graph : make FA compatible with MLA + add initial Metal kernels (#12953)

* graph : make mla compatible with FA

* metal : add exp FA kernels for DeepSeek models

ggml-ci

* llama : minor naming updates

ggml-ci

* ggml : disable FA for DS head sizes

* tests : add FA tests for MLA shapes

ggml-ci

.github/workflows/build.yml

@@ -1766,16 +1766,17 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
       run:
-       shell: bash -el {0}
-    runs-on: ubuntu-24.04-arm
+        shell: bash -el {0}
     strategy:
       matrix:
+        arch: [x86, aarch64]
         cann:
           - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10'
         device:
           - 'ascend910b3'
         build:
           - 'Release'
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     container: ascendai/cann:${{ matrix.cann }}
     steps:
       - name: Checkout

Makefile

@@ -780,10 +780,6 @@ ifdef GGML_HIP
 
 	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
 
-ifdef GGML_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # GGML_HIP_UMA
-
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
 	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas

common/common.cpp

@@ -830,7 +830,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {

convert_hf_to_gguf.py

@@ -4422,6 +4422,10 @@ class DeepseekV2Model(Model):
         self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
+
+        # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
+        self.hparams["num_key_value_heads"] = 1
+
         super().set_gguf_parameters()
         hparams = self.hparams
 
@@ -4430,8 +4434,13 @@ class DeepseekV2Model(Model):
         if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
             self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
         self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
-        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
-        self.gguf_writer.add_value_length(hparams["v_head_dim"])
+
+        # note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+        self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_mla(hparams["v_head_dim"])
+
         self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
         self.gguf_writer.add_expert_count(hparams["n_routed_experts"])
         self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"])
@@ -4500,6 +4509,26 @@ class DeepseekV2Model(Model):
             else:
                 return []
 
+        # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
+        if name.endswith("kv_b_proj.weight"):
+            name_kb = name.replace("kv_b_proj", "k_b_proj")
+            name_vb = name.replace("kv_b_proj", "v_b_proj")
+
+            n_head_kv = self.hparams["num_key_value_heads"]
+            v_head_dim = self.hparams["v_head_dim"]
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+
+            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
+
+            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            return [
+                (self.map_tensor_name(name_kb), k_b),
+                (self.map_tensor_name(name_vb), v_b)
+            ]
+
         return [(self.map_tensor_name(name), data_torch)]
 
     def prepare_tensors(self):

