arg: allow -kvu flag for llama-perplexity (#18117)

The -kvu (--kv-unified) flag is required for hellaswag and winogrande
benchmarks which use coupled sequences. Without unified KV cache,
these benchmarks fail with:

  split_equal: sequential split is not supported when there are
  coupled sequences in the input batch (you may need to use the -kvu flag)

This change adds LLAMA_EXAMPLE_PERPLEXITY to the allowed examples for
the -kvu argument, enabling its use with llama-perplexity.

.devops/cann.Dockerfile

@@ -107,7 +107,7 @@ ENTRYPOINT ["/app/tools.sh"]
 # ENTRYPOINT ["/app/llama-server"]
 
 ### Target: light
-# Lightweight image containing only llama-cli
+# Lightweight image containing only llama-cli and llama-completion
 # ==============================================================================
 FROM base AS light
 

.devops/llama-cli-cann.Dockerfile

@@ -23,11 +23,12 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 RUN echo "Building with static libs" && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
     cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF  && \
-    cmake --build build --config Release --target llama-cli
+    cmake --build build --config Release --target llama-cli && \
+    cmake --build build --config Release --target llama-completion
 
 # TODO: use image with NNRT
 FROM ascendai/cann:$ASCEND_VERSION AS runtime
-COPY --from=build /app/build/bin/llama-cli /llama-cli
+COPY --from=build /app/build/bin/llama-cli /app/build/bin/llama-completion /
 
 ENV LC_ALL=C.utf8
 

.devops/llama-cpp-cuda.srpm.spec

@@ -37,6 +37,7 @@ make -j GGML_CUDA=1
 %install
 mkdir -p %{buildroot}%{_bindir}/
 cp -p llama-cli %{buildroot}%{_bindir}/llama-cuda-cli
+cp -p llama-completion %{buildroot}%{_bindir}/llama-cuda-completion
 cp -p llama-server %{buildroot}%{_bindir}/llama-cuda-server
 cp -p llama-simple %{buildroot}%{_bindir}/llama-cuda-simple
 
@@ -68,6 +69,7 @@ rm -rf %{_builddir}/*
 
 %files
 %{_bindir}/llama-cuda-cli
+%{_bindir}/llama-cuda-completion
 %{_bindir}/llama-cuda-server
 %{_bindir}/llama-cuda-simple
 /usr/lib/systemd/system/llamacuda.service

.devops/llama-cpp.srpm.spec

@@ -39,6 +39,7 @@ make -j
 %install
 mkdir -p %{buildroot}%{_bindir}/
 cp -p llama-cli %{buildroot}%{_bindir}/llama-cli
+cp -p llama-completion %{buildroot}%{_bindir}/llama-completion
 cp -p llama-server %{buildroot}%{_bindir}/llama-server
 cp -p llama-simple %{buildroot}%{_bindir}/llama-simple
 
@@ -70,6 +71,7 @@ rm -rf %{_builddir}/*
 
 %files
 %{_bindir}/llama-cli
+%{_bindir}/llama-completion
 %{_bindir}/llama-server
 %{_bindir}/llama-simple
 /usr/lib/systemd/system/llama.service

common/arg.cpp

@@ -835,6 +835,19 @@ bool common_arg_utils::is_autoy(const std::string & value) {
 }
 
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+    // per-example default params
+    // we define here to make sure it's included in llama-gen-docs
+    if (ex == LLAMA_EXAMPLE_COMPLETION) {
+        params.use_jinja = false;   // disable jinja by default
+
+    } else if (ex == LLAMA_EXAMPLE_MTMD) {
+        params.use_jinja = false;   // disable jinja by default
+        params.sampling.temp = 0.2; // lower temp by default for better quality
+
+    } else if (ex == LLAMA_EXAMPLE_SERVER) {
+        params.n_parallel = -1;     // auto by default
+    }
+
     params.use_color = tty_can_use_colors();
 
     // load dynamic backends
@@ -1107,7 +1120,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_SWA_FULL"));
     add_opt(common_arg(
         {"--ctx-checkpoints", "--swa-checkpoints"}, "N",
-        string_format("max number of context checkpoints to create per slot (default: %d)\n"
+        string_format("max number of context checkpoints to create per slot (default: %d)"
             "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_ctx_checkpoints),
         [](common_params & params, int value) {
             params.n_ctx_checkpoints = value;
@@ -1115,7 +1128,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_CTX_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"--cache-ram", "-cram"}, "N",
-        string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)\n"
+        string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)"
             "[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)", params.cache_ram_mib),
         [](common_params & params, int value) {
             params.cache_ram_mib = value;
@@ -1123,12 +1136,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_CACHE_RAM").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"--kv-unified", "-kvu"},
-        string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
-            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.kv_unified ? "true" : "false"),
+        "use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)",
         [](common_params & params) {
             params.kv_unified = true;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED"));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},
@@ -1888,13 +1900,27 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             LOG_WRN("DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify\n");
         }
     ).set_env("LLAMA_ARG_DEFRAG_THOLD"));
-    add_opt(common_arg(
-        {"-np", "--parallel"}, "N",
-        string_format("number of parallel sequences to decode (default: %d)", params.n_parallel),
-        [](common_params & params, int value) {
-            params.n_parallel = value;
-        }
-    ).set_env("LLAMA_ARG_N_PARALLEL"));
+    if (ex == LLAMA_EXAMPLE_SERVER) {
+        // this is to make sure this option appears in the server-specific section of the help message
+        add_opt(common_arg(
+            {"-np", "--parallel"}, "N",
+            string_format("number of server slots (default: %d, -1 = auto)", params.n_parallel),
+            [](common_params & params, int value) {
+                if (value == 0) {
+                    throw std::invalid_argument("error: invalid value for n_parallel\n");
+                }
+                params.n_parallel = value;
+            }
+        ).set_env("LLAMA_ARG_N_PARALLEL").set_examples({LLAMA_EXAMPLE_SERVER}));
+    } else {
+        add_opt(common_arg(
+            {"-np", "--parallel"}, "N",
+            string_format("number of parallel sequences to decode (default: %d)", params.n_parallel),
+            [](common_params & params, int value) {
+                params.n_parallel = value;
+            }
+        ).set_env("LLAMA_ARG_N_PARALLEL"));
+    }
     add_opt(common_arg(
         {"-ns", "--sequences"}, "N",
         string_format("number of sequences to decode (default: %d)", params.n_sequences),

