vulkan: Convert FILL to distribute workgroups in 2D to avoid exceeding maxComputeWorkGroupCount (#28592 )

* divide workload to 2D This is to workaround FILL exceeding maxComputeWorkGroupCount for Intel GPUs on Qwen 3.8 flash next * minor change * Fixed comment
py : lower numpy to 2.2.6 (#28654 )
2026-09-09 17:17:09 +02:00 · 2026-09-09 16:54:15 +02:00 · 2026-09-09 15:56:27 +02:00 · 2026-09-09 14:59:37 +02:00 · 2026-09-09 15:27:25 +03:00 · 2026-09-09 13:25:54 +02:00
32 changed files with 257 additions and 236 deletions
@@ -872,17 +872,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
                    arg.c_str(), e.what(), opt.to_string().c_str()));
            }
        }
-
-        // TODO: remove this check after deprecating --mmap|mlock|dio
-        auto has_arg = [&](std::initializer_list<const char *> names) {
-            return std::any_of(names.begin(), names.end(), [&](const char * name) {
-                return seen_args.count(name);
-            });
-        };
-        if (has_arg({"-lm", "--load-mode"}) &&
-            has_arg({"--mlock", "--mmap", "--no-mmap", "-dio", "--direct-io", "-ndio", "--no-direct-io"})) {
-            LOG_WRN("DEPRECATED: `--load-mode` and `--mlock`/`--mmap`/`--direct-io` should not be combined; only the last flag on the command line will take effect\n");
-        }
    };

    // parse all CLI args now, so that -hf is available below for remote preset resolution
@@ -2694,32 +2683,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            }
        ).set_env("LLAMA_ARG_RPC"));
    }
-    add_opt(common_arg(
-        {"--mlock"},
-        "DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing",
-        [](common_params & params) {
-            LOG_WRN("DEPRECATED: --mlock is deprecated. use --load-mode mlock instead\n");
-            params.load_mode = LLAMA_LOAD_MODE_MLOCK;
-        }
-    ).set_env("LLAMA_ARG_MLOCK"));
-    add_opt(common_arg(
-        {"--mmap"},
-        {"--no-mmap"},
-        "DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)",
-        [](common_params & params, bool value) {
-            LOG_WRN("DEPRECATED: --mmap and --no-mmap are deprecated. use --load-mode mmap instead\n");
-            params.load_mode = value ? LLAMA_LOAD_MODE_MMAP : LLAMA_LOAD_MODE_NONE;
-        }
-    ).set_env("LLAMA_ARG_MMAP"));
-    add_opt(common_arg(
-        {"-dio", "--direct-io"},
-        {"-ndio", "--no-direct-io"},
-        "DEPRECATED in favor of `--load-mode`: use DirectIO if available",
-        [](common_params & params, bool value) {
-            LOG_WRN("DEPRECATED: --direct-io and --no-direct-io are deprecated. use --load-mode dio instead\n");
-            params.load_mode = value ? LLAMA_LOAD_MODE_DIRECT_IO : LLAMA_LOAD_MODE_NONE;
-        }
-    ).set_env("LLAMA_ARG_DIO"));
    add_opt(common_arg(
        {"-lm", "--load-mode"}, "MODE",
        "model loading mode (default: auto)\n"
@@ -167,6 +167,12 @@ value binary_expression::execute_impl(context & ctx) {
        }
        throw std::runtime_error("Cannot perform operation " + op.value + " on undefined values");
    } else if (is_val<value_none>(left_val) || is_val<value_none>(right_val)) {
+        if (!is_val<value_none>(right_val) && (op.value == "in" || op.value == "not in")) {
+            // case: none in {'low': 1}
+            // A null left operand is looked up like any other value.
+            bool member = test_is_in();
+            return mk_val<value_bool>(op.value == "in" ? member : !member);
+        }
        if (op.value == "+" || op.value == "~") {
            value res = mk_val<value_undefined>();
            if (workaround_concat_null_with_str(res)) {
@@ -300,7 +300,8 @@ The following compilation options are also available to tweak performance:
 |-------------------------------|------------------------|---------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
 | GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for V100, CDNA and RDNA4 which use FP32 compute type by default) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000).   |
-| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                                                                                                           |
+| GGML_CUDA_FA_QUANTS           | `all` or `type_K-type_V` list | q4_0-q4_0;q8_0-q8_0;f16-f16;bf16-bf16 | Select which K/V type combinations to compile the FlashAttention CUDA kernels for. `all` compiles every combination, but compilation takes much longer. Otherwise a `;`-separated list of `type_K-type_V` pairs; f16-f16 is always compiled. Combinations that were not compiled fall back to f16-f16 kernel with a warning. Legal types: f16, bf16, q4_0, q4_1, q5_0, q5_1, q8_0. |
+| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Deprecated alias for `GGML_CUDA_FA_QUANTS=all`.                                                                                                                                                                                                                                                                                                                               |

