Update ggml/src/ggml.c

ggml : fix n_threads_cur initialization with one thread
llama : use reserve/emplace_back in sampler_sample (#9534 )
2026-05-04 16:14:06 +00:00 · 2024-09-18 09:00:26 -07:00 · 2024-09-18 14:58:49 +02:00 · 2024-09-18 14:42:36 +03:00 · 2024-09-18 09:50:34 +03:00 · 2024-09-18 09:28:20 +03:00
27 changed files with 1037 additions and 714 deletions
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -82,11 +82,11 @@ set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})

 # change the default for these ggml options
 if (NOT DEFINED GGML_LLAMAFILE)
-    set(GGML_LLAMAFILE ON)
+    set(GGML_LLAMAFILE_DEFAULT ON)
 endif()

-if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
-    set(GGML_CUDA_USE_GRAPHS ON)
+if (NOT DEFINED GGML_CUDA_GRAPHS)
+    set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()

 # transition helpers
--- a/2
+++ b/2
@@ -619,7 +619,7 @@ ifdef GGML_CUDA
 			CUDA_PATH ?= /usr/local/cuda
 		endif

-		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 		MK_NVCCFLAGS += -use_fast_math
 	endif # GGML_MUSA
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -1312,7 +1312,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
        [](gpt_params & params, int value) {
            params.n_parallel = value;
        }
-    ));
+    ).set_env("LLAMA_ARG_N_PARALLEL"));
    add_opt(llama_arg(
        {"-ns", "--sequences"}, "N",
        format("number of sequences to decode (default: %d)", params.n_sequences),
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -132,12 +132,14 @@ class Model:
    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
        tensor_names_from_parts: set[str] = set()

-        if len(self.part_names) > 1:
+        index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
+        index_name += ".index.json"
+        index_file = self.dir_model / index_name
+
+        if index_file.is_file():
            self.tensor_names = set()
-            index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
-            index_name += ".index.json"
            logger.info(f"gguf: loading model weight map from '{index_name}'")
-            with open(self.dir_model / index_name, "r", encoding="utf-8") as f:
+            with open(index_file, "r", encoding="utf-8") as f:
                index: dict[str, Any] = json.load(f)
                weight_map = index.get("weight_map")
                if weight_map is None or not isinstance(weight_map, dict):
@@ -145,6 +147,7 @@ class Model:
                self.tensor_names.update(weight_map.keys())
        else:
            self.tensor_names = tensor_names_from_parts
+            weight_map = {}

        for part_name in self.part_names:
            logger.info(f"gguf: loading model part '{part_name}'")
@@ -171,9 +174,17 @@ class Model:
                            data = LazyTorchTensor.from_eager(data)
                    yield name, data

-        # only verify tensor name presence; it doesn't matter if they are not in the right files
-        if len(sym_diff := tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
-            raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
+        # verify tensor name presence and identify potentially missing files
+        if len(tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
+            missing = sorted(self.tensor_names.difference(tensor_names_from_parts))
+            extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
+            missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
+            if len(extra) == 0 and len(missing_files) > 0:
+                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+            else:
+                raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
+                                 f"Missing tensors: {missing}\n"
+                                 f"Extra tensors: {extra}")

    def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
        if key not in gguf.MODEL_TENSORS[self.model_arch]:
@@ -4069,6 +4080,36 @@ class ExaoneModel(Model):
        super().prepare_tensors()


+@Model.register("GraniteForCausalLM")
+class GraniteModel(LlamaModel):
+    """Conversion for IBM's GraniteForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE
+
+    def set_gguf_parameters(self):
+        """Granite uses standard llama parameters with the following differences:
+
+        - No head_dim support
+        - New multiplier params:
+            - attention_scale
+            - embedding_scale
+            - residual_scale
+        - logits_scaling
+        """
+        if head_dim := self.hparams.pop("head_dim", None):
+            logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim)
+        super().set_gguf_parameters()
+        # NOTE: Convert _multiplier params to _scale params for naming
+        #   consistency
+        if attention_scale := self.hparams.get("attention_multiplier"):
+            self.gguf_writer.add_attention_scale(attention_scale)
+        if embedding_scale := self.hparams.get("embedding_multiplier"):
+            self.gguf_writer.add_embedding_scale(embedding_scale)
+        if residual_scale := self.hparams.get("residual_multiplier"):
+            self.gguf_writer.add_residual_scale(residual_scale)
+        if logits_scaling := self.hparams.get("logits_scaling"):
+            self.gguf_writer.add_logit_scale(logits_scaling)
+
+
 ###### CONVERSION LOGIC ######

 # tree of lazy tensors
--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -636,6 +636,14 @@ use 1 SYCL GPUs: [0] with Max compute units:512

  It's same for other projects including llama.cpp SYCL backend.

+- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+
+  Device Memory is not enough.
+
+  |Reason|Solution|
+  |-|-|
+  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
+  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|

 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -439,6 +439,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                }
                types.push_back(gt);
            }
+            if (invalid_param) {
+                break;
+            }
            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
        } else if (arg == "-ctv" || arg == "--cache-type-v") {
            if (++i >= argc) {
@@ -455,6 +458,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                }
                types.push_back(gt);
            }
+            if (invalid_param) {
+                break;
+            }
            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
        } else if (arg == "-t" || arg == "--threads") {
            if (++i >= argc) {
@@ -520,6 +526,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                }
                modes.push_back(mode);
            }
+            if (invalid_param) {
+                break;
+            }
            params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
        } else if (arg == "-mg" || arg == "--main-gpu") {
            if (++i >= argc) {
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -87,7 +87,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
-| `-np, --parallel N` | number of parallel sequences to decode (default: 1) |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env:  LLAMA_ARG_N_PARALLEL) |
 | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
 | `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing |
@@ -501,7 +501,7 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte

    See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). While some OpenAI-specific features such as function calling aren't supported, llama.cpp `/completion`-specific features such as `mirostat` are supported.

-    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}`), similar to other OpenAI-inspired API providers.
+    The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.

