llama_cpp 0.3.3 → 0.3.4

data/ext/llama_cpp/src/ggml.c

@@ -31,11 +31,17 @@
 #include <unistd.h>
 #endif
 
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
 #ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
 #define static_assert(cond, msg) struct global_scope_noop_trick
 #endif
+#endif
 
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
@@ -112,10 +118,6 @@ typedef void * thread_ret_t;
 #endif
 #endif
 
-#ifdef __HAIKU__
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#endif
-
 /*#define GGML_PERF*/
 #define GGML_DEBUG 0
 #define GGML_GELU_FP16
@@ -4410,8 +4412,8 @@ void ggml_free(struct ggml_context * ctx) {
         if (&g_state.contexts[i].context == ctx) {
             g_state.contexts[i].used = false;
 
-            GGML_PRINT_DEBUG("%s: context %d with %d objects has been freed. memory used = %zu\n",
-                    __func__, i, ctx->n_objects, ctx->objects_end->offs + ctx->objects_end->size);
+            GGML_PRINT_DEBUG("%s: context %d has been freed. memory used = %zu\n",
+                    __func__, i, ggml_used_mem(ctx));
 
             if (ctx->mem_buffer_owned) {
                 GGML_ALIGNED_FREE(ctx->mem_buffer);
@@ -6955,6 +6957,8 @@ struct ggml_tensor * ggml_rope_impl(
         int                   n_dims,
         int                   mode,
         int                   n_ctx,
+        float                 freq_base,
+        float                 freq_scale,
         bool                  inplace) {
     GGML_ASSERT(n_past >= 0);
     bool is_node = false;
@@ -6967,12 +6971,14 @@ struct ggml_tensor * ggml_rope_impl(
 
     ggml_scratch_save(ctx);
 
-    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 4);
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 6);
 
     ((int32_t *) b->data)[0] = n_past;
     ((int32_t *) b->data)[1] = n_dims;
     ((int32_t *) b->data)[2] = mode;
     ((int32_t *) b->data)[3] = n_ctx;
+    memcpy((int32_t *) b->data + 4, &freq_base,  sizeof(float));
+    memcpy((int32_t *) b->data + 5, &freq_scale, sizeof(float));
 
     ggml_scratch_load(ctx);
 
@@ -6991,7 +6997,7 @@ struct ggml_tensor * ggml_rope(
         int                   n_dims,
         int                   mode,
         int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, false);
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, false);
 }
 
 struct ggml_tensor * ggml_rope_inplace(
@@ -7001,7 +7007,19 @@ struct ggml_tensor * ggml_rope_inplace(
         int                   n_dims,
         int                   mode,
         int                   n_ctx) {
-    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, true);
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, true);
+}
+
+struct ggml_tensor * ggml_rope_custom_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_past,
+        int                   n_dims,
+        int                   mode,
+        int                   n_ctx,
+        float                 freq_base,
+        float                 freq_scale) {
+    return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, true);
 }
 
 // ggml_rope_back
@@ -7011,7 +7029,8 @@ struct ggml_tensor * ggml_rope_back(
         struct ggml_tensor  * a,
         int                   n_past,
         int                   n_dims,
-        int                   mode) {
+        int                   mode,
+        int                   n_ctx) {
     GGML_ASSERT(n_past >= 0);
     GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
 
@@ -7025,12 +7044,13 @@ struct ggml_tensor * ggml_rope_back(
 
     ggml_scratch_save(ctx);
 
-    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 3);
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 4);
     ggml_set_name(b, "n_past, n_dims, mode");
 
     ((int32_t *) b->data)[0] = n_past;
     ((int32_t *) b->data)[1] = n_dims;
     ((int32_t *) b->data)[2] = mode;
+    ((int32_t *) b->data)[3] = n_ctx;
 
     ggml_scratch_load(ctx);
 
@@ -10684,6 +10704,8 @@ static void ggml_compute_forward_mul_mat(
 
     const enum ggml_type type = src0->type;
 
+    const bool src1_cont = ggml_is_contiguous(src1);
+
     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;
@@ -10747,7 +10769,7 @@ static void ggml_compute_forward_mul_mat(
                 float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
 
                 if (type != GGML_TYPE_F32) {
-                    float * const wdata = params->wdata;
+                            float * const wdata    = params->wdata;
                     ggml_to_float_t const to_float = type_traits[type].to_float;
 
                     size_t id = 0;
@@ -10805,7 +10827,7 @@ static void ggml_compute_forward_mul_mat(
     // src1 rows
     const int64_t nr1 = ne11*ne12*ne13;
 