 ## MUSA

@@ -204,6 +204,8 @@ option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copie
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
 option(GGML_CUDA_FA                         "ggml: compile ggml FlashAttention CUDA kernels"  ON)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
+set   (GGML_CUDA_FA_QUANTS "q4_0-q4_0;q8_0-q8_0;f16-f16;bf16-bf16" CACHE STRING
+                                            "ggml: FlashAttention K-V type combinations to compile, \"all\" or a list such as \"q8_0-q8_0;q8_0-q4_0\"")
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 option(GGML_CUDA_NCCL                       "ggml: use NVIDIA Collective Comm. Library"       ON)
 set   (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING
@@ -48,3 +48,74 @@ function(ggml_get_system_arch)
        set(GGML_SYSTEM_ARCH "UNKNOWN" PARENT_SCOPE)
    endif()
 endfunction()
+
+# Determines which FlashAttention vector kernel template instances to compile, returns them in OUT_SRCS.
+function(ggml_cuda_fattn_vec_instances DIR OUT_SRCS)
+    set(FA_TYPES q4_0 q4_1 q5_0 q5_1 q8_0 bf16 f16)
+
+    string(TOLOWER "${GGML_CUDA_FA_QUANTS}" FA_QUANTS)
+    string(STRIP   "${FA_QUANTS}" FA_QUANTS)
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        message(WARNING "GGML_CUDA_FA_ALL_QUANTS is deprecated, use GGML_CUDA_FA_QUANTS=all instead")
+        set(FA_QUANTS all)
+    endif()
+    if (NOT FA_QUANTS)
+        message(FATAL_ERROR "GGML_CUDA_FA_QUANTS must not be empty")
+    endif()
+
+    if (FA_QUANTS STREQUAL "all")
+        set(FA_COMBINATIONS "")
+        foreach (TYPE_V IN LISTS FA_TYPES)
+            foreach (TYPE_K IN LISTS FA_TYPES)
+                list(APPEND FA_COMBINATIONS ${TYPE_K}-${TYPE_V})
+            endforeach()
+        endforeach()
+    else()
+        set(FA_COMBINATIONS f16-f16)
+
+        string(REPLACE "," ";" FA_SELECTED "${FA_QUANTS}")
+        foreach (COMBINATION IN LISTS FA_SELECTED)
+            string(STRIP   "${COMBINATION}" COMBINATION)
+            if (NOT COMBINATION MATCHES "^([a-z0-9_]+)-([a-z0-9_]+)$")
+                message(FATAL_ERROR "GGML_CUDA_FA_QUANTS: \"${COMBINATION}\" is not \"all\" or a <type_K>-<type_V> combination")
+            endif()
+            set(TYPE_K ${CMAKE_MATCH_1})
+            set(TYPE_V ${CMAKE_MATCH_2})
+            foreach (TYPE ${TYPE_K} ${TYPE_V})
+                if (NOT TYPE IN_LIST FA_TYPES)
+                    message(FATAL_ERROR
+                        "GGML_CUDA_FA_QUANTS: unknown type \"${TYPE}\" in \"${COMBINATION}\", must be one of: ${FA_TYPES}")
+                endif()
+            endforeach()
+            list(APPEND FA_COMBINATIONS ${TYPE_K}-${TYPE_V})
+        endforeach()
+    endif()
+    list(REMOVE_DUPLICATES FA_COMBINATIONS)
+
+    string(REPLACE ";" "," FA_QUANTS_DEFINE "${FA_QUANTS}")
+    add_compile_definitions(GGML_CUDA_FA_QUANTS="${FA_QUANTS_DEFINE}")
+    foreach (TYPE_V IN LISTS FA_TYPES)
+        foreach (TYPE_K IN LISTS FA_TYPES)
+            if ("${TYPE_K}-${TYPE_V}" IN_LIST FA_COMBINATIONS)
+                set(COMPILED 1)
+            else()
+                set(COMPILED 0)
+            endif()
+            string(TOUPPER "GGML_CUDA_FA_${TYPE_K}_${TYPE_V}" COMBINATION_DEF)
+            add_compile_definitions(${COMBINATION_DEF}=${COMPILED})
+        endforeach()
+    endforeach()
+
+    message(STATUS "FlashAttention K-V type combinations: ${FA_COMBINATIONS}")
+
+    set(SRCS "")
+    foreach (COMBINATION IN LISTS FA_COMBINATIONS)
+        set(SRC "${DIR}/template-instances/fattn-vec-instance-${COMBINATION}.cu")
+        if (NOT EXISTS "${SRC}")
+            message(FATAL_ERROR "FlashAttention template instance \"${SRC}\" does not exist")
+        endif()
+        list(APPEND SRCS "${SRC}")
+    endforeach()
+
+    set(${OUT_SRCS} ${SRCS} PARENT_SCOPE)
+endfunction()
@@ -112,17 +112,8 @@ if (CUDAToolkit_FOUND)
    file(GLOB   SRCS "template-instances/mmf*.cu")
    list(APPEND GGML_SOURCES_CUDA ${SRCS})

-    if (GGML_CUDA_FA_ALL_QUANTS)
-        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-    else()
-        list(APPEND GGML_SOURCES_CUDA
-            template-instances/fattn-vec-instance-f16-f16.cu
-            template-instances/fattn-vec-instance-q4_0-q4_0.cu
-            template-instances/fattn-vec-instance-q8_0-q8_0.cu
-            template-instances/fattn-vec-instance-bf16-bf16.cu)
-    endif()
+    ggml_cuda_fattn_vec_instances(${CMAKE_CURRENT_SOURCE_DIR} SRCS)
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})

    ggml_add_backend_library(ggml-cuda
                             ${GGML_HEADERS_CUDA}
@@ -374,90 +374,101 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
    }
 }

