llama-cpp-bin 9840.0.0__tar.gz → 9843.0.0__tar.gz

{llama_cpp_bin-9840.0.0/src/llama_cpp_bin.egg-info → llama_cpp_bin-9843.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9840.0.0
+Version: 9843.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-9840.0.0 → llama_cpp_bin-9843.0.0}/llama.cpp/common/preset.cpp

@@ -7,6 +7,7 @@
 #include <fstream>
 #include <sstream>
 #include <filesystem>
+#include <regex>
 
 static std::string rm_leading_dashes(const std::string & str) {
     size_t pos = 0;
@@ -16,6 +17,23 @@ static std::string rm_leading_dashes(const std::string & str) {
     return str.substr(pos);
 }
 
+static std::string canonical_tag(const std::string & tag) {
+    static const std::regex re_tag("[-.]([A-Z0-9_]+)$", std::regex::icase);
+    std::smatch m;
+    if (std::regex_search(tag, m, re_tag)) {
+        std::string canon = m[1].str();
+        for (char & c : canon) {
+            c = (char) std::toupper((unsigned char) c);
+        }
+        return canon;
+    }
+    std::string upper = tag;
+    for (char & c : upper) {
+        c = (char) std::toupper((unsigned char) c);
+    }
+    return upper;
+}
+
 std::vector<std::string> common_preset::to_args(const std::string & bin_path) const {
     std::vector<std::string> args;
 
@@ -270,11 +288,18 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
 
     for (auto section : ini_data) {
         common_preset preset;
-        if (section.first.empty()) {
-            preset.name = COMMON_PRESET_DEFAULT_NAME;
-        } else {
-            preset.name = section.first;
+        std::string section_name = section.first.empty() ? std::string(COMMON_PRESET_DEFAULT_NAME) : section.first;
+        if (section_name != "*" && section_name != COMMON_PRESET_DEFAULT_NAME) {
+            auto colon_idx = section_name.rfind(':');
+            if (colon_idx != std::string::npos) {
+                std::string tag       = section_name.substr(colon_idx + 1);
+                std::string canon_tag = canonical_tag(tag);
+                if (canon_tag != tag) {
+                    section_name = section_name.substr(0, colon_idx + 1) + canon_tag;
+                }
+            }
         }
+        preset.name = section_name;
         LOG_DBG("loading preset: %s\n", preset.name.c_str());
         for (const auto & [key, value] : section.second) {
             if (key == "version") {

{llama_cpp_bin-9840.0.0 → llama_cpp_bin-9843.0.0}/llama.cpp/ggml/src/ggml-backend.cpp

@@ -1551,8 +1551,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
-        ggml_backend_synchronize(split_backend);
-
         // copy the input tensors to the split backend
         for (int input_id = 0; input_id < split->n_inputs; input_id++) {
             ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1563,15 +1561,15 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                } else if (!split_backend->iface.cpy_tensor_async) {
+                } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+                ggml_backend_tensor_copy(input, input_cpy);
             } else {
                 // wait for the split backend to finish using the input before overwriting it
                 if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                     ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
-                } else if (!split_backend->iface.cpy_tensor_async) {
+                } else {
                     ggml_backend_synchronize(split_backend);
                 }
 
@@ -1676,8 +1674,6 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
             }
         }
 
-        ggml_backend_synchronize(split_backend);
-
         if (!sched->callback_eval) {
             enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
             if (ec != GGML_STATUS_SUCCESS) {

{llama_cpp_bin-9840.0.0 → llama_cpp_bin-9843.0.0}/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3192,24 +3192,11 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
     ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
     ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
 