    *Examples:*

--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -2254,14 +2254,6 @@ static void log_server_request(const httplib::Request & req, const httplib::Resp
        return;
    }

-    //LOG_INFO("request", {
-    //    {"remote_addr", req.remote_addr},
-    //    {"remote_port", req.remote_port},
-    //    {"status",      res.status},
-    //    {"method",      req.method},
-    //    {"path",        req.path},
-    //    {"params",      req.params},
-    //});
    LOG_INF("request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);

    LOG_DBG("request:  %s\n", req.body.c_str());
@@ -2318,12 +2310,12 @@ int main(int argc, char ** argv) {
    std::unique_ptr<httplib::Server> svr;
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
    if (params.ssl_file_key != "" && params.ssl_file_cert != "") {
-        LOG_INFO("Running with SSL", {{"key", params.ssl_file_key}, {"cert", params.ssl_file_cert}});
+        LOG_INF("Running with SSL: key = %s, cert = %s\n", params.ssl_file_key.c_str(), params.ssl_file_cert.c_str());
        svr.reset(
            new httplib::SSLServer(params.ssl_file_cert.c_str(), params.ssl_file_key.c_str())
        );
    } else {
-        LOG_INFO("Running without SSL", {});
+        LOG_INF("Running without SSL\n");
        svr.reset(new httplib::Server());
    }
 #else
@@ -3108,7 +3100,6 @@ int main(int argc, char ** argv) {
    std::thread t([&]() { svr->listen_after_bind(); });
    svr->wait_until_ready();

-    //LOG_INFO("HTTP server is listening", log_data);
    LOG_INF("%s: HTTP server is listening, hostname: %s, port: %d, http threads: %d\n", __func__, params.hostname.c_str(), params.port, params.n_threads_http);

    // load the model
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@@ -331,6 +331,9 @@ static json oaicompat_completion_params_parse(
        std::string response_type = json_value(response_format, "type", std::string());
        if (response_type == "json_object") {
            llama_params["json_schema"] = json_value(response_format, "schema", json::object());
+        } else if (response_type == "json_schema") {
+            json json_schema = json_value(response_format, "json_schema", json::object());
+            llama_params["json_schema"] = json_value(json_schema, "schema", json::object());
        } else if (!response_type.empty() && response_type != "text") {
            throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type);
        }
--- a/examples/sycl/run-llama2.sh
+++ b/examples/sycl/run-llama2.sh
@@ -11,16 +11,17 @@ source /opt/intel/oneapi/setvars.sh
 #ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.

 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
-MODEL_FILE=llama-2-7b.Q4_0.gguf
+MODEL_FILE=models/llama-2-7b.Q4_0.gguf
 NGL=33
+CONEXT=8192

 if [ $# -gt 0 ]; then
    GGML_SYCL_DEVICE=$1
    echo "use $GGML_SYCL_DEVICE as main GPU"
    #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT} -mg $GGML_SYCL_DEVICE -sm none

 else
    #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m ${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONEXT}
 fi
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -56,6 +56,15 @@ else()
    set(GGML_NATIVE_DEFAULT ON)
 endif()

+# defaults
+if (NOT GGML_LLAMAFILE_DEFAULT)
+    set(GGML_LLAMAFILE_DEFAULT OFF)
+endif()
+
+if (NOT GGML_CUDA_GRAPHS_DEFAULT)
+    set(GGML_CUDA_GRAPHS_DEFAULT OFF)
+endif()
+
 # general
 option(GGML_STATIC "ggml: static link libraries"         OFF)
 option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
@@ -110,7 +119,7 @@ option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"
 option(GGML_BLAS                            "ggml: use BLAS"                                  ${GGML_BLAS_DEFAULT})
 set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
                                            "ggml: BLAS library vendor")
-option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             OFF)
+option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             ${GGML_LLAMAFILE_DEFAULT})

 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
@@ -127,7 +136,7 @@ set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
-option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
+option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})

 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -329,7 +329,7 @@ if (GGML_CUDA)
        add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})

-        if (GGML_CUDA_USE_GRAPHS)
+        if (GGML_CUDA_GRAPHS)
            add_compile_definitions(GGML_CUDA_USE_GRAPHS)
        endif()

@@ -1341,7 +1341,7 @@ list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
 find_library(MATH_LIBRARY m)
 if (MATH_LIBRARY)
    if (NOT WIN32 OR NOT GGML_SYCL)
-        target_link_libraries(ggml PRIVATE ${MATH_LIBRARY})
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
    endif()
 endif()

--- a/ggml/src/ggml-aarch64.c
+++ b/ggml/src/ggml-aarch64.c
@@ -4,6 +4,7 @@

 #include "ggml-quants.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"

 #include <math.h>
 #include <string.h>
--- a/ggml/src/ggml-cpu-impl.h
+++ b/ggml/src/ggml-cpu-impl.h
@@ -0,0 +1,614 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#include <sys/prctl.h>
+#endif
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+typedef __fp16 ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    ggml_fp16_internal_t tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    ggml_fp16_internal_t tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#if defined(__loongarch_asx)
+
+typedef union {
+    int32_t i;
+    float f;
+} ft_union;
+
+/* float type data load instructions */
+static __m128 __lsx_vreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+}
+
+static __m256 __lasx_xvreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+}
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+#ifdef __cplusplus
+}
+#endif
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -1,15 +1,17 @@
 #pragma once

-#include "ggml.h"
-
 // GGML internal header

+#include "ggml.h"
+
 #include <assert.h>
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-#include <stddef.h>
 #include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif

 #undef MIN
 #undef MAX
@@ -17,96 +19,6 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))

-#if defined(_MSC_VER)
-
-#define m512bh(p) p
-#define m512i(p) p
-
-#else
-
-#define m512bh(p) (__m512bh)(p)
-#define m512i(p) (__m512i)(p)
-
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
 // static_assert should be a #define, but if it's not,
 // fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
@@ -121,520 +33,6 @@ extern "C" {
 #endif
 #endif