-#define FATTN_VEC_CASE(D, type_K, type_V)                                                                        \
-    {                                                                                                            \
-        const bool type_K_okay = K->type == (type_K) || (K->type == GGML_TYPE_F32 && (type_K) == GGML_TYPE_F16); \
-        const bool type_V_okay = V->type == (type_V) || (V->type == GGML_TYPE_F32 && (type_V) == GGML_TYPE_F16); \
-        if (Q->ne[0] == (D) && type_K_okay && type_V_okay) {                                                     \
-            ggml_cuda_flash_attn_ext_vec_case<D, type_K, type_V>(ctx, dst);                                      \
-            return;                                                                                              \
-        }                                                                                                        \
-    }                                                                                                            \
+#define FATTN_VEC_CASE(D, type_K_case, type_V_case)                                                                                \
+    if constexpr (GGML_CUDA_FA_##type_K_case##_##type_V_case) {                                                                    \
+        const bool type_K_okay = type_K == GGML_TYPE_##type_K_case || (type_K == GGML_TYPE_F32 && GGML_TYPE_##type_K_case == GGML_TYPE_F16); \
+        const bool type_V_okay = type_V == GGML_TYPE_##type_V_case || (type_V == GGML_TYPE_F32 && GGML_TYPE_##type_V_case == GGML_TYPE_F16); \
+        if (head_size == (D) && type_K_okay && type_V_okay) {                                                                      \
+            return ggml_cuda_flash_attn_ext_vec_case<D, GGML_TYPE_##type_K_case, GGML_TYPE_##type_V_case>;                         \
+        }                                                                                                                          \
+    }                                                                                                                              \

-#define FATTN_VEC_CASES_ALL_D(type_K, type_V) \
-    FATTN_VEC_CASE( 64, type_K, type_V)       \
-    FATTN_VEC_CASE(128, type_K, type_V)       \
-    FATTN_VEC_CASE(256, type_K, type_V)       \
+#define FATTN_VEC_CASES_ALL_D(type_K_case, type_V_case) \
+    FATTN_VEC_CASE( 64, type_K_case, type_V_case)       \
+    FATTN_VEC_CASE(128, type_K_case, type_V_case)       \
+    FATTN_VEC_CASE(256, type_K_case, type_V_case)       \
+
+typedef void (* fattn_vec_case_t)(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+// Vector kernel for the given head size and K/V types, nullptr if its template instance was not compiled:
+static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, const ggml_type type_K, const ggml_type type_V) {
+    FATTN_VEC_CASES_ALL_D(F16,  F16)
+    FATTN_VEC_CASES_ALL_D(Q4_0, F16)
+    FATTN_VEC_CASES_ALL_D(Q4_1, F16)
+    FATTN_VEC_CASES_ALL_D(Q5_0, F16)
+    FATTN_VEC_CASES_ALL_D(Q5_1, F16)
+    FATTN_VEC_CASES_ALL_D(Q8_0, F16)
+    FATTN_VEC_CASES_ALL_D(BF16, F16)
+
+    FATTN_VEC_CASES_ALL_D(F16,  Q4_0)
+    FATTN_VEC_CASES_ALL_D(Q4_0, Q4_0)
+    FATTN_VEC_CASES_ALL_D(Q4_1, Q4_0)
+    FATTN_VEC_CASES_ALL_D(Q5_0, Q4_0)
+    FATTN_VEC_CASES_ALL_D(Q5_1, Q4_0)
+    FATTN_VEC_CASES_ALL_D(Q8_0, Q4_0)
+    FATTN_VEC_CASES_ALL_D(BF16, Q4_0)
+
+    FATTN_VEC_CASES_ALL_D(F16,  Q4_1)
+    FATTN_VEC_CASES_ALL_D(Q4_0, Q4_1)
+    FATTN_VEC_CASES_ALL_D(Q4_1, Q4_1)
+    FATTN_VEC_CASES_ALL_D(Q5_0, Q4_1)
+    FATTN_VEC_CASES_ALL_D(Q5_1, Q4_1)
+    FATTN_VEC_CASES_ALL_D(Q8_0, Q4_1)
+    FATTN_VEC_CASES_ALL_D(BF16, Q4_1)
+
+    FATTN_VEC_CASES_ALL_D(F16,  Q5_0)
+    FATTN_VEC_CASES_ALL_D(Q4_0, Q5_0)
+    FATTN_VEC_CASES_ALL_D(Q4_1, Q5_0)
+    FATTN_VEC_CASES_ALL_D(Q5_0, Q5_0)
+    FATTN_VEC_CASES_ALL_D(Q5_1, Q5_0)
+    FATTN_VEC_CASES_ALL_D(Q8_0, Q5_0)
+    FATTN_VEC_CASES_ALL_D(BF16, Q5_0)
+
+    FATTN_VEC_CASES_ALL_D(F16,  Q5_1)
+    FATTN_VEC_CASES_ALL_D(Q4_0, Q5_1)
+    FATTN_VEC_CASES_ALL_D(Q4_1, Q5_1)
+    FATTN_VEC_CASES_ALL_D(Q5_0, Q5_1)
+    FATTN_VEC_CASES_ALL_D(Q5_1, Q5_1)
+    FATTN_VEC_CASES_ALL_D(Q8_0, Q5_1)
+    FATTN_VEC_CASES_ALL_D(BF16, Q5_1)
+
+    FATTN_VEC_CASES_ALL_D(F16,  Q8_0)
+    FATTN_VEC_CASES_ALL_D(Q4_0, Q8_0)
+    FATTN_VEC_CASES_ALL_D(Q4_1, Q8_0)
+    FATTN_VEC_CASES_ALL_D(Q5_0, Q8_0)
+    FATTN_VEC_CASES_ALL_D(Q5_1, Q8_0)
+    FATTN_VEC_CASES_ALL_D(Q8_0, Q8_0)
+    FATTN_VEC_CASES_ALL_D(BF16, Q8_0)
+
+    FATTN_VEC_CASES_ALL_D(F16,  BF16)
+    FATTN_VEC_CASES_ALL_D(Q4_0, BF16)
+    FATTN_VEC_CASES_ALL_D(Q4_1, BF16)
+    FATTN_VEC_CASES_ALL_D(Q5_0, BF16)
+    FATTN_VEC_CASES_ALL_D(Q5_1, BF16)
+    FATTN_VEC_CASES_ALL_D(Q8_0, BF16)
+    FATTN_VEC_CASES_ALL_D(BF16, BF16)
+
+    return nullptr;
+}