docs/development/HOWTO-add-model.md

@@ -97,7 +97,7 @@ The model params and tensors layout must be defined in `llama.cpp` source files:
 1. Define a new `llm_arch` enum value in `src/llama-arch.h`.
 2. In `src/llama-arch.cpp`:
     - Add the architecture name to the `LLM_ARCH_NAMES` map.
-    - Add the tensor mappings to the `LLM_TENSOR_NAMES` map.
+    - Add the list of model tensors to `llm_get_tensor_names` (you may also need to update `LLM_TENSOR_NAMES`)
 3. Add any non-standard metadata loading in the `llama_model_loader` constructor in `src/llama-model-loader.cpp`.
 4. If the model has a RoPE operation, add a case for the architecture in `llama_model_rope_type` function in `src/llama-model.cpp`.
 

docs/docker.md

@@ -7,9 +7,9 @@
 ## Images
 We have three Docker images available for this project:
 
-1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
-2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
-3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the `llama-cli` and `llama-completion` executables and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the `llama-cli` and `llama-completion` executables. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the `llama-server` executable. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
 
 Additionally, there the following images, similar to the above:
 
@@ -44,13 +44,15 @@ docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --all-in-o
 On completion, you are ready to play!
 
 ```bash
-docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf
+docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --run-legacy -m /models/32B/ggml-model-q8_0.gguf -no-cnv -p "Building a mobile app can be done in 15 steps:" -n 512
 ```
 
 or with a light image:
 
 ```bash
-docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+docker run -v /path/to/models:/models --entrypoint /app/llama-cli ghcr.io/ggml-org/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf
+docker run -v /path/to/models:/models --entrypoint /app/llama-completion ghcr.io/ggml-org/llama.cpp:light -m /models/32B/ggml-model-q8_0.gguf -no-cnv -p "Building a mobile app can be done in 15 steps:" -n 512
 ```
 
 or with a server image:
@@ -59,6 +61,8 @@ or with a server image:
 docker run -v /path/to/models:/models -p 8080:8080 ghcr.io/ggml-org/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8080 --host 0.0.0.0 -n 512
 ```
 
+In the above examples, `--entrypoint /app/llama-cli` is specified for clarity, but you can safely omit it since it's the default entrypoint in the container.
+
 ## Docker With CUDA
 
 Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
@@ -80,9 +84,9 @@ The defaults are:
 
 The resulting images, are essentially the same as the non-CUDA images:
 
-1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
-3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
+1. `local/llama.cpp:full-cuda`: This image includes both the `llama-cli` and `llama-completion` executables and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-cuda`: This image only includes the `llama-cli` and `llama-completion` executables.
+3. `local/llama.cpp:server-cuda`: This image only includes the `llama-server` executable.
 
 ## Usage
 
@@ -114,9 +118,9 @@ The defaults are:
 
 The resulting images, are essentially the same as the non-MUSA images:
 
-1. `local/llama.cpp:full-musa`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `local/llama.cpp:light-musa`: This image only includes the main executable file.
-3. `local/llama.cpp:server-musa`: This image only includes the server executable file.
+1. `local/llama.cpp:full-musa`: This image includes both the `llama-cli` and `llama-completion` executables and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-musa`: This image only includes the `llama-cli` and `llama-completion` executables.
+3. `local/llama.cpp:server-musa`: This image only includes the `llama-server` executable.
 
 ## Usage
 

examples/gen-docs/gen-docs.cpp

@@ -48,7 +48,7 @@ static void write_table(std::ofstream & file, std::vector<common_arg *> & opts)
     }
 }
 
-static void export_md(std::string fname, llama_example ex) {
+static void export_md(std::string fname, llama_example ex, std::string name) {
     std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
 
     common_params params;
@@ -72,13 +72,14 @@ static void export_md(std::string fname, llama_example ex) {
     write_table(file, common_options);
     file << "\n\n**Sampling params**\n\n";
     write_table(file, sparam_options);
-    file << "\n\n**Example-specific params**\n\n";
+    file << "\n\n**" << name << "-specific params**\n\n";
     write_table(file, specific_options);
 }
 
 int main(int, char **) {
-    export_md("autogen-main.md", LLAMA_EXAMPLE_COMPLETION);
-    export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER);
+    // TODO: add CLI
+    export_md("autogen-completion.md", LLAMA_EXAMPLE_COMPLETION, "Tool");
+    export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER, "Server");
 
     return 0;
 }

examples/model-conversion/scripts/causal/modelcard.template

@@ -7,7 +7,7 @@ base_model:
 Recommended way to run this model:
 