docs/build.md

@@ -259,8 +259,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
       cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
       && cmake --build build --config Release -- -j 16
   ```
-  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
-  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
 
   To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system.
 
@@ -296,6 +294,10 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
 
+### Unified Memory
+
+On Linux it is possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
 ## Vulkan
 
 **Windows**

examples/llava/clip.cpp

@@ -323,8 +323,8 @@ struct clip_ctx {
     std::vector<ggml_backend_t> backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    ggml_backend_ptr backend;
-    ggml_backend_ptr backend_cpu;
+    ggml_backend_t backend;
+    ggml_backend_t backend_cpu;
     ggml_backend_buffer_ptr buf;
 
     ggml_backend_sched_ptr sched;
@@ -332,27 +332,34 @@ struct clip_ctx {
     clip_image_size load_image_size;
 
     clip_ctx(clip_context_params & ctx_params) {
-        backend_cpu = ggml_backend_ptr(ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr));
-        backend     = ggml_backend_ptr(ctx_params.use_gpu
+        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+        backend     = ctx_params.use_gpu
                         ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr)
-                        : nullptr);
+                        : nullptr;
 
         if (backend) {
-            LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend.get()));
-            backend_ptrs.push_back(backend.get());
-            backend_buft.push_back(ggml_backend_get_default_buffer_type(backend.get()));
+            LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend));
+            backend_ptrs.push_back(backend);
+            backend_buft.push_back(ggml_backend_get_default_buffer_type(backend));
         } else {
-            backend = std::move(backend_cpu);
+            backend = backend_cpu;
             LOG_INF("%s: CLIP using CPU backend\n", __func__);
         }
 
-        backend_ptrs.push_back(backend_cpu.get());
-        backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu.get()));
+        backend_ptrs.push_back(backend_cpu);
+        backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu));
 
         sched.reset(
             ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false)
         );
     }
+
+    ~clip_ctx() {
+        ggml_backend_free(backend);
+        if (backend != backend_cpu) {
+            ggml_backend_free(backend_cpu);
+        }
+    }
 };
 
 static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
@@ -1428,7 +1435,7 @@ struct clip_model_loader {
             }
 
             // alloc memory and offload data
-            ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend.get());
+            ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend);
             ctx_clip.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft));
             ggml_backend_buffer_set_usage(ctx_clip.buf.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
             for (auto & t : tensors_to_load) {
@@ -2610,7 +2617,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         }
     }
 
-    ggml_backend_cpu_set_n_threads(ctx->backend_cpu.get(), n_threads);
+    ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
 
     auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf);
     if (status != GGML_STATUS_SUCCESS) {

examples/llava/gemma3-cli.cpp

@@ -317,6 +317,6 @@ int main(int argc, char ** argv) {
             is_first_msg = false;
         }
     }
-
+    llama_perf_context_print(ctx.lctx);
     return 0;
 }

examples/quantize/quantize.cpp

@@ -9,6 +9,7 @@
 #include <fstream>
 #include <cmath>
 #include <cctype>
+#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -16,7 +17,7 @@ struct quant_option {
     std::string desc;
 };
 
-static const std::vector<struct quant_option> QUANT_OPTIONS = {
+static const std::vector<quant_option> QUANT_OPTIONS = {
     { "Q4_0",     LLAMA_FTYPE_MOSTLY_Q4_0,     " 4.34G, +0.4685 ppl @ Llama-3-8B",  },
     { "Q4_1",     LLAMA_FTYPE_MOSTLY_Q4_1,     " 4.78G, +0.4511 ppl @ Llama-3-8B",  },
     { "Q5_0",     LLAMA_FTYPE_MOSTLY_Q5_0,     " 5.21G, +0.1316 ppl @ Llama-3-8B",  },
@@ -105,7 +106,8 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable);
+    printf("       [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
@@ -114,6 +116,8 @@ static void usage(const char * executable) {
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
+    printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
+    printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
     printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
@@ -244,6 +248,107 @@ static ggml_type parse_ggml_type(const char * arg) {
     return GGML_TYPE_COUNT;
 }
 
+// Allowed tensors for arbitrary quantization with --tensor-type option
+static const std::vector<std::string> ALLOWED_TENSOR_TYPE = {
+    "attn_k",
+    "attn_kv_a_mqa",
+    "attn_kv_b",
+    "attn_o",
+    "attn_output",
+    "attn_q",
+    "attn_q_a",
+    "attn_q_b",
+    "attn_qkv",
+    "attn_v",
+    "channel_mix_key",
+    "channel_mix_receptance",
+    "channel_mix_value",
+    "cls",
+    "cls.output",
+    "cross_attn_k",
+    "cross_attn_o",
+    "cross_attn_q",
+    "cross_attn_v",
+    "ffn_act",
+    "ffn_down",
+    "ffn_down_exps",
+    "ffn_down_shexp",
+    "ffn_gate",
+    "ffn_gate_exps",
+    "ffn_gate_shexp",
+    "ffn_up",
+    "ffn_up_exps",
+    "ffn_up_shexp",
+    "ssm_in",
+    "ssm_out",
+    "time_mix_gate",
+    "time_mix_key",
+    "time_mix_output",
+    "time_mix_receptance",
+    "time_mix_value",
+};
+
+// changes to this struct must be replicated in llama-quant.cpp
+struct tensor_quantization {
+    std::string name;
+    ggml_type quant = GGML_TYPE_COUNT;
+};
+
+static bool parse_tensor_type(const char * data, std::vector<tensor_quantization> & tensor_type) {
+    const char * sep = strchr(data, '=');
+    if (sep == nullptr) {
+        printf("\n%s: malformed tensor type '%s'\n\n", __func__, data);
+        return false;
+    }
+
+    const size_t tn_len = sep - data;
+    if (tn_len == 0) {
+        printf("\n%s: missing tensor name\n\n", __func__);
+        return false;
+    }
+
+    if (const size_t qt_len = strlen(sep); qt_len == 1) {
+        printf("\n%s: missing quantization type\n\n", __func__);
+        return false;
+    }
+
+    std::string tn(data, tn_len);
+    std::transform(tn.begin(), tn.end(), tn.begin(), tolower);
+    sep++;
+    const std::string qt(sep);
+
+    bool found = false;
+    for (const auto & allowed : ALLOWED_TENSOR_TYPE) {
+        std::string tensor;
+        tensor = tn.rfind('.') != std::string::npos ? tn.substr(tn.rfind('.') + 1) : tn;
+        // handle special case of cls.output
+        std::string cls_output = "cls.output";
+        if (tn.find(cls_output) != std::string::npos) {
+            tensor = "cls.output";
+        }
+        // check if an allowed tensor exists and it's at the end of the kv string
+        if (tensor == allowed) {
+            found = true;
+            break;
+        }
+    }
+    if (!found) {
+        printf("\n%s: invalid tensor name '%s'\n\n", __func__, tn.c_str());
+        return false;
+    }
+
+    if (parse_ggml_type(qt.c_str()) == GGML_TYPE_COUNT) {
+        printf("\n%s: invalid quantization type '%s'\n\n", __func__, qt.c_str());
+        return false;
+    }
+
+    tensor_quantization tqz;
+    tqz.name = tn;
+    tqz.quant = parse_ggml_type(qt.c_str());
+    tensor_type.emplace_back(std::move(tqz));
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -255,6 +360,7 @@ int main(int argc, char ** argv) {
     std::string imatrix_file;
     std::vector<std::string> included_weights, excluded_weights;
     std::vector<llama_model_kv_override> kv_overrides;
+    std::vector<tensor_quantization> tensor_types;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -277,6 +383,10 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--tensor-type") == 0) {
+            if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
             if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
@@ -361,6 +471,9 @@ int main(int argc, char ** argv) {
         kv_overrides.back().key[0] = 0;
         params.kv_overrides = &kv_overrides;
     }
+    if (!tensor_types.empty()) {
+        params.tensor_types = &tensor_types;
+    }
 
     llama_backend_init();
 

examples/rpc/rpc-server.cpp

@@ -126,7 +126,7 @@ static std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {

examples/sycl/build.sh

@@ -8,10 +8,10 @@ cd build
 source /opt/intel/oneapi/setvars.sh
 
 #for FP16
-#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON # faster for long-prompt inference
+#cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DLLAMA_CURL=OFF # faster for long-prompt inference
 
 #for FP32
-cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=OFF
 
 #build example/main
 #cmake --build . --config Release --target main

ggml/CMakeLists.txt

@@ -170,7 +170,6 @@ option(GGML_HIP                             "ggml: use HIP"
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
+#include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
 #include <aclnnop/aclnn_exp.h>
@@ -713,6 +714,270 @@ void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst);
  */
 void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/*
+ * @brief A generic wrapper for ACL resources with custom deleter support.
+ */
+using any_acl_resource = std::unique_ptr<void, std::function<void(void*)>>;
+
+/**
+ * @brief Trait structure used to define how to destroy a given ACL resource type.
+ *
+ * @tparam T ACL resource type.
+ */
+template<typename T>
+struct acl_resource_traits;
+
+/**
+ * @brief Specialization for aclTensor, defines how to destroy an aclTensor resource.
+ */
+template<>
+struct acl_resource_traits<aclTensor> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensor(static_cast<aclTensor*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclIntArray, defines how to destroy an aclIntArray resource.
+ */
+template<>
+struct acl_resource_traits<aclIntArray> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyIntArray(static_cast<aclIntArray*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclScalar, defines how to destroy an aclScalar resource.
+ */
+template<>
+struct acl_resource_traits<aclScalar> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyScalar(static_cast<aclScalar*>(p)));
+    }
+};
+
+/**
+ * @brief Specialization for aclTensorList, defines how to destroy an aclTensorList resource.
+ */
+template<>
+struct acl_resource_traits<aclTensorList> {
+    static void destroy(void* p) {
+        ACL_CHECK(aclDestroyTensorList(static_cast<aclTensorList*>(p)));
+    }
+};
+
+/**
+ * @brief Creates a generic ACL resource wrapper with proper destruction logic.
+ *
+ * @tparam T ACL resource type.
+ * @param ptr Raw pointer to ACL resource.
+ * @return any_acl_resource Smart pointer that handles destruction.
+ */
+template<typename T>
+any_acl_resource make_acl_resource(T* ptr) {
+    return any_acl_resource(
+        static_cast<void*>(ptr),
+        [](void* p) {
+            acl_resource_traits<T>::destroy(p);
+        }
+    );
+}
+
+/**
+ * @brief Registers multiple ACL resources into a vector for lifetime management.
+ *
+ * @tparam Args Variadic list of ACL resource types.
+ * @param vec Target vector to hold ACL resources.
+ * @param args Raw pointers to ACL resources.
+ */
+template<typename... Args>
+void register_acl_resources(std::vector<any_acl_resource>& vec, Args*... args) {
+    (vec.emplace_back(make_acl_resource(args)), ...);
+}
+
+/**
+ * @brief Task class that wraps the execution of an aclnn function call.
+ */
+class aclnn_task : public cann_task {
+    public:
+        aclnn_task(aclnn_func_t aclnn_func, void * workspace_addr,
+                   uint64_t workspace_size, aclOpExecutor * executor,
+                   aclrtStream stream) :
+            aclnn_func_(aclnn_func),
+            workspace_addr_(workspace_addr),
+            workspace_size_(workspace_size),
+            executor_(executor),
+            stream_(stream) {}
+        virtual void run_task() override {
+            ACL_CHECK(aclnn_func_(workspace_addr_, workspace_size_, executor_, stream_));
+        }
+    private:
+        aclnn_func_t aclnn_func_;
+        void *          workspace_addr_;
+        uint64_t        workspace_size_;
+        aclOpExecutor * executor_;
+        aclrtStream     stream_;
+};
+
+/**
+ * @brief Task class that releases ACL resources after usage.
+ */
+class release_resource_task : public cann_task {
+public:
+    release_resource_task(std::vector<any_acl_resource>&& resources){
+        resource_ = std::move(resources);
+    }
+
+    virtual void run_task() override {
+        resource_.clear();
+    }
+private:
+    std::vector<any_acl_resource> resource_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory copy operations.
+ */
+class async_memcpy_task : public cann_task {
+public:
+    async_memcpy_task(void* dst, const void* src, size_t size,
+                      aclrtMemcpyKind kind, aclrtStream stream)
+        : dst_(dst), src_(src), size_(size), kind_(kind), stream_(stream) {}
+
+    virtual void run_task() override {
+        ACL_CHECK(aclrtMemcpyAsync(dst_, size_, src_, size_, kind_, stream_));
+    }
+private:
+    void* dst_;
+    const void* src_;
+    size_t size_;
+    aclrtMemcpyKind kind_;
+    aclrtStream stream_;
+};
+
+/**
+ * @brief Task class for performing asynchronous memory set operations.
+ */
+class async_memset_task : public cann_task {
+    public:
+    async_memset_task(void* buffer, size_t size, int32_t value, aclrtStream stream)
+            : buffer_(buffer), size_(size), value_(value), stream_(stream) {}
+
+        virtual void run_task() override {
+            ACL_CHECK(aclrtMemsetAsync(buffer_, size_, value_, size_, stream_));
+        }
+    private:
+        void* buffer_;
+        size_t size_;
+        int32_t value_;
+        aclrtStream stream_;
+};
+
+/**
+ * @brief Launches an asynchronous task using the memory allocator.
+ *
+ * This macro submit an asynchronous task on the specified stream.
+ * The task uses memory allocated by the allocator. It is guaranteed
+ * that the memory will not be accessed by other tasks until this task
+ * completes, due to the sequential execution order within the same stream.
+ *
+ * @param OP_NAME aclnn operator name.
+ * @param args Additional arguments required by the task.
+ *
+ * @note
+ * Memory from the allocator will be "freed" immediately and can be
+ * reallocated to other pointers. However, it won't be accessed by any
+ * other task before this asynchronous task ends, because all tasks in the
+ * same stream are executed in queue order.
+ */
+
+#define GGML_CANN_CALL_ACLNN_OP(CTX, OP_NAME, ...)                                          \
+    do {                                                                                    \
+        uint64_t        workspaceSize = 0;                                                  \
+        aclOpExecutor * executor;                                                           \
+        void *          workspaceAddr = nullptr;                                            \
+        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor));\
+        /* workspace should alloced in main thread to keep malloc order when using vmm. */  \
+        if (workspaceSize > 0) {                                                            \
+            ggml_cann_pool_alloc workspace_allocator(CTX.pool(), workspaceSize);            \
+            workspaceAddr = workspace_allocator.get();                                      \
+        }                                                                                   \
+        if (CTX.async_mode) {                                                               \
+            auto task =                                                                     \
+                std::make_unique<aclnn_task>(aclnn##OP_NAME, workspaceAddr, workspaceSize,  \
+                    executor, CTX.stream()); \
+            CTX.task_queue.submit_task(std::move(task));                                    \
+        } else {                                                                            \
+            ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, CTX.stream()));\
+        }                                                                                   \
+    } while (0)
+
+/**
+ * @brief Registers and releases multiple ACL resources, optionally deferring the release
+ *        using a task.
+ *
+ * @tparam Args Types of the ACL resources.
+ * @param ctx Backend context which manages task submission and async mode.
+ * @param args Pointers to ACL resources to be released.
+ */
+template <typename... Args>
+void ggml_cann_release_resources(ggml_backend_cann_context & ctx, Args &&... args) {
+    std::vector<any_acl_resource> resources;
+    register_acl_resources(resources, std::forward<Args>(args)...);
+    if(ctx.async_mode) {
+        auto task = std::make_unique<release_resource_task>(std::move(resources));
+        ctx.task_queue.submit_task(std::move(task));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory copy operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param dst Destination memory address.
+ * @param src Source memory address.
+ * @param len Size of memory to copy (in bytes).
+ * @param kind Type of memory copy (host-to-device, device-to-host, etc).
+ */
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context & ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx.stream()));
+    }
+}
+
+inline void ggml_cann_async_memcpy(ggml_backend_cann_context * ctx, void * dst,
+                                   const void * src, size_t len, aclrtMemcpyKind kind) {
+    if (ctx->async_mode) {
+        auto task = std::make_unique<async_memcpy_task>(dst, const_cast<void *>(src), len, kind, ctx->stream());
+        ctx->task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemcpyAsync(dst, len, src, len, kind, ctx->stream()));
+    }
+}
+
+/**
+ * @brief Performs an asynchronous memory set operation, optionally deferred via task submission.
+ *
+ * @param ctx Backend context containing stream and async configuration.
+ * @param buffer Memory buffer to be set.
+ * @param size Size of the memory buffer (in bytes).
+ * @param value Value to set in the buffer.
+ */
+inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffer,
+                                   size_t size, int value) {
+    if (ctx.async_mode) {
+        auto task = std::make_unique<async_memset_task>(buffer, size, value, ctx.stream());
+        ctx.task_queue.submit_task(std::move(task));
+    } else {
+        ACL_CHECK(aclrtMemsetAsync(buffer, size, value, size, ctx.stream()));
+    }
+}
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.
@@ -742,42 +1007,9 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst);
     binary_op(ctx, acl_src0, acl_src1, acl_dst);
 
-    ACL_CHECK(aclDestroyTensor(acl_src0));
-    ACL_CHECK(aclDestroyTensor(acl_src1));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src0, acl_src1, acl_dst);
 }
 
-/**
- * @brief Launches an asynchronous task using the memory allocator.
- *
- * This macro submit an asynchronous task on the specified stream.
- * The task uses memory allocated by the allocator. It is guaranteed
- * that the memory will not be accessed by other tasks until this task
- * completes, due to the sequential execution order within the same stream.
- *
- * @param OP_NAME aclnn operator name.
- * @param args Additional arguments required by the task.
- *
- * @note
- * Memory from the allocator will be "freed" immediately and can be
- * reallocated to other pointers. However, it won't be accessed by any
- * other task before this asynchronous task ends, because all tasks in the
- * same stream are executed in queue order.
- */
-#define GGML_CANN_CALL_ACLNN_OP(OP_NAME, ...)                                                \
-    do {                                                                                     \
-        uint64_t        workspaceSize = 0;                                                   \
-        aclOpExecutor * executor;                                                            \
-        void *          workspaceAddr = nullptr;                                             \
-                                                                                             \
-        ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \
-                                                                                             \
-        if (workspaceSize > 0) {                                                             \
-            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);             \
-            workspaceAddr = workspace_allocator.get();                                       \
-        }                                                                                    \
-        ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, ctx.stream()));     \
-    } while (0)
 
 /**
  * @brief Applies a unary operation to an input tensor using the CANN backend.
@@ -799,9 +1031,7 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
     unary_op(ctx, acl_src, acl_dst);
-
-    ACL_CHECK(aclDestroyTensor(acl_src));
-    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ggml_cann_release_resources(ctx, acl_src, acl_dst);
 }
 
 /**
@@ -832,7 +1062,7 @@ void ggml_cann_unary_op(
  *
  * Internally, the lambda will call:
  * @code
- * GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);
+ * GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
  * @endcode
  *
  * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
@@ -840,14 +1070,14 @@ void ggml_cann_unary_op(
  * @see ggml_cann_unary_op
  * @see GGML_CANN_CALL_ACLNN_OP
  */
-#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                         \
-    do {                                                         \
-        auto lambda = [](ggml_backend_cann_context& ctx,         \
-            aclTensor* acl_src,                                  \
-            aclTensor* acl_dst) {                                \
-            GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst);  \
-        };                                                       \
-        ggml_cann_unary_op(lambda, ctx, dst);                    \
-    }                                                            \
+#define GGML_CANN_CALL_UNARY_OP(OP_NAME)                              \
+    do {                                                              \
+        auto lambda = [](ggml_backend_cann_context& ctx,              \
+            aclTensor* acl_src,                                       \
+            aclTensor* acl_dst) {                                     \
+            GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);  \
+        };                                                            \
+        ggml_cann_unary_op(lambda, ctx, dst);                         \
+    }                                                                 \
     while (0)
 #endif  // CANN_ACLNN_OPS

ggml/src/ggml-cann/common.h

@@ -31,9 +31,16 @@
 #include <memory>
 #include <string>
 #include <vector>
+#include <atomic>
+#include <condition_variable>
+#include <mutex>
+#include <thread>
+#include <unistd.h>
+#include <functional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