 static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * Q = dst->src[0];
-    ggml_tensor * K = dst->src[1];
-    ggml_tensor * V = dst->src[2];
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];

-#ifdef GGML_CUDA_FA_ALL_QUANTS
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_F16)
-
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q4_0)
-
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q4_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q4_1)
-
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q5_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q5_0)
-
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q5_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q5_1)
-
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q8_0)
-
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_BF16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_BF16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_BF16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_BF16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_BF16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_BF16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_BF16)
-#else
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_F16)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_BF16)
-#endif // GGML_CUDA_FA_ALL_QUANTS
-
-    GGML_ABORT("fatal error");
+    fattn_vec_case_t vec_case = ggml_cuda_get_fattn_vec_case(Q->ne[0], K->type, V->type);
+    if (vec_case == nullptr) {
+        static bool warned = false;
+        if (!warned) {
+            GGML_LOG_WARN("%s: no FlashAttention vector kernel compiled for K/V types %s-%s, converting K and V to f16 instead (slow). "
+                "Add \"%s-%s\" to GGML_CUDA_FA_QUANTS to compile it.\n",
+                __func__, ggml_type_name(K->type), ggml_type_name(V->type), ggml_type_name(K->type), ggml_type_name(V->type));
+            warned = true;
+        }
+        vec_case = ggml_cuda_get_fattn_vec_case(Q->ne[0], GGML_TYPE_F16, GGML_TYPE_F16);
+    }
+    GGML_ASSERT(vec_case != nullptr);
+    vec_case(ctx, dst);
 }

 // Best FlashAttention kernel for a specific GPU:
@@ -468,20 +479,17 @@ enum best_fattn_kernel {
    BEST_FATTN_KERNEL_MMA_F16 = 400,
 };

-static bool ggml_cuda_fattn_kv_type_supported(ggml_type type) {
+// K/V types for which there is a vector kernel template instance, other kernels convert these to f16:
+static bool ggml_cuda_fattn_kv_type_supported(const ggml_type type) {
    switch (type) {
        case GGML_TYPE_F32:
        case GGML_TYPE_F16:
-            return true;
+        case GGML_TYPE_BF16:
+        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
        case GGML_TYPE_Q5_1:
-#ifndef GGML_CUDA_FA_ALL_QUANTS
-            return false;
-#endif // GGML_CUDA_FA_ALL_QUANTS
-        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q8_0:
-        case GGML_TYPE_BF16:
            return true;
        default:
            return false;
@@ -572,12 +580,6 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
            return BEST_FATTN_KERNEL_NONE;
    }

-#ifndef GGML_CUDA_FA_ALL_QUANTS
-    if (K->type != V->type) {
-        return BEST_FATTN_KERNEL_NONE;
-    }
-#endif // GGML_CUDA_FA_ALL_QUANTS
-
    if (!ggml_cuda_fattn_kv_type_supported(K->type) || !ggml_cuda_fattn_kv_type_supported(V->type)) {
        return BEST_FATTN_KERNEL_NONE;
    }
@@ -669,6 +671,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 size_t ggml_cuda_flash_attn_ext_get_alloc_size(int device, const ggml_tensor * dst) {
    GGML_ASSERT(dst->op == GGML_OP_FLASH_ATTN_EXT);

+    const ggml_tensor * Q = dst->src[0];
    const ggml_tensor * K = dst->src[1];
    const ggml_tensor * V = dst->src[2];

@@ -686,10 +689,11 @@ size_t ggml_cuda_flash_attn_ext_get_alloc_size(int device, const ggml_tensor * d
            need_f16_K = true;
            need_f16_V = true;
            break;
-        case BEST_FATTN_KERNEL_VEC:
-            need_f16_K = K->type == GGML_TYPE_F32;
-            need_f16_V = V->type == GGML_TYPE_F32;
-            break;
+        case BEST_FATTN_KERNEL_VEC: {
+            const bool f16_fallback = ggml_cuda_get_fattn_vec_case(Q->ne[0], K->type, V->type) == nullptr;
+            need_f16_K = K->type == GGML_TYPE_F32 || f16_fallback;
+            need_f16_V = V->type == GGML_TYPE_F32 || f16_fallback;
+        } break;
        case BEST_FATTN_KERNEL_NONE:
            break;
    }
@@ -5640,8 +5640,8 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
        features.push_back({ "USE_GRAPHS", "1" });
    #endif

