whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf

@@ -0,0 +1,38 @@
+[Version]
+Signature   = "$WINDOWS NT$"
+Class       = ComputeAccelerator
+ClassGuid   = {F01A9D53-3FF6-48D2-9F97-C8A7004BE10C}
+Provider    = %GGML%
+DriverVer   = 01/01/2026,1.0.0.0
+CatalogFile = libggml-htp.cat
+PnpLockDown = 1
+
+[DestinationDirs]
+Drivers_Dir = 6
+
+[SourceDisksNames]
+1 = %DiskId%
+
+[SourceDisksFiles]
+libggml-htp-v68.so = 1
+libggml-htp-v69.so = 1
+libggml-htp-v73.so = 1
+libggml-htp-v75.so = 1
+libggml-htp-v81.so = 1
+
+[ControlFlags]
+ExcludeFromSelect = *
+
+[DefaultInstall.NTarm64]
+CopyFiles=Drivers_Dir
+
+[Drivers_Dir]
+libggml-htp-v68.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v69.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v73.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v75.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v81.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+
+[Strings]
+GGML   = 'GGML'
+DiskId = 'GGML HTP library'

data/ext/sources/ggml/src/ggml-hexagon/op-desc.h

@@ -0,0 +1,153 @@
+#ifndef OP_DESC_H
+#define OP_DESC_H
+
+#define GGML_COMMON_IMPL_CPP
+#include "ggml-backend-impl.h"
+#include "ggml-common.h"
+
+#include <string>
+#include <stdio.h>
+
+struct op_desc {
+    char strides[64 * GGML_MAX_SRC];
+    char dims[64 * GGML_MAX_SRC];
+    char types[16 * GGML_MAX_SRC];
+    char buffs[64 * GGML_MAX_SRC];
+    char names[64 * GGML_MAX_SRC];
+
+    int format_tensor_dims(char * str, const struct ggml_tensor * t) {
+        if (t->ne[2] == 1 && t->ne[3] == 1) {
+            return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
+        } else {
+            return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
+        }
+    }
+
+    void format_op_dims(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += format_tensor_dims(p, t->src[0]);
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += format_tensor_dims(p, t->src[i]);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        // format self dims separately for better visual alignment
+        char self[64];
+        format_tensor_dims(self, t);
+
+        p += sprintf(p, "%s", self);
+    }
+
+    int format_tensor_strides(char * str, const struct ggml_tensor * t) {
+        const char * c = ggml_is_contiguous(t) ? "" : "!";
+
+        if (t->ne[2] == 1 && t->ne[3] == 1) {
+            return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
+        } else {
+            return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
+        }
+    }
+
+    void format_op_strides(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += format_tensor_strides(p, t->src[0]);
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += format_tensor_strides(p, t->src[i]);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        // format self dims separately for better visual alignment
+        char self[64];
+        format_tensor_strides(self, t);
+
+        p += sprintf(p, "%s", self);
+    }
+
+    void format_op_types(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += sprintf(p, "%s", ggml_type_name(t->src[0]->type));
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", ggml_type_name(t->src[i]->type));
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", ggml_type_name(t->type));
+    }
+
+    const char * tensor_buff_name(const struct ggml_tensor * t) {
+        if (t->buffer) {
+            return ggml_backend_buffer_name(t->buffer);
+        }
+        return "NONE";
+    }
+
+    void format_op_buffs(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += sprintf(p, "%s", tensor_buff_name(t->src[0]));
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", tensor_buff_name(t->src[i]));
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", tensor_buff_name(t));
+    }
+
+    void format_op_names(char * str, const struct ggml_tensor * t) {
+        char * p = str;
+
+        // append src0 and src1 (if any)
+        if (t->src[0]) {
+            p += sprintf(p, "%s", t->src[0]->name);
+
+            for (int i = 1; i < GGML_MAX_SRC && t->src[i]; i++) {
+                p += sprintf(p, " x ");
+                p += sprintf(p, "%s", t->src[i]->name);
+            }
+
+            p += sprintf(p, " -> ");
+        }
+
+        p += sprintf(p, "%s", t->name);
+    }
+
+    void format(const ggml_tensor * op) {
+        format_op_dims(dims, op);
+        format_op_strides(strides, op);
+        format_op_types(types, op);
+        format_op_buffs(buffs, op);
+        format_op_names(names, op);
+    }
+
+    op_desc() {}
+    op_desc(const ggml_tensor * op) { format(op); }
+};
+
+#endif // OP_DESC_H

data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt

@@ -11,6 +11,10 @@ endif()
 list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
 list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
 