+#include "../ggml-impl.h"
 
 #define MATRIX_ROW_PADDING 512
 #define GGML_CANN_MAX_STREAMS 8
@@ -205,6 +212,127 @@ struct ggml_cann_pool_alloc {
     ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
 };
 
+/**
+ * @brief Function pointer type for ACLNN operator calls.
+ */
+using aclnn_func_t = aclnnStatus (*)(void*, uint64_t, aclOpExecutor*, aclrtStream);
+
+/**
+ * @brief Base class for all CANN tasks to be submitted to the task queue.
+ *
+ * Users should override the run_task() method with actual task logic.
+ */
+class cann_task {
+public:
+    virtual void run_task() {}
+};
+
+/**
+ * @brief A lock-free ring-buffer based task queue for asynchronously executing cann_task instances.
+ */
+class cann_task_queue {
+public:
+    /**
+     * @brief Constructs a task queue with a fixed power-of-two capacity for a specific device.
+     *
+     * @param capacity Queue capacity. Must be a power of 2.
+     * @param device Target device ID (used for context setting).
+     */
+    explicit cann_task_queue(size_t capacity, int32_t device)
+        : buffer_(capacity), capacity_(capacity), head_(0), tail_(0),
+          running_(false), device_(device) {
+        GGML_ASSERT((capacity & (capacity - 1)) == 0 && "capacity must be power of 2");
+        mask_ = capacity_ - 1;
+    }
+
+    /**
+     * @brief Attempts to enqueue a task into the queue.
+     *
+     * @param item Unique pointer to the task.
+     * @return true if the task was successfully enqueued, false if the queue was full.
+     */
+    bool enqueue(std::unique_ptr<cann_task>&& item) {
+        size_t next_tail = (tail_ + 1) & mask_;
+
+        if (next_tail == head_) {
+            return false;
+        }
+
+        buffer_[tail_] = std::move(item);
+        std::atomic_thread_fence(std::memory_order_release);
+        tail_ = next_tail;
+
+        return true;
+    }
+
+    /**
+     * @brief Submits a task to the queue, and starts the worker thread if not already running.
+     *
+     * @param task Task to be submitted.
+     */
+    void submit_task(std::unique_ptr<cann_task>&& task) {
+        while(!enqueue(std::move(task))) {
+            std::this_thread::yield();
+            continue;
+        }
+
+        if (!running_) {
+            running_ = true;
+            thread_ = std::thread(&cann_task_queue::execute, this);
+        }
+
+    }
+
+    /**
+     * @brief Waits until the queue is completely empty and no tasks are being processed.
+     */
+    void wait() {
+        while (running_ && head_ != tail_) {
+            std::this_thread::yield();
+            continue;
+        }
+    }
+
+    /**
+     * @brief Stops the task queue and joins the worker thread.
+     */
+    void stop() {
+        running_ = false;
+        if (thread_.joinable()) {
+            thread_.join();
+        }
+    }
+
+private:
+    /**
+     * @brief Worker thread function that continuously dequeues and executes tasks.
+     */
+    void execute() {
+        ggml_cann_set_device(device_);
+
+        while (running_) {
+            if(head_ == tail_) {
+                std::this_thread::yield();
+                continue;
+            }
+
+            std::atomic_thread_fence(std::memory_order_acquire);
+            buffer_[head_]->run_task();
+            buffer_[head_].reset();
+            head_ = (head_ + 1) & mask_;
+        }
+    }
+
+    std::vector<std::unique_ptr<cann_task>> buffer_;
+    const size_t capacity_;
+    size_t mask_;
+    size_t head_;
+    size_t tail_;
+    bool running_;
+    std::thread thread_;
+    int32_t device_;
+};
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -213,6 +341,8 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+    cann_task_queue task_queue;
+    bool async_mode;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -221,9 +351,12 @@ struct ggml_backend_cann_context {
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {
+        : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
+        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+        GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
+            device, async_mode ? "ON" : "OFF");
     }
 
     /**
@@ -231,6 +364,7 @@ struct ggml_backend_cann_context {
      */
     ~ggml_backend_cann_context() {
         ggml_cann_set_device(device);
+        task_queue.stop();
         if (copy_event != nullptr) {
             ACL_CHECK(aclrtDestroyEvent(copy_event));
         }

ggml/src/ggml-cann/ggml-cann.cpp

@@ -29,6 +29,8 @@
 #include <cstdio>
 #include <cstring>
 #include <mutex>
+#include <queue>
+#include <chrono>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -119,9 +121,10 @@ static ggml_cann_device_info ggml_cann_init() {
         prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
         prop.location.id = id;
         prop.reserve = 0;
-        ACL_CHECK(aclrtMemGetAllocationGranularity(
+        err = aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
-            &info.devices[id].vmm_granularity));
+            &info.devices[id].vmm_granularity);
+        info.devices[id].vmm = err == ACL_SUCCESS;
 
         size_t free, total;
         ggml_backend_cann_get_device_memory(id, &free, &total);
@@ -148,11 +151,223 @@ const ggml_cann_device_info& ggml_cann_info() {
 
 //#define DEBUG_CANN_MALLOC
 /**
- * @brief A pool of CANN buffers(legacy).
+ * @brief A pool of CANN buffers(priority segment buffer).
  *
  * This class manages a pool of CANN buffers for a specific device.
  */
-struct ggml_cann_pool_leg : public ggml_cann_pool {
+struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer.
+        size_t size = 0;      ///< Size of the buffer.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
+
+        bool operator>(const ggml_cann_buffer& other) const {
+            return size > other.size;
+        }
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    std::unordered_map<void*, size_t> buffer_pool;
+    std::priority_queue<ggml_cann_buffer,
+                        std::vector<ggml_cann_buffer>,
+                        std::greater<>> free_buffers ;
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_buf_prio(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_buf_prio() {
+        ggml_cann_set_device(device);
+        for (auto& [b_ptr, b_size] : buffer_pool) {
+            aclrtFree(b_ptr);
+            pool_size -= b_size;
+        }
+        buffer_pool.clear();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        std::vector<ggml_cann_buffer> free_buffers_rest;
+        free_buffers_rest.reserve(free_buffers.size());
+        while (!free_buffers.empty()) {
+            auto b = free_buffers.top();
+            free_buffers.pop();
+
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    ptr = b.ptr;
+#ifdef DEBUG_CANN_MALLOC
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
+#endif
+                    break;
+                }
+            }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+                buffer_pool.erase(b.ptr);
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                continue;
+            }
+            free_buffers_rest.push_back(b);
+        }
+        for (ggml_cann_buffer &b : free_buffers_rest) {
+            free_buffers.push(std::move(b));
+        }
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d] free pool_size = %5u MB\n\n", device, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+        if (ptr != nullptr) {
+            return ptr;
+        }
+
+        // allocate a new buffer if no buffer can be reused
+        ggml_cann_set_device(device);
+        ACL_CHECK(aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: allocate %p, "
+            "pool_size = %5u MB, "
+            "size = %5u MB\n",
+            device, ptr, (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+            (uint32_t)(GGML_PAD(size, 1048576) / 1048576));
+#endif
+        buffer_pool.emplace(ptr, size);
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
+        auto it = buffer_pool.find(ptr);
+        if (it == buffer_pool.end()) {
+            GGML_ABORT("cann pool[%d]: buffer %p not found in pool\n", device, ptr);
+        }
+
+        auto now = std::chrono::steady_clock::now();
+        free_buffers.emplace(ggml_cann_buffer{ptr, it->second, now});
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "cann pool[%d]: return   %p, "
+            "pool_size = %5u MB\n",
+            device, ptr,
+            (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers(segment buffer).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_buf : public ggml_cann_pool {
+    /**
+     * @brief The maximum reuse margin for a buffer.
+     */
+    static const size_t max_reuse_margin = 1ull << 22;  // 4MB
+
+    /**
+     * @brief The minimum free margin for a buffer.
+     */
+    static const size_t min_free_margin = 1ull << 20;   // 1MB
+
+    /**
+     * @brief The alignment for buffer allocation.
+     */
+    static const size_t alignment = 128;
+
     /**
      * @brief The maximum number of buffers in the pool.
      */
@@ -163,12 +378,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      */
     int device;
 
+    /**
+     * @brief Whether to disable clean during buffer allocation.
+     */
+    bool disable_clean = false;
+
     /**
      * @brief Structure representing a CANN buffer.
      */
     struct ggml_cann_buffer {
         void* ptr = nullptr;  ///< Pointer to the buffer memory.
         size_t size = 0;      ///< Size of the buffer.
+        bool used = false;    ///< Whether the buffer is currently in use.
+        std::chrono::steady_clock::time_point last_used;  ///< Last used time.
     };
 
     /**
@@ -186,17 +408,19 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      *
      * @param device The device ID to associate with this buffer pool.
      */
-    explicit ggml_cann_pool_leg(int device) : device(device) {}
+    explicit ggml_cann_pool_buf(int device) : device(device) {
+        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+    }
 
     /**
      * @brief Destructor to free all buffers in the pool.
      */
-    ~ggml_cann_pool_leg() {
+    ~ggml_cann_pool_buf() {
         ggml_cann_set_device(device);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
             if (b.ptr != nullptr) {
-                ACL_CHECK(aclrtFree(b.ptr));
+                aclrtFree(b.ptr);
                 pool_size -= b.size;
             }
         }
@@ -212,63 +436,93 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
-        const size_t alignment = 128;
         size = GGML_PAD(size, alignment);
         if (size == 0) {
             size = alignment;
         }
-#ifdef DEBUG_CANN_MALLOC
-        int nnz = 0;
-        size_t max_size = 0;
-#endif
-        size_t best_diff = 1ull << 36;
-        int ibest = -1;
-        for (int i = 0; i < MAX_BUFFERS; ++i) {
+
+        void* ptr = nullptr;
+        auto now = std::chrono::steady_clock::now();
+
+        int i = 0;
+        for (; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr != nullptr) {
+            if (b.ptr == nullptr) {
+                break;
+            }
+            if (b.used) {
+                continue;
+            }
+            if (b.size >= size) {
+                // reuse the buffer if the size is enough
+                const size_t margin = b.size - size;
+                if (margin <= max_reuse_margin) {
+                    *actual_size = b.size;
+                    b.used = true;
+                    ptr = b.ptr;
 #ifdef DEBUG_CANN_MALLOC
-                ++nnz;
-                if (b.size > max_size) max_size = b.size;
+                    GGML_LOG_INFO(
+                        "cann pool[%d]: reused   %p, "
+                        "pool_size = %5u MB, "
+                        "size = %5u MB, "
+                        "margin = %5u MB\n",
+                        device, b.ptr,
+                        (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(size, 1048576) / 1048576),
+                        (uint32_t)(GGML_PAD(margin, 1048576) / 1048576));
 #endif
-                if (b.size >= size) {
-                    size_t diff = b.size - size;
-                    if (diff < best_diff) {
-                        best_diff = diff;
-                        ibest = i;
-                        if (!best_diff) {
-                            void* ptr = b.ptr;
-                            *actual_size = b.size;
-                            b.ptr = nullptr;
-                            b.size = 0;
-                            return ptr;
-                        }
-                    }
+                    break;
                 }
             }
+
+            bool should_clean = !disable_clean &&
+                                b.size > min_free_margin &&
+                                std::chrono::duration_cast<std::chrono::milliseconds>(now - b.last_used).count() > 100;
+            if (should_clean) {
+                // free the buffer if the size is needed to be freed
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+#ifdef DEBUG_CANN_MALLOC
+                GGML_LOG_INFO(
+                    "cann pool[%d]: clean    %p, "
+                    "pool_size = %5u MB, "
+                    "size = %5u MB\n",
+                    device, b.ptr,
+                    (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                    (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
+#endif
+                b.ptr = nullptr;
+            }
         }
-        if (ibest >= 0) {
-            ggml_cann_buffer& b = buffer_pool[ibest];
-            void* ptr = b.ptr;
-            *actual_size = b.size;
-            b.ptr = nullptr;
-            b.size = 0;
+        if (ptr != nullptr) {
             return ptr;
         }
-        void* ptr;
-        ggml_cann_set_device(device);
-        ACL_CHECK(
-            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = size;
-        pool_size += size;
+
+        if (i < MAX_BUFFERS) {
+            // allocate a new buffer if no buffer can be reused
+            ggml_cann_buffer& b = buffer_pool[i];
+            ggml_cann_set_device(device);
+            ACL_CHECK(aclrtMalloc(&b.ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+            pool_size += size;
+            *actual_size = size;
+            b.size = size;
+            b.used = true;
+            if (i >= MAX_BUFFERS - 8) {
+                GGML_LOG_WARN("cann pool[%d]: slots almost full\n", device);
+            }
 #ifdef DEBUG_CANN_MALLOC
-        GGML_LOG_INFO(
-            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
-            "requested %u MB\n",
-            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
-            (uint32_t)(pool_size / 1024 / 1024),
-            (uint32_t)(size / 1024 / 1024));
+            GGML_LOG_INFO(
+                "cann pool[%d]: allocate %p, "
+                "pool_size = %5u MB, "
+                "size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576),
+                (uint32_t)(GGML_PAD(b.size, 1048576) / 1048576));
 #endif
-        return ptr;
+            return b.ptr;
+        }
+
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 
     /**
@@ -278,18 +532,24 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @param size Size of the buffer to free.
      */
     void free(void* ptr, size_t size) override {
+        GGML_UNUSED(size);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cann_buffer& b = buffer_pool[i];
-            if (b.ptr == nullptr) {
-                b.ptr = ptr;
-                b.size = size;
-                return;
+            if (b.ptr != ptr) {
+                continue;
             }
+            b.used = false;
+            b.last_used = std::chrono::steady_clock::now();
+#ifdef DEBUG_CANN_MALLOC
+            GGML_LOG_INFO(
+                "cann pool[%d]: return   %p, "
+                "pool_size = %5u MB\n",
+                device, b.ptr,
+                (uint32_t)(GGML_PAD(pool_size, 1048576) / 1048576));
+#endif
+            return;
         }
-        // memory should always buffered. these memory may still needed by
-        // tasks in stream.
-        // TODO, fix me.
-        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+        GGML_ABORT("cann pool[%d]: slots full\n", device);
     }
 };
 
@@ -347,8 +607,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_vmm(int device)
-        : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+    : device(device) {
         auto dev = ggml_cann_info().devices[device];
         granularity = dev.vmm_granularity;
         max_size = dev.total_vram;
@@ -471,7 +730,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
+    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
+        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    }
+    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
+    if (enable_buf_prio) {
+        GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
+    }
+    GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
+    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }
 
 // cann buffer
@@ -1020,8 +1290,11 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
 
     ggml_cann_set_device(buft_ctx->device);
 
-    size = std::max(size, (size_t)1);
-
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
     void* dev_ptr;
     aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
     if (err != ACL_SUCCESS) {
@@ -1333,7 +1606,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
                     auto lambda = [](ggml_backend_cann_context& ctx,
                         aclTensor* acl_src,
                         aclTensor* acl_dst) {
-                        GGML_CANN_CALL_ACLNN_OP(GeluV2, acl_src, 0, acl_dst);
+                        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
                     };
                     ggml_cann_unary_op(lambda, ctx, dst);
                 } break;
@@ -1516,12 +1789,11 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
     delete backend;
 }
 
+
 /**
  * @brief Sets tensor data asynchronously in the CANN backend.
  *
- * This function asynchronously sets tensor data in the CANN backend. Depending
- * on the tensor type, it may perform data transformations before copying data
- * to the device.
+ * This function asynchronously sets tensor data in the CANN backend.
  *
  * @param backend Pointer to the CANN backend structure.
  * @param tensor Pointer to the tensor structure to set data for.
@@ -1536,23 +1808,28 @@ static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
                                                size_t size) {
     ggml_backend_cann_context *cann_ctx =
         (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
-                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+        "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
 
-        ACL_CHECK(aclrtMemcpyAsync(
-            (char *)tensor->data + offset, size, transform_buffer, size,
-            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        free(transform_buffer);
-    }
+    ggml_cann_async_memcpy(cann_ctx, (char *)tensor->data + offset, data, size,
+        ACL_MEMCPY_HOST_TO_DEVICE);
 }
 
+/**
+ * @brief Gets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously gets tensor data in the CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to get data from.
+ * @param data Pointer to the host data to copy from the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
 static void ggml_backend_cann_get_tensor_async(
     ggml_backend_t backend, const ggml_tensor *tensor, void *data,
     size_t offset, size_t size) {
@@ -1563,20 +1840,11 @@ static void ggml_backend_cann_get_tensor_async(
 
     GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
                 "unsupported buffer type");
+    GGML_ASSERT(!ggml_is_quantized(tensor->type));
+
+    ggml_cann_async_memcpy(cann_ctx, data, (char *)tensor->data + offset, size,
+        ACL_MEMCPY_DEVICE_TO_HOST);
 
-    if (!need_transform(tensor->type)) {
-        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
-                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
-                                   cann_ctx->stream()));
-    } else {
-        void *transform_buffer = malloc(size);
-        ACL_CHECK(aclrtMemcpyAsync(
-            transform_buffer, size, (char *)tensor->data + offset, size,
-            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
-        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
-        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
-        free(transform_buffer);
-    }
 }
 
 /**
@@ -1636,6 +1904,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
 
+        // wait for task_queue empty to keep task order.
+        cann_ctx_src->task_queue.wait();
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
@@ -1663,9 +1933,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
 static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
-
+    cann_ctx->task_queue.wait();
     ggml_cann_set_device(cann_ctx->device);
-
     ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
 }
 
@@ -1749,6 +2018,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return true;
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
                     // only support contiguous for quantized types.
                     return ggml_is_contiguous(op->src[0]) &&
                             ggml_is_contiguous(op->src[1]);
@@ -1816,6 +2089,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 return false;
             }
 
+            if(!ggml_is_contiguous(op->src[0])){
+                return false;
+            }
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -1831,6 +2107,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_POOL_2D: {
             const int32_t * opts = (const int32_t *) op->op_params;
+#ifdef ASCEND_310P
+            enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
+            if(opt == GGML_OP_POOL_MAX){
+                return false;
+            }
+#endif
             const int       k0   = opts[1];
             const int       k1   = opts[2];
             const int       p0   = opts[5];

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -425,6 +425,8 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
         }
         case GGML_OP_IM2COL_BACK:
             return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
+        case GGML_OP_GET_ROWS_BACK:
+            return src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16;
         case GGML_OP_OUT_PROD:
             return (src0->type == GGML_TYPE_F32 || (ggml_is_quantized(src0->type) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
                 src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;

ggml/src/ggml-cuda/cpy.cu

@@ -551,7 +551,7 @@ static void ggml_cpy_f16_f16_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -588,7 +588,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char ** dest_ptrs_d = nullptr;
     int graph_cpynode_index = -1;
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
@@ -636,7 +636,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
     }
 #if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)
-    if(ctx.cuda_graph->use_cpy_indirection) {
+    if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
     }
 #endif
@@ -645,7 +645,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
-    ggml_cuda_cpy(ctx, src0, dst);
+    bool disable_indirection = true;
+    ggml_cuda_cpy(ctx, src0, dst, disable_indirection);
 }
 
 void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {

ggml/src/ggml-cuda/cpy.cuh

@@ -2,7 +2,7 @@
 