-    #ifdef GGML_CUDA_FA_ALL_QUANTS
-        features.push_back({ "FA_ALL_QUANTS", "1" });
+    #ifdef GGML_CUDA_FA_QUANTS
+        features.push_back({ "FA_QUANTS", GGML_CUDA_FA_QUANTS });
    #endif

    {
@@ -171,7 +171,7 @@ void ggml_cuda_mul_mat_q(
            ne00, ne01, ne1, s01, ne11, s1,
            ne02, ne12, s02, s12, s2,
            ne03, ne13, s03, s13, s3,
-            ne1};
+            ne1, ne1};
        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
        return;
    }
@@ -244,6 +244,13 @@ void ggml_cuda_mul_mat_q(
                                         ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
    const int64_t s13 = ne12*s12;

+    // Each expert only sees ne12*n_expert_used/ne02 tokens on average.
+    // On RDNA3 and RDNA4 it is faster to pick the tile size against this value instead of ne12.
+    int64_t ncols_opt = ne12;
+    if (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+        ncols_opt = (ne12*n_expert_used + ne02 - 1) / ne02;
+    }
+
    // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
    const mmq_args args = {
        src0_d, src0->type, (const int *) src1_q8_1.get(), ids_dst.get(), expert_bounds.get(), dst_d,
@@ -251,7 +258,7 @@ void ggml_cuda_mul_mat_q(
        ne00, ne01, ne_get_rows, s01, ne_get_rows, s1,
        ne02, ne02, s02, s12, s2,
        ne03, ne13, s03, s13, s3,
-        ne12};
+        ne12, ncols_opt};

    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 }
@@ -1376,6 +1376,7 @@ struct mmq_args {
    int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
    int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
    int64_t ncols_max;
+    int64_t ncols_opt; // value to optimize the tile size against, launch grid still uses ncols_max
 };

 static size_t mmq_get_nbytes_shared(const ggml_cuda_mmq_config & config, const int cc) {
@@ -1486,7 +1487,7 @@ void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args,
            continue;
        }

-        const int ntiles_x = (args.ncols_max + config.J - 1) / config.J;
+        const int ntiles_x = (args.ncols_opt + config.J - 1) / config.J;

        if (ntiles_x < ntiles_J_best) {
            J_best = J;
@@ -70,17 +70,8 @@ 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")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-else()
-    list(APPEND GGML_SOURCES_ROCM
-        ../ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu
-        ../ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu
-        ../ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu
-        ../ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu)
-endif()
+ggml_cuda_fattn_vec_instances(${CMAKE_CURRENT_SOURCE_DIR}/../ggml-cuda SRCS)
+list(APPEND GGML_SOURCES_ROCM ${SRCS})

 ggml_add_backend_library(ggml-hip
                         ${GGML_HEADERS_ROCM}
@@ -43,17 +43,8 @@ if (MUSAToolkit_FOUND)
        add_compile_definitions(GGML_MUSA_MUDNN_COPY)
    endif()

-    if (GGML_CUDA_FA_ALL_QUANTS)
-        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
-        list(APPEND GGML_SOURCES_MUSA ${SRCS})
-        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-    else()
-        list(APPEND GGML_SOURCES_MUSA
-            ../ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu
-            ../ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu
-            ../ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu
-            ../ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu)
-    endif()
+    ggml_cuda_fattn_vec_instances(${CMAKE_CURRENT_SOURCE_DIR}/../ggml-cuda SRCS)
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})

    set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
    foreach(SOURCE ${GGML_SOURCES_MUSA})
@@ -13748,9 +13748,11 @@ static void ggml_vk_arange(ggml_backend_vk_context * ctx, vk_context& subctx, gg

 static void ggml_vk_fill(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
    VK_LOG_DEBUG("ggml_vk_fill(dst=" << dst << ", ne=" << ggml_nelements(dst) << ")");
+    const uint64_t n = ggml_nelements(dst);
+    GGML_ASSERT(n > 0);

    vk_op_push_constants pc = {
-        (uint32_t)ggml_nelements(dst),
+        (uint32_t)n,
        1,
        ggml_get_op_params_f32(dst, 0),
        0.0f,
@@ -13760,11 +13762,16 @@ static void ggml_vk_fill(ggml_backend_vk_context * ctx, vk_context& subctx, ggml
    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, nullptr, nullptr, nullptr, dst, GGML_OP_FILL);
    GGML_ASSERT(pipeline != nullptr);

+    // Split the task distribution to 2D to avoid exceeding maxComputeWorkGroupCount
+    const uint32_t total_wg = CEIL_DIV(n, pipeline->wg_denoms[0]);
+    const uint32_t wg_x = std::min(total_wg, ctx->device->properties.limits.maxComputeWorkGroupCount[0]);
+    const uint32_t wg_y = CEIL_DIV(total_wg, wg_x);
+    GGML_ASSERT(wg_y <= ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+
    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst, false);

-    std::array<uint32_t, 3> elements = { (uint32_t)ggml_nelements(dst), 1, 1 };
-
+    std::array<uint32_t, 3> elements = { wg_x * pipeline->wg_denoms[0], wg_y * pipeline->wg_denoms[1], 1 };
    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { dst_buf }, pc, elements);
 }

