tests: add gradient tests for all backends (ggml/932)

* tests: add gradient checking to test-backend-ops

* remove old comment

* reorder includes

* adjust SIN/COS parameters

* add documentation, use supports_op if possible

ggml/include/ggml.h

@@ -1272,7 +1272,7 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     GGML_API struct ggml_tensor * ggml_set_inplace(
@@ -1282,19 +1282,19 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     GGML_API struct ggml_tensor * ggml_set_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     GGML_API struct ggml_tensor * ggml_set_1d_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set_2d(
@@ -1302,7 +1302,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     GGML_API struct ggml_tensor * ggml_set_2d_inplace(
@@ -1310,7 +1310,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(

ggml/src/ggml-backend.c

@@ -827,6 +827,10 @@ GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const
                 op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         default:
             return true;
     }

ggml/src/ggml-cuda.cu

@@ -27,6 +27,7 @@
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/scale.cuh"
 #include "ggml-cuda/softmax.cuh"
+#include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
@@ -2180,6 +2181,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             ggml_cuda_dup(ctx, dst);
             break;
         case GGML_OP_ADD:
+        case GGML_OP_ADD1: // TODO: more efficient implementation
             ggml_cuda_op_add(ctx, dst);
             break;
         case GGML_OP_SUB:
@@ -2196,6 +2198,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             break;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_NEG:
+                    ggml_cuda_op_neg(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_GELU:
                     ggml_cuda_op_gelu(ctx, dst);
                     break;
@@ -2304,6 +2309,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
+        case GGML_OP_SUM:
+            ggml_cuda_op_sum(ctx, dst);
+            break;
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
@@ -2748,6 +2756,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_NEG:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
@@ -2877,6 +2886,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_TRANSPOSE:
         case GGML_OP_NORM:
         case GGML_OP_ADD:
+        case GGML_OP_ADD1:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -2896,7 +2906,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_ROPE:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_IM2COL:
+            return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_2D:
+        case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:

ggml/src/ggml-cuda/cross-entropy-loss.cu

@@ -1,6 +1,6 @@
 #include "common.cuh"
 #include "cross-entropy-loss.cuh"
-#include "sumrows.cuh"
+#include "sum.cuh"
 
 #include <cmath>
 #include <cstdint>
@@ -102,5 +102,5 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     cross_entropy_loss_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
 
     // Combine results from individual blocks:
-    sum_rows_f32_cuda(dst_tmp.ptr, dst_d, blocks_num.x, 1, stream);
+    sum_f32_cuda(pool, dst_tmp.ptr, dst_d, blocks_num.x, stream);
 }

ggml/src/ggml-cuda/sum.cu

@@ -0,0 +1,41 @@
+#include "sumrows.cuh"
+#include "sum.cuh"
+
+#include <cstdint>
+
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+#include <cub/cub.cuh>
+using namespace cub;
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+
+void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+    size_t tmp_size = 0;
+    DeviceReduce::Sum(nullptr,       tmp_size, x, dst, ne, stream);
+    ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
+    DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, x, dst, ne, stream);
+#else
+    // Use (inefficient) sum_rows implementation as a fallback.
+    // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
+    sum_rows_f32_cuda(x, dst, ne, 1, stream);
+    GGML_UNUSED(pool);
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+}
+
+void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const float * src0_d = (const float *) src0->data;
+    float * dst_d = (float *) dst->data;
+
+    const int64_t ne = ggml_nelements(src0);
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t stream = ctx.stream();
+
+    sum_f32_cuda(pool, src0_d, dst_d, ne, stream);
+}

ggml/src/ggml-cuda/sum.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream);
+
+void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/unary.cu

@@ -1,5 +1,15 @@
 #include "unary.cuh"
 
+static __global__ void neg_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = -x[i];
+}
+
 static __global__ void gelu_f32(const float * x, float * dst, const int k) {
     const float GELU_COEF_A    = 0.044715f;
     const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
@@ -119,6 +129,11 @@ static __global__ void cos_f32(const float * x, float * dst, const int k) {
     dst[i] = cosf(x[i]);
 }
 