 #define CUDA_CPY_BLOCK_SIZE 64
 
-void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1,  bool disable_indirection = false);
 
 void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -96,31 +96,32 @@ int ggml_cuda_get_device() {
 
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
-    auto res = hipMallocManaged(ptr, size);
-    if (res == hipSuccess) {
-        // if error we "need" to know why...
-        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
-    }
-    return res;
-#else
-
-#if !defined(GGML_USE_HIP)
     cudaError_t err;
     if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
     {
         err = cudaMallocManaged(ptr, size);
+#if defined(GGML_USE_HIP)
+        if (err == hipSuccess) {
+            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+        }
+
+        // fall back to cudaMalloc if not supported (e.g. on Windows)
+        if (err == hipErrorNotSupported) {
+            static bool warned_unsupported = false;
+            if (!warned_unsupported) {
+                GGML_LOG_WARN("hipMallocManaged unsupported, falling back to hipMalloc.\n");
+                warned_unsupported = true;
+            }
+
+            err = cudaMalloc(ptr, size);
+        }
+#endif // defined(GGML_USE_HIP)
     }
     else
     {
         err = cudaMalloc(ptr, size);
     }
     return err;
-#else
-    return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIP)
-
-#endif
 }
 
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@@ -2491,7 +2492,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
         if (node->op == GGML_OP_MUL_MAT_ID) {
             use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
         }
 
@@ -3236,6 +3237,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek MLA
+                return false;
+            }
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

ggml/src/ggml-cuda/vendors/hip.h

@@ -71,6 +71,8 @@
 #define cudaLaunchHostFunc hipLaunchHostFunc
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMallocManaged hipMallocManaged
+#define cudaMemAdvise hipMemAdvise
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpyAsync hipMemcpyAsync
 #define cudaMemcpyPeerAsync hipMemcpyPeerAsync

ggml/src/ggml-hip/CMakeLists.txt

@@ -89,10 +89,6 @@ endif()
 
 add_compile_definitions(GGML_USE_HIP)
 
-if (GGML_HIP_UMA)
-    add_compile_definitions(GGML_HIP_UMA)
-endif()
-
 if (GGML_CUDA_FORCE_MMQ)
     add_compile_definitions(GGML_CUDA_FORCE_MMQ)
 endif()

ggml/src/ggml-metal/ggml-metal.m

@@ -354,6 +354,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,
@@ -362,6 +363,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,
@@ -370,6 +372,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,
@@ -378,6 +381,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,
@@ -386,6 +390,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,
@@ -394,6 +399,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,
@@ -402,6 +408,14 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,
@@ -430,6 +444,13 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,
     GGML_METAL_KERNEL_TYPE_SET_I32,
     GGML_METAL_KERNEL_TYPE_SET_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
@@ -1011,6 +1032,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192,         flash_attn_ext_f16_h192,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128,  flash_attn_ext_f16_hk192_hv128,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,         flash_attn_ext_f16_h256,         has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512,  flash_attn_ext_f16_hk576_hv512,  has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64,         flash_attn_ext_bf16_h64,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80,         flash_attn_ext_bf16_h80,         has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96,         flash_attn_ext_bf16_h96,         has_simdgroup_mm && use_bfloat);
@@ -1019,6 +1041,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192,        flash_attn_ext_bf16_h192,        has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128, flash_attn_ext_bf16_hk192_hv128, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256,        flash_attn_ext_bf16_h256,        has_simdgroup_mm && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512, flash_attn_ext_bf16_hk576_hv512, has_simdgroup_mm && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64,         flash_attn_ext_q4_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80,         flash_attn_ext_q4_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96,         flash_attn_ext_q4_0_h96,         has_simdgroup_mm);
@@ -1027,6 +1050,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192,        flash_attn_ext_q4_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128, flash_attn_ext_q4_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256,        flash_attn_ext_q4_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512, flash_attn_ext_q4_0_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64,         flash_attn_ext_q4_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80,         flash_attn_ext_q4_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96,         flash_attn_ext_q4_1_h96,         has_simdgroup_mm);
@@ -1035,6 +1059,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192,        flash_attn_ext_q4_1_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128, flash_attn_ext_q4_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256,        flash_attn_ext_q4_1_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512, flash_attn_ext_q4_1_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64,         flash_attn_ext_q5_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80,         flash_attn_ext_q5_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96,         flash_attn_ext_q5_0_h96,         has_simdgroup_mm);
@@ -1043,6 +1068,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192,        flash_attn_ext_q5_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128, flash_attn_ext_q5_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256,        flash_attn_ext_q5_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512, flash_attn_ext_q5_0_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64,         flash_attn_ext_q5_1_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80,         flash_attn_ext_q5_1_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96,         flash_attn_ext_q5_1_h96,         has_simdgroup_mm);
@@ -1051,6 +1077,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192,        flash_attn_ext_q5_1_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128, flash_attn_ext_q5_1_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256,        flash_attn_ext_q5_1_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512, flash_attn_ext_q5_1_hk576_hv512, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64,         flash_attn_ext_q8_0_h64,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80,         flash_attn_ext_q8_0_h80,         has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96,         flash_attn_ext_q8_0_h96,         has_simdgroup_mm);
@@ -1059,6 +1086,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192,        flash_attn_ext_q8_0_h192,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,        flash_attn_ext_q8_0_h256,        has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,      flash_attn_ext_vec_f16_h96,      has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,     flash_attn_ext_vec_bf16_h96,     has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,     flash_attn_ext_vec_q4_0_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96,     flash_attn_ext_vec_q4_1_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96,     flash_attn_ext_vec_q5_0_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96,     flash_attn_ext_vec_q5_1_h96,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96,     flash_attn_ext_vec_q8_0_h96,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,     flash_attn_ext_vec_f16_h128,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128,    flash_attn_ext_vec_bf16_h128,    has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128,    flash_attn_ext_vec_q4_0_h128,    has_simdgroup_reduction);
@@ -1087,6 +1122,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,    flash_attn_ext_vec_q5_0_h256,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,    flash_attn_ext_vec_q5_1_h256,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,    flash_attn_ext_vec_q8_0_h256,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512,     flash_attn_ext_vec_f16_hk576_hv512,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512,    flash_attn_ext_vec_bf16_hk576_hv512,    has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512,    flash_attn_ext_vec_q4_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512,    flash_attn_ext_vec_q4_1_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512,    flash_attn_ext_vec_q5_0_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512,    flash_attn_ext_vec_q5_1_hk576_hv512,    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512,    flash_attn_ext_vec_q8_0_hk576_hv512,    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32,                         set_f32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32,                         set_i32,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                     cpy_f32_f32,                     true);
@@ -1351,6 +1393,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 // TODO: not sure if it is worth adding kernels for this size
                 return false;
             }
+            if (op->src[0]->ne[0] == 576) {
+                // DeepSeek sizes
+                // TODO: disabled for now, until optmized
+                return false;
+            }
             if (op->src[1]->type != op->src[2]->type) {
                 return false;
             }
@@ -3843,12 +3890,14 @@ static void ggml_metal_encode_node(
                 // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
                 //       for now avoiding mainly to keep the number of templates/kernels a bit lower
                 //       these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
-                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 192)) {
+                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16:
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
@@ -3871,6 +3920,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break;
@@ -3893,6 +3944,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64 ].pipeline; break;
@@ -3915,6 +3968,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64 ].pipeline; break;
@@ -3937,6 +3992,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64 ].pipeline; break;
@@ -3959,6 +4016,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64 ].pipeline; break;
@@ -3981,6 +4040,8 @@ static void ggml_metal_encode_node(
                             {
                                 if (ne00 == 192 && ne20 == 128) {
                                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128].pipeline;
+                                } else if (ne00 == 576 && ne20 == 512) {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512].pipeline;
                                 } else {
                                     switch (ne00) {
                                         case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64 ].pipeline; break;
@@ -4010,6 +4071,24 @@ static void ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
+                        case 96:
+                            {
+                                switch (src1->type) {
+                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96].pipeline; break;
+                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96].pipeline; break;
+                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H96].pipeline; break;
+                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H96].pipeline; break;
+                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H96].pipeline; break;
+                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H96].pipeline; break;
+                                    default:
+                                        {
+                                            GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                            GGML_LOG_ERROR("add template specialization for this type\n");
+                                            GGML_ABORT("add template specialization for this type");
+                                        }
+                                }
+                            } break;
                         case 128:
                             {
                                 switch (src1->type) {
@@ -4082,12 +4161,36 @@ static void ggml_metal_encode_node(
                                         }
                                 }
                             } break;
+                        case 576:
+                            {
+                                if (ne20 == 512) {
+                                    switch (src1->type) {
+                                        case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK576_HV512].pipeline; break;
+                                        case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK576_HV512].pipeline; break;
+                                        default:
+                                            {
+                                                GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                                GGML_LOG_ERROR("add template specialization for this type\n");
+                                                GGML_ABORT("add template specialization for this type");
+                                            }
+                                    }
+                                } else {
+                                    GGML_LOG_ERROR("unsupported size: %lld\n", ne20);
+                                    GGML_LOG_ERROR("add template specialization for this size\n");
+                                    GGML_ABORT("add template specialization for this size");
+                                }
+                            } break;
                         default:
-                                  {
-                                      GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
-                                      GGML_LOG_ERROR("add template specialization for this size\n");
-                                      GGML_ABORT("add template specialization for this size");
-                                  }
+                            {
+                                GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                GGML_LOG_ERROR("add template specialization for this size\n");
+                                GGML_ABORT("add template specialization for this size");
+                            }
                     }
                 }
 

ggml/src/ggml-metal/ggml-metal.metal

@@ -3546,6 +3546,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h128")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_h192")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 192>;
 template [[host_name("kernel_flash_attn_ext_f16_hk192_hv128")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  192, 128>;
 template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  256, 256>;
+template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
@@ -3556,6 +3557,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3566,6 +3568,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
@@ -3575,6 +3578,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q4_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q4_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
@@ -3584,6 +3588,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
@@ -3593,6 +3598,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q5_1_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q5_1_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 576, 512>;
 
 template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
@@ -3602,6 +3608,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h128")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 192>;
 template [[host_name("kernel_flash_attn_ext_q8_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 192, 128>;
 template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
 
@@ -3959,6 +3966,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  96, 96, 4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 96, 96, 4>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 96, 96, 4>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 96, 96, 4>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  128, 128, 4>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 128, 128, 4>;
@@ -3999,6 +4016,16 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_
 template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 256, 256, 4>;
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 256, 256, 4>;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_hk576_hv512")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,      1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  576, 512, 2>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,    1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 576, 512, 2>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1, 8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0, 8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1, 8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 576, 512, 2>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk576_hv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 576, 512, 2>;
+
 #undef FA_TYPES
 
 template<typename T>

ggml/src/ggml-opencl/CMakeLists.txt

@@ -54,16 +54,41 @@ function(ggml_opencl_add_kernel KNAME)
 endfunction()
 
 set(GGML_OPENCL_KERNELS
-    ggml-opencl
-    ggml-opencl_mm
-    ggml-opencl_cvt
-    ggml-opencl_gemv_noshuffle
-    ggml-opencl_gemv_noshuffle_general
-    ggml-opencl_mul_mat_Ab_Bi_8x4
-    ggml-opencl_transpose_16
-    ggml-opencl_transpose_32
-    ggml-opencl_transpose_32_16
-    ggml-opencl_im2col
+    add
+    clamp
+    cpy
+    cvt
+    diag_mask_inf
+    gelu
+    gemv_noshuffle_general
+    gemv_noshuffle
+    get_rows
+    im2col_f32
+    im2col_f16
+    mul_mat_Ab_Bi_8x4
+    mul_mv_f16_f16
+    mul_mv_f16_f32_1row
+    mul_mv_f16_f32_l4
+    mul_mv_f16_f32
+    mul_mv_f32_f32
+    mul_mv_q4_0_f32
+    mul_mv_q4_0_f32_v
+    mul_mv_q4_0_f32_8x_flat
+    mul_mv_q4_0_f32_1d_8x_flat
+    mul_mv_q4_0_f32_1d_16x_flat
+    mul_mv_q6_k
+    mul
+    norm
+    relu
+    rms_norm
+    rope
+    scale
+    silu
+    softmax_4_f32
+    softmax_4_f16
+    softmax_f32
+    softmax_f16
+    transpose
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

ggml/src/ggml-opencl/kernels/add.cl

@@ -0,0 +1,83 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add
+//------------------------------------------------------------------------------
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
+kernel void kernel_add(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global char * dst,
+        ulong  offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int   ne10,
+        int   ne11,
+        int   ne12,
+        int   ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int   ne0,
+        int   ne1,
+        int   ne2,
+        int   ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) + *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_add_row(
+        global float4 * src0,
+        ulong  offset0,
+        global float4 * src1,
+        ulong  offset1,
+        global float4 * dst,
+        ulong  offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] + src1[idx1];
+}

ggml/src/ggml-opencl/kernels/clamp.cl

@@ -0,0 +1,20 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// clamp
+//------------------------------------------------------------------------------
+kernel void kernel_clamp(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float min,
+        float max
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = src0[get_global_id(0)] < min ?
+        min :
+        (src0[get_global_id(0)] > max ? max : src0[get_global_id(0)]);
+}

ggml/src/ggml-opencl/kernels/cpy.cl

@@ -0,0 +1,184 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// cpy
+//------------------------------------------------------------------------------
+
+kernel void kernel_cpy_f16_f16(
+        global half * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f16_f32(
+        global half * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f16(
+        global float * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}

ggml/src/ggml-opencl/kernels/cvt.cl

@@ -1,39 +1,20 @@
 //------------------------------------------------------------------------------
-// This file is contains additional kernels for data conversion.
+// This file is contains kernels for data conversion.
 // These kernels are used when loading the model, so its performance is less
 // important.
 //------------------------------------------------------------------------------
-#ifdef cl_khr_fp16
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
-#elif defined(cl_amd_fp16)
-#pragma OPENCL EXTENSION cl_amd_fp16 : enable
-#else
-#error "Half precision floating point not supportedby OpenCL implementation on your device."
-#endif
-
-#ifdef cl_khr_subgroups
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#elif defined(cl_intel_subgroups)
-#pragma OPENCL EXTENSION cl_intel_subgroups : enable
-#else
-#error "Subgroup not supported on your device."
-#endif
 
 #ifdef cl_intel_required_subgroup_size
-// Always use subgroup size of 32 on Intel.
 #pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
 #define INTEL_GPU 1
 #define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
 #define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
 #elif defined(cl_qcom_reqd_sub_group_size)
-// Always use subgroups size of 64 on Adreno.
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
 #define ADRENO_GPU 1
 #define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
 #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#else
-// TODO: do not know how to choose subgroup size on other GPUs.
-#error "Selecting subgroup size is not supported on your device."
 #endif
 
 #define QK4_0                   32
@@ -66,13 +47,44 @@ struct block_q4_0
 };
 
 //------------------------------------------------------------------------------
-// mul_vec_q_n_f32_flat_noshuffle
-//
-// This variation uses flat arrays (struct of arrays, SOA) representation for
-// quant tensors. It also uses non shuffled bit order for weights.
-//
-// The shuffled version is kept in the original file because moving it here
-// seems to result in worse performance for adreno.
+// kernel_convert_block_q4_0
+// Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_0(
+    global struct block_q4_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK4_0/2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q4_0(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_0_noshuffle
+// Flatten q4_0 weights and unshuffle the bits
 //------------------------------------------------------------------------------
 
 kernel void kernel_convert_block_q4_0_noshuffle(

ggml/src/ggml-opencl/kernels/diag_mask_inf.cl

@@ -0,0 +1,58 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// diag_mask_inf kernels