 ```sh
-llama-server -hf {namespace}/{model_name}-GGUF -c 0 -fa
+llama-server -hf {namespace}/{model_name}-GGUF -c 0
 ```
 
 Then, access http://localhost:8080

ggml/src/ggml-cuda/argmax.cu

@@ -21,7 +21,7 @@ static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __rest
     }
 
 #pragma unroll
-    for (int offset = 16; offset > 0; offset >>= 1) {
+    for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
         const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
         const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
         if (val > maxval) {
@@ -50,7 +50,7 @@ static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __rest
                 argmax = shared_argmax[lane_id];
             }
 #pragma unroll
-            for (int offset = 16; offset > 0; offset >>= 1) {
+            for (int offset = WARP_SIZE/2; offset > 0; offset >>= 1) {
                 const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
                 const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
                 if (val > maxval) {

gguf-py/gguf/utility.py

@@ -288,7 +288,7 @@ class LocalTensor:
     data_range: LocalTensorRange
 
     def mmap_bytes(self) -> np.ndarray:
-        return np.memmap(self.data_range.filename, offset=self.data_range.offset, shape=self.data_range.size)
+        return np.memmap(self.data_range.filename, mode='r', offset=self.data_range.offset, shape=self.data_range.size)
 
 
 class SafetensorsLocal:

src/llama-arch.h

@@ -3,6 +3,7 @@
 #include "ggml.h" // ggml_op
 
 #include <string>
+#include <set>
 
 //
 // gguf constants (sync with gguf.py)
@@ -316,6 +317,7 @@ enum llm_tensor {
     LLM_TENSOR_DENSE_3_OUT,
     LLM_TENSOR_OUTPUT,
     LLM_TENSOR_OUTPUT_NORM,
+    LLM_TENSOR_OUTPUT_NORM_LFM2, // fix for wrong tensor name
     LLM_TENSOR_ROPE_FREQS,
     LLM_TENSOR_ROPE_FACTORS_LONG,
     LLM_TENSOR_ROPE_FACTORS_SHORT,
@@ -526,6 +528,10 @@ struct LLM_TN_IMPL {
     const int bid;
     const int xid;
 
+    const std::set<llm_tensor> model_tensors;
+
+    LLM_TN_IMPL(llm_arch arch, llm_tensor tensor, const char * suffix, int bid, int xid);
+
     std::string str() const;
 
     operator std::string() const {
@@ -547,11 +553,11 @@ struct LLM_TN {
     llm_arch arch;
 
     LLM_TN_IMPL operator()(llm_tensor tensor, const char * suffix, int bid = -1, int xid = -1) const {
-        return { arch, tensor, suffix, bid, xid };
+        return LLM_TN_IMPL(arch, tensor, suffix, bid, xid);
     }
 
     LLM_TN_IMPL operator()(llm_tensor tensor, int bid = -1, int xid = -1) const {
-        return { arch, tensor, nullptr, bid, xid };
+        return LLM_TN_IMPL(arch, tensor, nullptr, bid, xid);
     }
 };
 

src/llama-model.cpp

@@ -6236,8 +6236,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
-                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM_LFM2, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,           "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
 
                     if (output == NULL) {
                         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);

src/llama.cpp

@@ -71,8 +71,9 @@ static std::vector<llama_device_memory_data> llama_get_device_memory_data(
     }, &ud);
 
     llama_model_params mparams_copy = *mparams;
-    mparams_copy.no_alloc = true;
-    mparams_copy.use_mmap = false;
+    mparams_copy.no_alloc  = true;
+    mparams_copy.use_mmap  = false;
+    mparams_copy.use_mlock = false;
 
     llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
     if (model == nullptr) {
@@ -180,11 +181,12 @@ static void llama_params_fit_impl(
         }
     }
 
-    int64_t sum_total          = 0;
-    int64_t sum_projected_free = 0;
-    int64_t min_projected_free = INT64_MAX;
-    int64_t sum_projected_used = 0;
-    int64_t sum_projected_ctx  = 0;
+    int64_t sum_total           = 0;
+    int64_t sum_projected_free  = 0;
+    int64_t min_projected_free  = INT64_MAX;
+    int64_t sum_projected_used  = 0;
+    int64_t sum_projected_model = 0;
+    int64_t sum_projected_ctx   = 0;
 
     if (nd > 1) {
         LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
@@ -195,11 +197,12 @@ static void llama_params_fit_impl(
         const int64_t projected_used = dmd.mb.total();
         const int64_t projected_free = dmd.free - projected_used;
 
-        sum_total          += dmd.total;
-        sum_projected_used += projected_used;
-        sum_projected_free += projected_free;
-        min_projected_free  = std::min(min_projected_free, projected_free);
-        sum_projected_ctx  += dmd.mb.context;
+        sum_total           += dmd.total;
+        sum_projected_used  += projected_used;
+        sum_projected_free  += projected_free;
+        min_projected_free   = std::min(min_projected_free, projected_free);
+        sum_projected_model += dmd.mb.model;
+        sum_projected_ctx   += dmd.mb.context;
 
         if (nd > 1) {
             LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
@@ -234,13 +237,34 @@ static void llama_params_fit_impl(
             if (cparams->n_ctx == 0) {
                 if (hp_nct > n_ctx_min) {
                     const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
-                    const uint32_t ctx_reduction = std::min(
-                        uint32_t((-global_surplus + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
+
+                    int64_t memory_reduction = -global_surplus;
+                    if (nd > 1) {
+                        // for multiple devices we need to be more conservative in terms of how much context we think can fit:
+                        //   - for dense models only whole layers can be assigned to devices
+                        //   - for MoE models only whole tensors can be assigned to devices, which we estimate to be <= 1/3 of a layer
+                        //   - on average we expect a waste of 0.5 layers/tensors per device
+                        //   - use slightly more than the expected average for nd devices to be safe
+                        const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
+                        memory_reduction += (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
+                    }
+
+                    uint32_t ctx_reduction = std::min(uint32_t((memory_reduction + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
                     cparams->n_ctx = hp_nct - ctx_reduction;
-                    const int64_t memory_reduction = ctx_reduction * bytes_per_ctx;
+                    cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
+
+                    ctx_reduction = hp_nct - cparams->n_ctx;
+                    memory_reduction = ctx_reduction * bytes_per_ctx;
                     global_surplus += memory_reduction;
                     LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
                         __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                    if (global_surplus >= 0) {
+                        if (nd == 1) {
+                            LLAMA_LOG_INFO("%s: entire model can be fit by reducing context\n", __func__);
+                            return;
+                        }
+                        LLAMA_LOG_INFO("%s: entire model should be fit across devices by reducing context\n", __func__);
+                    }
                 } else {
                     LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
                         __func__, hp_nct, n_ctx_min);
@@ -249,10 +273,6 @@ static void llama_params_fit_impl(
                 LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
             }
         }
-        if (global_surplus >= 0) {