-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
-#include <sys/prctl.h>
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddlvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddlvq_s16(int16x8_t v) {
-    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
-    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#if defined(__loongarch64)
-#if defined(__loongarch_asx)
-#include <lasxintrin.h>
-#endif
-#if defined(__loongarch_sx)
-#include <lsxintrin.h>
-#endif
-#endif
-
-#if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
-/* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
-}
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-enum ggml_cgraph_eval_order {
-    GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
-    GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
-    GGML_CGRAPH_EVAL_ORDER_COUNT
-};
-
 // bitset

 typedef uint32_t ggml_bitset_t;
@@ -761,6 +159,12 @@ static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct g

 // computation graph

+enum ggml_cgraph_eval_order {
+    GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+    GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+    GGML_CGRAPH_EVAL_ORDER_COUNT
+};
+
 struct ggml_cgraph {
    int size;
    int n_nodes;
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -3,6 +3,7 @@

 #include "ggml-quants.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"


 #include <math.h>
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -2,6 +2,7 @@
 #define _USE_MATH_DEFINES // For M_PI on MSVC

 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 #include "ggml.h"
 #include "ggml-aarch64.h"
@@ -2012,10 +2013,11 @@ struct ggml_threadpool {
    // these are atomic as an annotation for thread-sanitizer
    atomic_bool stop;         // Used for stopping the threadpool altogether
    atomic_bool pause;        // Used for pausing the threadpool or individual threads
+    atomic_bool abort;        // Used for aborting processing of a graph

    struct ggml_compute_state * workers;   // per thread state
    int          n_threads_max; // number of threads in the pool
-    int          n_threads_cur; // number of threads used in the current graph
+    atomic_int   n_threads_cur; // number of threads used in the current graph

    int32_t      prio;        // Scheduling priority
    uint32_t     poll;        // Polling level (0 - no polling)
@@ -3177,41 +3179,36 @@ inline static void ggml_critical_section_start(void) {
    }
 }

+static void ggml_barrier(struct ggml_threadpool * tp) {
+    int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
+    if (n_threads == 1) {
+        return;
+    }
+
 #ifdef GGML_USE_OPENMP
-static void ggml_barrier(struct ggml_threadpool * threadpool) {
-    if (threadpool->n_threads_cur == 1) {
-        return;
-    }
-
    #pragma omp barrier
-}
 #else
-static void ggml_barrier(struct ggml_threadpool * threadpool) {
-    if (threadpool->n_threads_cur == 1) {
-        return;
-    }
+    int n_passed = atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed);

-    atomic_int * n_barrier = &threadpool->n_barrier;
-    atomic_int * n_barrier_passed = &threadpool->n_barrier_passed;
+    // enter barrier (full seq-cst fence)
+    int n_barrier = atomic_fetch_add_explicit(&tp->n_barrier, 1, memory_order_seq_cst);

-    int n_threads = threadpool->n_threads_cur;
-    int passed_old = atomic_load_explicit(n_barrier_passed, memory_order_relaxed);
-
-    if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) {
+    int last = 0;
+    if (n_barrier == (n_threads - 1)) {
        // last thread
-        atomic_store(n_barrier, 0);
-        atomic_fetch_add_explicit(n_barrier_passed, 1, memory_order_relaxed);
+        atomic_store_explicit(&tp->n_barrier, 0, memory_order_relaxed);
+        last = 1;
    } else {
        // wait for other threads
-        while (true) {
-            if (atomic_load_explicit(n_barrier_passed, memory_order_relaxed) != passed_old) {
-                return;
-            }
+        while (atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed) == n_passed) {
            ggml_thread_cpu_relax();
        }
    }
-}
+
+    // exit barrier (full seq-cst fence)
+    atomic_fetch_add_explicit(&tp->n_barrier_passed, last, memory_order_seq_cst);
 #endif
+}

 // TODO: make this somehow automatically executed
 //       some sort of "sentry" mechanism
@@ -19932,34 +19929,33 @@ struct ggml_cplan ggml_graph_plan(

 static thread_ret_t ggml_graph_compute_thread(void * data) {
    struct ggml_compute_state * state = (struct ggml_compute_state *) data;
+    struct ggml_threadpool    * tp    = state->threadpool;

-    const struct ggml_cgraph * cgraph = state->threadpool->cgraph;
-    const struct ggml_cplan  * cplan  = state->threadpool->cplan;
+    const struct ggml_cgraph * cgraph = tp->cgraph;
+    const struct ggml_cplan  * cplan  = tp->cplan;

    set_numa_thread_affinity(state->ith);

    struct ggml_compute_params params = {
        /*.ith       =*/ state->ith,
-        /*.nth       =*/ state->threadpool->n_threads_cur,
+        /*.nth       =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
        /*.wsize     =*/ cplan->work_size,
        /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ state->threadpool,
+        /*.threadpool=*/ tp,
    };

-    for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
+    for (int node_n = 0; node_n < cgraph->n_nodes && !tp->abort; node_n++) {
        struct ggml_tensor * node = cgraph->nodes[node_n];

        ggml_compute_forward(&params, node);

-        if (state->ith == 0 && cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
-            state->threadpool->ec = GGML_STATUS_ABORTED;
+        if (state->ith == 0 && cplan->abort_callback &&
+                cplan->abort_callback(cplan->abort_callback_data)) {
+            tp->abort = true;
+            tp->ec    = GGML_STATUS_ABORTED;
        }

        ggml_barrier(state->threadpool);
-
-        if (state->threadpool->ec != GGML_STATUS_SUCCESS) {
-            break;
-        }
    }

    return 0;
@@ -19967,7 +19963,15 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {

 #ifndef GGML_USE_OPENMP

-static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) {
+// check if thread is active
+static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+    int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
+    return (state->ith < n_threads);
+}
+
+// check if thread is ready to proceed (exit from polling or sleeping)
+static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
    struct ggml_threadpool * threadpool = state->threadpool;

    if (state->pending || threadpool->stop || threadpool->pause) { return true; }
@@ -19975,21 +19979,34 @@ static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) {
    // check for new graph/work
    int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
    if (new_graph != state->last_graph) {
-        state->pending    = (state->ith < threadpool->n_threads_cur);
+        state->pending    = ggml_graph_compute_thread_active(state);
        state->last_graph = new_graph;
    }

    return state->pending;
 }

+// sync thread state after polling
+static inline void ggml_graph_compute_thread_sync(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+    // this should just be atomic_thread_fence(seq_cst) but it confuses thread-sanitizer
+    // so instead we just use a dummy read-modify-write
+    atomic_fetch_add_explicit(&threadpool->n_graph, 0, memory_order_seq_cst);
+}
+
 static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) {
    struct ggml_threadpool * threadpool = state->threadpool;

+    // Skip polling for unused threads
+    if (!ggml_graph_compute_thread_active(state)) {
+        return state->pending;
+    }
+
    // This seems to make 0 ... 100 a decent range for polling level across modern processors.
    // Perhaps, we can adjust it dynamically based on load and things.
    const uint64_t n_rounds = 1024UL * 128 * threadpool->poll;

-    for (uint64_t i=0; !ggml_graph_compute_ready(state) && i<n_rounds; i++) {
+    for (uint64_t i=0; !ggml_graph_compute_thread_ready(state) && i < n_rounds; i++) {
        // No new work. Keep polling.
        ggml_thread_cpu_relax();
    }
@@ -20001,13 +20018,14 @@ static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state *
    struct ggml_threadpool * threadpool = state->threadpool;

    if (ggml_graph_compute_poll_for_work(state)) {
+        ggml_graph_compute_thread_sync(state);
        return state->pending;
    }

    ggml_mutex_lock_shared(&threadpool->mutex);
-    while (!ggml_graph_compute_ready(state)) {
+    while (!ggml_graph_compute_thread_ready(state)) {
        // No new work. Wait for the signal.
-        GGML_PRINT_DEBUG("thread #%d waiting for work\n", state->ith);
+        GGML_PRINT_DEBUG("thread #%d waiting for work (sleeping)\n", state->ith);
        ggml_cond_wait(&threadpool->cond, &threadpool->mutex);
    }
    ggml_mutex_unlock_shared(&threadpool->mutex);
@@ -20054,13 +20072,20 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
 }

 // Start processing new graph
-static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool)
+static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int n_threads)
 {
-    // always take the mutex here because the worker threads are doing hybrid poll/wait
+    // Always take the mutex here because the worker threads are doing hybrid poll/wait

    ggml_mutex_lock(&threadpool->mutex);

-    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_relaxed);
+    GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);
+
+    // Update the number of active threads
+    atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+
+    // Indicate the graph is ready to be processed
+    // We need the full seq-cst fence here because of the polling threads (used in thread_sync)
+    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_seq_cst);

    if (threadpool->pause) {
       // Update main thread prio and affinity to match the threadpool settings
@@ -20119,6 +20144,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
        threadpool->current_chunk    = 0;
        threadpool->stop             = false;
        threadpool->pause            = tpp->paused;
+        threadpool->abort            = false;
        threadpool->workers          = NULL;
        threadpool->n_threads_max    = tpp->n_threads;
        threadpool->n_threads_cur    = tpp->n_threads;
@@ -20194,15 +20220,11 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
        // No worker threads should be accessing the parameters below at this stage
        threadpool->cgraph           = cgraph;
        threadpool->cplan            = cplan;
-        threadpool->n_threads_cur    = n_threads;
        threadpool->current_chunk    = 0;
+        threadpool->abort            = false;
        threadpool->ec               = GGML_STATUS_SUCCESS;
    }

-    if (n_threads > threadpool->n_threads_max) {
-        GGML_PRINT("WARNING: cplan is requesting more threads than the threadpool contains. Expect a bad time!\n");
-    }
-
 #ifdef GGML_USE_OPENMP
    if (n_threads > 1) {
        #pragma omp parallel num_threads(n_threads)
@@ -20211,17 +20233,23 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
            {
                // update the number of threads from the actual number of threads that we got from OpenMP
                n_threads = omp_get_num_threads();
-                threadpool->n_threads_cur = n_threads;
+                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
            }

            ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
        }
    } else {
+        atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
        ggml_graph_compute_thread(&threadpool->workers[0]);
    }
 #else
+    if (n_threads > threadpool->n_threads_max) {
+        GGML_PRINT("WARNING: cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
+        n_threads = threadpool->n_threads_max;
+    }
+
    // Kick all threads to start the new graph
-    ggml_graph_compute_kickoff(threadpool);
+    ggml_graph_compute_kickoff(threadpool, n_threads);

    // This is a work thread too
    ggml_graph_compute_thread(&threadpool->workers[0]);
--- a/ggml/src/llamafile/sgemm.cpp
+++ b/ggml/src/llamafile/sgemm.cpp
@@ -50,6 +50,7 @@

 #include "sgemm.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 #include "ggml-quants.h"

 #ifdef _MSC_VER
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -97,6 +97,8 @@ class Keys:
        RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
        TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
        TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
+        RESIDUAL_SCALE                    = "{arch}.residual_scale"
+        EMBEDDING_SCALE                   = "{arch}.embedding_scale"

    class Attention:
        HEAD_COUNT        = "{arch}.attention.head_count"
@@ -112,6 +114,7 @@ class Keys:
        KV_LORA_RANK      = "{arch}.attention.kv_lora_rank"
        REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count"
        SLIDING_WINDOW    = "{arch}.attention.sliding_window"
+        SCALE             = "{arch}.attention.scale"

    class Rope:
        DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -231,6 +234,7 @@ class MODEL_ARCH(IntEnum):
    JAIS         = auto()
    NEMOTRON     = auto()
    EXAONE       = auto()
+    GRANITE      = auto()


 class MODEL_TENSOR(IntEnum):
@@ -387,6 +391,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.JAIS:           "jais",
    MODEL_ARCH.NEMOTRON:       "nemotron",
    MODEL_ARCH.EXAONE:         "exaone",
+    MODEL_ARCH.GRANITE:        "granite",
 }

