llama_cpp 0.9.3 → 0.9.5

data/ext/llama_cpp/src/ggml.c

@@ -4826,7 +4826,17 @@ struct ggml_tensor * ggml_diag_mask_zero_inplace(
 static struct ggml_tensor * ggml_soft_max_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
+        struct ggml_tensor  * mask,
+        float                 scale,
         bool                  inplace) {
+    GGML_ASSERT(ggml_is_contiguous(a));
+    if (mask) {
+        GGML_ASSERT(ggml_is_contiguous(mask));
+        GGML_ASSERT(mask->ne[2] == 1);
+        GGML_ASSERT(mask->ne[3] == 1);
+        GGML_ASSERT(ggml_can_repeat_rows(mask, a));
+    }
+
     bool is_node = false;
 
     if (a->grad) {
@@ -4835,9 +4845,13 @@ static struct ggml_tensor * ggml_soft_max_impl(
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
+    float params[] = { scale };
+    ggml_set_op_params(result, params, sizeof(params));
+
     result->op   = GGML_OP_SOFT_MAX;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
+    result->src[1] = mask;
 
     return result;
 }
@@ -4845,13 +4859,21 @@ static struct ggml_tensor * ggml_soft_max_impl(
 struct ggml_tensor * ggml_soft_max(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, false);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, false);
 }
 
 struct ggml_tensor * ggml_soft_max_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, true);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, true);
+}
+
+struct ggml_tensor * ggml_soft_max_ext(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * mask,
+        float                 scale) {
+    return ggml_soft_max_impl(ctx, a, mask, scale, false);
 }
 
 // ggml_soft_max_back
@@ -9373,7 +9395,7 @@ static bool ggml_compute_forward_mul_mat_use_blas(
     // TODO: find the optimal values for these
     if (ggml_is_contiguous(src0) &&
         ggml_is_contiguous(src1) &&
-        src0->type == GGML_TYPE_F32 &&
+      //src0->type == GGML_TYPE_F32 &&
         src1->type == GGML_TYPE_F32 &&
         (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
 
@@ -9611,10 +9633,12 @@ static void ggml_compute_forward_out_prod_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
+    GGML_ASSERT(ne0  == ne00);
+    GGML_ASSERT(ne1  == ne10);
+    GGML_ASSERT(ne2  == ne02);
     GGML_ASSERT(ne02 == ne12);
-    GGML_ASSERT(ne03 == ne13);
-    GGML_ASSERT(ne2  == ne12);
     GGML_ASSERT(ne3  == ne13);
+    GGML_ASSERT(ne03 == ne13);
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == sizeof(float));
@@ -9625,18 +9649,25 @@ static void ggml_compute_forward_out_prod_f32(
     // GGML_ASSERT(nb1 <= nb2);
     // GGML_ASSERT(nb2 <= nb3);
 
-    GGML_ASSERT(ne0 == ne00);
-    GGML_ASSERT(ne1 == ne10);
-    GGML_ASSERT(ne2 == ne02);
-    GGML_ASSERT(ne3 == ne03);
-
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
     // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
-    // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
+    // TODO: #if defined(GGML_USE_CLBLAST)
+
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
+    bool use_blas = ggml_is_matrix(src0) &&
+        ggml_is_matrix(src1) &&
+        ggml_is_contiguous(src0) &&
+        (ggml_is_contiguous(src1) || ggml_is_transposed(src1));
+#endif
 
     if (params->type == GGML_TASK_INIT) {
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) // gemm beta will zero dst
+        if (use_blas) {
+            return;
+        }
+#endif
         ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
         return;
     }
@@ -9645,6 +9676,50 @@ static void ggml_compute_forward_out_prod_f32(
         return;
     }
 