-            LLAMA_LOG_INFO("%s: entire model can be fit across devices by reducing context\n", __func__);
-            return;
-        }
     }
 
     if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
@@ -478,8 +498,13 @@ static void llama_params_fit_impl(
     } else {
         LLAMA_LOG_INFO("%s: filling dense-only layers back-to-front:\n", __func__);
     }
-    uint32_t n_unassigned = hp_ngl;
     for (int id = nd - 1; id >= 0; id--) {
+        uint32_t n_unassigned = hp_ngl;
+        for (size_t jd = id + 1; jd < nd; ++jd) {
+            assert(n_unassigned >= ngl_per_device[jd].n_layer);
+            n_unassigned -= ngl_per_device[jd].n_layer;
+        }
+
         std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
         ngl_per_device_high[id].n_layer = n_unassigned;
         if (hp_nex > 0) {
@@ -488,7 +513,9 @@ static void llama_params_fit_impl(
         if (ngl_per_device_high[id].n_layer > 0) {
             std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
             if (mem_high[id] > targets[id]) {
+                assert(ngl_per_device_high[id].n_layer > ngl_per_device[id].n_layer);
                 uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                LLAMA_LOG_DEBUG("%s: start filling device %" PRIu32 ", delta=%" PRIu32 "\n", __func__, id, delta);
                 while (delta > 1) {
                     uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
                     step_size = std::max(step_size, uint32_t(1));
@@ -502,20 +529,19 @@ static void llama_params_fit_impl(
                     const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
 
                     if (mem_test[id] <= targets[id]) {
-                        ngl_per_device  = ngl_per_device_test;
-                        mem             = mem_test;
-                        n_unassigned   -= ngl_per_device[id].n_layer;
+                        ngl_per_device = ngl_per_device_test;
+                        mem            = mem_test;
                         LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
                     } else {
                         ngl_per_device_high = ngl_per_device_test;
                         mem_high            = mem_test;
-                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device_high[id].n_layer);
                     }
                     delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
                 }
             } else {
-                ngl_per_device  = ngl_per_device_high;
-                n_unassigned   -= ngl_per_device[id].n_layer;
+                assert(ngl_per_device_high[id].n_layer == n_unassigned);
+                ngl_per_device = ngl_per_device_high;
                 LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
             }
         }

src/models/models.h

@@ -441,23 +441,13 @@ private:
                 ggml_tensor * cur,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    ggml_tensor * build_delta_net_recurrent(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                        int   il);
-
     ggml_tensor * build_delta_net_chunking(
                 ggml_tensor * q,
                 ggml_tensor * k,
@@ -467,8 +457,18 @@ private:
                 ggml_tensor * state,
                 ggml_tensor * causal_mask,
                 ggml_tensor * identity,
+                ggml_tensor * diag_mask,
                         int   il);
 
+    ggml_tensor * build_delta_net_autoregressive(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * beta,
+                ggml_tensor * state,
+                int           il);
+
     ggml_tensor * build_norm_gated(
                 ggml_tensor * input,
                 ggml_tensor * weights,

src/models/qwen3next.cpp

@@ -17,13 +17,15 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     ggml_tensor * causal_mask =
-        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens, ubatch.n_seq_tokens), 1.0f),
+        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
                     GGML_TRI_TYPE_LOWER);
 
-    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, ubatch.n_seq_tokens), 1.0f));
+    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
+    ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
 
     ggml_build_forward_expand(gf, causal_mask);
     ggml_build_forward_expand(gf, identity);
+    ggml_build_forward_expand(gf, diag_mask);
 
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
@@ -34,7 +36,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
-            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, il);
+            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
         } else {
             // Full attention layer
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
@@ -93,14 +95,8 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
         ggml_tensor * state,
         ggml_tensor * causal_mask,
         ggml_tensor * identity,
+        ggml_tensor * diag_mask,
         int           il) {
-    GGML_ASSERT(ggml_is_contiguous(q));
-    GGML_ASSERT(ggml_is_contiguous(k));
-    GGML_ASSERT(ggml_is_contiguous(v));
-    GGML_ASSERT(ggml_is_contiguous(g));
-    GGML_ASSERT(ggml_is_contiguous(beta));
-    GGML_ASSERT(ggml_is_contiguous(state));
-
     const int64_t S_k      = q->ne[0];
     const int64_t H_k      = q->ne[1];
     const int64_t n_tokens = q->ne[2];
@@ -120,15 +116,10 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
 
     GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
 
-    // TODO: can this ever be false?
-    const bool use_qk_l2norm = true;
+    const float eps_norm = hparams.f_norm_rms_eps;
 
-    if (use_qk_l2norm) {
-        const float eps_norm = hparams.f_norm_rms_eps;
-
-        q = ggml_l2_norm(ctx0, q, eps_norm);
-        k = ggml_l2_norm(ctx0, k, eps_norm);
-    }
+    q = ggml_l2_norm(ctx0, q, eps_norm);
+    k = ggml_l2_norm(ctx0, k, eps_norm);
 
     const float scale = 1.0f / sqrtf(S_v);
 
@@ -136,8 +127,6 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
 
     beta = ggml_sigmoid(ctx0, beta);
 
-    ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity);
-
     cb(q, "q_in", il);
     cb(k, "k_in", il);
     cb(v, "v_in", il);
@@ -188,36 +177,21 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
     cb(v_beta, "v_beta", il);
     cb(k_beta, "k_beta", il);
 
-    ggml_tensor * chunked_mask =
-        ggml_view_4d(ctx0, causal_mask, chunk_size,
-                chunk_size,         causal_mask->ne[2], causal_mask->ne[3],
-                causal_mask->nb[1], causal_mask->nb[2], causal_mask->nb[3], 0);
+    q      = ggml_reshape_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
+    k      = ggml_reshape_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
+    k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
+    v      = ggml_reshape_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
+    v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
 
-    ggml_tensor * chunked_diag_mask =
-        ggml_view_4d(ctx0, causal_diag_mask, chunk_size,
-                chunk_size,              causal_diag_mask->ne[2], causal_diag_mask->ne[3],
-                causal_diag_mask->nb[1], causal_diag_mask->nb[2], causal_diag_mask->nb[3], 0);
-
-    ggml_tensor * chunked_identity =
-        ggml_view_4d(ctx0, identity, chunk_size,
-            chunk_size,      identity->ne[2], identity->ne[3],
-            identity->nb[1], identity->nb[2], identity->nb[3], 0);
-
-    q      = ggml_cont_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k      = ggml_cont_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k_beta = ggml_cont_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-    v      = ggml_cont_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