+static void neg_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
+    neg_f32<<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
     gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -184,6 +199,20 @@ static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t
     cos_f32<<<num_blocks, CUDA_COS_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    neg_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;

ggml/src/ggml-cuda/unary.cuh

@@ -1,5 +1,6 @@
 #include "common.cuh"
 
+#define CUDA_NEG_BLOCK_SIZE 256
 #define CUDA_GELU_BLOCK_SIZE 256
 #define CUDA_SILU_BLOCK_SIZE 256
 #define CUDA_TANH_BLOCK_SIZE 256
@@ -12,6 +13,8 @@
 #define CUDA_SIN_BLOCK_SIZE 256
 #define CUDA_COS_BLOCK_SIZE 256
 
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml.c

@@ -5267,6 +5267,7 @@ struct ggml_tensor * ggml_concat(
     bool is_node = false;
 
     if (a->grad || b->grad) {
+        GGML_ABORT("fatal error"); // TODO: implement
         is_node = true;
     }
 
@@ -5388,6 +5389,7 @@ struct ggml_tensor * ggml_leaky_relu(
     bool is_node = false;
 
     if (!inplace && (a->grad)) {
+        GGML_ABORT("fatal error"); // TODO: not implemented
         is_node = true;
     }
 
@@ -5826,6 +5828,7 @@ static struct ggml_tensor * ggml_set_impl(
     // make a view of the destination
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
+    GGML_ASSERT(offset < (size_t)(1 << 30));
     int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 };
     ggml_set_op_params(result, params, sizeof(params));
 
@@ -6783,14 +6786,12 @@ struct ggml_tensor * ggml_rope_back(
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
     GGML_ASSERT(a->ne[2] == b->ne[0]);
-    GGML_ASSERT(c == NULL && "freq factors not implemented yet");
-
-    GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
 
     bool is_node = false;
 
     if (a->grad) {
-        is_node = false; // TODO: implement backward
+        GGML_ASSERT(false && "backwards pass not implemented");
+        is_node = false;
     }
 
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
@@ -6808,6 +6809,7 @@ struct ggml_tensor * ggml_rope_back(
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
+    result->src[2] = c;
 
     return result;
 }
@@ -7361,6 +7363,11 @@ struct ggml_tensor * ggml_argsort(
         enum ggml_sort_order  order) {
     bool is_node = false;
 
+    if (a->grad) {
+        GGML_ABORT("fatal error"); // TODO: not implemented
+        is_node = true;
+    }
+
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -10953,9 +10960,6 @@ static void ggml_compute_forward_sum_f32(
         return;
     }
 
-    assert(ggml_is_scalar(dst));
-
-
     assert(ggml_is_scalar(dst));
     assert(src0->nb[0] == sizeof(float));
 
@@ -18356,14 +18360,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad || src1->grad) {
                     GGML_ASSERT(src0->type == tensor->type);
                     GGML_ASSERT(tensor->grad->type == tensor->type);
-                    GGML_ASSERT(tensor->grad->type == src1->grad->type);
+                    GGML_ASSERT(!src1->grad || src1->grad->type == tensor->grad->type);
 
                     tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad,
-                        src1->grad->ne[0],
-                        src1->grad->ne[1],
-                        src1->grad->ne[2],
-                        src1->grad->ne[3],
+                        tensor->grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                         nb1, nb2, nb3, offset);
                 }
 
@@ -18432,9 +18432,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
 
                     memcpy(&offset, tensor->op_params, sizeof(offset));
 
-                    size_t nb1     = tensor->nb[1];
-                    size_t nb2     = tensor->nb[2];
-                    size_t nb3     = tensor->nb[3];
+                    size_t nb1 = tensor->nb[1];
+                    size_t nb2 = tensor->nb[2];
+                    size_t nb3 = tensor->nb[3];
 
                     if (src0->type != src0->grad->type) {
                         // gradient is typically F32, but src0 could be other type