+if (NOT DEFINED CMAKE_HIP_FLAGS_DEBUG)
+    set(CMAKE_HIP_FLAGS_DEBUG "-g -O2")
+endif()
+
 # CMake on Windows doesn't support the HIP language yet
 if (WIN32)
     set(CXX_IS_HIPCC TRUE)
@@ -28,9 +32,12 @@ if (CXX_IS_HIPCC)
                 " Prefer setting the HIP compiler directly. See README for details.")
     endif()
 else()
-    # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
-    if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-        set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+    # Forward (AMD)GPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+    if(AMDGPU_TARGETS AND NOT GPU_TARGETS)
+        set(GPU_TARGETS ${AMDGPU_TARGETS})
+    endif()
+    if(GPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+        set(CMAKE_HIP_ARCHITECTURES ${GPU_TARGETS})
     endif()
     cmake_minimum_required(VERSION 3.21)
     enable_language(HIP)
@@ -39,12 +46,6 @@ endif()
 find_package(hip     REQUIRED)
 find_package(hipblas REQUIRED)
 find_package(rocblas REQUIRED)
-if (GGML_HIP_ROCWMMA_FATTN)
-    CHECK_INCLUDE_FILE_CXX("rocwmma/rocwmma.hpp" FOUND_ROCWMMA)
-    if (NOT ${FOUND_ROCWMMA})
-        message(FATAL_ERROR "rocwmma has not been found")
-    endif()
-endif()
 
 if (${hip_VERSION} VERSION_LESS 6.1)
     message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
@@ -59,10 +60,14 @@ file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
 list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
 
 file(GLOB   GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
+file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-tile*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
 list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
 list(APPEND GGML_SOURCES_ROCM ${SRCS})
+file(GLOB   SRCS "../ggml-cuda/template-instances/mmf*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
 
 if (GGML_CUDA_FA_ALL_QUANTS)
     file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
@@ -117,10 +122,6 @@ if (NOT GGML_HIP_MMQ_MFMA)
     add_compile_definitions(GGML_HIP_NO_MMQ_MFMA)
 endif()
 
-if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
-    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
-endif()
-
 if (GGML_HIP_EXPORT_METRICS)
     set(CMAKE_HIP_FLAGS "${CMAKE_HIP_FLAGS} -Rpass-analysis=kernel-resource-usage --save-temps")
 endif()

data/ext/sources/ggml/src/ggml-impl.h

@@ -24,10 +24,6 @@
 #include <arm_neon.h>
 #endif
 
-#if defined(__F16C__)
-#include <immintrin.h>
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -102,6 +98,13 @@ static bool ggml_op_is_empty(enum ggml_op op) {
     }
 }
 
+static inline bool ggml_impl_is_view(const struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
+static inline float ggml_compute_softplus_f32(float input) {
+    return (input > 20.0f) ? input : logf(1 + expf(input));
+}
 //
 // logging
 //
@@ -488,6 +491,61 @@ static inline float ggml_e8m0_to_fp32_half(uint8_t x) {
 #define GGML_E8M0_TO_FP32(x) ggml_e8m0_to_fp32(x)
 #define GGML_E8M0_TO_FP32_HALF(x) ggml_e8m0_to_fp32_half(x)
 