-    v_beta = ggml_cont_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
-
-    g    = ggml_cont_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-    beta = ggml_cont_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
+    g    = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
+    beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
 
     ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
 
     cb(g_cumsum, "g_cumsum", il);
 
-    ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
-    ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
+    ggml_tensor * gcs_i = ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
+    ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
 
     ggml_tensor * gcs_j_broadcast =
         ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
@@ -226,23 +200,23 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
 
     cb(decay_mask, "decay_mask", il);
 
-    decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask);
+    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
     decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, chunked_diag_mask);
+    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
 
     ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
 
     ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, chunked_mask));
+    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
 
     cb(attn, "attn_pre_solve", il);
 
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, chunked_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, chunked_identity, attn_lower), attn_lower);
+    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
+    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
 
     ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, chunked_mask);
-    attn                     = ggml_add(ctx0, attn, chunked_identity);
+    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
+    attn                     = ggml_add(ctx0, attn, identity);
 
     cb(attn, "attn_solved", il);
 
@@ -291,7 +265,7 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
         // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
         attn = ggml_mul_mat(ctx0, k_chunk, q_chunk);
         attn = ggml_mul(ctx0, attn, decay_mask_chunk);
-        attn = ggml_mul(ctx0, attn, ggml_add(ctx0, chunked_identity, chunked_mask));
+        attn = ggml_mul(ctx0, attn, diag_mask);
 
         ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
 
@@ -361,23 +335,14 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_chunking(
     return ggml_concat(ctx0, flat_output, flat_state, 0);
 }
 
-ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
+ggml_tensor * llm_build_qwen3next::build_delta_net_autoregressive(
         ggml_tensor * q,
         ggml_tensor * k,
         ggml_tensor * v,
         ggml_tensor * g,
         ggml_tensor * beta,
         ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
         int           il) {
-    GGML_ASSERT(ggml_is_contiguous(q));
-    GGML_ASSERT(ggml_is_contiguous(k));
-    GGML_ASSERT(ggml_is_contiguous(v));
-    GGML_ASSERT(ggml_is_contiguous(g));
-    GGML_ASSERT(ggml_is_contiguous(beta));
-    GGML_ASSERT(ggml_is_contiguous(state));
-
     const int64_t S_k      = q->ne[0];
     const int64_t H_k      = q->ne[1];
     const int64_t n_tokens = q->ne[2];
@@ -386,6 +351,7 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
     const int64_t S_v = v->ne[0];
     const int64_t H_v = v->ne[1];
 
+    GGML_ASSERT(n_tokens == 1);  // This function is optimized for single token processing
     GGML_ASSERT(v->ne[2] == n_tokens);
     GGML_ASSERT(k->ne[2] == n_tokens);
     GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
@@ -397,215 +363,65 @@ ggml_tensor * llm_build_qwen3next::build_delta_net_recurrent(
 
     GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
 
-    // TODO: can this ever be false?
-    const bool use_qk_l2norm = true;
+    const float eps_norm = hparams.f_norm_rms_eps;
 
-    if (use_qk_l2norm) {
-        const float eps_norm = hparams.f_norm_rms_eps;
-
-        q = ggml_l2_norm(ctx0, q, eps_norm);
-        k = ggml_l2_norm(ctx0, k, eps_norm);
-    }
+    q = ggml_l2_norm(ctx0, q, eps_norm);
+    k = ggml_l2_norm(ctx0, k, eps_norm);
 
     const float scale = 1.0f / sqrtf(S_v);
 
-    q = ggml_scale(ctx0, q, scale);
-
+    q    = ggml_scale(ctx0, q, scale);
     beta = ggml_sigmoid(ctx0, beta);
 
-    ggml_tensor * causal_diag_mask = ggml_add(ctx0, causal_mask, identity);
-
     cb(q, "q_in", il);
     cb(k, "k_in", il);
     cb(v, "v_in", il);
     cb(beta, "beta_in", il);
     cb(g, "g_in", il);
 
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
-
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
     state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
 
-    cb(q, "q_perm", il);
-    cb(k, "k_perm", il);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(g, "g_perm", il);
-    cb(state, "state_in", il);
+    ggml_tensor * g_t    = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
+    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
 
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
+    // Apply exponential to g_t
+    g_t = ggml_exp(ctx0, g_t);
 
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
+    // Apply the gated delta rule for the single timestep
+    // last_recurrent_state = last_recurrent_state * g_t
+    state = ggml_mul(ctx0, state, g_t);
 
-    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
+    // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
+    ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
+    ggml_tensor * kv_mem         = ggml_mul(ctx0, state, k_t_unsqueezed);
+    // we need to sum over dim=-2, so we transpose, sum, then transpose again
+    kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
 
-    cb(k_beta, "k_beta", il);
-    cb(v_beta, "v_beta", il);
-    cb(g_cumsum, "g_cumsum", il);
+    // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
+    ggml_tensor * v_t    = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
+    // delta = (v_t - kv_mem) * beta_t
+    ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);  // both should be [S_v, 1, H_v, n_seqs]
+    ggml_tensor * delta  = ggml_mul(ctx0, v_diff, beta_t);
 
-    ggml_tensor * gcs_i = ggml_cont_4d(ctx0, g_cumsum, n_tokens, 1, H_v, n_seqs);  // [chunk_size, 1, n_tokens, n_seqs]
-    ggml_tensor * gcs_j = ggml_cont_4d(ctx0, g_cumsum, 1, n_tokens, H_v, n_seqs);  // [1, chunk_size, n_tokens, n_seqs]
+    // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
+    ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
+    state                   = ggml_add(ctx0, state, k_t_delta);
 
-    // Broadcast both tensors to [chunk_size, chunk_size, H_v, n_seqs]
-    // ggml_tensor * gcs_i_broadcast =
-    //     ggml_repeat_4d(ctx0, gcs_i, GGML_DELTA_NET_CHUNK, GGML_DELTA_NET_CHUNK, num_chunks * H_v,
-    //                     n_seqs);  // [chunk_size, 1, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs]
-    // Don't need this, this one will get auto-broadcast
-    ggml_tensor * gcs_j_broadcast =
-        ggml_repeat_4d(ctx0, gcs_j, n_tokens, n_tokens, H_v, n_seqs);  // [1, chunk_size, H_v, n_seqs] -> [chunk_size, chunk_size, H_v, n_seqs]
-
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
-
-    // Apply lower triangular mask to ensure attention is causal (only past tokens influence current)
-    decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask);
-    // Apply exponential to get the decay mask values
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    // Apply lower triangular mask again to ensure only lower triangular values remain