@@ -8,7 +8,9 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) writeonly buffer D {D_TYPE data_d[];};

 void main() {
-    const uint i = gl_GlobalInvocationID.x;
+    // 2D grid flattening: each x workgroup covers gl_WorkGroupSize.x elements,
+    // each y workgroup covers gl_NumWorkGroups.x * gl_WorkGroupSize.x elements.
+    const uint i = (gl_GlobalInvocationID.y * gl_NumWorkGroups.x * gl_WorkGroupSize.x) + gl_GlobalInvocationID.x;

    if (i >= p.KX) {
        return;
@@ -7,7 +7,7 @@ dynamic = ["classifiers"]
 readme = "README.md"
 authors = [{name = "GGML", email = "ggml@ggml.ai"}]
 requires-python = '>=3.10'
-dependencies = ['numpy (>=1.17)', 'tqdm (>=4.27)', 'pyyaml (>=5.1)', 'requests (>=2.25)']
+dependencies = ['numpy (>=2.2.6)', 'tqdm (>=4.27)', 'pyyaml (>=5.1)', 'requests (>=2.25)']
 classifiers = [
    "Programming Language :: Python :: 3",
    "License :: OSI Approved :: MIT License",
@@ -1357,7 +1357,7 @@ extern "C" {
    LLAMA_API struct llama_sampler * llama_sampler_chain_get(      struct llama_sampler * chain, int32_t i);

    // the total number of samplers in the chain
-    LLAMA_API int                    llama_sampler_chain_n  (const struct llama_sampler * chain);
+    LLAMA_API int32_t                llama_sampler_chain_n  (const struct llama_sampler * chain);

    // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed
    LLAMA_API struct llama_sampler * llama_sampler_chain_remove(   struct llama_sampler * chain, int32_t i);
@@ -8,7 +8,7 @@ readme = "README.md"
 authors = [{name = "GGML", email = "ggml@ggml.ai"}]
 requires-python = '>=3.10,<3.15'
 dependencies = [
-    'numpy (>=1.26.4,<3.0.0)',
+    'numpy (>=2.2.6,<3.0.0)',
    'sentencepiece (>=0.1.98,<0.3.0)',
    'transformers (==4.57.6)',
    'protobuf (>=4.21.0,<5.0.0)',
@@ -1,4 +1,4 @@
-numpy~=1.26.4
+numpy~=2.2.6
 sentencepiece>=0.1.98,<0.3.0

 transformers==4.57.6
@@ -1,3 +1,3 @@
-numpy~=1.26.4
+numpy~=2.2.6
 PySide6~=6.9.0
 gguf>=0.17.0
@@ -1,5 +1,5 @@
 datasets~=4.8.0
 matplotlib~=3.10.0
-numpy~=1.26.4
+numpy~=2.2.6
 requests~=2.32.3
 tqdm~=4.67.1
@@ -1,7 +1,7 @@
 aiohttp~=3.9.3
 pytest~=8.3.3
 matplotlib~=3.10.0
-numpy~=1.26.4
+numpy~=2.2.6
 openai~=2.14.0
 pandas~=2.2.3
 prometheus-client~=0.20.0
@@ -935,6 +935,7 @@ const char * llm_type_name(llm_type type) {
        case LLM_TYPE_17B_16E:       return "17Bx16E (Scout)";
        case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
        case LLM_TYPE_A13B:          return "A13B";
+        case LLM_TYPE_1B_A400M:      return "1B.A400M";
        case LLM_TYPE_7B_A1B:        return "7B.A1B";
        case LLM_TYPE_8B_A1B:        return "8B.A1B";
        case LLM_TYPE_7_9B_A1_3B:    return "7.9B.A1.3B";
@@ -116,6 +116,7 @@ enum llm_type {
    LLM_TYPE_17B_16E, // llama4 Scout
    LLM_TYPE_17B_128E, // llama4 Maverick
    LLM_TYPE_A13B,
+    LLM_TYPE_1B_A400M, // Granite3 MoE
    LLM_TYPE_7B_A1B,
    LLM_TYPE_8B_A1B, // lfm2moe
    LLM_TYPE_7_9B_A1_3B, // Ling-3.0-tiny
@@ -1006,7 +1006,7 @@ struct llama_sampler * llama_sampler_chain_remove(struct llama_sampler * chain,
    return result;
 }

-int llama_sampler_chain_n(const struct llama_sampler * chain) {
+int32_t llama_sampler_chain_n(const struct llama_sampler * chain) {
    const auto * p = (const llama_sampler_chain *) chain->ctx;

    return p->samplers.size();
@@ -8,6 +8,7 @@ void llama_model_granite_moe::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);

    switch (hparams.n_layer()) {
+        case 24: type = LLM_TYPE_1B_A400M; break;
        case 32: type = LLM_TYPE_3B; break;
        case 40: type = LLM_TYPE_3B; break;
        // Add additional layer/vocab/etc checks here for other model sizes
@@ -374,6 +374,24 @@ static void test_expressions(testing & t) {
        "42"
    );