+// UE4M3: unsigned, 4 exp bits (bias=7), 3 mantissa bits
+// Returns value * 0.5 to match kvalues_mxfp4 convention (kvalues = 2 * E2M1_float)
+static inline float ggml_ue4m3_to_fp32(uint8_t x) {
+    if (x == 0 || x == 0x7F) {
+        return 0.0f;
+    }
+    int   exp = (x >> 3) & 0xF;
+    int   man = x & 0x7;
+    float raw;
+    if (exp == 0) {
+        raw = ldexpf((float) man, -9);
+    } else {
+        raw = ldexpf(1.0f + (float) man / 8.0f, exp - 7);
+    }
+    return raw * 0.5f;
+}
+
+static inline uint8_t ggml_fp32_to_ue4m3(float x) {
+    if (!(x > 0.0f)) {
+        return 0;
+    }
+    if (x > 448.0f) {
+        x = 448.0f;
+    }
+    uint32_t bits;
+    memcpy(&bits, &x, 4);
+    int fp32_exp  = ((bits >> 23) & 0xFF) - 127;
+    int fp32_man  = (bits >> 20) & 0x7;
+    int ue4m3_exp = fp32_exp + 7;
+    if (ue4m3_exp <= 0) {
+        // subnormal: value = man * 2^-9, man = round(x * 2^9)
+        int man = (int) (x * 512.0f + 0.5f);
+        if (man > 7) {
+            man = 7;
+        }
+        if (man < 1) {
+            return 0;
+        }
+        return (uint8_t) man;
+    }
+    if (ue4m3_exp >= 15) {
+        return 0x7E;
+    }
+    int round_bit = (bits >> 19) & 1;
+    int ue4m3_man = fp32_man + round_bit;
+    if (ue4m3_man > 7) {
+        ue4m3_man = 0;
+        ue4m3_exp++;
+        if (ue4m3_exp >= 15) {
+            return 0x7E;
+        }
+    }
+    return (uint8_t) ((ue4m3_exp << 3) | ue4m3_man);
+}
+
 /**
  * Converts brain16 to float32.
  *
@@ -562,14 +620,23 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
 #define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
 #define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
+static inline int32_t ggml_node_get_use_count(const struct ggml_cgraph * cgraph, int node_idx) {
+    const struct ggml_tensor * node = cgraph->nodes[node_idx];
+
+    size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+    if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        return 0;
+    }
+    return cgraph->use_counts[hash_pos];
+}
+
 // return true if the node's results are only used by N other nodes
 // and can be fused into their calculations.
 static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int node_idx, int32_t n_uses) {
     const struct ggml_tensor * node = cgraph->nodes[node_idx];
 
     // check the use count against how many we're replacing
-    size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
-    if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos) || cgraph->use_counts[hash_pos] != n_uses) {
+    if (ggml_node_get_use_count(cgraph, node_idx) != n_uses) {
         return false;
     }
 
@@ -603,6 +670,9 @@ static inline bool ggml_can_fuse_ext(const struct ggml_cgraph * cgraph, const in
         if (node->op != ops[i]) {
             return false;
         }
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            return false;
+        }
         if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idxs[i], 1)) {
             return false;
         }
@@ -635,11 +705,42 @@ static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
     return ggml_can_fuse_ext(cgraph, idxs, ops, num_ops);
 }
 
+GGML_API bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
+                                         const int *                node_idxs,
+                                         int                        count,
+                                         const enum ggml_op *       ops,
+                                         const int *                outputs,
+                                         int                        num_outputs);
+
+// Returns true if the subgraph formed by {node_idxs} can be fused
+// checks whethers all nodes which are not part of outputs can be elided
+// by checking if their num_uses are confined to the subgraph
+static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
+                                          int                        node_idx,
+                                          int                        count,
+                                          const enum ggml_op *       ops,
+                                          const int *                outputs,
+                                          int                        num_outputs) {
+    GGML_ASSERT(count < 32);
+    if (node_idx + count > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+
+    for (int i = 0; i < count; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return ggml_can_fuse_subgraph_ext(cgraph, idxs, count, ops, outputs, num_outputs);
+}
+
 #ifdef __cplusplus
 }
 #endif
 
 #ifdef __cplusplus
+#include <array>
 #include <initializer_list>
 #include <vector>
 
@@ -648,6 +749,28 @@ inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::
     return ggml_can_fuse(cgraph, node_idx, ops.begin(), (int)ops.size());
 }
 
+inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph *          cgraph,
+                                   int                                 start_idx,
+                                   std::initializer_list<enum ggml_op> ops,
+                                   std::initializer_list<int>          outputs = {}) {
+    return ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
+}
+
+// Return true if the edges in the graph match expectations.
+inline bool ggml_check_edges(const struct ggml_cgraph *                cgraph,
+                             int                                       start_idx,
+                             std::initializer_list<std::array<int, 3>> edges) {
+    for (const auto & edge : edges) {
+        int dst_node = edge[0];
+        int src_idx  = edge[1];
+        int src_node = edge[2];
+        if (cgraph->nodes[start_idx + dst_node]->src[src_idx] != cgraph->nodes[start_idx + src_node]) {
+            return false;
+        }
+    }
+    return true;
+}
+
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);
 GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);

data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt

@@ -23,11 +23,6 @@ if (GGML_METAL_NDEBUG)
     add_compile_definitions(GGML_METAL_NDEBUG)
 endif()
 