-    decay_mask = ggml_mul(ctx0, decay_mask, causal_diag_mask);
-
-    cb(decay_mask, "decay_mask", il);
-
-    // attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0)
-    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
-
-    cb(kmulkbeta, "kmulkbeta", il);
-
-    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
-
-    cb(attn, "attn_pre_rec", il);
-
-    // for i in range(1, chunk_size):
-    //          row = attn[..., i, :i].clone()
-    //          sub = attn[..., :i, :i].clone()
-    //          attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
-    // attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
-    //
-    // We reduce this to a linear triangular solve: AX = B, where B = attn, A = I - tril(A)
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
-
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, identity);
-
-    // value = attn @ v_beta
-    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
-
-    cb(v, "value_beta", il);
-
-    // k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
-    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
-    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
-
-    cb(gexp, "g_cum_exp", il);
-
-    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
-
-    cb(kbeta_gexp, "kbeta_gexp", il);
-
-    ggml_tensor * k_cumdecay =
-        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
-
-    cb(k_cumdecay, "k_cumdecay", il);
-
-    // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
-    attn = ggml_mul_mat(ctx0, k, q);
-    attn = ggml_mul(ctx0, attn, decay_mask);
-    attn = ggml_mul(ctx0, attn, ggml_add(ctx0, identity, causal_mask));
-
-    cb(attn, "attn_decay_key", il);
-
-    ggml_tensor * state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
-
-    // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
-    ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay);
-
-    cb(v_prime, "v_prime", il);
-
-    // v_new = v_i - v_prime
-    ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v, v_prime), v_prime);
-
-    ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-
-    cb(v_new, "v_new", il);
-
-    // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-    ggml_tensor * q_g_exp    = ggml_mul(ctx0, q, gexp);
-    ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
-
-    cb(attn_inter, "attn_inter", il);
-
-    // core_attn_out[:, :, i] = attn_inter + attn @ v_new
-    ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn);
-
-    cb(v_attn, "v_attn", il);
-
-    ggml_tensor * core_attn_out = ggml_add(ctx0, attn_inter, v_attn);
-
-    cb(core_attn_out, "core_attn_out", il);
-
-    // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
-    // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
-    // key_gdiff = key * g_diff.unsqueeze(-1)
-    // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-    // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-
-    ggml_tensor * g_cum_last =
-        ggml_cont(ctx0, ggml_view_4d(ctx0, g_cumsum_t, g_cumsum_t->ne[0], 1, g_cumsum_t->ne[2], g_cumsum_t->ne[3],
-                                    g_cumsum_t->nb[1], g_cumsum_t->nb[2], g_cumsum_t->nb[3],
-                                    g_cumsum_t->nb[0] * (g_cumsum_t->ne[1] - 1)));
-
-    cb(g_cum_last, "g_cum_last", il);
-
-    ggml_tensor * gexp_last =
-        ggml_reshape_4d(ctx0, ggml_exp(ctx0, g_cum_last), 1, 1, g_cum_last->ne[0] * g_cum_last->ne[2], g_cum_last->ne[3]);
-
-    cb(gexp_last, "gexp_last", il);
-
-    ggml_tensor * g_cum_last_3d =
-        ggml_reshape_3d(ctx0, g_cum_last, g_cum_last->ne[0], g_cum_last->ne[2], g_cum_last->ne[3]);
-
-    cb(g_cum_last_3d, "g_cum_last_3d", il);
-
-    ggml_tensor * g_cumsum_3d = ggml_reshape_3d(ctx0, g_cumsum, g_cumsum->ne[0], g_cumsum->ne[2], g_cumsum->ne[3]);
-
-    cb(g_cumsum_3d, "g_cumsum_3d", il);
-
-    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum_3d, g_cum_last_3d));
-
-    cb(g_diff, "g_diff", il);
-
-    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-
-    cb(g_diff_exp, "g_diff_exp", il);
-
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k,
-                                    ggml_reshape_4d(ctx0, g_diff_exp, 1, g_diff_exp->ne[0], g_diff_exp->ne[1],
-                                                    g_diff_exp->ne[2] * g_diff_exp->ne[3]));
-
-    cb(key_gdiff, "key_gdiff", il);
-
-    ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)));
-
-    cb(kgdmulvnew, "kgdmulvnew", il);
-
-    state = ggml_add(ctx0, ggml_mul(ctx0, state, gexp_last), kgdmulvnew);
+    // Compute the attention output
+    // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
+    ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);  // unsqueeze q_t
+    ggml_tensor * state_q        = ggml_mul(ctx0, state, q_t_unsqueezed);
+    // again, since it's over dim = -2, transpose, sum, transpose back
+    ggml_tensor * core_attn_out =
+        ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
 
+    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
+    cb(core_attn_out, "output_tokens", il);
     cb(state, "new_state", il);
 
-    // flatten output
-    ggml_tensor * flat_output =
-        ggml_cont_1d(ctx0, ggml_permute(ctx0, core_attn_out, 0, 2, 1, 3), S_v * H_v * n_tokens * n_seqs);
-
-    ggml_tensor * flat_state = ggml_cont_1d(ctx0, state, S_v * S_v * H_v * n_seqs);
+    // flatten output, no need to permute since n_tokens is 1 so [S_v, 1, H_v, n_seqs] and [S_v, H_v, 1, n_seqs] are equivalent memory-layout wise
+    ggml_tensor * flat_output = ggml_reshape_1d(ctx0, core_attn_out, S_v * H_v * n_tokens * n_seqs);
+    ggml_tensor * flat_state  = ggml_reshape_1d(ctx0, state, S_v * S_v * H_v * n_seqs);
 
     return ggml_concat(ctx0, flat_output, flat_state, 0);
 }
@@ -712,6 +528,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
         ggml_tensor *        cur,
         ggml_tensor *        causal_mask,
         ggml_tensor *        identity,
+        ggml_tensor *        diag_mask,
         int                  il) {
     const auto * mctx_cur = inp->mctx;
 
@@ -737,11 +554,11 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(mixed_ba, "linear_attn_mixed_ba", il);
 
     int64_t       qkvz_new_dim        = 2 * head_k_dim + 2 * head_v_dim * (num_v_heads / num_k_heads);
-    ggml_tensor * mixed_qkvz_reshaped = ggml_cont_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * mixed_qkvz_reshaped = ggml_reshape_4d(ctx0, mixed_qkvz, qkvz_new_dim, num_k_heads, n_seq_tokens, n_seqs);
 
     // Reshape mixed_ba: [batch, seq_len, hidden_size] -> [batch, seq_len, num_k_heads, 2*num_v_heads/num_k_heads]
     int64_t       ba_new_dim        = 2 * num_v_heads / num_k_heads;
-    ggml_tensor * mixed_ba_reshaped = ggml_cont_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);
+    ggml_tensor * mixed_ba_reshaped = ggml_reshape_4d(ctx0, mixed_ba, ba_new_dim, num_k_heads, n_seq_tokens, n_seqs);
 
     // Split mixed_ba into b and a (beta and alpha parameters)
     int64_t split_sizes_ba[2] = {
@@ -762,8 +579,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * beta  = ggml_cont_3d(ctx0, b, num_v_heads, n_seq_tokens, n_seqs);
     ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
 
-    GGML_ASSERT(ggml_nelements(beta) + ggml_nelements(alpha) == ggml_nelements(mixed_ba));
-
     ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
     ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
     cb(alpha_softplus, "a_softplus", il);
@@ -799,9 +614,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
                                    (split_sizes_qkvz[0] + split_sizes_qkvz[1] + split_sizes_qkvz[2]) * sizeof(float));