-    // Enables async copies from CPU to CUDA, instead of only CUDA-to-CUDA
-    // Excluding this path for HIP and MUSA as a precaution.
-    // According to the summary in https://github.com/ggml-org/llama.cpp/pull/20793#issuecomment-4275794315, this change is not beneficial for hip anyways.
-    // Additionally, there is a lot of anectodal evidence that hip/musa stream behavior might not always 1:1 match CUDA behavior.
-    // e.g. https://github.com/ROCm/rocm-systems/issues/5109
-    // It thus makes sense to exclude this path for HIP and MUSA. This PR was not aimed these backends, the majority of testing happened on CUDA.
-    // This can be revisited in the future if enabling copy_from_host benefits hip/MUSA, and if the PR author can extensively test on these backends.
-#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
-    const bool copy_from_host = false;
-#else
-    const bool copy_from_host = ggml_backend_buffer_is_host(buf_src) && ggml_backend_dev_type(backend_src->device) == GGML_BACKEND_DEVICE_TYPE_CPU;
-#endif
-
-    if (!(copy_from_host || ggml_backend_is_cuda(backend_src)) || !ggml_backend_is_cuda(backend_dst)) {
+    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
         return false;
     }
 
-    if (!(copy_from_host || ggml_backend_buffer_is_cuda(buf_src)) || !ggml_backend_buffer_is_cuda(buf_dst)) {
+    if (!ggml_backend_buffer_is_cuda(buf_src) || !ggml_backend_buffer_is_cuda(buf_dst)) {
         return false;
     }
 
@@ -3220,17 +3207,14 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
     ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *) buf_src->context;
     ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *) buf_dst->context;
 
-    if ((copy_from_host && cuda_ctx_dst->device != buf_ctx_dst->device) ||
-        !copy_from_host && (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device)) {
+    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
 #endif // NDEBUG
         return false;
     }
 
-    if (copy_from_host) {
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyHostToDevice, cuda_ctx_dst->stream()));
-    } else if (backend_src != backend_dst) {
+    if (backend_src != backend_dst) {
         // copy on src stream
         if (cuda_ctx_src->device == cuda_ctx_dst->device) {
             CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));

{llama_cpp_bin-9840.0.0 → llama_cpp_bin-9843.0.0}/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1907,6 +1907,38 @@ static bool vk_enable_sync_logger = false;
 static uint32_t vk_perf_logger_frequency = 1;
 static std::string vk_pipeline_stats_filter;
 
+static uint64_t ggml_vk_get_node_flops(const ggml_tensor * node) {
+    if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
+        const uint64_t m     = node->ne[0];
+        const uint64_t n     = node->ne[1];
+        const uint64_t k     = node->src[1]->ne[0];
+        const uint64_t batch = node->ne[2] * node->ne[3];
+        return m * n * (k + (k - 1)) * batch;
+    }
+    if (node->op == GGML_OP_CONV_2D || node->op == GGML_OP_CONV_TRANSPOSE_2D) {
+        const ggml_tensor * knl = node->src[0];
+        const uint64_t Cout  = node->ne[2];
+        const uint64_t size_K = node->src[1]->ne[2] * knl->ne[0] * knl->ne[1];
+        const uint64_t size_N = node->ne[3] * node->ne[0] * node->ne[1];
+        return Cout * size_N * (size_K + (size_K - 1));
+    }
+    if (node->op == GGML_OP_CONV_3D) {
+        const ggml_tensor * knl = node->src[0];
+        const uint64_t OC     = ggml_get_op_params_i32(node, 11);
+        const uint64_t IC     = ggml_get_op_params_i32(node, 9);
+        const uint64_t size_K = IC * knl->ne[0] * knl->ne[1] * knl->ne[2];
+        const uint64_t size_N = node->ne[3] / OC * node->ne[0] * node->ne[1] * node->ne[2];
+        return OC * size_N * (size_K + (size_K - 1));
+    }
+    if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+        const ggml_tensor * q = node->src[0];
+        const ggml_tensor * k = node->src[1];
+        const ggml_tensor * v = node->src[2];
+        return 2ull * q->ne[1] * q->ne[2] * (k->ne[0] + v->ne[0]) * k->ne[1] * q->ne[3];
+    }
+    return 0;
+}
+
 class vk_perf_logger {
   public:
     void print_timings(bool force = false) {
@@ -1955,7 +1987,7 @@ class vk_perf_logger {
     }
 
