer_type(lm_ggml_backend_buffer_type_t buft) {
-    for (auto extra : lm_ggml_backend_cpu_get_extra_buffers_type()) {
-        if (extra && extra == buft) return true;
+    for (auto * extra : lm_ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra == buft) {
+            return true;
+        }
     }
     return false;
 }
@@ -330,9 +334,18 @@ static const char * lm_ggml_backend_cpu_device_get_description(lm_ggml_backend_d
 }
 
 static void lm_ggml_backend_cpu_device_get_memory(lm_ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
+#ifdef _WIN32
+    MEMORYSTATUSEX status;
+    status.dwLength = sizeof(status);
+    GlobalMemoryStatusEx(&status);
+    *total = status.ullTotalPhys;
+    *free = status.ullAvailPhys;
+#else
+    long pages = sysconf(_SC_PHYS_PAGES);
+    long page_size = sysconf(_SC_PAGE_SIZE);
+    *total = pages * page_size;
+    *free = *total;
+#endif
 
     LM_GGML_UNUSED(dev);
 }