 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -1224,6 +1229,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.GRANITE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
    # TODO
 }

--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -679,6 +679,12 @@ class GGUFWriter:
    def add_time_decay_extra_dim(self, dim: int) -> None:
        self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)

+    def add_residual_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value)
+
+    def add_embedding_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
+
    def add_wkv_head_size(self, size: int) -> None:
        self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)

@@ -703,6 +709,9 @@ class GGUFWriter:
    def add_sliding_window(self, value: int) -> None:
        self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)

+    def add_attention_scale(self, value: float) -> None:
+        self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value)
+
    def add_pooling_type(self, value: PoolingType) -> None:
        self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)

--- a/grammars/README.md
+++ b/grammars/README.md
@@ -120,7 +120,7 @@ You can use GBNF grammars:

 - In [llama-server](../examples/server):
    - For any completion endpoints, passed as the `json_schema` body field
-    - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}`)
+    - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}` or `{ type: "json_schema", json_schema: {"schema": ...} }`)
 - In [llama-cli](../examples/main), passed as the `--json` / `-j` flag
 - To convert to a grammar ahead of time:
    - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py)
--- a/include/llama.h
+++ b/include/llama.h
@@ -441,6 +441,7 @@ extern "C" {
    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);
+    LLAMA_API int32_t llama_n_head     (const struct llama_model * model);

    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);

--- a/src/llama-sampling.cpp
+++ b/src/llama-sampling.cpp
@@ -236,9 +236,10 @@ llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_conte
    const int n_vocab = llama_n_vocab(llama_get_model(ctx));

    // TODO: do not allocate each time
-    std::vector<llama_token_data> cur(n_vocab);
+    std::vector<llama_token_data> cur;
+    cur.reserve(n_vocab);
    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
    }

    llama_token_data_array cur_p = {
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -214,6 +214,7 @@ enum llm_arch {
    LLM_ARCH_NEMOTRON,
    LLM_ARCH_EXAONE,
    LLM_ARCH_RWKV6,
+    LLM_ARCH_GRANITE,
    LLM_ARCH_UNKNOWN,
 };

@@ -264,6 +265,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_NEMOTRON,        "nemotron"     },
    { LLM_ARCH_EXAONE,          "exaone"       },
    { LLM_ARCH_RWKV6,           "rwkv6"        },
+    { LLM_ARCH_GRANITE,         "granite"      },
    { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };

@@ -303,6 +305,8 @@ enum llm_kv {
    LLM_KV_RESCALE_EVERY_N_LAYERS,
    LLM_KV_TIME_MIX_EXTRA_DIM,
    LLM_KV_TIME_DECAY_EXTRA_DIM,
+    LLM_KV_RESIDUAL_SCALE,
+    LLM_KV_EMBEDDING_SCALE,

    LLM_KV_ATTENTION_HEAD_COUNT,
    LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -317,6 +321,7 @@ enum llm_kv {
    LLM_KV_ATTENTION_KV_LORA_RANK,
    LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
    LLM_KV_ATTENTION_SLIDING_WINDOW,
+    LLM_KV_ATTENTION_SCALE,

    LLM_KV_ROPE_DIMENSION_COUNT,
    LLM_KV_ROPE_FREQ_BASE,
@@ -407,6 +412,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
    { LLM_KV_TIME_MIX_EXTRA_DIM,                "%s.time_mix_extra_dim"                },
    { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
+    { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
+    { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },

    { LLM_KV_ATTENTION_HEAD_COUNT,             "%s.attention.head_count"             },
    { LLM_KV_ATTENTION_HEAD_COUNT_KV,          "%s.attention.head_count_kv"          },
@@ -421,6 +428,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_ATTENTION_KV_LORA_RANK,           "%s.attention.kv_lora_rank"           },
    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
    { LLM_KV_ATTENTION_SLIDING_WINDOW,         "%s.attention.sliding_window"         },
+    { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },

    { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
    { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
@@ -1454,6 +1462,22 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
            { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,    "blk.%d.channel_mix_receptance" },
        },
    },
+    {
+        LLM_ARCH_GRANITE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
    {
        LLM_ARCH_UNKNOWN,
        {
@@ -2372,6 +2396,11 @@ struct llama_hparams {
    float f_max_alibi_bias = 0.0f;
    float f_logit_scale    = 0.0f;

+    // Additional scale factors (Granite)
+    float f_residual_scale  = 0.0f;
+    float f_embedding_scale = 0.0f;
+    float f_attention_scale = 0.0f;
+
    bool causal_attn   = true;
    bool use_alibi     = false;
    bool attn_soft_cap = false;
@@ -2434,6 +2463,9 @@ struct llama_hparams {
        if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
        if (!is_float_close(this->expert_weights_scale,  other.expert_weights_scale,  EPSILON)) return true;
        if (!is_float_close(this->rope_yarn_log_mul,     other.rope_yarn_log_mul,     EPSILON)) return true;
+        if (!is_float_close(this->f_residual_scale,      other.f_residual_scale,      EPSILON)) return true;
+        if (!is_float_close(this->f_embedding_scale,     other.f_embedding_scale,     EPSILON)) return true;
+        if (!is_float_close(this->f_attention_scale,     other.f_attention_scale,     EPSILON)) return true;

        return false;
    }
@@ -6019,6 +6051,20 @@ static void llm_load_hparams(
                    default: model.type = e_model::MODEL_UNKNOWN;
                }
            } break;
+        case LLM_ARCH_GRANITE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
+                ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale);
+                ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
+                ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale);
+
+                switch (hparams.n_layer) {
+                    case 40: model.type = e_model::MODEL_3B; break;
+                    // Add additional layer/vocab/etc checks here for other model sizes
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
        default: (void)0;
    }

@@ -6061,8 +6107,15 @@ static void llm_load_vocab(
            vocab.special_mask_id = -1;
            vocab.linefeed_id     = -1;