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
+    if (use_blas) {
+        if (params->ith != 0) { // All threads other than the first do no work.
+            return;
+        }
+        // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
+        // src0: (k,n)
+        // src1: (k,m)
+        // dst:  (m,n)
+        //
+        // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+        // Also expressed as (major,minor)
+        // a: (m,k): so src1 transposed
+        // b: (k,n): so src0
+        // c: (m,n)
+        //
+        // However, if ggml_is_transposed(src1) is true, then
+        // src1->data already contains a transposed version, so sgemm mustn't
+        // transpose it further.
+
+        int n = src0->ne[0];
+        int k = src0->ne[1];
+        int m = src1->ne[0];
+
+        int transposeA, lda;
+
+        if (!ggml_is_transposed(src1)) {
+            transposeA = CblasTrans;
+            lda = m;
+        } else {
+            transposeA = CblasNoTrans;
+            lda = k;
+        }
+
+        float * a = (float *) ((char *) src1->data);
+        float * b = (float *) ((char *) src0->data);
+        float * c = (float *) ((char *) dst->data);
+
+        cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+        return;
+    }
+#endif
+
     // dst[:,:,:,:] = 0
     // for i2,i3:
     //   for i1:
@@ -10498,20 +10573,25 @@ static void ggml_compute_forward_diag_mask_zero(
 static void ggml_compute_forward_soft_max_f32(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
+    assert(ggml_is_contiguous(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
         return;
     }
 
+    float scale = 1.0f;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+
     // TODO: handle transposed/permuted matrices
 
     const int ith = params->ith;
     const int nth = params->nth;
 
+    const int64_t ne11 = src1 ? src1->ne[1] : 1;
+
     const int nc = src0->ne[0];
     const int nr = ggml_nrows(src0);
 
@@ -10522,29 +10602,40 @@ static void ggml_compute_forward_soft_max_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
+    float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
+
     for (int i1 = ir0; i1 < ir1; i1++) {
-        float *sp = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float *dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+        float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
+        float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+
+        // broadcast the mask across rows
+        float * mp = src1 ? (float *)((char *) src1->data + (i1%ne11)*src1->nb[1]) : NULL;
+
+        ggml_vec_cpy_f32  (nc, wp, sp);
+        ggml_vec_scale_f32(nc, wp, scale);
+        if (mp) {
+            ggml_vec_acc_f32(nc, wp, mp);
+        }
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
             //printf("p[%d] = %f\n", i, p[i]);
-            assert(!isnan(sp[i]));
+            assert(!isnan(wp[i]));
         }
 #endif
 
         float max = -INFINITY;
-        ggml_vec_max_f32(nc, &max, sp);
+        ggml_vec_max_f32(nc, &max, wp);
 
         ggml_float sum = 0.0;
 
         uint16_t scvt;
         for (int i = 0; i < nc; i++) {
-            if (sp[i] == -INFINITY) {
+            if (wp[i] == -INFINITY) {
                 dp[i] = 0.0f;
             } else {
-                // const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max);
-                ggml_fp16_t s = GGML_FP32_TO_FP16(sp[i] - max);
+                // const float val = (wp[i] == -INFINITY) ? 0.0 : exp(wp[i] - max);
+                ggml_fp16_t s = GGML_FP32_TO_FP16(wp[i] - max);
                 memcpy(&scvt, &s, sizeof(scvt));
                 const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
                 sum += (ggml_float)val;
@@ -10569,11 +10660,12 @@ static void ggml_compute_forward_soft_max_f32(
 static void ggml_compute_forward_soft_max(
         const struct ggml_compute_params * params,
         const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
+        const struct ggml_tensor * src1,
+              struct ggml_tensor * dst) {
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_soft_max_f32(params, src0, dst);
+                ggml_compute_forward_soft_max_f32(params, src0, src1, dst);
             } break;
         default:
             {
@@ -13810,7 +13902,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                ggml_compute_forward_soft_max(params, tensor->src[0], tensor);
+                ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
@@ -15636,13 +15728,14 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = 1;
             } break;
-        case GGML_OP_COUNT:
-            {
-                GGML_ASSERT(false);
-            } break;
         default:
             {
-                printf("%s: op %s not implemented\n", __func__, ggml_op_name(node->op));
+                fprintf(stderr, "%s: op not implemented: ", __func__);
+                if (node->op < GGML_OP_COUNT) {
+                    fprintf(stderr, "%s\n", ggml_op_name(node->op));
+                } else {
+                    fprintf(stderr, "%d\n", node->op);
+                }
                 GGML_ASSERT(false);
             } break;
     }
@@ -15845,6 +15938,12 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
+            case GGML_OP_SOFT_MAX:
+                {
+                    n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
+
+                    cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
+                } break;
             case GGML_OP_CONV_TRANSPOSE_1D:
                 {
                     GGML_ASSERT(node->src[0]->ne[3] == 1);
@@ -18399,24 +18498,29 @@ int gguf_find_key(const struct gguf_context * ctx, const char * key) {
 }
 