-    void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+    const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
     const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
     for (int64_t ir1 = 0; ir1 < nr1; ++ir1) {
@@ -10828,7 +10850,15 @@ static void ggml_compute_forward_mul_mat(
         const int64_t i3 = i13;
 
         const char * src0_row = (const char *) src0->data + (  0 + i02*nb02 + i03*nb03     );
-        const char * src1_col = (const char *)      wdata + (i11 + i12*ne11 + i13*ne12*ne11)*row_size;
+
+        // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+        //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+        //       the original src1 data pointer, so we should index using the indices directly
+        // TODO: this is a bit of a hack, we should probably have a better way to handle this
+        const char * src1_col = (const char *) wdata +
+            (src1_cont || src1->type != vec_dot_type
+             ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
+             : (i11*nb11 + i12*nb12 + i13*nb13));
 
         float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
 
@@ -12062,16 +12092,21 @@ static void ggml_compute_forward_rope_f32(
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(ggml_nelements(src1) == 4);
+    GGML_ASSERT(ggml_nelements(src1) == 6);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
+    float freq_base;
+    float freq_scale;
+
     const int n_past = ((int32_t *) src1->data)[0];
     const int n_dims = ((int32_t *) src1->data)[1];
     const int mode   = ((int32_t *) src1->data)[2];
     const int n_ctx  = ((int32_t *) src1->data)[3];
+    memcpy(&freq_base,  (int32_t *) src1->data + 4, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) src1->data + 5, sizeof(float));
 
     assert(n_past >= 0);
 
@@ -12100,7 +12135,7 @@ static void ggml_compute_forward_rope_f32(
     // row index used to determine which thread to use
     int ir = 0;
 
-    const float theta_scale = powf(10000.0, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
 
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
@@ -12112,7 +12147,7 @@ static void ggml_compute_forward_rope_f32(
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                float theta = (float)p;
+                float theta = freq_scale * (float)p;
 
                 if (is_glm) {
                     theta = MIN(p, n_ctx - 2);
@@ -12189,16 +12224,21 @@ static void ggml_compute_forward_rope_f16(
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(ggml_nelements(src1) == 4);
+    GGML_ASSERT(ggml_nelements(src1) == 6);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
+    float freq_base;
+    float freq_scale;
+
     const int n_past = ((int32_t *) src1->data)[0];
     const int n_dims = ((int32_t *) src1->data)[1];
     const int mode   = ((int32_t *) src1->data)[2];
     const int n_ctx  = ((int32_t *) src1->data)[3];
+    memcpy(&freq_base,  (int32_t *) src1->data + 4, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) src1->data + 5, sizeof(float));
 
     assert(n_past >= 0);
 
@@ -12227,7 +12267,7 @@ static void ggml_compute_forward_rope_f16(
     // row index used to determine which thread to use
     int ir = 0;
 
-    const float theta_scale = powf(10000.0, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
 
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
@@ -12239,7 +12279,7 @@ static void ggml_compute_forward_rope_f16(
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                float theta = (float)p;
+                float theta = freq_scale * (float)p;
 
                 if (is_glm) {
                     theta = MIN(p, n_ctx - 2);
@@ -12300,7 +12340,7 @@ static void ggml_compute_forward_rope_f16(
                             const float x0 = GGML_FP16_TO_FP32(src[0]);
                             const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
 
-                            dst_data[0]     = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
                             dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     }
@@ -12339,7 +12379,7 @@ static void ggml_compute_forward_rope_back_f32(
         const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
     assert(src1->type == GGML_TYPE_I32);
-    assert(ggml_nelements(src1) == 3);
+    assert(ggml_nelements(src1) == 4);
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
@@ -12982,12 +13022,13 @@ static void ggml_compute_forward_conv_1d(
     };
 }
 
-// ggml_compute_forward_conv_2d_sk_p0
+// ggml_compute_forward_conv_2d
 
-static void ggml_compute_forward_conv_2d_sk_p0_f16_f32(
+static void ggml_compute_forward_conv_2d_f16_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
+        const struct ggml_tensor * opt0,
               struct ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
@@ -13007,28 +13048,37 @@ static void ggml_compute_forward_conv_2d_sk_p0_f16_f32(
     // size of the convolution row - the kernel size unrolled across all channels
     const int ew0 = nk0*nk1*ne02;
 