+            // read vocab size from metadata
+            if (!ml.get_key(LLM_KV_VOCAB_SIZE, vocab.n_vocab, false)) {
+                vocab.n_vocab = 0;
+                LLAMA_LOG_WARN("%s: there is no vocab_size in metadata, vocab.n_vocab will be set to %u\n", __func__, vocab.n_vocab);
+            }
            return;
-        } else if (tokenizer_model == "llama") {
+        }
+
+        if (tokenizer_model == "llama") {
            vocab.type = LLAMA_VOCAB_TYPE_SPM;

            // default special tokens
@@ -6717,6 +6770,12 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
        LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
        LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
    }
+
+    if (model.arch == LLM_ARCH_GRANITE) {
+        LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
+        LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
+        LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
+    }
 }

 // Returns false if cancelled by progress_callback
@@ -6885,6 +6944,7 @@ static bool llm_load_tensors(
            case LLM_ARCH_LLAMA:
            case LLM_ARCH_REFACT:
            case LLM_ARCH_MINICPM:
+            case LLM_ARCH_GRANITE:
                {
                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});

@@ -8868,6 +8928,11 @@ static struct ggml_tensor * llm_build_inp_embd(
        ggml_set_input(lctx.inp_embd);
    }

+    // For Granite architecture
+    if (hparams.f_embedding_scale != 0.0f) {
+        inpL = ggml_scale(ctx, inpL, hparams.f_embedding_scale);
+    }
+
    cb(inpL, "inp_embd", -1);

    return inpL;
@@ -9571,7 +9636,7 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
        struct ggml_tensor * cur,
        struct ggml_tensor * x_prev,
        struct ggml_tensor ** wkv_state) {
-    size_t n_embed      = cur->ne[0];
+    size_t n_embd       = cur->ne[0];
    size_t n_seq_tokens = cur->ne[1];
    size_t n_seqs       = cur->ne[2];

@@ -9582,8 +9647,8 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(

    struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur);

-    sx  = ggml_reshape_2d(ctx, sx,  n_embed, n_tokens);
-    cur = ggml_reshape_2d(ctx, cur, n_embed, n_tokens);
+    sx  = ggml_reshape_2d(ctx, sx,  n_embd, n_tokens);
+    cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);

    struct ggml_tensor * xxx = ggml_add(ctx, ggml_mul(ctx, sx, layer->time_mix_lerp_x), cur);

@@ -9608,11 +9673,11 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
        xxx
    );

-    struct ggml_tensor *mw = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], 0);
-    struct ggml_tensor *mk = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * sizeof(float));
-    struct ggml_tensor *mv = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 2 * sizeof(float));
-    struct ggml_tensor *mr = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 3 * sizeof(float));
-    struct ggml_tensor *mg = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 4 * sizeof(float));
+    struct ggml_tensor *mw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0);
+    struct ggml_tensor *mk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
+    struct ggml_tensor *mv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
+    struct ggml_tensor *mr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
+    struct ggml_tensor *mg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));

    struct ggml_tensor * xw = ggml_add(
        ctx,
@@ -9681,7 +9746,7 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
        )
    );

-    w = ggml_add(ctx, w, ggml_reshape_1d(ctx, layer->time_mix_decay, n_embed));
+    w = ggml_add(ctx, w, ggml_reshape_1d(ctx, layer->time_mix_decay, n_embd));
    w = ggml_exp(ctx, ggml_neg(ctx, ggml_exp(ctx, w)));
    w = ggml_reshape_4d(ctx, w, 1, head_size, head_count, n_tokens);

@@ -9690,21 +9755,21 @@ static struct ggml_tensor * llm_build_rwkv6_time_mix(
    r = ggml_transpose(ctx, r);

    struct ggml_tensor * wkv_output = ggml_rwkv_wkv(ctx, k, v, r, layer->time_mix_first, w, *wkv_state);
-    cur = ggml_view_1d(ctx, wkv_output, n_embed * n_tokens, 0);
-    *wkv_state = ggml_view_1d(ctx, wkv_output, n_embed * head_size * n_seqs, n_embed * n_tokens * sizeof(float));
+    cur = ggml_view_1d(ctx, wkv_output, n_embd * n_tokens, 0);
+    *wkv_state = ggml_view_1d(ctx, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));

    // group norm with head_count groups
-    cur = ggml_reshape_3d(ctx, cur, n_embed / head_count, head_count, n_tokens);
+    cur = ggml_reshape_3d(ctx, cur, n_embd / head_count, head_count, n_tokens);
    cur = ggml_norm(ctx, cur, 64e-5f);

    // Convert back to regular vectors.
-    cur = ggml_reshape_2d(ctx, cur, n_embed, n_tokens);
+    cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);
    cur = ggml_add(ctx, ggml_mul(ctx, cur, layer->time_mix_ln), layer->time_mix_ln_b);

    cur = ggml_mul(ctx, cur, g);
    cur = llm_build_lora_mm(lctx, ctx, layer->time_mix_output, cur);

-    return ggml_reshape_3d(ctx, cur, n_embed, n_seq_tokens, n_seqs);
+    return ggml_reshape_3d(ctx, cur, n_embd, n_seq_tokens, n_seqs);
 }

 static struct ggml_tensor * llm_build_rwkv6_channel_mix(
@@ -10146,6 +10211,7 @@ struct llm_build_context {
        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();

+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
        for (int il = 0; il < n_layer; ++il) {
            struct ggml_tensor * inpSA = inpL;

@@ -10198,7 +10264,7 @@ struct llm_build_context {

                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
            }

            if (il == n_layer - 1) {
@@ -10209,6 +10275,11 @@ struct llm_build_context {
                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
            }

+            // For Granite architecture
+            if (hparams.f_residual_scale) {
+                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            }
+
            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
            cb(ffn_inp, "ffn_inp", il);

@@ -10245,6 +10316,11 @@ struct llm_build_context {
                cb(cur, "ffn_moe_out", il);
            }

+            // For Granite architecture
+            if (hparams.f_residual_scale) {
+                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            }
+
            cur = ggml_add(ctx0, cur, ffn_inp);
            cb(cur, "ffn_out", il);