+//------------------------------------------------------------------------------
+kernel void kernel_diag_mask_inf(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i02 = get_global_id(2);
+    int i01 = get_global_id(1);
+    int i00 = get_global_id(0);
+
+    if (i00 > n_past + i01) {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
+    } else {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+    }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    int i = 2*get_global_id(0);
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int i4 = 4*i;
+    int i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    int i01 = i4/(ne00);      i4 -= i01*ne00;
+    int i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        (&dst[i+1])[k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            (&dst[i])[k] = -INFINITY;
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/gelu.cl

@@ -0,0 +1,62 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// gelu
+//------------------------------------------------------------------------------
+#define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
+#define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+
+kernel void kernel_gelu(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_quick(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_gelu_quick_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}

ggml/src/ggml-opencl/kernels/get_rows.cl

@@ -0,0 +1,163 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+#define QK4_0                   32
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+
+//------------------------------------------------------------------------------
+// dequantize_q4_0_f32, dequantize_q4_0_f16
+//------------------------------------------------------------------------------
+void dequantize_q4_0_f32(global struct block_q4_0 * xb, short il, float16 * reg) {
+    global ushort * qs = ((global ushort *)xb + 1);
+    float d1 = il ? (xb->d / 16.h) : xb->d;
+    float d2 = d1 / 256.f;
+    float md = -8.h * xb->d;
+    ushort mask0 = il ? 0x00F0 : 0x000F;
+    ushort mask1 = mask0 << 8;
+
+    reg->s0 = d1 * (qs[0] & mask0) + md;
+    reg->s1 = d2 * (qs[0] & mask1) + md;
+
+    reg->s2 = d1 * (qs[1] & mask0) + md;
+    reg->s3 = d2 * (qs[1] & mask1) + md;
+
+    reg->s4 = d1 * (qs[2] & mask0) + md;
+    reg->s5 = d2 * (qs[2] & mask1) + md;
+
+    reg->s6 = d1 * (qs[3] & mask0) + md;
+    reg->s7 = d2 * (qs[3] & mask1) + md;
+
+    reg->s8 = d1 * (qs[4] & mask0) + md;
+    reg->s9 = d2 * (qs[4] & mask1) + md;
+
+    reg->sa = d1 * (qs[5] & mask0) + md;
+    reg->sb = d2 * (qs[5] & mask1) + md;
+
+    reg->sc = d1 * (qs[6] & mask0) + md;
+    reg->sd = d2 * (qs[6] & mask1) + md;
+
+    reg->se = d1 * (qs[7] & mask0) + md;
+    reg->sf = d2 * (qs[7] & mask1) + md;
+}
+
+
+//------------------------------------------------------------------------------
+// get_rows
+//------------------------------------------------------------------------------
+kernel void kernel_get_rows_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        ((global float *) ((global char *) dst + i11*nb2 + i10*nb1))[ind] =
+            ((global half *) ((global char *) src0 + r*nb01 + i02*nb02))[ind];
+    }
+}
+
+kernel void kernel_get_rows_q4_0(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb1,
+        ulong nb2
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    const int NL = 2;
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+
+    int r = ((global int32_t *) ((global char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+
+    for (int ind = get_local_id(0); ind < ne00/16; ind += get_local_size(0)) {
+        float16 temp;
+        dequantize_q4_0_f32(
+            ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02)) + ind/NL, ind%NL, &temp);
+        *(((global float16 *) ((global char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}

ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl

@@ -1,146 +0,0 @@
-#ifdef cl_khr_fp16
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-#elif defined(cl_amd_fp16)
-#pragma OPENCL EXTENSION cl_amd_fp16 : enable
-#else
-#error "Half precision floating point not supportedby OpenCL implementation on your device."
-#endif
-
-#ifdef cl_khr_subgroups
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#elif defined(cl_intel_subgroups)
-#pragma OPENCL EXTENSION cl_intel_subgroups : enable
-#else
-#error "Subgroup not supported on your device."
-#endif
-
-#ifdef cl_intel_required_subgroup_size
-// Always use subgroup size of 32 on Intel.
-#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
-#define INTEL_GPU 1
-#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
-#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
-#elif defined(cl_qcom_reqd_sub_group_size)
-// Always use subgroups size of 64 on Adreno.
-#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
-#define ADRENO_GPU 1
-#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
-#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
-#else
-// TODO: do not know how to choose subgroup size on other GPUs.
-#error "Selecting subgroup size is not supported on your device."
-#endif
-
-kernel void kernel_im2col_f32(
-        global float * src1,
-        ulong offset1,
-        global float * dst,
-        ulong offsetd,
-        ulong batch_offset,
-        ulong delta_offset,
-        long IW,
-        long IH,
-        long IC,
-        long OW,
-        long OH,
-        long KW,
-        long KH,
-        long pelements,
-        long CHW,
-        int  s0,
-        int  s1,
-        int  p0,
-        int  p1,
-        int  d0,
-        int  d1
-) {
-    // threadIdx.x + blockIdx.x * blockDim.x
-    long i = get_global_id(0);
-    if (i >= pelements) {
-        return;
-    }
-
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
-
-    long  ksize = OW * (KH > 1 ? KW : 1);
-    long  kx = i / ksize;
-    long  kd = kx * ksize;
-    long  ky = (i - kd) / OW;
-    long  ix = i % OW;
-
-    long  oh = get_group_id(1);
-    long  batch = get_group_id(2) / IC;
-    long  ic = get_group_id(2) % IC;
-
-    long iiw = ix * s0 + kx * d0 - p0;
-    long iih = oh * s1 + ky * d1 - p1;
-
-    long offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        long offset_src = ic * delta_offset + batch * batch_offset;
-        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
-    }
-}
-
-kernel void kernel_im2col_f16(
-        global float * src1,
-        ulong offset1,
-        global half  * dst,
-        ulong offsetd,
-        ulong batch_offset,
-        ulong delta_offset,
-        long IW,
-        long IH,
-        long IC,
-        long OW,
-        long OH,
-        long KW,
-        long KH,
-        long pelements,
-        long CHW,
-        int  s0,
-        int  s1,
-        int  p0,
-        int  p1,
-        int  d0,
-        int  d1
-) {
-    long i = get_global_id(0);
-
-    if (i >= pelements) {
-        return;
-    }
-
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global half*)((global char*)dst + offsetd);
-
-    long  ksize = OW * (KH > 1 ? KW : 1);
-    long  kx = i / ksize;
-    long  kd = kx * ksize;
-    long  ky = (i - kd) / OW;
-    long  ix = i % OW;
-
-    long  oh = get_group_id(1);
-    long  batch = get_group_id(2) / IC;
-    long  ic = get_group_id(2) % IC;
-
-    long iiw = ix * s0 + kx * d0 - p0;
-    long iih = oh * s1 + ky * d1 - p1;
-
-    long offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        long offset_src = ic * delta_offset + batch * batch_offset;
-        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
-    }
-}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl

@@ -1,26 +0,0 @@
-// 16-bit transpose, loading/storing a 4x4 tile of elements
-
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-
-kernel void kernel_transpose_16(
-    __read_only image1d_buffer_t input,
-    __write_only image1d_buffer_t output,
-    const uint rows,
-    const uint cols
-) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-
-    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
-    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
-    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
-    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);
-
-    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
-    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_32.cl

@@ -1,25 +0,0 @@
-// 32-bit transpose, loading/storing a 4x4 tile of elements
-
-kernel void kernel_transpose_32(
-    __read_only image1d_buffer_t input,
-    __write_only image1d_buffer_t output,
-    const uint rows,
-    const uint cols
-) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-
-    float4 temp0 = read_imagef(input, (j_2+0)*cols+i);
-    float4 temp1 = read_imagef(input, (j_2+1)*cols+i);
-    float4 temp2 = read_imagef(input, (j_2+2)*cols+i);
-    float4 temp3 = read_imagef(input, (j_2+3)*cols+i);
-
-    write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
-    write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-
-}

ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_32_16.cl

@@ -1,35 +0,0 @@
-// 32-bit transpose, loading/storing a 4x4 tile of elements
-// Only used for activations
-// converts to FP16
-// also adds zero padding for non multiple of 8 prompt lengths
-#pragma OPENCL EXTENSION cl_khr_fp16 : enable
-
-kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) {
-
-    const int i = get_global_id(0);
-    const int j = get_global_id(1);
-    const int i_2 = i<<2;
-    const int j_2 = j<<2;
-    half4 temp0 = {0,0,0,0}; // initialize outputs to 0
-    half4 temp1 = {0,0,0,0};
-    half4 temp2 = {0,0,0,0};
-    half4 temp3 = {0,0,0,0};
-
-    if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0
-        temp0 = read_imageh(input, (j_2+0)*cols+i);
-    }
-    if((j_2+1)*cols+i*4+3 < rows*cols*16){
-        temp1 = read_imageh(input, (j_2+1)*cols+i);
-    }
-    if((j_2+2)*cols+i*4+3 < rows*cols*16){
-        temp2 = read_imageh(input, (j_2+2)*cols+i);
-    }
-    if((j_2+3)*cols+i*4+3 < rows*cols*16){
-        temp3 = read_imageh(input, (j_2+3)*cols+i);
-    }
-
-    write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding
-    write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
-    write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
-    write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
-}

ggml/src/ggml-opencl/kernels/im2col_f16.cl

@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f16(
+        global float * src1,
+        ulong offset1,
+        global half  * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

ggml/src/ggml-opencl/kernels/im2col_f32.cl

@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f32(
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}

ggml/src/ggml-opencl/kernels/mul.cl

@@ -0,0 +1,79 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// mul
+//------------------------------------------------------------------------------
+kernel void kernel_mul(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) * *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_mul_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F16 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3)
+{
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F16;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half * y = (global half *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (half) x[i] * (half) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half  * y  = (global half  *) (src1 + offset_src1);
+            global half4 * y4 = (global half4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (half) x4[i].s0 * y4[i].s0;
+                sumf += (half) x4[i].s1 * y4[i].s1;
+                sumf += (half) x4[i].s2 * y4[i].s2;
+                sumf += (half) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (half) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += convert_float(x[i]) * y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += convert_float(x4[i].s0) * y4[i].s0;
+                sumf += convert_float(x4[i].s1) * y4[i].s1;
+                sumf += convert_float(x4[i].s2) * y4[i].s2;
+                sumf += convert_float(x4[i].s3) * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl

@@ -0,0 +1,94 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_1row(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global half  * x = (global half  *) (src0 + offset_src0);
+    global float * y = (global float *) (src1 + offset_src1);
+
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        global half4  * x4 = (global half4  *) x;
+        global float4 * y4 = (global float4 *) y;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += (float) x4[i].s0 * y4[i].s0;
+            sumf += (float) x4[i].s1 * y4[i].s1;
+            sumf += (float) x4[i].s2 * y4[i].s2;
+            sumf += (float) x4[i].s3 * y4[i].s3;
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) {
+                all_sum += (float) x[i] * y[i];
+            }
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+
+}

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl

@@ -0,0 +1,84 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Assumes row size (ne00) is a multiple of 4
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nrows = ne11;
+    int r0 = get_group_id(0);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half4 * x4 = (global half4 *) (src0 + offset_src0);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+        global float4 * y4 = (global float4 *) (src1 + offset_src1);
+
+        float sumf = 0;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += convert_float(x4[i].s0) * y4[i].s0;
+            sumf += convert_float(x4[i].s1) * y4[i].s1;
+            sumf += convert_float(x4[i].s2) * y4[i].s2;
+            sumf += convert_float(x4[i].s3) * y4[i].s3;
+        }
+
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl

@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F32_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f32_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F32_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global float * x = (global float *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global float4 * x4 = (global float4 *)x;
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (float) x4[i].s0 * y4[i].s0;
+                sumf += (float) x4[i].s1 * y4[i].s1;
+                sumf += (float) x4[i].s2 * y4[i].s2;
+                sumf += (float) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl

@@ -0,0 +1,192 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//------------------------------------------------------------------------------
+// mul_vec_q_n_f32
+//------------------------------------------------------------------------------
+// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_4_0_dot_y(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        private float * yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float2 acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+    for (int i = 0; i < 8; i+=2) {
+        acc.s0 += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc.s1 += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
+    }
+    return d * (sumy * -8.f + acc.s0 + acc.s1);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];       // src1 vector cache
+    float sumf[N_DST]={0.f};
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+        for (int i = 0; i < 8; i += 2) {
+            sumy += yb[i] + yb[i+1];
+            yl[i+0] = yb[i+ 0];
+            yl[i+1] = yb[i+ 1]/256.f;
+            sumy += yb[i+16] + yb[i+17];
+            yl[i+8] = yb[i+16]/16.f;
+            yl[i+9] = yb[i+17]/4096.f;
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            sumf[row] += block_q_4_0_dot_y(x+ib+row*nb, sumy, yl, il);
+        }
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float tot[N_DST] = {
+        sub_group_reduce_add(sumf[0]), sub_group_reduce_add(sumf[1]),
+        sub_group_reduce_add(sumf[2]), sub_group_reduce_add(sumf[3])};
+    for (int row = 0; row < N_DST; ++row) {
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot[row];
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl

@@ -0,0 +1,307 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 16 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 16
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 16x output.
+// Eeach simdgroup outputs 16 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float16 sumf = (float16)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f,
+                             0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  0*nb*QK4_0/2, d + ib +  0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  1*nb*QK4_0/2, d + ib +  1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  2*nb*QK4_0/2, d + ib +  2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  3*nb*QK4_0/2, d + ib +  3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  4*nb*QK4_0/2, d + ib +  4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  5*nb*QK4_0/2, d + ib +  5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  6*nb*QK4_0/2, d + ib +  6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  7*nb*QK4_0/2, d + ib +  7*nb, sumy, yl, il);
+
+        sumf.s8 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  8*nb*QK4_0/2, d + ib +  8*nb, sumy, yl, il);
+        sumf.s9 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  9*nb*QK4_0/2, d + ib +  9*nb, sumy, yl, il);
+        sumf.sa += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 10*nb*QK4_0/2, d + ib + 10*nb, sumy, yl, il);
+        sumf.sb += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 11*nb*QK4_0/2, d + ib + 11*nb, sumy, yl, il);
+
+        sumf.sc += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 12*nb*QK4_0/2, d + ib + 12*nb, sumy, yl, il);
+        sumf.sd += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 13*nb*QK4_0/2, d + ib + 13*nb, sumy, yl, il);
+        sumf.se += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 14*nb*QK4_0/2, d + ib + 14*nb, sumy, yl, il);
+        sumf.sf += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 15*nb*QK4_0/2, d + ib + 15*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float16 tot = (float16)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7),
+
+        sub_group_reduce_add(sumf.s8), sub_group_reduce_add(sumf.s9),
+        sub_group_reduce_add(sumf.sa), sub_group_reduce_add(sumf.sb),
+        sub_group_reduce_add(sumf.sc), sub_group_reduce_add(sumf.sd),
+        sub_group_reduce_add(sumf.se), sub_group_reduce_add(sumf.sf)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+
+        if (first_row + 8 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 8] = tot.s8;
+        }
+        if (first_row + 9 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 9] = tot.s9;
+        }
+        if (first_row + 10 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 10] = tot.sa;
+        }
+        if (first_row + 11 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 11] = tot.sb;
+        }
+
+        if (first_row + 12 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 12] = tot.sc;
+        }
+        if (first_row + 13 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 13] = tot.sd;
+        }
+        if (first_row + 14 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 14] = tot.se;
+        }
+        if (first_row + 15 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 15] = tot.sf;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_16x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl

@@ -0,0 +1,265 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 8x output.
+// Eeach simdgroup outputs 8 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = (float8)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl

@@ -0,0 +1,272 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 32
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl

@@ -0,0 +1,254 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//
+// This variant unrolls the loops and uses vector types instead of pointers.
+// It improves performance on Adreno but not so much on Intel.
+//
+inline float block_q_4_0_dot_y_v(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_v(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;       // src1 vector cache
+    float4 sumf = (float4)(0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_v(x+ib+0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_v(x+ib+1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_v(x+ib+2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_v(x+ib+3*nb, sumy, yl, il);
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_v(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_v(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}

ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl

@@ -0,0 +1,190 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q6_K
+//------------------------------------------------------------------------------
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;             // super-block scale
+} block_q6_K;
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32
+//------------------------------------------------------------------------------
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 1 // number of rows each SIMD group works on
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 1
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    uchar kmask1 = 0x03;
+    uchar kmask2 = 0x0C;
+    uchar kmask3 = 0x30;
+    uchar kmask4 = 0xC0;
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int row = N_SIMDGROUP * r0 + get_sub_group_id();
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global block_q6_K * x = (global block_q6_K *) src0 + row*nb + offset_src0;
+    global float      * yy = (global float     *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float sumf = 0;
+
+    // For Q6_K quantization, 16 values forms a subblock, 16 subblock forms a
+    // block. Values in a subblock shares a scale that is quantized with 8 bits;
+    // the entire block shares a single floating point scale.
+    // For work distribution, each thread processes a subblock (16 weights), hence
+    // 16 threads process a (super) block -- a subgroup thus handles SIMDWIDTH/16
+    // (super) blocks -- this is the block stride.
+    // The 16 threads that process a (super) block are split into 2 portions, each has
+    // 8 threads; each portion works on 8 subblocks.
+    // For subgroup of 16 threads, the entire subgroup works on a single (super) block
+    // before moving to the next (super) block. Thread0 - thread7 work on the
+    // first 8 subblocks; thread8 - thread15 works on the last 8 subblocks.
+    // Thread0 - thread3 work on subblocks 0, 2, 4, 6; thread4 - thread7 work on
+    // subblocks 1, 3, 5, 7. Each thread does not work on an entire subblock, but
+    // works on a total of 16 weight values.
+    int tid  = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix   = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip   = tid/8;   // first or second half of (super) block (0 or 1)
+    int il   = tid%8;   // each half has 8 parts, one per scale
+    int n    = 4;       // 4 scales at a time (and 4 sums)
+    int l0   = n*il;    // offset into half-block, 0..28
+    int is   = 8*ip + l0/16; // 0, 1, 8, 9
+
+    int y_offset = 128*ip + l0;
+    int q_offset_l = 64*ip + l0;
+    int q_offset_h = 32*ip + l0;
+
+    for (int i = ix; i < nb; i += BLOCK_STRIDE) {
+
+        global uint8_t * q1 = x[i].ql + q_offset_l;
+        global uint8_t * q2 = q1 + QK_K/8;
+        global uint8_t * qh = x[i].qh + q_offset_h;