+    const int32_t s0 = ((const int32_t*)(opt0->data))[0];
+    const int32_t s1 = ((const int32_t*)(opt0->data))[1];
+    const int32_t p0 = ((const int32_t*)(opt0->data))[2];
+    const int32_t p1 = ((const int32_t*)(opt0->data))[3];
+    const int32_t d0 = ((const int32_t*)(opt0->data))[4];
+    const int32_t d1 = ((const int32_t*)(opt0->data))[5];
+
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == GGML_TASK_INIT) {
-        // TODO: fix this memset (wsize is overestimated)
         memset(params->wdata, 0, params->wsize);
 
         // prepare source data (src1)
         {
             ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0;
 
-            for (int i13 = 0; i13 < ne13; i13++) {
-                for (int i12 = 0; i12 < ne12; i12++) {
-                    const float * const src = (float *)((char *) src1->data + i13*nb13 + i12*nb12);
-                    ggml_fp16_t * dst_data = wdata + i13*(ne1*ne0*ew0);
+            for (int i12 = 0; i12 < ne12; i12++) {
+                const float * const src = (float *)((char *) src1->data + i12*nb12);
+                ggml_fp16_t * dst_data = wdata;
+
+                for (int i1 = 0; i1 < ne1; i1++) {
+                    for (int i0 = 0; i0 < ne0; i0++) {
+                        for (int ik1 = 0; ik1 < nk1; ik1++) {
+                            for (int ik0 = 0; ik0 < nk0; ik0++) {
+                                const int idx0 = i0*s0 + ik0*d0 - p0;
+                                const int idx1 = i1*s1 + ik1*d1 - p1;
 
-                    for (int i1 = 0; i1 < ne1; i1++) {
-                        for (int i0 = 0; i0 < ne0; i0++) {
-                            for (int ik1 = 0; ik1 < nk1; ik1++) {
-                                for (int ik0 = 0; ik0 < nk0; ik0++) {
+                                if (!(idx1 < 0 || idx1 >= ne11 || idx0 < 0 || idx0 >= ne10)) {
                                     dst_data[(i1*ne0 + i0)*ew0 + i12*(nk0*nk1) + ik1*nk0 + ik0] =
-                                        GGML_FP32_TO_FP16(src[(i1*nk1 + ik1)*ne10 + (i0*nk0 + ik0)]);
+                                        GGML_FP32_TO_FP16(src[idx1*ne10 + idx0]);
                                 }
                             }
                         }
@@ -13071,19 +13121,21 @@ static void ggml_compute_forward_conv_2d_sk_p0_f16_f32(
     }
 }
 
-static void ggml_compute_forward_conv_2d_sk_p0(
+static void ggml_compute_forward_conv_2d(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
         const struct ggml_tensor * src1,
-        struct ggml_tensor * dst) {
+        const struct ggml_tensor * opt0,
+        struct ggml_tensor * dst
+        ) {
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_conv_2d_sk_p0_f16_f32(params, src0, src1, dst);
+                ggml_compute_forward_conv_2d_f16_f32(params, src0, src1, opt0, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                //ggml_compute_forward_conv_2d_sk_p0_f32(params, src0, src1, dst);
+                //ggml_compute_forward_conv_2d_f32(params, src0, src1, opt0, dst);
                 GGML_ASSERT(false);
             } break;
         default:
@@ -13093,32 +13145,6 @@ static void ggml_compute_forward_conv_2d_sk_p0(
     }
 }
 
-// ggml_compute_forward_conv_2d
-
-static void ggml_compute_forward_conv_2d(
-    const struct ggml_compute_params* params,
-    const struct ggml_tensor* src0,
-    const struct ggml_tensor* src1,
-    const struct ggml_tensor* opt0,
-    struct ggml_tensor* dst) {
-    const int32_t s0 = ((const int32_t*)(opt0->data))[0];
-    const int32_t s1 = ((const int32_t*)(opt0->data))[1];
-    const int32_t p0 = ((const int32_t*)(opt0->data))[2];
-    const int32_t p1 = ((const int32_t*)(opt0->data))[3];
-    const int32_t d0 = ((const int32_t*)(opt0->data))[4];
-    const int32_t d1 = ((const int32_t*)(opt0->data))[5];
-    GGML_ASSERT(d0 == 1); // dilation not supported
-    GGML_ASSERT(d1 == 1);
-    GGML_ASSERT(p0 == 0); // padding not supported
-    GGML_ASSERT(p1 == 0);
-
-    if (s0 == src0->ne[0] && s1 == src0->ne[1]) {
-        ggml_compute_forward_conv_2d_sk_p0(params, src0, src1, dst);
-    } else {
-        GGML_ASSERT(false); // only stride equal to kernel size is supported
-    }
-}
-
 // ggml_compute_forward_pool_1d_sk_p0
 