 const char * gguf_get_key(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     return ctx->kv[key_id].key.data;
 }
 
 enum gguf_type gguf_get_kv_type(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     return ctx->kv[key_id].type;
 }
 
 enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
     return ctx->kv[key_id].value.arr.type;
 }
 
 const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
     return ctx->kv[key_id].value.arr.data;
 }
 
 const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
     struct gguf_kv * kv = &ctx->kv[key_id];
     struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[i];
@@ -18424,70 +18528,90 @@ const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i
 }
 
 int gguf_get_arr_n(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY);
     return ctx->kv[key_id].value.arr.n;
 }
 
 uint8_t gguf_get_val_u8(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT8);
     return ctx->kv[key_id].value.uint8;
 }
 
 int8_t gguf_get_val_i8(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT8);
     return ctx->kv[key_id].value.int8;
 }
 
 uint16_t gguf_get_val_u16(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT16);
     return ctx->kv[key_id].value.uint16;
 }
 
 int16_t gguf_get_val_i16(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT16);
     return ctx->kv[key_id].value.int16;
 }
 
 uint32_t gguf_get_val_u32(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT32);
     return ctx->kv[key_id].value.uint32;
 }
 
 int32_t gguf_get_val_i32(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT32);
     return ctx->kv[key_id].value.int32;
 }
 
 float gguf_get_val_f32(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT32);
     return ctx->kv[key_id].value.float32;
 }
 
 uint64_t gguf_get_val_u64(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT64);
     return ctx->kv[key_id].value.uint64;
 }
 
 int64_t gguf_get_val_i64(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT64);
     return ctx->kv[key_id].value.int64;
 }
 
 double gguf_get_val_f64(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT64);
     return ctx->kv[key_id].value.float64;
 }
 
 bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_BOOL);
     return ctx->kv[key_id].value.bool_;
 }
 
 const char * gguf_get_val_str(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
     GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_STRING);
     return ctx->kv[key_id].value.str.data;
 }
 
+const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id) {
+    GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx));
+    GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_ARRAY);
+    GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_STRING);
+    return &ctx->kv[key_id].value;
+}
+
 int gguf_get_n_tensors(const struct gguf_context * ctx) {
     return ctx->header.n_tensors;
 }

data/ext/llama_cpp/src/ggml.h

@@ -244,11 +244,10 @@
 #define GGML_ASSERT(x) \
     do { \
         if (!(x)) { \
-            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            fflush(stderr); \
             fflush(stdout); \
+            fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
             ggml_print_backtrace(); \
-            exit(1); \
+            abort(); \
         } \
     } while (0)
 
@@ -1283,6 +1282,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // fused soft_max(a*scale + mask)
+    // mask is optional
+    GGML_API struct ggml_tensor * ggml_soft_max_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * mask,
+            float                 scale);
+
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2045,6 +2052,7 @@ extern "C" {
     GGML_API double       gguf_get_val_f64 (const struct gguf_context * ctx, int key_id);
     GGML_API bool         gguf_get_val_bool(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id);
+    GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id);
     GGML_API int          gguf_get_arr_n   (const struct gguf_context * ctx, int key_id);
     GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id);
     GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i);