+        global int8_t  * sc = x[i].scales + is;
+
+        global float * y = yy + i * QK_K + y_offset;
+
+        float dall = x[i].d;
+
+        float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+        sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & kmask4) >> 2)) - 32.f);
+
+        sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+    }
+
+    float tot = sub_group_reduce_add(sumf);
+    if (get_sub_group_local_id() == 0) {
+        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
+    }
+}

ggml/src/ggml-opencl/kernels/norm.cl

@@ -0,0 +1,81 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// norm
+//------------------------------------------------------------------------------
+kernel void kernel_norm(
+        global void * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        float eps,
+        local float * sum
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    dst = (global void*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
+
+    // MEAN
+    // parallel sum
+    sum[get_local_id(0)] = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        sum[get_local_id(0)] += x[i00];
+    }
+    // reduce
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint i = get_local_size(0)/2; i > 0; i /= 2) {
+        if (get_local_id(0) < i) {
+            sum[get_local_id(0)] += sum[get_local_id(0) + i];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    float mean  = sum[0] / ne00;
+
+    // recenter and VARIANCE
+    barrier(CLK_LOCAL_MEM_FENCE);
+    global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    sum[get_local_id(0)] = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        y[i00] = x[i00] - mean;
+        sum[get_local_id(0)] += y[i00] * y[i00];
+    }
+
+    // reduce
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint i = get_local_size(0)/2; i > 0; i /= 2) {
+        if (get_local_id(0) < i) {
+            sum[get_local_id(0)] += sum[get_local_id(0) + i];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    float variance = sum[0] / ne00;
+
+    float scale = 1.0f/sqrt(variance + eps);
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        y[i00] = y[i00] * scale;
+    }
+}

ggml/src/ggml-opencl/kernels/relu.cl

@@ -0,0 +1,16 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// relu
+//------------------------------------------------------------------------------
+kernel void kernel_relu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = fmax(0.0f, src0[get_global_id(0)]);
+}

ggml/src/ggml-opencl/kernels/rms_norm.cl

@@ -0,0 +1,96 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// rms_norm
+//------------------------------------------------------------------------------
+// This kernel depends on subgroup size.
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_rms_norm(
+        global void * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        float eps,
+        local float * sum // Note, the size depends on number of subgroups
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global float * x_scalar = (global float *) x;
+    float4 sumf = 0;
+    float all_sum = 0;
+
+    // parallel sum
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        sumf += x[i00] * x[i00];
+    }
+    all_sum = sumf.s0 + sumf.s1 + sumf.s2 + sumf.s3;
+    all_sum = sub_group_reduce_add(all_sum);
+    if (get_sub_group_local_id() == 0) {
+        sum[get_sub_group_id()] = all_sum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+    // broadcast
+    for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
+       if (get_local_id(0) < i) {
+           sum[get_local_id(0)] += sum[get_local_id(0) + i];
+       }
+    }
+    if (get_local_id(0) == 0) {
+        for (int i = 4 * (ne00 / 4); i < ne00; i++) {
+            sum[0] += x_scalar[i];
+        }
+        sum[0] /= ne00;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float mean  = sum[0];
+    const float scale = 1.0f/sqrt(mean + eps);
+
+    global float4 * y = (global float4 *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global float * y_scalar = (global float *) y;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        y[i00] = x[i00] * scale;
+    }
+    if (get_local_id(0) == 0) {
+        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
+            y_scalar[i00] = x_scalar[i00] * scale;
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/rope.cl

@@ -0,0 +1,721 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// kernel_rope
+//------------------------------------------------------------------------------
+float rope_yarn_ramp(float low, float high, int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+float2 rope_yarn(
+    float theta_extrap, float freq_scale, float2 corr_dims, int i0, float ext_factor, float mscale
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.s0, corr_dims.s1, i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    return (float2)(cos(theta) * mscale, sin(theta) * mscale);
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+float2 rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow
+) {
+    // start and end correction dims
+    return (float2)(
+        max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base))),
+        min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)))
+    );
+}
+
+kernel void kernel_rope_norm_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src       = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            float x0 = src[0];
+            float x1 = src[1];
+
+            dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_norm_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            float x0 = src[0];
+            float x1 = src[1];
+
+            dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_neox_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_neox_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (sector < sections.s0) {
+                theta_base = pos[i2];
+            }
+            else if (sector >= sections.s0 && sector < sec_w) {
+                theta_base = pos[i2 + ne2 * 1];
+            }
+            else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 2];
+            }
+            else if (sector >= sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 3];
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (sector < sections.s0) {
+                theta_base = pos[i2];
+            }
+            else if (sector >= sections.s0 && sector < sec_w) {
+                theta_base = pos[i2 + ne2 * 1];
+            }
+            else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 2];
+            }
+            else if (sector >= sec_w + sections.s2) {
+                theta_base = pos[i2 + ne2 * 3];
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_vision_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}
+
+kernel void kernel_rope_vision_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}

ggml/src/ggml-opencl/kernels/scale.cl

@@ -0,0 +1,16 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// scale
+//------------------------------------------------------------------------------
+kernel void kernel_scale(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        float scale
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale;
+}

ggml/src/ggml-opencl/kernels/silu.cl

@@ -0,0 +1,30 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// silu
+//------------------------------------------------------------------------------
+kernel void kernel_silu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_silu_4(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}

ggml/src/ggml-opencl/kernels/softmax_4_f16.cl

@@ -0,0 +1,87 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4_f16(
+        global float * src0,
+        ulong offset0,
+        global half * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global half *)((global char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global half4  * pmask = (global char *)src1 != (global char *)src0 ? (global half4 *)(src1 + i01*ne00) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/softmax_4_f32.cl

@@ -0,0 +1,87 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4(
+        global float * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i01*ne00) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/softmax_f16.cl

@@ -0,0 +1,86 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_f16(
+        global float * src0,
+        ulong offset0,
+        global half * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global half *)((global char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    global half  * pmask = (global char *)src1 != (global char *)src0 ? src1 + i01*ne00 : 0;
+    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/softmax_f32.cl

@@ -0,0 +1,86 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max(
+        global float * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    global float * pmask = src1 != src0 ? src1 + i01*ne00 : 0;
+    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}

ggml/src/ggml-opencl/kernels/transpose.cl

@@ -0,0 +1,84 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+// 16-bit transpose, loading/storing a 4x4 tile of elements
+kernel void kernel_transpose_16(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
+    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
+    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
+    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);
+
+    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}
+
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+kernel void kernel_transpose_32(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    float4 temp0 = read_imagef(input, (j_2+0)*cols+i);
+    float4 temp1 = read_imagef(input, (j_2+1)*cols+i);
+    float4 temp2 = read_imagef(input, (j_2+2)*cols+i);
+    float4 temp3 = read_imagef(input, (j_2+3)*cols+i);
+
+    write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+
+}
+
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+// Only used for activations
+// converts to FP16
+// also adds zero padding for non multiple of 8 prompt lengths
+kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+    half4 temp0 = {0,0,0,0}; // initialize outputs to 0
+    half4 temp1 = {0,0,0,0};
+    half4 temp2 = {0,0,0,0};
+    half4 temp3 = {0,0,0,0};
+
+    if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0
+        temp0 = read_imageh(input, (j_2+0)*cols+i);
+    }
+    if((j_2+1)*cols+i*4+3 < rows*cols*16){
+        temp1 = read_imageh(input, (j_2+1)*cols+i);
+    }
+    if((j_2+2)*cols+i*4+3 < rows*cols*16){
+        temp2 = read_imageh(input, (j_2+2)*cols+i);
+    }
+    if((j_2+3)*cols+i*4+3 < rows*cols*16){
+        temp3 = read_imageh(input, (j_2+3)*cols+i);
+    }
+
+    write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding
+    write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -1,6 +1,7 @@
 #include "ggml-rpc.h"
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
+#include "ggml-cpp.h"
 
 #include <cinttypes>
 #include <string>
@@ -853,12 +854,13 @@ bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_
         /*.no_alloc   =*/ true,
     };
 
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
 
     if (tensor == nullptr) {
         GGML_LOG_ERROR("Null tensor pointer passed to server get_alloc_size function.\n");
-        ggml_free(ctx);
         return false;
     }
 
@@ -871,7 +873,6 @@ bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_
 
     response.alloc_size = ggml_backend_buft_get_alloc_size(buft,tensor);
 
-    ggml_free(ctx);
     return true;
 }
 
@@ -985,11 +986,12 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
@@ -1016,7 +1018,6 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         printf("[%s] saved to '%s'\n", __func__, cache_file.c_str());
     }
     ggml_backend_tensor_set(tensor, data, offset, size);
-    ggml_free(ctx);
     return true;
 }
 
@@ -1060,11 +1061,12 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash);
@@ -1080,7 +1082,6 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
     }
     ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
     response.result = 1;
-    ggml_free(ctx);
     return true;
 }
 
@@ -1090,11 +1091,12 @@ bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("Null tensor pointer passed to server init_tensor function.\n");
-        ggml_free(ctx);
         return false;
     }
 
@@ -1110,11 +1112,9 @@ bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {
         // This pointer can either be passed around client/server, or probably better stored server-side and kept track of.
         // Currently unimplemented.
         GGML_LOG_ERROR("tensor->extra populated by the backend, this is currently unsupported.\n");
-        ggml_free(ctx);
         return false;
     }
 
-    ggml_free(ctx);
     return true;
 }
 
@@ -1124,11 +1124,12 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
     if (tensor == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__);
-        ggml_free(ctx);
         return false;
     }
     GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, request.offset, request.size);
@@ -1147,7 +1148,6 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<
 
     response.resize(request.size, 0);
     ggml_backend_tensor_get(tensor, response.data(), request.offset, request.size);
-    ggml_free(ctx);
     return true;
 }
 
@@ -1157,12 +1157,14 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
+
     ggml_tensor * src = deserialize_tensor(ctx, &request.src);
     ggml_tensor * dst = deserialize_tensor(ctx, &request.dst);
     if (src == nullptr || dst == nullptr) {
         GGML_LOG_ERROR("[%s] error deserializing tensors\n", __func__);
-        ggml_free(ctx);
         return false;
     }
 
@@ -1180,7 +1182,6 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
                          dst_data + src_size,
                          dst_base,
                          dst_base + dst_buf_sz);
-        ggml_free(ctx);
         return false;
     }
 
@@ -1188,7 +1189,6 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_co
                      __func__, (void*) src->buffer, (void*) dst->buffer);
 
     response.result = ggml_backend_buffer_copy_tensor(src, dst);
-    ggml_free(ctx);
     return true;
 }
 
@@ -1242,7 +1242,9 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
         /*.mem_buffer =*/ NULL,
         /*.no_alloc   =*/ true,
     };
-    struct ggml_context * ctx = ggml_init(params);
+    ggml_context_ptr ctx_ptr { ggml_init(params) };
+    GGML_ASSERT(ctx_ptr != nullptr);
+    ggml_context * ctx = ctx_ptr.get();
     struct ggml_cgraph * graph = ggml_new_graph_custom(ctx, n_nodes, false);
     graph->n_nodes = n_nodes;
     std::unordered_map<uint64_t, const rpc_tensor*> tensor_ptrs;
@@ -1257,7 +1259,6 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
     }
     ggml_status status = ggml_backend_graph_compute(backend, graph);
     response.result = status;
-    ggml_free(ctx);
     return true;
 }
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4009,17 +4009,14 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_ROPE:
             {
                 const int mode = ((const int32_t *) op->op_params)[2];
-                if (mode & GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                if (mode & GGML_ROPE_TYPE_VISION) {
+                // mode is not used as a bitmask in practice, the various rope type modes are independent implementations
+                if (mode == GGML_ROPE_TYPE_MROPE) {
                     return false;
                 }
                 return ggml_is_contiguous(op->src[0]);
             }
         case GGML_OP_IM2COL:
-            // TODO: add support for the new F32 operations
-            return op->src[0]->type == GGML_TYPE_F16;
+            return true;
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
         case GGML_OP_POOL_2D:

ggml/src/ggml-sycl/im2col.cpp

@@ -12,110 +12,125 @@
 
 #include "im2col.hpp"
 
+#include <sycl/sycl.hpp>
+#include <type_traits>  // For std::is_same_v
+
+#include "ggml.h"
+
 template <typename T>
-static void im2col_kernel(
-        const float *x, T *dst, int64_t batch_offset, int64_t offset_delta,
-        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
-        int64_t pelements, int64_t CHW, int s0, int s1, int p0, int p1, int d0, int d1,
-        const sycl::nd_item<3> &item_ct1) {
+static void im2col_kernel(const float * x, T * dst, int64_t batch_offset, int64_t offset_delta, int64_t IC, int64_t IW,
+                          int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+                          int s0, int s1, int p0, int p1, int d0, int d1, const sycl::nd_item<3> & item_ct1) {
     const int64_t work_group_size = item_ct1.get_local_range(2);
-    const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
+    const int64_t global_id       = item_ct1.get_local_id(2) + (work_group_size * item_ct1.get_group(2));
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
-    for (int64_t i = global_id; i < pelements; i += work_group_size * item_ct1.get_group_range(2)) {
-
+    for (int64_t i = global_id; i < pelements; i += (work_group_size * item_ct1.get_group_range(2))) {
         const int64_t ksize = OW * (KH > 1 ? KW : 1);
-        const int64_t kx = i / ksize;
-        const int64_t kd = kx * ksize;
-        const int64_t ky = (i - kd) / OW;
-        const int64_t ix = i % OW;
+        const int64_t kx    = i / ksize;
+        const int64_t kd    = kx * ksize;
+        const int64_t ky    = (i - kd) / OW;
+        const int64_t ix    = i % OW;
 
-        const int64_t  oh = item_ct1.get_group(1);
-        const int64_t  batch = item_ct1.get_group(0) / IC;
-        const int64_t  ic = item_ct1.get_group(0) % IC;
+        const int64_t oh    = item_ct1.get_group(1);
+        const int64_t batch = item_ct1.get_group(0) / IC;
+        const int64_t ic    = item_ct1.get_group(0) % IC;
 
-        const int64_t iiw = ix * s0 + kx * d0 - p0;
-        const int64_t iih = oh * s1 + ky * d1 - p1;
+        const int64_t iiw = (ix * s0) + (kx * d0) - p0;
+        const int64_t iih = (oh * s1) + (ky * d1) - p1;
 
-        const int64_t offset_dst =
-            ((batch * OH + oh) * OW + ix) * CHW +
-            (ic * (KW * KH) + ky * KW + kx);
+        const int64_t offset_dst = (((batch * OH + oh) * OW + ix) * CHW) + (ic * (KW * KH) + ky * KW + kx);
 
-        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-            dst[offset_dst] =
-                sycl::vec<float, 1>(0.0f)
-                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
-        } else {
-            const int64_t offset_src = ic * offset_delta + batch * batch_offset;
-            dst[offset_dst] =
-                sycl::vec<float, 1>(x[offset_src + iih * IW + iiw])
-                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+        const int64_t offset_src_base = (ic * offset_delta) + (batch * batch_offset);
+        const int64_t offset_src      = offset_src_base + (iih * IW) + iiw;
+
+        const bool  out_of_bounds = (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW);
+        const float src_val       = out_of_bounds ? 0.0f : x[offset_src];
+
+        if constexpr (std::is_same_v<T, sycl::half>) {
+            dst[offset_dst] = sycl::half(src_val);
+        } else if constexpr (std::is_same_v<T, float>) {
+            dst[offset_dst] = src_val;
         }
     }
 }
 
 template <typename T>
-static void im2col_sycl(
-        const float *x, T *dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
-        int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
-        int s0, int s1, int p0, int p1, int d0, int d1,
-        queue_ptr stream) {
+static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
+                                 int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
+                                 int s0, int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
     const int64_t parallel_elements = OW * KW * KH;
-    const int64_t num_blocks = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
+    const int64_t num_blocks        = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
 
     // decrease global range when it exceeds the max int
     int64_t local_size = downsample_sycl_global_range(batch * IC * OH * num_blocks, SYCL_IM2COL_BLOCK_SIZE);
+
     sycl::range<3> block_nums(batch * IC, OH, num_blocks);
     sycl::range<3> local_range(1, 1, local_size);
 
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
+    const int64_t CHW = IC * KH * KW;
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * local_range, local_range),
-            [=](sycl::nd_item<3> item_ct1) {
-                im2col_kernel(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH,
-                               parallel_elements, (IC * KH * KW), s0, s1, p0,
-                               p1, d0, d1, item_ct1);
-            });
-    }
+    stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
+        im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
+                         p0, p1, d0, d1, item_ct1);
+    });
 }
 
-void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+static void im2col_sycl_f16(const float * x, sycl::half * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH,
+                            int64_t KW, int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset,
+                            int64_t offset_delta, int s0, int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
+    if (!stream->get_device().has(sycl::aspect::fp16)) {
+        throw sycl::exception(sycl::make_error_code(sycl::errc::kernel_not_supported),
+                              "Device does not support half precision (fp16) operations!");
+    }
+    im2col_sycl_internal<sycl::half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0,
+                                     p1, d0, d1, stream);
+}
+
+static void im2col_sycl_f32(const float * x, float * dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
+                            int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta, int s0,
+                            int s1, int p0, int p1, int d0, int d1, queue_ptr stream) {
+    im2col_sycl_internal<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1,
+                                d0, d1, stream);
+}
+
+void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
 
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+    const int32_t s0 = ((const int32_t *) (dst->op_params))[0];
+    const int32_t s1 = ((const int32_t *) (dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *) (dst->op_params))[2];
+    const int32_t p1 = ((const int32_t *) (dst->op_params))[3];
+    const int32_t d0 = ((const int32_t *) (dst->op_params))[4];
+    const int32_t d1 = ((const int32_t *) (dst->op_params))[5];
 
-    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+    const bool is_2D = ((const int32_t *) (dst->op_params))[6] == 1;
 
     const int64_t IC = src1->ne[is_2D ? 2 : 1];
     const int64_t IH = is_2D ? src1->ne[1] : 1;
-    const int64_t IW =         src1->ne[0];
+    const int64_t IW = src1->ne[0];
 
     const int64_t KH = is_2D ? src0->ne[1] : 1;
-    const int64_t KW =         src0->ne[0];
+    const int64_t KW = src0->ne[0];
 
     const int64_t OH = is_2D ? dst->ne[2] : 1;
-    const int64_t OW =         dst->ne[1];
+    const int64_t OW = dst->ne[1];
 
-    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-    const int64_t batch = src1->ne[3];