     cb(z, "z", il);
 
-    GGML_ASSERT(ggml_nelements(query) + ggml_nelements(key) + ggml_nelements(value) + ggml_nelements(z) ==
-                ggml_nelements(mixed_qkvz));
-
     // After creating query, key, and value_reshaped, reshape each to flatten the head dimensions
     // query: [head_k_dim, num_k_heads, n_tokens, n_seqs] -> [head_k_dim * num_k_heads, n_tokens, n_seqs]
     ggml_tensor * query_flat = ggml_cont_3d(ctx0, query, head_k_dim * num_k_heads, n_seq_tokens, n_seqs);
@@ -925,10 +737,13 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
-    ggml_tensor * attn_out = n_seq_tokens > CHUNK_SIZE ?
-        build_delta_net_chunking (q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il) :
-        build_delta_net_recurrent(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, il);
+    // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
+    ggml_tensor * attn_out;
+    if (n_seq_tokens == 1) {
+        attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
+    } else {
+        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
+    }
     cb(attn_out, "attn_out", il);
 
     // The tensors were concatenated 1d, so we need to extract them 1d as well

tools/completion/completion.cpp

@@ -87,9 +87,6 @@ int main(int argc, char ** argv) {
     common_params params;
     g_params = &params;
 
-    // disable jinja by default
-    params.use_jinja = false;
-
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMPLETION, print_usage)) {
         return 1;
     }

tools/fit-params/fit-params.cpp

@@ -4,7 +4,11 @@
 #include "common.h"
 #include "log.h"
 
-#include <iostream>
+#include <chrono>
+#include <cinttypes>
+#include <thread>
+
+using namespace std::chrono_literals;
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -22,13 +26,17 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
     auto mparams = common_model_params_to_llama(params);
     auto cparams = common_context_params_to_llama(params);
-    llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+    const bool success = llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
         params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
         params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+    if (!success) {
+        LOG_ERR("%s: failed to fit CLI arguments to free memory, exiting...\n", __func__);
+        exit(1);
+    }
 
-    LOG_INF("Printing fitted CLI arguments to stdout...\n");
-    std::cout << "-c "    << cparams.n_ctx;
-    std::cout << " -ngl " << mparams.n_gpu_layers;
+    LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__);
+    std::this_thread::sleep_for(10ms); // to avoid a race between stderr and stdout
+    printf("-c %" PRIu32 " -ngl %" PRIu32, cparams.n_ctx, mparams.n_gpu_layers);
 
     size_t nd = llama_max_devices();
     while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) {
@@ -37,26 +45,22 @@ int main(int argc, char ** argv) {
     if (nd > 1) {
         for (size_t id = 0; id < nd; id++) {
             if (id == 0) {
-                std::cout << " -ts ";
+                printf(" -ts ");
             }
-            if (id > 0) {
-                std::cout << ",";
-            }
-            std::cout << mparams.tensor_split[id];
+            printf("%s%" PRIu32, id > 0 ? "," : "", uint32_t(mparams.tensor_split[id]));
         }
     }
 
     const size_t ntbo = llama_max_tensor_buft_overrides();
+    bool any_tbo = false;
     for (size_t itbo = 0; itbo < ntbo && mparams.tensor_buft_overrides[itbo].pattern != nullptr; itbo++) {
         if (itbo == 0) {
-            std::cout << " -ot ";
+            printf(" -ot \"");
         }
-        if (itbo > 0) {
-            std::cout << ",";
-        }
-        std::cout << mparams.tensor_buft_overrides[itbo].pattern << "=" << ggml_backend_buft_name(mparams.tensor_buft_overrides[itbo].buft);
+        printf("%s%s=%s", itbo > 0 ? "," : "", mparams.tensor_buft_overrides[itbo].pattern, ggml_backend_buft_name(mparams.tensor_buft_overrides[itbo].buft));
+        any_tbo = true;
     }
-    std::cout << "\n";
+    printf("%s\n", any_tbo ? "\"" : "");
 
     return 0;
 }

tools/mtmd/mtmd-cli.cpp

@@ -270,8 +270,6 @@ int main(int argc, char ** argv) {
     ggml_time_init();
 
     common_params params;
-    params.use_jinja = false;   // disable jinja by default
-    params.sampling.temp = 0.2; // lower temp by default for better quality
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MTMD, show_additional_info)) {
         return 1;

tools/server/README.md

@@ -52,7 +52,6 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
 | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
 | `--swa-full` | use full-size SWA cache (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)<br/>(env: LLAMA_ARG_SWA_FULL) |
-| `--kv-unified, -kvu` | use single unified KV buffer for the KV cache of all sequences (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)<br/>(env: LLAMA_ARG_KV_UNIFIED) |
 | `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `--perf, --no-perf` | whether to enable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_PERF) |
 | `-e, --escape, --no-escape` | whether to process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
@@ -67,11 +66,10 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
-| `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_HOST) |
+| `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
-| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--mmap, --no-mmap` | whether to memory-map model (if disabled, slower load but may reduce pageouts if not using mlock) (default: enabled)<br/>(env: LLAMA_ARG_MMAP) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
@@ -150,19 +148,20 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
 
 
-**Example-specific params**
+**Server-specific params**
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--ctx-checkpoints, --swa-checkpoints N` | max number of context checkpoints to create per slot (default: 8)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_CTX_CHECKPOINTS) |
-| `--cache-ram, -cram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
+| `--ctx-checkpoints, --swa-checkpoints N` | max number of context checkpoints to create per slot (default: 8)[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_CTX_CHECKPOINTS) |
+| `--cache-ram, -cram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
+| `--kv-unified, -kvu` | use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)<br/>(env: LLAMA_ARG_KV_UNIFIED) |
 | `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
 | `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode<br/> |
 | `-sp, --special` | special tokens output enabled (default: false) |
 | `--warmup, --no-warmup` | whether to perform warmup with an empty run (default: enabled) |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
 | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified<br/>(env: LLAMA_ARG_POOLING) |
-| `-cb, --cont-batching, -nocb, --no-cont-batching` | whether to enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
+| `-np, --parallel N` | number of server slots (default: -1, -1 = auto)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `-cb, --cont-batching, -nocb, --no-cont-batching` | whether to enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-mm, --mmproj FILE` | path to a multimodal projector file. see tools/mtmd/README.md<br/>note: if -hf is used, this argument can be omitted<br/>(env: LLAMA_ARG_MMPROJ) |
 | `-mmu, --mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
@@ -1430,7 +1429,7 @@ Model presets allow advanced users to define custom configurations using an `.in
 llama-server --models-preset ./my-models.ini
 ```
 
