Sync master with upstream release b5468

ggml/include/ggml.h

@@ -528,15 +528,15 @@ extern "C" {
         GGML_UNARY_OP_STEP,
         GGML_UNARY_OP_TANH,
         GGML_UNARY_OP_ELU,
+        GGML_UNARY_OP_RELU,
         GGML_UNARY_OP_SIGMOID,
         GGML_UNARY_OP_GELU,
-        GGML_UNARY_OP_GELU_ERF,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
         GGML_UNARY_OP_HARDSWISH,
         GGML_UNARY_OP_HARDSIGMOID,
         GGML_UNARY_OP_EXP,
-        GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_GELU_ERF,
 
         GGML_UNARY_OP_COUNT,
     };

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2697,14 +2697,10 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         }
     }
 
-    // GroupedMatmulV2 required tensor_list.size < 128
     size_t GROUP_SIZE = 128;
-    std::vector<std::vector<aclTensor*>> src0_tensor_vec_vec;
-    std::vector<std::vector<aclTensor*>> src1_tensor_vec_vec;
-    std::vector<std::vector<aclTensor*>> dst_tensor_vec_vec;
-
-    // split and call GroupedMatmulV2
+    // GroupedMatmulV2 required tensor_list.size < 128
     for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
+        // split and call GroupedMatmulV2
         size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
         std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
         std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2722,13 +2718,144 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
     return;
 }
 
+/**
+ * @brief Performs expert-specific matrix multiplication (MoE) with
+ * quantized precision using the CANN backend.
+ *
+ * This function executes a matrix multiplication operation tailored for
+ * Mixture of Experts (MoE) models, where the input tensor is multiplied
+ * with expert-specific quantized weight matrices. It leverages the CANN
+ * backend to perform efficient low-precision computations and stores the
+ * quantized result in the destination tensor `dst`.
+ *
+ * Quantization techniques reduce memory footprint and improve performance
+ * by using lower-bit representations (e.g., int8) instead of floating-point.
+ * This function is designed to work with such formats and may incorporate
+ * optimizations like identity-based fast paths or routing masks for sparse
+ * expert selection.
+ *
+ * @param ctx The context for executing CANN backend operations.
+ * @param dst The destination tensor where the quantized MoE multiplication result
+ * will be stored.
+ *
+ * @note This function assumes quantized data types and is designed for
+ * MoE architectures with potential sparse expert routing.
+ */
+static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    // TODO: Use aclnnGroupedMatMul
+    //dst   [M, K, N, 1]
+    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
+    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
+    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    // copy index from npu to cpu
+    int64_t n_as = ne02; // A
+    int64_t n_ids = ids->ne[0]; // K
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
+        ACL_MEMCPY_DEVICE_TO_HOST);
+    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+
+    char * src0_original = (char *) src0->data;
+    char * src1_original = (char *) src1->data;
+    char * dst_original  = (char *)  dst->data;
+
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    const enum ggml_type type = dst->src[0]->type;
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    } else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    } else {
+        GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = weight_elem_size;
+    src0_row.nb[1] = weight_elem_size * ne00;
+    src0_row.nb[2] = weight_elem_size * ne00;
+    src0_row.nb[3] = weight_elem_size * ne00;
+    size_t weight_stride = ne00 * ne01 * weight_elem_size;
+    size_t weight_size = weight_stride * ne02 * ne03;
+
+    // scale [D, M, 1, 1] -> scale && permute
+    size_t scale_elem_size = sizeof(uint16_t);
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
+
+    // src1_row [D, 1, 1, 1] -> input
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    // dst_row [M, 1, 1, 1] -> out
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    //create weight for one row
+    ggml_cann_pool_alloc weight_allocator(ctx.pool());
+    void* weight_buffer = weight_allocator.alloc(nb02);
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*weight_stride;
+            void* scale_tmp_ptr = src0_original + weight_size + i02*scale_stride;
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            // mem cpy
+            ggml_cann_async_memcpy(ctx, weight_buffer, src0_tmp_ptr, weight_stride,
+                ACL_MEMCPY_DEVICE_TO_DEVICE);
+            void* scale_buffer = (char*)weight_buffer + weight_stride;
+            ggml_cann_async_memcpy(ctx, scale_buffer, scale_tmp_ptr, scale_stride,
+                ACL_MEMCPY_DEVICE_TO_DEVICE);
+
+            src0_row.data = weight_buffer;
+            src1_row.data = src1_tmp_ptr;
+            dst_row.data = dst_tmp_ptr;
+            dst_row.src[0] = &src0_row;
+            dst_row.src[1] = &src1_row;
+
+            ggml_cann_mul_mat(ctx, &dst_row);
+        }
+    }
+    return;
+}
+
 void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     const enum ggml_type type = dst->src[0]->type;
     switch (type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
             ggml_cann_mul_mat_id_fp(ctx, dst);
             break;
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            ggml_cann_mul_mat_id_quant(ctx, dst);
+            break;
         default:
             GGML_ABORT("Unsupported type for mul_mat_id");
             break;