-    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+    const size_t  delta_offset = src1->nb[is_2D ? 2 : 1] / sizeof(float);
+    const int64_t batch        = src1->ne[is_2D ? 3 : 2];
+    const size_t  batch_offset = src1->nb[is_2D ? 3 : 2] / sizeof(float);
+
+    queue_ptr stream = ctx.stream();
 
     if (dst->type == GGML_TYPE_F16) {
-        im2col_sycl((const float *) src1->data, (sycl::half *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream());
+        im2col_sycl_f16((const float *) src1->data, (sycl::half *) dst->data, IW, IH, OW, OH, KW, KH, IC, batch,
+                        batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
     } else {
-        im2col_sycl((const float *) src1->data, (float *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream());
+        im2col_sycl_f32((const float *) src1->data, (float *) dst->data, IW, IH, OW, OH, KW, KH, IC, batch,
+                        batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
     }
 }

ggml/src/ggml-sycl/rope.cpp

@@ -1,9 +1,15 @@
 #include "rope.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml.h"
 
 struct rope_corr_dims {
     float v[2];
 };
 
+struct mrope_sections {
+    int v[4];
+};
+
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low);
     return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y));
@@ -114,6 +120,48 @@ static void rope_neox(
     dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }
 
+template <typename T, bool has_ff>
+static void rope_vision(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                        const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
+                        const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
+                        const float theta_scale, const float * freq_factors, const mrope_sections sections,
+                        const sycl::nd_item<3> & item_ct1) {
+    // get index pos
+    const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1));
+    if (i0 >= ne0) {
+        return;
+    }
+    const int    row_dst   = (item_ct1.get_group(2) * item_ct1.get_local_range(2)) + item_ct1.get_local_id(2);
+    const int    row_x     = row_dst % ne1;
+    const int    channel_x = row_dst / ne1;
+    const int    idst      = (row_dst * ne0) + (i0 / 2);
+    const size_t ix        = ((size_t) channel_x * s2) + ((size_t) row_x * s1) + (i0 / 2);
+
+    const int sect_dims = sections.v[0] + sections.v[1];
+    const int sector    = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0f;
+    if (sector < sections.v[0]) {
+        const int p = sector;
+        theta_base  = pos[channel_x] * sycl::pow(theta_scale, (float) p);
+    } else {
+        // Simplified from CUDA backend code: if (sector >= sections.v[0] && sector < sec_w) which is just sector >= sections.v[0]
+        const int p = sector - sections.v[0];
+        theta_base  = pos[channel_x + ne2] * sycl::pow(theta_scale, (float) p);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    float       cos_theta;
+    float       sin_theta;
+    rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+    const float x0 = x[ix + 0];
+    const float x1 = x[ix + n_dims];
+
+    // store results in dst
+    dst[idst + 0]      = x0 * cos_theta - x1 * sin_theta;
+    dst[idst + n_dims] = x0 * sin_theta + x1 * cos_theta;
+}
+
 template <typename T>
 static void rope_norm_sycl(
     const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows,
@@ -192,21 +240,58 @@ static void rope_neox_sycl(
     }
 }
 
+// rope vision
+template <typename T>
+static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                             const size_t s2, const int n_dims, const int nr, const int32_t * pos,
+                             const float freq_scale, const float freq_base, const float ext_factor,
+                             const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors,
+                             const mrope_sections sections, queue_ptr stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
+    const int               n_blocks_y = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const sycl::range<3>    grid_dims(1, n_blocks_y, nr);
+    const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims);
+
+    const float theta_scale = std::pow(freq_base, -2.0f / n_dims);
+    // Add FP16 capability check if T could be sycl::half
+    if constexpr (std::is_same_v<T, sycl::half>) {
+        dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+    }
+    // launch kernel
+    if (freq_factors == nullptr) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_vision<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    } else {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_vision<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                 corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    }
+}
+
 void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
     GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
     GGML_ASSERT(dst->src[0]->type == dst->type);
-
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t ne01 = dst->src[0]->ne[1];
+    const int64_t ne00 = dst->src[0]->ne[0]; // head dims
+    const int64_t ne01 = dst->src[0]->ne[1]; // num heads
+    const int64_t ne02 = dst->src[0]->ne[2]; // num heads
     const int64_t nr = ggml_nrows(dst->src[0]);
 
+    const size_t s01 = dst->src[0]->nb[1] / ggml_type_size(dst->src[0]->type);
+    const size_t s02 = dst->src[0]->nb[2] / ggml_type_size(dst->src[0]->type);
+
+
     //const int n_past      = ((int32_t *) dst->op_params)[0];
     const int n_dims      = ((int32_t *) dst->op_params)[1];
     const int mode        = ((int32_t *) dst->op_params)[2];
     //const int n_ctx       = ((int32_t *) dst->op_params)[3];
     const int n_ctx_orig  = ((int32_t *) dst->op_params)[4];
+    mrope_sections sections;
 
     // RoPE alteration for extended context
     float freq_base;
@@ -222,8 +307,10 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections.v,  (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     const int32_t * pos = (const int32_t *) dst->src[1]->data;
 
@@ -240,6 +327,7 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
     // compute
     if (is_neox) {
+        GGML_SYCL_DEBUG("%s: neox path\n", __func__);
         if (dst->src[0]->type == GGML_TYPE_F32) {
             rope_neox_sycl(
                 (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
@@ -253,7 +341,19 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
         } else {
             GGML_ABORT("fatal error");
         }
+    } else if (is_vision) {
+        GGML_SYCL_DEBUG("%s: vision path\n", __func__);
+        if (dst->src[0]->type == GGML_TYPE_F16) {
+            rope_vision_sycl((const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, main_stream);
+        } else if (dst->src[0]->type == GGML_TYPE_F32) {
+            rope_vision_sycl((const float *) dst->src[0]->data, (float *)dst->data, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, main_stream);
+        } else {
+            GGML_ABORT("Fatal error: Tensor type unsupported!");
+        }
     } else {
+        GGML_SYCL_DEBUG("%s: norm path\n", __func__);
         if (dst->src[0]->type == GGML_TYPE_F32) {
             rope_norm_sycl(
                 (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -5531,7 +5531,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     uint32_t workgroups_y = (uint32_t)neq2;
     uint32_t workgroups_z = (uint32_t)neq3;
 
-    if (N == 1 && qk_ratio > 1 && is_pow2(qk_ratio) && gqa_ratio <= flash_attention_num_small_rows &&
+    if (N == 1 && qk_ratio > 1 && gqa_ratio <= flash_attention_num_small_rows &&
         qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
         // grouped query attention - make the N dimension equal to gqa_ratio, reduce
         // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
@@ -5544,8 +5544,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     uint32_t split_kv = KV;
     uint32_t split_k = 1;
 
-    if (gqa_ratio > 1 && ctx->device->shader_core_count > 0) {
-        GGML_ASSERT(workgroups_x == 1);
+    // Try to use split_k when KV is large enough to be worth the overhead
+    if (workgroups_x == 1 && ctx->device->shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
         split_k = ctx->device->shader_core_count * 2 / workgroups_y;
         if (split_k > 1) {
@@ -9261,6 +9261,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 case 112:
                 case 128:
                 case 256:
+                case 575: // DeepSeek MLA
                     break;
                 default:
                     return false;

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -131,7 +131,7 @@ ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in A
 // Load the slope matrix, indexed by Q's dimension 2.
 ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
 {
-    const uint32_t h = iq2 + (r & (p.gqa_ratio - 1));
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
 
     const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
     const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
@@ -201,6 +201,11 @@ void main() {
     uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
     uint32_t k_stride = p.nb11;
     uint32_t v_stride = p.nb21;
+    // When using grouped query attention, all rows use the same mask (stride 0).
+    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
+    // that prevents the compiler from folding the "&" through the select
+    // and breaking the alignment detection.
+    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
     {
@@ -209,6 +214,7 @@ void main() {
         k_stride &= ~7;
         v_stride &= ~7;
 #endif
+        m_stride &= ~7;
     }
     tensorLayoutQ = setTensorLayoutStrideNV(tensorLayoutQ, q_stride, 1);
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
@@ -261,10 +267,7 @@ void main() {
         if (p.mask != 0) {
             tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
             tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
-            // When using grouped query attention, all rows use the same mask.
-            if (p.gqa_ratio > 1) {
-                tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, 0, 1);
-            }
+            tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
 
             coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
 

gguf-py/gguf/constants.py

@@ -139,6 +139,8 @@ class Keys:
         REL_BUCKETS_COUNT            = "{arch}.attention.relative_buckets_count"
         SLIDING_WINDOW               = "{arch}.attention.sliding_window"
         SCALE                        = "{arch}.attention.scale"
+        KEY_LENGTH_MLA               = "{arch}.attention.key_length_mla"
+        VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
 
     class Rope:
         DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -382,6 +384,8 @@ class MODEL_TENSOR(IntEnum):
     ATTN_Q_B             = auto()
     ATTN_KV_A_MQA        = auto()
     ATTN_KV_B            = auto()
+    ATTN_K_B             = auto()
+    ATTN_V_B             = auto()
     ATTN_Q_A_NORM        = auto()
     ATTN_KV_A_NORM       = auto()
     FFN_SUB_NORM         = auto()
@@ -590,6 +594,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ATTN_Q_B:                  "blk.{bid}.attn_q_b",
     MODEL_TENSOR.ATTN_KV_A_MQA:             "blk.{bid}.attn_kv_a_mqa",
     MODEL_TENSOR.ATTN_KV_B:                 "blk.{bid}.attn_kv_b",
+    MODEL_TENSOR.ATTN_K_B:                  "blk.{bid}.attn_k_b",
+    MODEL_TENSOR.ATTN_V_B:                  "blk.{bid}.attn_v_b",
     MODEL_TENSOR.ATTN_Q_A_NORM:             "blk.{bid}.attn_q_a_norm",
     MODEL_TENSOR.ATTN_KV_A_NORM:            "blk.{bid}.attn_kv_a_norm",
     MODEL_TENSOR.ATTN_SUB_NORM:             "blk.{bid}.attn_sub_norm",
@@ -1517,6 +1523,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_Q_B,
         MODEL_TENSOR.ATTN_KV_A_MQA,
         MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
         MODEL_TENSOR.ATTN_Q_A_NORM,
         MODEL_TENSOR.ATTN_KV_A_NORM,
         MODEL_TENSOR.ATTN_OUT,

gguf-py/gguf/gguf_writer.py

@@ -689,6 +689,12 @@ class GGUFWriter:
     def add_value_length(self, length: int) -> None:
         self.add_uint32(Keys.Attention.VALUE_LENGTH.format(arch=self.arch), length)
 
+    def add_key_length_mla(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.KEY_LENGTH_MLA.format(arch=self.arch), length)
+
+    def add_value_length_mla(self, length: int) -> None:
+        self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length)
+
     def add_max_alibi_bias(self, bias: float) -> None:
         self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias)
 

gguf-py/gguf/tensor_mapping.py

@@ -677,6 +677,14 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.kv_b_proj", # deepseek2
         ),
 
+        MODEL_TENSOR.ATTN_K_B: (
+            "model.layers.{bid}.self_attn.k_b_proj",  # deepseek2
+        ),
+
+        MODEL_TENSOR.ATTN_V_B: (
+            "model.layers.{bid}.self_attn.v_b_proj",  # deepseek2
+        ),
+
         MODEL_TENSOR.ATTN_Q_A_NORM: (
             "model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2
         ),

include/llama.h

@@ -367,17 +367,18 @@ extern "C" {
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // token embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int32_t nthread;                      // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;               // quantize to this llama_ftype
+        enum ggml_type output_tensor_type;    // output tensor type
+        enum ggml_type token_embedding_type;  // token embeddings tensor type
+        bool allow_requantize;                // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor;          // quantize output.weight
+        bool only_copy;                       // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                            // quantize all tensors to the default type
+        bool keep_split;                      // quantize to the same number of shards
+        void * imatrix;                       // pointer to importance matrix data
+        void * kv_overrides;                  // pointer to vector containing overrides
+        void * tensor_types;                  // pointer to vector containing tensor types
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {

scripts/sync-ggml.last

@@ -1 +1 @@
-2abf606f098844faebee578996cae9c6d63a40e2
+f71d538ece3fb32a04824dc6d1e73e360be9d22f

src/llama-arch.cpp

@@ -140,6 +140,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { 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_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -1103,6 +1105,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_Q_B,           "blk.%d.attn_q_b" },
             { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
             { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_K_B,           "blk.%d.attn_k_b" },
+            { LLM_TENSOR_ATTN_V_B,           "blk.%d.attn_v_b" },
             { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
@@ -1563,23 +1567,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_K,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP_SHEXP,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_A,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},

src/llama-arch.h

@@ -144,6 +144,8 @@ enum llm_kv {
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -306,6 +308,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_K_B,
+    LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,

src/llama-context.cpp

@@ -10,6 +10,7 @@
 #include <cstring>
 #include <stdexcept>
 #include <cinttypes>
+#include <cmath>
 
 //
 // llama_context
@@ -473,7 +474,6 @@ ggml_tensor * llama_context::build_rope_shift(
     const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
 
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
-    const auto & yarn_attn_factor = cparams.yarn_attn_factor;
     const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
 
@@ -482,6 +482,10 @@ ggml_tensor * llama_context::build_rope_shift(
     const auto & n_rot     = hparams.n_rot;
     const auto & rope_type = hparams.rope_type;
 
+    // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
+    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
+    const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;
+
     ggml_tensor * tmp;
 
     if (ggml_is_quantized(cur->type)) {

src/llama-graph.cpp

@@ -1188,6 +1188,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_tensor * v,
          ggml_tensor * kq_b,
          ggml_tensor * kq_mask,
+         ggml_tensor * v_mla,
              bool      v_trans,
              float     kq_scale) const {
   //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
@@ -1199,8 +1200,6 @@ ggml_tensor * llm_graph_context::build_attn_mha(
   //const auto & n_embd_head_k = hparams.n_embd_head_k;
   //const auto & n_embd_head_v = hparams.n_embd_head_v;
 
-    const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0];
-
     const auto n_tokens = q->ne[1];
     const auto n_head   = q->ne[2];
     const auto n_kv     = k->ne[1];
@@ -1229,7 +1228,12 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
-        cur = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens);
+        if (v_mla) {
+            cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
+            cur = ggml_mul_mat(ctx0, v_mla, cur);
+        }
+
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
 
@@ -1267,9 +1271,14 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
 
-        ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+        // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
+        if (v_mla) {
+            kqv = ggml_mul_mat(ctx0, v_mla, kqv);
+        }
 
-        cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+        cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
 
         if (!cparams.offload_kqv) {
             // all nodes between the KV store and the attention output are run on the CPU
@@ -1304,6 +1313,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     GGML_UNUSED(n_tokens);
@@ -1325,7 +1335,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
     //cb(k, "v", il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
 
     cb(cur, "kqv_out", il);
 
@@ -1379,6 +1389,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
@@ -1464,7 +1475,7 @@ ggml_tensor * llm_graph_context::build_attn(
                 ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v,
                 0);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, v_trans, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1504,6 +1515,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * k_cur,
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
+        ggml_tensor * v_mla,
             float     kq_scale,
             int       il) const {
     // these nodes are added to the graph together so that they are not reordered
@@ -1523,7 +1535,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
     //cb(k, "v", il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale);
 
     cb(cur, "kqv_out", il);
 
@@ -1692,4 +1704,3 @@ void llm_graph_context::build_pooling(
 
     ggml_build_forward_expand(gf, cur);
 }
-

src/llama-graph.h

@@ -505,11 +505,12 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn_mha(
              ggml_cgraph * gf,
-             ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q]
-             ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k]
-             ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
+             ggml_tensor * q,     // [n_embd_head_q, n_tokens, n_head_q]
+             ggml_tensor * k,     // [n_embd_head_k, n_tokens, n_head_k]
+             ggml_tensor * v,     // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
              ggml_tensor * kq_b,
              ggml_tensor * kq_mask,
+             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                     bool   v_trans,
                    float   kq_scale) const;
 
@@ -524,6 +525,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
@@ -538,6 +540,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
@@ -552,6 +555,7 @@ struct llm_graph_context {
             ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 

src/llama-hparams.h

@@ -43,6 +43,10 @@ struct llama_hparams {
     uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
 
+    // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+    uint32_t n_embd_head_k_mla = 0;
+    uint32_t n_embd_head_v_mla = 0;
+
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;

src/llama-kv-cache.cpp

@@ -27,7 +27,7 @@ bool llama_kv_cache_unified::init(
 
     recurrent = llama_model_is_recurrent(&model);
     v_trans   = !recurrent && !cparams.flash_attn;
-    can_shift = !recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+    can_shift = !recurrent;
 
     LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n",
             __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift);

src/llama-model.cpp

@@ -1156,6 +1156,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
                 }
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla, false);
+                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale);
@@ -3205,8 +3207,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 {
                     const bool is_lite = (hparams.n_layer == 27);
 
+                    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+                    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+                    const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+                    const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
                     const int64_t n_embd_head_qk_rope = hparams.n_rot;
-                    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+                    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
 
                     const int64_t q_lora_rank  = hparams.n_lora_q;
                     const int64_t kv_lora_rank = hparams.n_lora_kv;
@@ -3232,14 +3240,22 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                         if (!is_lite) {
                             layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
-                            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
+                            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
                         } else {
-                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
                         }
 
-                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
-                        layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
-                        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
+                        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+
+                        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+                        if (is_mla) {
+                            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+                            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+                        } else {
+                            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0);
+                        }
+
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -4290,6 +4306,8 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
         LLAMA_LOG_INFO("%s: n_lora_kv            = %d\n",     __func__, hparams.n_lora_kv);
+        LLAMA_LOG_INFO("%s: n_embd_head_k_mla    = %d\n",     __func__, hparams.n_embd_head_k_mla);