+    test_template(t, "none in object",
+        "{{ x in {'low': 1, 'high': 2} }}",
+        {{"x", nullptr}},
+        "False"
+    );
+
+    test_template(t, "none not in object",
+        "{{ x not in {'low': 1, 'high': 2} }}",
+        {{"x", nullptr}},
+        "True"
+    );
+
+    test_template(t, "none in array",
+        "{{ x in [1, none, 3] }}",
+        {{"x", nullptr}},
+        "True"
+    );
+
    test_template(t, "dot notation",
        "{{ user.name }}",
        {{"user", {{"name", "Bob"}}}},
@@ -55,9 +55,6 @@
 | `-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) |
 | `--rpc SERVERS` | comma-separated list of RPC servers (host:port)<br/>(env: LLAMA_ARG_RPC) |
-| `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
-| `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
 | `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `-lzm, --lazy-mode MODE` | on-demand reading of certain tensors, for example per-layer embeddings (default: auto)<br/>- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)<br/>- auto: on, but only for tensors larger than 4 GiB<br/>- off: always keep them resident<br/>(env: LLAMA_ARG_LAZY_MODE) |
 | `--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) |
@@ -138,9 +138,6 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-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) |
 | `--rpc SERVERS` | comma-separated list of RPC servers (host:port)<br/>(env: LLAMA_ARG_RPC) |
-| `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
-| `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
 | `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `-lzm, --lazy-mode MODE` | on-demand reading of certain tensors, for example per-layer embeddings (default: auto)<br/>- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)<br/>- auto: on, but only for tensors larger than 4 GiB<br/>- off: always keep them resident<br/>(env: LLAMA_ARG_LAZY_MODE) |
 | `--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) |
@@ -476,8 +476,6 @@ static void print_usage(int /* argc */, char ** argv) {
    printf("  -dev, --device <dev0/dev1/...>                    (default: auto)\n");
    printf("  -lm, --load-mode <auto|none|mmap|mlock|mmap+mlock|dio> (default: %s)\n", join(transform_to_str(cmd_params_defaults.load_mode, llama_load_mode_name), ",").c_str());
    printf("  -lzm, --lazy-mode <on|auto|off>                   (default: %s)\n", join(transform_to_str(cmd_params_defaults.lazy_mode, lazy_mode_str), ",").c_str());
-    printf("  -mmp, --mmap <0|1>                                (DEPRECATED IN FAVOUR OF --load-mode)\n");
-    printf("  -dio, --direct-io <0|1>                           (DEPRECATED IN FAVOUR OF --load-mode)\n");
    printf("  -embd, --embeddings <0|1>                         (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
    printf("  -ts, --tensor-split <ts0/ts1/..>                  (default: 0)\n");
    printf("  -ot --override-tensor <tensor name pattern>=<buffer type>;...\n");
@@ -883,44 +881,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                    break;
                }
                params.flash_attn.insert(params.flash_attn.end(), types.begin(), types.end());
-            } else if (arg == "-mmp" || arg == "--mmap") {
-                if (++i >= argc) {
-                    invalid_param = true;
-                    break;
-                }
-                LOG_WRN("DEPRECATED: -mmp and --mmap are deprecated in favour of --load-mode. Please use --load-mode mmap instead.\n");
-                auto p = string_split<bool>(argv[i], split_delim);
-
-                std::vector<llama_load_mode> modes;
-                for (const auto & m : p) {
-                    llama_load_mode mode;
-                    if (m) {
-                        mode = LLAMA_LOAD_MODE_MMAP;
-                    } else {
-                        mode = LLAMA_LOAD_MODE_NONE;
-                    }
-                    modes.push_back(mode);
-                }
-                params.load_mode.insert(params.load_mode.end(), modes.begin(), modes.end());
-            } else if (arg == "-dio" || arg == "--direct-io") {
-                if (++i >= argc) {
-                    invalid_param = true;
-                    break;
-                }
-                LOG_WRN("DEPRECATED: -dio and --direct-io are deprecated in favour of --load-mode. Please use --load-mode dio instead.\n");
-                auto p = string_split<bool>(argv[i], split_delim);
-
-                std::vector<llama_load_mode> modes;
-                for (const auto & m : p) {
-                    llama_load_mode mode;
-                    if (m) {
-                        mode = LLAMA_LOAD_MODE_DIRECT_IO;
-                    } else {
-                        mode = LLAMA_LOAD_MODE_NONE;
-                    }
-                    modes.push_back(mode);
-                }
-                params.load_mode.insert(params.load_mode.end(), modes.begin(), modes.end());
            } else if (arg == "-embd" || arg == "--embeddings") {
                if (++i >= argc) {
                    invalid_param = true;
@@ -371,6 +371,7 @@ static bool is_webp_file(const unsigned char * buf, size_t len) {
 #ifdef MTMD_VIDEO
 static mtmd_bitmap * decode_webp_with_ffmpeg(const mtmd_context * mctx, const unsigned char * buf, size_t len, bool placeholder,
                                             const mtmd_helper_video_init_params & params);
+static void mtmd_helper_video_set_id(mtmd_helper_video * vctx, const std::string & id);
 #endif

 mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(const mtmd_context * ctx, const unsigned char * buf, size_t len, bool placeholder,
@@ -436,6 +437,7 @@ mtmd_helper_bitmap_wrapper mtmd_helper_bitmap_init_from_buf(const mtmd_context *
            LOG_ERR("%s: failed to decode buffer as either image/audio/video\n", __func__);
            return {nullptr, nullptr};
        }
+        mtmd_helper_video_set_id(video_ctx, id); // propagate the hash to the frames
        result = mtmd_bitmap_init_lazy(ctx,
            id.empty() ? nullptr : id.c_str(),
            video_ctx,
@@ -527,6 +529,7 @@ struct mtmd_helper_video {
    std::string ffprobe_bin;
    float fps_target = 0.0f;
    mtmd_helper_video_info info = {};
+    std::string id; // hash of the input video