@@ -10264,6 +10340,12 @@ struct llm_build_context {

        // lm_head
        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        // For Granite architecture
+        if (hparams.f_logit_scale) {
+            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        }
+
        cb(cur, "result_output", -1);

        ggml_build_forward_expand(gf, cur);
@@ -15789,6 +15871,7 @@ static struct ggml_cgraph * llama_build_graph(

    switch (model.arch) {
        case LLM_ARCH_LLAMA:
+        case LLM_ARCH_GRANITE:
            {
                result = llm.build_llama();
            } break;
@@ -16577,7 +16660,7 @@ static int llama_decode_internal(
    const uint32_t n_tokens_all = batch_all.n_tokens;

    if (n_tokens_all == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
        return -1;
    }

@@ -16590,7 +16673,7 @@ static int llama_decode_internal(
    if (batch_all.token) {
        for (uint32_t i = 0; i < n_tokens_all; ++i) {
            if (batch_all.token[i] < 0 || (uint32_t)batch_all.token[i] >= model.vocab.n_vocab) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch_all.token[i]);
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch_all.token[i]);
                return -1;
            }
        }
@@ -16878,7 +16961,7 @@ static int llama_encode_internal(
    const uint32_t n_tokens = batch.n_tokens;

    if (n_tokens == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0", __func__);
+        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
        return -1;
    }

@@ -16891,7 +16974,7 @@ static int llama_encode_internal(
    if (batch.token) {
        for (uint32_t i = 0; i < n_tokens; ++i) {
            if (batch.token[i] < 0 || (uint32_t)batch.token[i] >= model.vocab.n_vocab) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d", __func__, i, batch.token[i]);
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                return -1;
            }
        }
@@ -19047,6 +19130,10 @@ int32_t llama_n_layer(const struct llama_model * model) {
    return model->hparams.n_layer;
 }

+int32_t llama_n_head(const struct llama_model * model) {
+    return model->hparams.n_head();
+}
+
 const struct llama_model * llama_get_model(const struct llama_context * ctx) {
    return &ctx->model;
 }
@@ -19085,6 +19172,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
        case LLM_ARCH_ARCTIC:
        case LLM_ARCH_DEEPSEEK2:
        case LLM_ARCH_CHATGLM:
+        case LLM_ARCH_GRANITE:
            return LLAMA_ROPE_TYPE_NORM;

        // the pairs of head values are offset by n_rot/2
--- a/src/unicode.cpp
+++ b/src/unicode.cpp
@@ -5,6 +5,7 @@
 #include "unicode.h"
 #include "unicode-data.h"

+#include <algorithm>
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -119,6 +119,7 @@ llama_target_and_test(test-grammar-parser.cpp)
 llama_target_and_test(test-llama-grammar.cpp)
 llama_target_and_test(test-grammar-integration.cpp)
 llama_target_and_test(test-grad0.cpp)
+llama_target_and_test(test-barrier.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)