-# copy metal files to bin directory
-configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
-configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
-
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
@@ -37,12 +32,12 @@ if (GGML_METAL_EMBED_LIBRARY)
     set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
     set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
-    file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+    file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
     # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
     add_custom_command(
         OUTPUT "${METALLIB_EMBED_ASM}"
@@ -62,6 +57,11 @@ if (GGML_METAL_EMBED_LIBRARY)
 
     target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
 else()
+    # copy metal files to bin directory
+    configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
+    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
+    configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
+
     if (GGML_METAL_SHADER_DEBUG)
         # custom command to do the following:
         #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
@@ -71,7 +71,7 @@ else()
         #       disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+        #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
         #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)

data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp

@@ -112,7 +112,7 @@ static bool ggml_mem_ranges_add_dst(ggml_mem_ranges_t mrs, const ggml_tensor * t
 }
 
 bool ggml_mem_ranges_add(ggml_mem_ranges_t mrs, const ggml_tensor * tensor) {
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (tensor->src[i]) {
             ggml_mem_ranges_add_src(mrs, tensor->src[i]);
         }
@@ -173,7 +173,7 @@ static bool ggml_mem_ranges_check_dst(ggml_mem_ranges_t mrs, const ggml_tensor *
 }
 
 bool ggml_mem_ranges_check(ggml_mem_ranges_t mrs, const ggml_tensor * tensor) {
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (tensor->src[i]) {
             if (!ggml_mem_ranges_check_src(mrs, tensor->src[i])) {
                 return false;
@@ -264,15 +264,26 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
             case GGML_OP_NORM:
             case GGML_OP_RMS_NORM:
             case GGML_OP_GROUP_NORM:
+            case GGML_OP_L2_NORM:
             case GGML_OP_SUM_ROWS:
+            case GGML_OP_SSM_CONV:
+            case GGML_OP_SSM_SCAN:
+            case GGML_OP_CLAMP:
+            case GGML_OP_TRI:
+            case GGML_OP_DIAG:
             case GGML_OP_MUL:
             case GGML_OP_ADD:
+            case GGML_OP_SUB:
             case GGML_OP_DIV:
             case GGML_OP_GLU:
             case GGML_OP_SCALE:
+            case GGML_OP_UNARY:
             case GGML_OP_GET_ROWS:
-            case GGML_OP_CPY:
             case GGML_OP_SET_ROWS:
+            case GGML_OP_SET:
+            case GGML_OP_CPY:
+            case GGML_OP_CONT:
+            case GGML_OP_REPEAT:
                 return true;
             default:
                 return ggml_op_is_empty(op);
@@ -312,7 +323,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
             h_add(mrs1, node0);
 
             // that many nodes forward to search for a concurrent node
-            constexpr int N_FORWARD = 8;
+            constexpr int N_FORWARD = 64;
 
             for (int i1 = i0 + 1; i1 < i0 + N_FORWARD && i1 < n; i1++) {
                 if (used[i1]) {

data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h

@@ -15,14 +15,22 @@ typedef struct ggml_metal * ggml_metal_t;
 ggml_metal_t ggml_metal_init(ggml_metal_device_t dev);
 void ggml_metal_free(ggml_metal_t ctx);
 
+const char * ggml_metal_get_name(ggml_metal_t ctx);
+
 void ggml_metal_synchronize(ggml_metal_t ctx);
 
 void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
 void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
 enum ggml_status ggml_metal_graph_compute (ggml_metal_t ctx, struct ggml_cgraph * gf);
 void             ggml_metal_graph_optimize(ggml_metal_t ctx, struct ggml_cgraph * gf);
 
+void ggml_metal_event_record(ggml_metal_t ctx, ggml_metal_event_t ev);
+void ggml_metal_event_wait  (ggml_metal_t ctx, ggml_metal_event_t ev);
+
+ggml_metal_event_t ggml_metal_get_ev_cpy(ggml_metal_t ctx);
+
 void ggml_metal_set_n_cb            (ggml_metal_t ctx, int n_cb);
 void ggml_metal_set_abort_callback  (ggml_metal_t ctx, ggml_abort_callback abort_callback, void * user_data);
 bool ggml_metal_supports_family     (ggml_metal_t ctx, int family);