 static void ggml_compute_forward_pool_1d_sk_p0(
@@ -15712,17 +15738,19 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // necessary for llama
                 if (src0->grad) {
                     assert(src1->type == GGML_TYPE_I32);
-                    assert(ggml_nelements(src1) == 4);
+                    assert(ggml_nelements(src1) == 6);
                     const int n_past = ((int32_t *) src1->data)[0];
                     const int n_dims = ((int32_t *) src1->data)[1];
                     const int mode   = ((int32_t *) src1->data)[2];
+                    const int n_ctx  = ((int32_t *) src1->data)[3];
                     src0->grad = ggml_add_impl(ctx,
                             src0->grad,
                             ggml_rope_back(ctx,
                                 tensor->grad,
                                 n_past,
                                 n_dims,
-                                mode),
+                                mode,
+                                n_ctx),
                             inplace);
                 }
                 if (src1->grad) {
@@ -16293,8 +16321,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
                 if (GGML_OP_HAS_FINALIZE[node->op]) {
                     params.nth = n_tasks_arr[node_n];
                     ggml_compute_forward(&params, node);
-                    ggml_graph_compute_perf_stats_node(node, state->shared);
                 }
+                ggml_graph_compute_perf_stats_node(node, state->shared);
             }
 
             // distribute new work or execute it direct if 1T
@@ -16324,8 +16352,9 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
                     if (GGML_OP_HAS_FINALIZE[node->op]) {
                         params.type = GGML_TASK_FINALIZE;
                         ggml_compute_forward(&params, node);
-                        ggml_graph_compute_perf_stats_node(node, state->shared);
                     }
+
+                    ggml_graph_compute_perf_stats_node(node, state->shared);
                 } else {
                     break;
                 }
@@ -16575,19 +16604,22 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     const int64_t ne11 = node->src[1]->ne[1]; // H
                     const int64_t ne12 = node->src[1]->ne[2]; // C
 
+                    const int64_t ne0 = node->ne[0];
+                    const int64_t ne1 = node->ne[1];
+                    const int64_t ne2 = node->ne[2];
                     const int64_t nk = ne00*ne01;
+                    const int64_t ew0 = nk * ne02;
 
-                    UNUSED(ne02);
                     UNUSED(ne03);
-                    UNUSED(nk);
+                    UNUSED(ne2);
 
                     size_t cur = 0;
 
                     if (node->src[0]->type == GGML_TYPE_F16 &&
-                            node->src[1]->type == GGML_TYPE_F32) {
-                        cur = sizeof(ggml_fp16_t)*(ne10*ne11*ne12);
+                        node->src[1]->type == GGML_TYPE_F32) {
+                        cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0);
                     } else if (node->src[0]->type == GGML_TYPE_F32 &&
-                            node->src[1]->type == GGML_TYPE_F32) {
+                               node->src[1]->type == GGML_TYPE_F32) {
                         cur = sizeof(float)*      (ne10*ne11*ne12);
                     } else {
                         GGML_ASSERT(false);
@@ -16864,9 +16896,6 @@ static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char
 }
 
 void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
-    //assert(cgraph->work      == NULL);
-    //assert(cgraph->work_size == 0);
-
     uint64_t size_eval = 0;
 
     // compute size of intermediate results
@@ -17305,9 +17334,6 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
 
     GGML_PRINT("=== GRAPH ===\n");
 
-    GGML_PRINT_DEBUG("n_threads       = %d\n",        cgraph->n_threads);
-    GGML_PRINT_DEBUG("total work size = %zu bytes\n", cgraph->work_size);
-
     GGML_PRINT("n_nodes = %d\n", cgraph->n_nodes);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];

data/ext/llama_cpp/src/ggml.h

@@ -1121,6 +1121,17 @@ extern "C" {
             int                   mode,
             int                   n_ctx);
 
+    // custom RoPE, in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            float                 freq_base,
+            float                 freq_scale);
+
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
     GGML_API struct ggml_tensor * ggml_rope_back(
@@ -1128,7 +1139,8 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_dims,
-            int                   mode);
+            int                   mode,
+            int                   n_ctx);
 
     // alibi position embedding
     // in-place, returns view(a)

data/ext/llama_cpp/src/k_quants.h

@@ -15,6 +15,14 @@
 #define K_SCALE_SIZE 12
 #endif
 
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+
 //
 // Super-block quantization structures
 //