    // RAII wrapper for managing subprocess
    struct subprocess_handle {
@@ -785,9 +788,14 @@ struct mtmd_helper_video {
        }

        LOG_DBG("%s: frame %d read OK\n", __func__, current_frame);
-        current_frame++;
        mtmd_bitmap * frame = mtmd_bitmap_init(info.width, info.height, frame_buf.data());
        mtmd_bitmap_set_mergeable(frame, true);
+        if (!id.empty()) {
+            // each frame gets a unique id in the form of {hash}+{frame}, so that it can be identified in cache
+            std::string frame_id = id + "+" + std::to_string(current_frame);
+            mtmd_bitmap_set_id(frame, frame_id.c_str());
+        }
+        current_frame++;
        return frame;
    }

@@ -886,6 +894,10 @@ static std::string video_resolve_bin(const char * bin_dir, const char * name) {
 }

 #ifdef MTMD_VIDEO
+static void mtmd_helper_video_set_id(mtmd_helper_video * vctx, const std::string & id) {
+    vctx->id = id;
+}
+
 static mtmd_bitmap * decode_webp_with_ffmpeg(const mtmd_context * mctx, const unsigned char * buf, size_t len, bool placeholder,
                                             const mtmd_helper_video_init_params & params) {
    mtmd_helper_video vctx;
@@ -72,9 +72,6 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `-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) |
 | `--rpc SERVERS` | comma-separated list of RPC servers (host:port)<br/>(env: LLAMA_ARG_RPC) |
-| `--mlock` | DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
-| `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
-| `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
 | `-lm, --load-mode MODE` | model loading mode (default: auto)<br/>- auto: mmap, unless a device does not support it<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
 | `-lzm, --lazy-mode MODE` | on-demand reading of certain tensors, for example per-layer embeddings (default: auto)<br/>- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)<br/>- auto: on, but only for tensors larger than 4 GiB<br/>- off: always keep them resident<br/>(env: LLAMA_ARG_LAZY_MODE) |
 | `--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) |
@@ -2,7 +2,7 @@ aiohttp~=3.9.3
 pytest~=8.3.3
 pytest-xdist~=3.6
 filelock~=3.16
-numpy~=1.26.4
+numpy~=2.2.6
 openai~=2.14.0
 prometheus-client~=0.20.0
 requests~=2.32.3
Author	SHA1	Message	Date
Masato NakasakaandGitHub	22397c31a0	vulkan: Convert FILL to distribute workgroups in 2D to avoid exceeding maxComputeWorkGroupCount (#28592 ) * divide workload to 2D This is to workaround FILL exceeding maxComputeWorkGroupCount for Intel GPUs on Qwen 3.8 flash next * minor change * Fixed comment	2026-09-09 16:54:15 +02:00
Sigbjørn SkjæretandGitHub	4b98ab805a	py : lower numpy to 2.2.6 (#28654 ) * Revert "py : bump numpy to 2.4.6 (#28649)" This reverts commit `9cf3bf256b`. * bump numpy to 2.2.6	2026-09-09 15:56:27 +02:00
Sigbjørn SkjæretandGitHub	9cf3bf256b	py : bump numpy to 2.4.6 (#28649 )	2026-09-09 14:59:37 +02:00
linsen458-spec GitHub linsen	4850c7727f	llama : use int32_t for llama_sampler_chain_n return type (#28631 ) Contributes to #4574 Co-authored-by: linsen <linsen@insta360.com>	2026-09-09 15:27:25 +03:00
Piotr Wilkin (ilintar)GitHub Carl Philipp Klemm ravel7524	d4abd573f6	CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#28552 ) Recreated from #24546 --------- Co-authored-by: Carl Philipp Klemm <carl@uvos.xyz> * CUDA: pick MMQ tile size against ncols_opt set on the host side Assisted-by: Claude Fable 5.1 Claude-Session: https://claude.ai/code/session_011SYPfRhKoUpU3gMsGxq6go --------- Co-authored-by: ravel7524 <58877666+ravel7524@users.noreply.github.com> Co-authored-by: Carl Philipp Klemm <carl@uvos.xyz>	2026-09-09 13:25:54 +02:00
Piotr Wilkin (ilintar)GitHub Johannes Gäßler	5a4d0fecae	CUDA: replace GGML_FA_ALL_QUANTS with GGML_FA_QUANTS, more control over what is compiled (#28079 ) * CUDA: add configurable FA quant combinations Assisted-by: Codex * remove all flags but , add runtime fallback with warning for uncompiled combination * Update docs/build.md Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * apply code review comments --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>	2026-09-09 12:50:08 +02:00
Aaron TeoandGitHub	14a9d09f75	args: officially deprecate --mmap\|mlock\|dio (#28334 ) Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2026-09-09 18:36:27 +08:00
Aaron TeoandGitHub	e2d2c0d6aa	model: fix granite3 moe unknown parameter count (#28632 ) Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2026-09-09 16:46:22 +08:00
Xuan-Son NguyenandGitHub	6de9cdb26b	mtmd: propagate video ID to bitmap (#28601 )	2026-09-09 10:37:36 +02:00
PascalandGitHub	b31b71f3a0	jinja: treat a null left operand of in as a plain lookup (#28620 ) Templates that default an optional variable to none and then test its membership in a map hit an error, while the same expression is a normal lookup returning false in Jinja. The undefined counterpart of this case was already handled just above.	2026-09-09 10:08:27 +03:00