--- a/tests/test-barrier.cpp
+++ b/tests/test-barrier.cpp
@@ -0,0 +1,93 @@
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <chrono>
+#include <iostream>
+#include <cstdio>
+#include <cstdlib>
+#include <cassert>
+#include <vector>
+
+#define MAX_NARGS 2
+
+int main(int argc, char *argv[]) {
+
+    int n_threads = 4;
+    int n_rounds  = 100;
+
+    if (argc > 1) {
+        n_threads = std::atoi(argv[1]);
+    }
+
+    if (argc > 2) {
+        n_rounds  = std::atoi(argv[2]);
+    }
+
+    struct ggml_init_params params = {
+        /* .mem_size   = */ 1024*1024*1024,
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ false,
+    };
+
+    struct ggml_context * ctx = ggml_init(params);
+
+    // Create graph
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+
+    // Lots of small, parallel ops where barriers in between will dominate
+    struct ggml_tensor * out = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,  64);
+    for (int i = 0; i < 1000; i++) {
+        struct ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, 64, 128);
+        out = ggml_mul_mat(ctx, a, out);
+
+        struct ggml_tensor * d = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, 128, 64);
+        out = ggml_mul_mat(ctx, d, out);
+    }
+
+    ggml_build_forward_expand(gf, out);
+    int n_nodes = ggml_graph_n_nodes(gf);
+
+    // Create threadpool
+    struct ggml_threadpool_params tpp  = ggml_threadpool_params_default(n_threads);
+    struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
+    if (!threadpool) {
+        fprintf(stderr, "threadpool create failed : n_threads %d\n", n_threads);
+        exit(1);
+    }
+
+    // Create compute plan
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads, threadpool);
+
+    std::vector<uint8_t> work_data(cplan.work_size);
+    cplan.work_data = work_data.data();
+
+    std::cerr << "graph-compute with"
+              << "\n n_threads: " << n_threads
+              << "\n   n_nodes: " << n_nodes
+              << "\n  n_rounds: " << n_rounds
+              << "\n";
+    // ggml_graph_print(gf);
+
+    // Warmup
+    ggml_graph_compute(gf, &cplan);
+
+    auto t0 = std::chrono::high_resolution_clock::now();
+
+    for (int i=0; i < n_rounds; i++) {
+        ggml_graph_compute(gf, &cplan);
+    }
+
+    auto t1 = std::chrono::high_resolution_clock::now();
+
+    auto usec = std::chrono::duration_cast<std::chrono::microseconds>(t1-t0).count();
+    auto nsec = std::chrono::duration_cast<std::chrono::nanoseconds>(t1-t0).count();
+    std::cerr << "graph-compute took " << usec << " usec "
+              << "\n " << (float) usec / n_rounds << " usec per-iter"
+              << "\n " << (float) nsec / (n_rounds * n_nodes) << " nsec per-node"
+              << "\n";
+
+    ggml_threadpool_free(threadpool);
+    ggml_free(ctx);
+
+    return 0;
+}
Author	SHA1	Message	Date
Max Krasnyansky	6b0248c29a	Update ggml/src/ggml.c	2024-09-18 09:00:26 -07:00
slaren	f9196c9174	ggml : fix n_threads_cur initialization with one thread	2024-09-18 14:58:49 +02:00
Daniel Bevenius	6443ddd985	llama : use reserve/emplace_back in sampler_sample (#9534 ) This commit updates the llama_sampler_sample function to use reserve and emplace_back for the vector of llama_token_data structs. The motivation for this change is to avoid the creation of n_vocab default-constructed llama_token_data structs which are then immediately overwritten.	2024-09-18 14:42:36 +03:00
Vinesh Janarthanan	8a308354f6	server : match OAI structured output response (#9527 )	2024-09-18 09:50:34 +03:00
Eric Zhang	f799155ab8	server : fix OpenSSL build (remove obsolete `LOG_INFO`) (#9529 )	2024-09-18 09:28:20 +03:00
Neo Zhang Jianyu	faf67b3de4	[SYCL]set context default value to avoid memory issue, update guide (#9476 ) * set context default to avoid memory issue, update guide * Update docs/backend/SYCL.md Co-authored-by: Meng, Hengyu <hengyu.meng@intel.com> --------- Co-authored-by: arthw <14088817+arthw@users.noreply.github.com> Co-authored-by: Meng, Hengyu <hengyu.meng@intel.com>	2024-09-18 08:30:31 +08:00
Michael Podvitskiy	7be099fa81	llama-bench: correct argument parsing error message (#9524 )	2024-09-17 22:41:38 +02:00
Bert Wagner	8b836ae731	arg : add env variable for parallel (#9513 ) * add env variable for parallel * Update README.md with env: LLAMA_ARG_N_PARALLEL	2024-09-17 16:35:38 +03:00
Michael Podvitskiy	8344ef58f8	llama : fix n_vocab init for 'no_vocab' case (#9511 ) * llama: fixed n_vocab for `no_vocab` models * llama: updated error output for `llama_decode_internal` and `llama_encode_internal` * llama: log warning if there's no vocab_size in metadata * llama: correct vocab size for logging Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>	2024-09-17 13:18:22 +03:00
Max Krasnyansky	0226613853	threadpool : skip polling for unused threads (#9461 ) * threadpool: skip polling for unused threads Currently all threads do N polling rounds even if only 1 thread is active (n_threads_cur == 1). This commit adds a check to skip the polling for unused threads (ith >= n_threads_cur). n_threads_cur is now an atomic_int to explicitly tell thread sanitizer that it is written from one thread and read from other threads (not a race conditions). * threadpool: further simplify and improve ggml_barrier Avoid using strict memory order while polling, yet make sure that all threads go through full memory barrier (memory fence) on ggml_barrier entrace and exit. * threads: add simple barrier test This test does lots of small, parallel matmul ops where the barriers in between dominate the overhead. * threadpool: improve thread sync for new-graphs Using the same tricks as ggml_barrier. All the polling is done with relaxed memory order to keep it efficient, once the new graph is detected we do full fence using read-modify-write with strict memory order. * threadpool: improve abort handling Do not use threadpool->ec (exit code) to decide whether to exit the compute loop. threadpool->ec is not atomic which makes thread-sanitizer rightfully unhappy about it. Instead introduce atomic threadpool->abort flag used for this. This is consistent with how we handle threadpool->stop or pause. While at it add an explicit atomic_load for n_threads_cur for consistency. * test-barrier: release threadpool before releasing the context fixes use-after-free detected by gcc thread-sanitizer on x86-64 for some reason llvm sanitizer is not detecting this issue.	2024-09-17 11:19:46 +03:00
Yuri Khrustalev	503147a9f9	unicode : add <algorithm> (#9508 )	2024-09-17 09:51:15 +03:00
Gabe Goodhart	0d2ec43833	llama : support IBM Granite architecture (#9412 ) * feat(gguf-py): Add Granite model and params to gguf-py Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * feat(convert_hf_to_gguf): Add registration and param setup for Granite Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * feat(llama.cpp): Add config parsing for Granite multiplier params Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * feat(llama.cpp): First pass at full port of granite deviations from llama Something is still not working right since the results are mostly terrible, but on occasion it's producing relevant results at this point, so _something_ is working. Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(llama.cpp): Determine granite language 3b instruct by vocab size Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(convert_hf_to_gguf): Use LlamaModel as base for GraniteModel The defaults in LlamaModel are needed for Granite as well Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(llama.cpp): Switch Granite param names to use _scale for consistency Other scalar multipliers are called _scale, so this provides a more consistent naming convention. Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> fix(convert_hf_to_gguf/gguf-py): _multiplier -> _scale The transformers names with _multiplier will now be converted to the _scale equivalent during conversion. Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(llama.cpp): Use separate switch clause for granite in llm_load_hparams Branch: GraniteLM Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> --------- Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>	2024-09-17 09:44:58 +03:00
Michael Podvitskiy	37f3a3810e	llama : add llama_n_head() (#9512 )	2024-09-17 09:23:30 +03:00
slaren	23e0d70bac	ggml : move common CPU backend impl to new header (#9509 )	2024-09-16 16:22:07 +02:00
Daniel Bevenius	acb2c32c33	llama : rename n_embed to n_embd in rwkv6_time_mix (#9504 ) This commit renames n_embed to n_embd in llm_build_rwkv6_time_mix. The motivation for this change is consistency with the other rwkv6 functions like build_rwkv6 (and other parts of the code base).	2024-09-16 14:07:13 +03:00
Michael Podvitskiy	a6a3a5c531	ggml : link MATH_LIBRARY not by its full path (#9339 )	2024-09-16 14:06:50 +03:00
compilade	d54c21df7e	convert : identify missing model files (#9397 )	2024-09-16 10:30:22 +03:00
Georgi Gerganov	19514d632e	cmake : do not hide GGML options + rename option (#9465 ) * cmake : do not hide GGML options ggml-ci * build : rename flag GGML_CUDA_USE_GRAPHS -> GGML_CUDA_GRAPHS for consistency ggml-ci	2024-09-16 10:27:50 +03:00