+        LLAMA_LOG_INFO("%s: n_embd_head_v_mla    = %d\n",     __func__, hparams.n_embd_head_v_mla);
         LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
         LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
@@ -4496,7 +4514,7 @@ struct llm_build_llama : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
                 cb(cur, "attn_out", il);
             }
 
@@ -4709,7 +4727,7 @@ struct llm_build_deci : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -4851,7 +4869,7 @@ struct llm_build_baichuan : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -4966,7 +4984,7 @@ struct llm_build_xverse : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5091,7 +5109,7 @@ struct llm_build_falcon : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5221,7 +5239,7 @@ struct llm_build_grok : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -5372,7 +5390,7 @@ struct llm_build_dbrx : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5486,7 +5504,7 @@ struct llm_build_starcoder : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5585,7 +5603,7 @@ struct llm_build_refact : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5739,7 +5757,7 @@ struct llm_build_bert : public llm_graph_context {
 
             cur = build_attn(inp_attn, gf,
                     model.layers[il].wo, model.layers[il].bo,
-                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             cb(cur, "kqv_out", il);
 
             if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
@@ -5856,7 +5874,7 @@ struct llm_build_bloom : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -5997,7 +6015,7 @@ struct llm_build_mpt : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6143,7 +6161,7 @@ struct llm_build_stablelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6266,7 +6284,7 @@ struct llm_build_qwen : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6386,7 +6404,7 @@ struct llm_build_qwen2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6507,7 +6525,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6634,7 +6652,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6787,7 +6805,7 @@ struct llm_build_qwen3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -6908,7 +6926,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7048,7 +7066,7 @@ struct llm_build_phi2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -7177,7 +7195,7 @@ struct llm_build_phi3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -7312,7 +7330,7 @@ struct llm_build_plamo : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
             ggml_tensor * sa_out = cur;
 
@@ -7419,7 +7437,7 @@ struct llm_build_gpt2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7535,7 +7553,7 @@ struct llm_build_codeshell : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7664,7 +7682,7 @@ struct llm_build_orion : public llm_graph_context {
 
             cur = build_attn(inp_attn, gf,
                     model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
         }
 
         if (il == n_layer - 1) {
@@ -7791,7 +7809,7 @@ struct llm_build_internlm2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -7988,7 +8006,7 @@ struct llm_build_minicpm3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -8118,7 +8136,7 @@ struct llm_build_gemma : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             if (il == n_layer - 1) {
@@ -8240,7 +8258,7 @@ struct llm_build_gemma2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             cur = build_norm(cur,
@@ -8381,7 +8399,7 @@ struct llm_build_gemma3 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
             }
 
             cur = build_norm(cur,
@@ -8521,7 +8539,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8856,7 +8874,7 @@ struct llm_build_command_r : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -8991,7 +9009,7 @@ struct llm_build_cohere2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9122,7 +9140,7 @@ struct llm_build_olmo : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9242,7 +9260,7 @@ struct llm_build_olmo2 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             cur = build_norm(cur,
@@ -9375,7 +9393,7 @@ struct llm_build_olmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9508,7 +9526,7 @@ struct llm_build_openelm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9622,7 +9640,7 @@ struct llm_build_gptneox : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9772,7 +9790,7 @@ struct llm_build_arctic : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -9927,7 +9945,7 @@ struct llm_build_deepseek : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, kq_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -10017,15 +10035,22 @@ struct llm_build_deepseek2 : public llm_graph_context {
     llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         bool is_lite = (hparams.n_layer == 27);
 
+        const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+
+        // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+        const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
+        const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+
+        const int64_t n_embd_head_qk_rope = hparams.n_rot;
+        const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
         // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
         // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
         const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
-        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
-        const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
-
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
+        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k));
+        const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
 
         ggml_tensor * cur;
         ggml_tensor * inpL;
@@ -10051,16 +10076,14 @@ struct llm_build_deepseek2 : public llm_graph_context {
             {
                 ggml_tensor * q = NULL;
                 if (!is_lite) {
-                    // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
                     q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
                     cb(q, "q", il);
 
                     q = build_norm(q,
-                            model.layers[il].attn_q_a_norm, NULL,
+                            model.layers[il].attn_q_a_norm, nullptr,
                             LLM_NORM_RMS, il);
                     cb(q, "q", il);
 
-                    // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
                     q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
                     cb(q, "q", il);
                 } else {
@@ -10068,96 +10091,125 @@ struct llm_build_deepseek2 : public llm_graph_context {
                     cb(q, "q", il);
                 }
 
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                ggml_tensor * q_nope = ggml_view_3d(ctx0, q,
+                        n_embd_head_qk_nope, n_head, n_tokens,
+                        ggml_row_size(q->type, n_embd_head_k),
+                        ggml_row_size(q->type, n_embd_head_k) * n_head,
                         0);
                 cb(q_nope, "q_nope", il);
 
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
+                // and {n_embd_head_qk_rope, n_head, n_tokens}
+                ggml_tensor * q_pe = ggml_view_3d(ctx0, q,
+                        n_embd_head_qk_rope, n_head, n_tokens,
+                        ggml_row_size(q->type, n_embd_head_k),
+                        ggml_row_size(q->type, n_embd_head_k) * n_head,
                         ggml_row_size(q->type, n_embd_head_qk_nope));
                 cb(q_pe, "q_pe", il);
 
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
+                ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
+                cb(kv_cmpr_pe, "kv_cmpr_pe", il);
 
                 // split into {kv_lora_rank, n_tokens}
-                ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
+                ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe,
+                        kv_lora_rank, n_tokens,
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
                         0);
-                cb(kv_compressed, "kv_compressed", il);
+                cb(kv_cmpr, "kv_cmpr", il);
 
-                // and {n_embd_head_qk_rope, n_tokens}
-                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
+                // and {n_embd_head_qk_rope, 1, n_tokens}
+                ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe,
+                        n_embd_head_qk_rope, 1, n_tokens,
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                        ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
                 cb(k_pe, "k_pe", il);
 
-                // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont
-                kv_compressed = ggml_cont(ctx0, kv_compressed);
-                kv_compressed = build_norm(kv_compressed,
-                        model.layers[il].attn_kv_a_norm, NULL,
-                        LLM_NORM_RMS, il);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
-
-                // and {n_head * n_embd_head_v, n_tokens}
-                ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                        ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
-                        0);
-                cb(v_states, "v_states", il);
-
-                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                q_pe = ggml_rope_ext(
-                        ctx0, q_pe, inp_pos, nullptr,
+                q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                        );
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                );
                 cb(q_pe, "q_pe", il);
 
-                // shared RoPE key
-                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                k_pe = ggml_rope_ext(
-                        ctx0, k_pe, inp_pos, nullptr,
+                k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
                         n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                        );
+                        ext_factor, attn_factor, beta_fast, beta_slow
+                );
                 cb(k_pe, "k_pe", il);
 
-                ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
+                kv_cmpr = build_norm(kv_cmpr,
+                        model.layers[il].attn_kv_a_norm, nullptr,
+                        LLM_NORM_RMS, il);
+                cb(kv_cmpr, "kv_cmpr", il);
 
-                ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
+                if (is_mla) {
+                    // {n_embd_head_qk_nope, n_tokens, n_head}
+                    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+                    cb(q_nope, "q_nope_perm", il);
 
-                cur = build_attn(inp_attn, gf,
-                        model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                    // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+                    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                    cb(q_nope_absorbed, "q_nope_absorbed", il);
+
+                    // {kv_lora_rank, n_head, n_tokens}
+                    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+                    cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
+
+                    // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+                    // note: rope must go first for in-place context shifting in build_rope_shift()
+                    ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                    cb(Qcur, "Qcur", il);
+
+                    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+                    cb(kv_cmpr, "kv_cmpr_reshape", il);
+
+                    // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+                    ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                    cb(Kcur, "Kcur", il);
+
+                    // {kv_lora_rank, 1, n_tokens}
+                    ggml_tensor * Vcur = kv_cmpr;
+                    cb(Vcur, "Vcur", il);
+
+                    // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
+                    cur = build_attn(inp_attn, gf,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il);
+                } else {
+                    ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
+                    cb(kv, "kv", il);
+
+                    // split into {n_embd_head_qk_nope, n_head, n_tokens}
+                    ggml_tensor * k_nope = ggml_view_3d(ctx0, kv,
+                            n_embd_head_qk_nope, n_head, n_tokens,
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                            0);
+                    cb(k_nope, "k_nope_view", il);
+
+                    // and {n_embd_head_v, n_head, n_tokens}
+                    ggml_tensor * Vcur = ggml_view_3d(ctx0, kv,
+                            n_embd_head_v, n_head, n_tokens,
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
+                            ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
+                            ggml_row_size(kv->type, n_embd_head_qk_nope));
+                    cb(Vcur, "Vcur_view", il);
+
+                    Vcur = ggml_cont(ctx0, Vcur);
+                    cb(Vcur, "Vcur_cont", il);
+
+                    // note: rope must go first for in-place context shifting in build_rope_shift()
+                    ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                    cb(Qcur, "Qcur", il);
+
+                    ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                    cb(Kcur, "Kcur", il);
+
+                    // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
+                    cur = build_attn(inp_attn, gf,
+                            model.layers[il].wo, NULL,
+                            Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                }
             }
 
             if (il == n_layer - 1) {
@@ -10323,7 +10375,7 @@ struct llm_build_bitnet : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         NULL, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 cur = build_norm(cur,
                         model.layers[il].attn_sub_norm, NULL,
@@ -10446,7 +10498,7 @@ struct llm_build_t5_enc : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo_enc, nullptr,
-                        Qcur, Kcur, Vcur, kq_b, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -10552,7 +10604,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, kq_b, 1.0f, il);
+                        Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
             }
 
@@ -10584,7 +10636,7 @@ struct llm_build_t5_dec : public llm_graph_context {
 
                 cur = build_attn(inp_attn_cross, gf,
                         model.layers[il].wo_cross, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
                 cb(cur, "kqv_out", il);
 
                 //ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
@@ -10717,7 +10769,7 @@ struct llm_build_jais : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il);
             }
 
             if (il == n_layer - 1) {
@@ -10849,7 +10901,7 @@ struct llm_build_chatglm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -10982,7 +11034,7 @@ struct llm_build_glm4 : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -11126,7 +11178,7 @@ struct llm_build_nemotron : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -11257,7 +11309,7 @@ struct llm_build_exaone : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
             if (il == n_layer - 1) {
@@ -12159,7 +12211,7 @@ struct llm_build_chameleon : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
 
                 if (hparams.swin_norm) {
                     cur = build_norm(cur,
@@ -12515,7 +12567,7 @@ struct llm_build_plm : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        q_states, k_states, v_states, nullptr, kq_scale, il);
+                        q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
             if (il == n_layer - 1) {
@@ -12638,7 +12690,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_rot)), il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
             }
 
             if (il == n_layer - 1) {

src/llama-model.h

@@ -171,6 +171,8 @@ struct llama_layer {
     struct ggml_tensor * wq_b      = nullptr;
     struct ggml_tensor * wkv_a_mqa = nullptr;
     struct ggml_tensor * wkv_b     = nullptr;
+    struct ggml_tensor * wk_b      = nullptr;
+    struct ggml_tensor * wv_b      = nullptr;
     struct ggml_tensor * wq_cross  = nullptr;
     struct ggml_tensor * wk_cross  = nullptr;
     struct ggml_tensor * wv_cross  = nullptr;

src/llama-quant.cpp

@@ -10,6 +10,7 @@
 #include <cinttypes>
 #include <fstream>
 #include <mutex>
+#include <regex>
 #include <thread>
 #include <unordered_map>
 
@@ -47,8 +48,14 @@ struct quantize_state_impl {
         {}
 };
 
+// changes to this struct must be replicated in quantize.cpp
+struct tensor_quantization {
+    std::string name;
+    ggml_type quant = GGML_TYPE_COUNT;
+};
+
 static void llama_tensor_dequantize_impl(
-    struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
+    ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
     const size_t nelements, const int nthread
 ) {
     if (output.size() < nelements) {
@@ -536,7 +543,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     model.load_hparams(ml);
     model.load_stats  (ml);
 
-    struct quantize_state_impl qs(model, params);
+    quantize_state_impl qs(model, params);
 
     if (params->only_copy) {
         ftype = ml.ftype;
@@ -661,7 +668,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     // populate the original tensors so we get an initial meta data
     for (const auto * it : tensors) {
         uint16_t i_split = params->keep_split ? it->idx : 0;
-        struct ggml_tensor * tensor = it->tensor;
+        ggml_tensor * tensor = it->tensor;
         if (!ctx_outs[i_split]) {
             ctx_outs[i_split].reset(gguf_init_empty());
         }
@@ -710,7 +717,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     new_ofstream(0);
     for (const auto * it : tensors) {
         const auto & weight = *it;
-        struct ggml_tensor * tensor = weight.tensor;
+        ggml_tensor * tensor = weight.tensor;
         if (weight.idx != cur_split && params->keep_split) {
             close_ofstream();
             new_ofstream(weight.idx);
@@ -776,7 +783,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
 
-        enum ggml_type new_type;
+        ggml_type new_type;
         void * new_data;
         size_t new_size;
 
@@ -786,6 +793,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             // get more optimal quantization type based on the tensor shape, layer, etc.
             if (!params->pure && ggml_is_quantized(default_type)) {
                 new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
+                // unless the user specifies a type
+                if (params->tensor_types) {
+                    const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
+                    for (const auto & [tname, qtype] : tensor_types) {
+                        if (std::regex pattern(tname); std::regex_search(tensor->name, pattern)) {
+                            if (qtype != new_type) {
+                                LLAMA_LOG_DEBUG("(overriding %s -> %s), ", ggml_type_name(new_type), ggml_type_name(qtype));
+                            }
+                            new_type = qtype;
+                            break;
+                        }
+                    }
+                }
             }
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;
@@ -910,8 +930,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 // interface implementation
 //
 
-struct llama_model_quantize_params llama_model_quantize_default_params() {
-    struct llama_model_quantize_params result = {
+llama_model_quantize_params llama_model_quantize_default_params() {
+    llama_model_quantize_params result = {
         /*.nthread                     =*/ 0,
         /*.ftype                       =*/ LLAMA_FTYPE_MOSTLY_Q5_1,
         /*.output_tensor_type          =*/ GGML_TYPE_COUNT,
@@ -923,6 +943,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
         /*.keep_split                  =*/ false,
         /*.imatrix                     =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
+        /*.tensor_type                 =*/ nullptr,
     };
 
     return result;

src/llama-vocab.cpp

@@ -1841,6 +1841,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_prefix|>"  // Qwen
                         || t.first == "<fim-prefix>"
+                        || t.first == "<fim_prefix>"    // Granite
                         || t.first == "<｜fim▁begin｜>" // DeepSeek
                         || t.first == "<PRE>"
                         || t.first == "▁<PRE>"          // CodeLlama
@@ -1859,6 +1860,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_suffix|>" // Qwen
                         || t.first == "<fim-suffix>"
+                        || t.first == "<fim_suffix>"   // Granite
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == "<SUF>"
                         || t.first == "▁<SUF>"         // CodeLlama
@@ -1877,6 +1879,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_middle|>" // Qwen
                         || t.first == "<fim-middle>"
+                        || t.first == "<fim_middle>"   // Granite
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == "<MID>"
                         || t.first == "▁<MID>"         // CodeLlama
@@ -1895,6 +1898,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 if (false
                         || t.first == "<|fim_pad|>" // Qwen
                         || t.first == "<fim-pad>"
+                        || t.first == "<fim_pad>"   // Granite
                         || t.first == "<PAD>"
                         ) {
                     special_fim_pad_id = t.second;
@@ -1913,6 +1917,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|repo_name|>"
                         || t.first == "<fim-repo>"
                         || t.first == "<REPO>"
+                        || t.first == "<reponame>"    // Granite
                         ) {
                     special_fim_rep_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {

tests/test-backend-ops.cpp

@@ -4428,10 +4428,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-    for (int hsk : { 64, 80, 128, 192, 256, }) {
-        for (int hsv : { 64, 80, 128, 192, 256, }) {
-            if (hsk != 192 && hsk != hsv) continue;
+    for (int hsk : { 64, 80, 128, 192, 256, 576 }) {
+        for (int hsv : { 64, 80, 128, 192, 256, 512 }) {
+            if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
             if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
+            if (hsk == 576 && hsv != 512) continue; // DeepSeek MLA
 
             for (bool mask : { true, false } ) {
                 for (float max_bias : { 0.0f, 8.0f }) {
@@ -4532,7 +4533,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
 
     for (int kv : { 4096, 8192, 16384, }) {
         for (int hs : { 64, 128, }) {
-            test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, 4, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+            for (int nr : { 1, 4, }) {
+                test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, nr, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+            }
         }
     }