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2035,6 +2035,15 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_F16:
                 case GGML_TYPE_F32:
                     return true;
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
+                    // only support contiguous for quantized types.
+                    return ggml_is_contiguous(op->src[0]) &&
+                            ggml_is_contiguous(op->src[1]);
                 default:
                     return false;
             }

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -2804,23 +2804,29 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 pipeline_robustness = true;
             } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
                 device->subgroup_size_control = true;
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
             } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
                        !getenv("GGML_VK_DISABLE_COOPMAT")) {
                 device->coopmat_support = true;
                 device->coopmat_m = 0;
                 device->coopmat_n = 0;
                 device->coopmat_k = 0;
+#endif
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
             } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
                        !getenv("GGML_VK_DISABLE_COOPMAT2")) {
                 coopmat2_support = true;
+#endif
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
             } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 &&
                        !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) {
                 device->integer_dot_product = true;
 #endif
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
             } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 &&
                        !getenv("GGML_VK_DISABLE_BFLOAT16")) {
                 bfloat16_support = true;
+#endif
             }
         }
 
@@ -4670,6 +4676,19 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
         }
     }
 
+    if (src->type == to) {
+        // Copy two or four bytes at a time, depending on block size.
+        // For quantized types, we scale by block size/type size. But
+        // this path is also used for bf16->bf16 for example, where the
+        // type size must be exactly 2 or 4.
+        GGML_ASSERT(ggml_is_quantized(to) || ggml_type_size(src->type) == 2 || ggml_type_size(src->type) == 4);
+        if ((ggml_type_size(src->type) % 4) == 0) {
+            return ctx->device->pipeline_contig_cpy_f32_f32;
+        } else {
+            return ctx->device->pipeline_contig_cpy_f16_f16;
+        }
+    }
+
     std::cerr << "Missing CPY op for types: " << ggml_type_name(src->type) << " " << ggml_type_name(to) << std::endl;
     GGML_ABORT("fatal error");
 }
@@ -6731,7 +6750,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_UNARY:
     case GGML_OP_CONV_2D_DW:
         {
-            const uint32_t ne = ggml_nelements(dst);
+            uint32_t ne = ggml_nelements(dst);
+            if (op == GGML_OP_CPY && ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
+                // Convert from number of logical elements to 2- or 4-byte units.
+                ne /= ggml_blck_size(src0->type);
+                if ((ggml_type_size(src0->type) % 4) == 0) {
+                    ne *= ggml_type_size(src0->type) / 4;
+                } else {
+                    ne *= ggml_type_size(src0->type) / 2;
+                }
+            }
             if (ne > 262144) {
                 elements = { 512, 512, CEIL_DIV(ne, 262144) };
             } else if (ne > 512) {
@@ -7281,8 +7309,19 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
+    uint32_t ne = (uint32_t)ggml_nelements(src0);
+    if (ggml_is_quantized(src0->type) && ggml_is_quantized(dst->type)) {
+        // Convert from number of logical elements to 2- or 4-byte units.
+        ne /= ggml_blck_size(src0->type);
+        if ((ggml_type_size(src0->type) % 4) == 0) {
+            ne *= ggml_type_size(src0->type) / 4;
+        } else {
+            ne *= ggml_type_size(src0->type) / 2;
+        }
+    }
+
     ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
-        (uint32_t)ggml_nelements(src0),
+        ne,
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
@@ -9264,8 +9303,7 @@ static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_
     try {
         ptr = ggml_vk_host_malloc(vk_instance.devices[0], size);
     } catch (vk::SystemError& e) {
-        std::cerr << "ggml_vulkan: Failed to allocate pinned memory." << std::endl;
-        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        GGML_LOG_WARN("ggml_vulkan: Failed to allocate pinned memory (%s)\n", e.what());
         // fallback to cpu buffer
         return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
     }
@@ -9867,6 +9905,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
                     return true;
                 }
+
+                // We can handle copying from a type to the same type if it's
+                // contiguous (memcpy). We use f16 or f32 shaders to do the copy,
+                // so the type/block size must be a multiple of 4.
+                if (src0_type == src1_type &&
+                    ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op) &&
+                    (ggml_type_size(src0_type) % 2) == 0) {
+                    return true;
+                }
                 return false;
             } break;
         case GGML_OP_REPEAT:

src/llama-hparams.cpp

@@ -2,6 +2,22 @@
 
 #include "ggml.h"
 
+void llama_hparams::set_swa_pattern(uint32_t n_pattern) {
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        swa_layers[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));
+    }
+}
+
+bool llama_hparams::is_swa_any() const {
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (swa_layers[il]) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
 uint32_t llama_hparams::n_head(uint32_t il) const {
     if (il < n_layer) {
         return n_head_arr[il];
@@ -72,7 +88,7 @@ uint32_t llama_hparams::n_embd_v_s() const {
 
 bool llama_hparams::is_swa(uint32_t il) const {
     if (il < n_layer) {
-        return n_swa_pattern == 0 || (il % n_swa_pattern < (n_swa_pattern - 1));
+        return swa_layers[il];
     }
 
     GGML_ABORT("fatal error");

src/llama-hparams.h

@@ -102,20 +102,12 @@ struct llama_hparams {
 
     // Sliding Window Attention (SWA)
     llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
-
-    uint32_t n_swa = 0;         // the size of the sliding window (0 - no SWA)
-    uint32_t n_swa_pattern = 1; // this value n means that every nth layer is dense (i.e. non-SWA)
-                                // by default n == 1, all layers are dense
-                                // note that if n_swa_pattern == 0, all layers are SWA
-                                // example: n_swa_pattern = 3
-                                //   il == 0: swa
-                                //   il == 1: swa
-                                //   il == 2: dense
-                                //   il == 3: swa
-                                //   il == 4: swa
-                                //   il == 5: dense
-                                //   il == 6: swa
-                                //   etc ...
+    // the size of the sliding window (0 - no SWA)
+    uint32_t n_swa = 0;
+    // if swa_layers[il] == true, then layer il is SWA
+    // if swa_layers[il] == false, then layer il is dense (i.e. non-SWA)
+    // by default, all layers are dense
+    std::array<bool, LLAMA_MAX_LAYERS> swa_layers;
 
     // for State Space Models
     uint32_t ssm_d_conv  = 0;
@@ -153,6 +145,23 @@ struct llama_hparams {
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
     enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
 
+    // this value n_pattern means that every nth layer is dense (i.e. non-SWA)
+    // note that if n_pattern == 0, all layers are SWA
+    //           if n_pattern == 1, all layers are dense
+    // example: n_pattern = 3
+    //   il == 0: swa
+    //   il == 1: swa
+    //   il == 2: dense
+    //   il == 3: swa
+    //   il == 4: swa
+    //   il == 5: dense
+    //   il == 6: swa
+    //   etc ...
+    void set_swa_pattern(uint32_t n_pattern);
+
+    // return true if one of the layers is SWA
+    bool is_swa_any() const;
+
     uint32_t n_head(uint32_t il = 0) const;
 
     uint32_t n_head_kv(uint32_t il = 0) const;