     std::string get_node_fusion_name(const ggml_tensor * node, const char *fusion_name, uint64_t *n_flops) {
-        *n_flops = 0;
+        *n_flops = ggml_vk_get_node_flops(node);
         std::string fusion_str;
         if (fusion_name) {
             fusion_str = fusion_name + std::string(" ");
@@ -1982,35 +2014,22 @@ class vk_perf_logger {
             if (batch > 1) {
                 name += " batch=" + std::to_string(batch);
             }
-            name = fusion_str + name;
-            *n_flops = m * n * (k + (k - 1)) * batch;
-            return name;
+            return fusion_str + name;
         }
         if (node->op == GGML_OP_CONV_2D || node->op == GGML_OP_CONV_TRANSPOSE_2D) {
             std::string   name    = ggml_op_name(node->op);
-            ggml_tensor * knl     = node->src[0];
-            uint64_t      OW      = node->ne[0];
-            uint64_t      OH      = node->ne[1];
-            uint64_t      N       = node->ne[3];
+            const ggml_tensor * knl = node->src[0];
             uint64_t      Cout    = node->ne[2];
-            uint64_t      KW      = knl->ne[0];
-            uint64_t      KH      = knl->ne[1];
-            uint64_t      Cin     = node->src[1]->ne[2];
-            // KxCRS @ CRSxNPQ = KxNPQ -> M=K, K=CRS, N=NPQ
-            uint64_t      size_M  = Cout;
-            uint64_t      size_K  = Cin * KW * KH;
-            uint64_t      size_N  = N * OW * OH;
-            *n_flops = size_M * size_N * (size_K + (size_K - 1));
-            name += " M=Cout=" + std::to_string(size_M) + ", K=Cin*KW*KH=" + std::to_string(size_K) +
+            uint64_t      size_K  = node->src[1]->ne[2] * knl->ne[0] * knl->ne[1];
+            uint64_t      size_N  = node->ne[3] * node->ne[0] * node->ne[1];
+            name += " M=Cout=" + std::to_string(Cout) + ", K=Cin*KW*KH=" + std::to_string(size_K) +
                     ", N=N*OW*OH=" + std::to_string(size_N);
-            name = fusion_str + name;
-            return name;
+            return fusion_str + name;
         }
         if (node->op == GGML_OP_RMS_NORM) {
             std::string   name    = ggml_op_name(node->op);
             name += "(" + std::to_string(node->ne[0]) + "," + std::to_string(node->ne[1]) + "," + std::to_string(node->ne[2]) + "," + std::to_string(node->ne[3]) + ")";
-            name = fusion_str + name;
-            return name;
+            return fusion_str + name;
         }
         if (node->op == GGML_OP_FLASH_ATTN_EXT) {
             const ggml_tensor * dst = node;
@@ -2026,7 +2045,6 @@ class vk_perf_logger {
                 " k(" << k->ne[0] << "," << k->ne[1] << "," << k->ne[2] << "," << k->ne[3] << "), " <<
                 " v(" << v->ne[0] << "," << v->ne[1] << "," << v->ne[2] << "," << v->ne[3] << "), " <<
                 " m(" << (m?m->ne[0]:0) << "," << (m?m->ne[1]:0) << "," << (m?m->ne[2]:0) << "," << (m?m->ne[3]:0) << ")";
-            *n_flops = 2ull * q->ne[1] * q->ne[2] * (k->ne[0] + v->ne[0]) * k->ne[1] * q->ne[3];
             return name.str();
         }
         if (node->op == GGML_OP_TOP_K) {
@@ -2090,7 +2108,7 @@ struct ggml_backend_vk_context {
     bool do_add_rms_partials_offset_calculation;
     bool do_add_rms_partials;
 
-    uint64_t last_total_mul_mat_bytes {};
+    uint64_t last_total_flops {UINT64_MAX};
 
     // Cache most recent tensor that was converted into prealloc_y, and what pipeline it used to convert.
     vk_pipeline_struct * prealloc_y_last_pipeline_used {};
@@ -16188,22 +16206,23 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     }
 
     // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
-    // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
-    // (and scaled down based on model size, so smaller models submit earlier).
-    int submitted_nodes = 0;
-    int submit_count = 0;
-    uint64_t mul_mat_bytes = 0;
-    uint64_t total_mul_mat_bytes = 0;
-    uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), ctx->last_total_mul_mat_bytes / 40u);
+    // Estimate the amount of compute work using flops, and submit every 200 GFLOP
+    // (and scaled down based on total graph flops, so smaller models submit earlier).
+    // Also submit at least every 100 nodes, in case there are workloads without heavy compute.
+    uint32_t submitted_nodes = 0;
+    uint32_t submit_count = 0;
+    uint64_t batch_flops = 0;
+    uint64_t total_flops = 0;
+    uint64_t flops_per_submit = std::min(uint64_t(200'000'000'000), ctx->last_total_flops / 40u);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
 
-        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
-            auto bytes = ggml_nbytes(cgraph->nodes[i]->src[0]);
-            mul_mat_bytes += bytes;
-            total_mul_mat_bytes += bytes;
+        {
+            auto node_flops = ggml_vk_get_node_flops(cgraph->nodes[i]);
+            batch_flops += node_flops;
+            total_flops += node_flops;
         }
 
         // op_srcs_fused_elementwise indicates whether an op's srcs all contribute to
@@ -16415,8 +16434,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
         bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5;
-        bool submit = ((uint32_t)submitted_nodes >= ctx->device->max_nodes_per_submit) ||
-                      (mul_mat_bytes_per_submit != 0 && mul_mat_bytes >= mul_mat_bytes_per_submit) ||
+        bool submit = (submitted_nodes >= ctx->device->max_nodes_per_submit) ||
+                      (flops_per_submit != 0 && batch_flops >= flops_per_submit) ||
                       (i + ctx->num_additional_fused_ops >= last_node) ||
                       (almost_ready && !ctx->almost_ready_fence_pending);
 
@@ -16450,9 +16469,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         if (submit && enqueued) {
             first_node_in_batch = true;
             submitted_nodes = 0;
-            mul_mat_bytes = 0;
+            batch_flops = 0;
             if (submit_count < 3) {
-                mul_mat_bytes_per_submit *= 2;
+                flops_per_submit *= 2;
             }
             submit_count++;
         }
@@ -16461,7 +16480,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->fused_ops_write_mask = 0;
     }
 
-    ctx->last_total_mul_mat_bytes = total_mul_mat_bytes;
+    ctx->last_total_flops = total_flops;
 
     if (vk_perf_logger_enabled) {
         // End the command buffer and submit/wait

{llama_cpp_bin-9840.0.0 → llama_cpp_bin-9843.0.0}/src/llama_cpp_bin/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '9840.0.0'
-__version_tuple__ = version_tuple = (9840, 0, 0)
+__version__ = version = '9843.0.0'
+__version_tuple__ = version_tuple = (9843, 0, 0)
 
-__commit_id__ = commit_id = 'g1543836a2'
+__commit_id__ = commit_id = 'g13a00fa11'

{llama_cpp_bin-9840.0.0 → llama_cpp_bin-9843.0.0/src/llama_cpp_bin.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 9840.0.0
+Version: 9843.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

llama_cpp_bin-9843.0.0/src/llama_cpp_bin.egg-info/scm_version.json

@@ -0,0 +1,8 @@
+{
+  "tag": "9843.0.0",
+  "distance": 0,
+  "node": "g13a00fa11c51863b21f0ed1a83fff2bbe10c48a8",
+  "dirty": false,
+  "branch": "HEAD",
+  "node_date": "2026-06-30"
+}

llama_cpp_bin-9840.0.0/src/llama_cpp_bin.egg-info/scm_version.json

@@ -1,8 +0,0 @@
-{
-  "tag": "9840.0.0",
-  "distance": 0,
-  "node": "g1543836a23283b0880eb36c528b8cb970518d252",
-  "dirty": false,
-  "branch": "HEAD",
-  "node_date": "2026-06-29"
-}