-Each section in the file defines a new preset. Keys within a section correspond to command-line arguments (without leading dashes). For example, the argument `--n-gpu-layer 123` is written as `n-gpu-layer = 123`.
+Each section in the file defines a new preset. Keys within a section correspond to command-line arguments (without leading dashes). For example, the argument `--n-gpu-layers 123` is written as `n-gpu-layers = 123`.
 
 Short argument forms (e.g., `c`, `ngl`) and environment variable names (e.g., `LLAMA_ARG_N_GPU_LAYERS`) are also supported as keys.
 
@@ -1445,7 +1444,7 @@ version = 1
 ; string value
 chat-template = chatml
 ; numeric value
-n-gpu-layer = 123
+n-gpu-layers = 123
 ; flag value (for certain flags, you need to use the "no-" prefix for negation)
 jinja = true
 ; shorthand argument (for example, context size)

tools/server/server.cpp

@@ -73,12 +73,17 @@ int main(int argc, char ** argv, char ** envp) {
         return 1;
     }
 
-    // TODO: should we have a separate n_parallel parameter for the server?
-    //       https://github.com/ggml-org/llama.cpp/pull/16736#discussion_r2483763177
-    // TODO: this is a common configuration that is suitable for most local use cases
-    //       however, overriding the parameters is a bit confusing - figure out something more intuitive
-    if (params.n_parallel == 1 && params.kv_unified == false && !params.has_speculative()) {
-        LOG_WRN("%s: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)\n", __func__);
+    // validate batch size for embeddings
+    // embeddings require all tokens to be processed in a single ubatch
+    // see https://github.com/ggml-org/llama.cpp/issues/12836
+    if (params.embedding && params.n_batch > params.n_ubatch) {
+        LOG_WRN("%s: embeddings enabled with n_batch (%d) > n_ubatch (%d)\n", __func__, params.n_batch, params.n_ubatch);
+        LOG_WRN("%s: setting n_batch = n_ubatch = %d to avoid assertion failure\n", __func__, params.n_ubatch);
+        params.n_batch = params.n_ubatch;
+    }
+
+    if (params.n_parallel < 0) {
+        LOG_INF("%s: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true\n", __func__);
 
         params.n_parallel = 4;
         params.kv_unified = true;