diff --git a/.clang-format b/.clang-format
index 47d96b6b4..117e6986f 100644
--- a/.clang-format
+++ b/.clang-format
@@ -22,7 +22,7 @@ AllowShortIfStatementsOnASingleLine: Never
 AllowShortLambdasOnASingleLine: Inline
 AllowShortLoopsOnASingleLine: false
 AlwaysBreakBeforeMultilineStrings: true
-BinPackArguments: false
+BinPackArguments: true
 BinPackParameters: false # OnePerLine
 BitFieldColonSpacing: Both
 BreakBeforeBraces: Custom # Attach
diff --git a/common/arg.cpp b/common/arg.cpp
index 5746728c5..c5f115c77 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -1550,11 +1550,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-fa", "--flash-attn"}, "FA",
         string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')", llama_flash_attn_type_name(params.flash_attn_type)),
         [](common_params & params, const std::string & value) {
-            if (value == "on" || value == "enabled") {
+            if (value == "on" || value == "enabled" || value == "1") {
                 params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
-            } else if (value == "off" || value == "disabled") {
+            } else if (value == "off" || value == "disabled" || value == "0") {
                 params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
-            } else if (value == "auto") {
+            } else if (value == "auto" || value == "-1") {
                 params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
             } else {
                 throw std::runtime_error(string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
@@ -2964,13 +2964,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.endpoint_metrics = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_METRICS"));
-    add_opt(common_arg(
-        {"--slots"},
-        string_format("enable slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"),
-        [](common_params & params) {
-            params.endpoint_slots = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_SLOTS"));
     add_opt(common_arg(
         {"--props"},
         string_format("enable changing global properties via POST /props (default: %s)", params.endpoint_props ? "enabled" : "disabled"),
@@ -2978,6 +2971,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.endpoint_props = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_PROPS"));
+    add_opt(common_arg(
+        {"--slots"},
+        string_format("enable slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.endpoint_slots = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_SLOTS"));
     add_opt(common_arg(
         {"--no-slots"},
         "disables slots monitoring endpoint",
diff --git a/common/common.h b/common/common.h
index 3a03cf0cd..0ba1c7624 100644
--- a/common/common.h
+++ b/common/common.h
@@ -440,7 +440,7 @@ struct common_params {
 
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
-    bool endpoint_slots   = false;
+    bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
diff --git a/common/sampling.cpp b/common/sampling.cpp
index 9c04d35fd..c710ee173 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -426,8 +426,29 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 
 // helpers
 
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
-    return &gsmpl->cur_p;
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    auto * res = &gsmpl->cur_p;
+
+    if (do_sort && !res->sorted) {
+        // remember the selected token before sorting
+        const llama_token id = res->data[res->selected].id;
+
+        std::sort(res->data, res->data + res->size, [](const llama_token_data & a, const llama_token_data & b) {
+            return a.p > b.p;
+        });
+
+        // restore the selected token after sorting
+        for (size_t i = 0; i < res->size; ++i) {
+            if (res->data[i].id == id) {
+                res->selected = i;
+                break;
+            }
+        }
+
+        res->sorted = true;
+    }
+
+    return res;
 }
 
 llama_token common_sampler_last(const struct common_sampler * gsmpl) {
diff --git a/common/sampling.h b/common/sampling.h
index 2064421db..e198eecda 100644
--- a/common/sampling.h
+++ b/common/sampling.h
@@ -86,7 +86,9 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 // helpers
 
 // access the internal list of current candidate tokens
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
+// if do_sort == true, the candidates are guaranteed to be sorted afterwards (in descending order of probability)
+// the .sorted flag of the result indicates whether the returned candidates are sorted
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort);
 
 // get the last accepted token
 llama_token common_sampler_last(const struct common_sampler * gsmpl);
diff --git a/common/speculative.cpp b/common/speculative.cpp
index 262b2c23e..3e83b0964 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -317,7 +317,7 @@ llama_tokens common_speculative_gen_draft(
 
         common_sampler_sample(smpl, ctx_dft, 0, true);
 
-        const auto * cur_p = common_sampler_get_candidates(smpl);
+        const auto * cur_p = common_sampler_get_candidates(smpl, true);
 
         for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
             LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index df37c4a6e..908435d7e 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -302,10 +302,6 @@ class ModelBase:
                 # data = data_torch.squeeze().numpy()
                 data = data_torch.numpy()
 
-                # if data ends up empty, it means data_torch was a scalar tensor -> restore
-                if len(data.shape) == 0:
-                    data = data_torch.numpy()
-
                 n_dims = len(data.shape)
                 data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
diff --git a/examples/model-conversion/scripts/utils/curl-embedding-server.sh b/examples/model-conversion/scripts/utils/curl-embedding-server.sh
new file mode 100755
index 000000000..b6898665f
--- /dev/null
+++ b/examples/model-conversion/scripts/utils/curl-embedding-server.sh
@@ -0,0 +1,6 @@
+#!/bin/bash
+curl --request POST \
+    --url http://localhost:8080/embedding \
+    --header "Content-Type: application/json" \
+    --data '{"input": "Hello world today"}' \
+    --silent
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index a2977ea2e..4f246f6cc 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -307,6 +307,9 @@ extern "C" {
     GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
     GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
+    // Split graph without allocating it
+    GGML_API void                 ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+
     // Allocate and compute graph on the backend scheduler
     GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); // returns success
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index fa3efe976..3a13f8447 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -651,7 +651,7 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #endif
 
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
-#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
+#define GGML_SCHED_MAX_SPLIT_INPUTS 30
 #endif
 
 #ifndef GGML_SCHED_MAX_COPIES
@@ -908,7 +908,7 @@ static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, stru
 }
 
 // assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
-static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     // reset splits
     sched->n_splits = 0;
     sched->n_graph_inputs = 0;
@@ -1693,6 +1693,8 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
     GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
+    ggml_backend_sched_reset(sched);
+
     ggml_backend_sched_synchronize(sched);
 
     ggml_backend_sched_split_graph(sched, measure_graph);
diff --git a/ggml/src/ggml-cpu/arch/riscv/quants.c b/ggml/src/ggml-cpu/arch/riscv/quants.c
index 6c74417c9..ee41a3502 100644
--- a/ggml/src/ggml-cpu/arch/riscv/quants.c
+++ b/ggml/src/ggml-cpu/arch/riscv/quants.c
@@ -1270,29 +1270,40 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
             const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
-            int tmp, tmp2, sumi;
+            float ftmp, ft2;
+            const uint8_t * restrict q40;
+            const uint8_t * restrict q41;
+            const uint8_t * restrict q42;
+            const uint8_t * restrict q43;
+            const int8_t  * restrict q80;
+            const int8_t  * restrict q81;
+            const int8_t  * restrict q82;
+            const int8_t  * restrict q83;
+            int s0, s1, s2, s3;
+
             __asm__ __volatile__(
-                "vsetivli zero, 12, e8, m1\n\t"
-                "vle8.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
-                "vsetivli zero, 4, e32, m1\n\t"
+                "li %[s1], 8\n\t"
+                "vsetivli zero, 4, e32, m1, ta, ma\n\t"
+                "vle32.v v1, (%[s6b])\n\t"
+                "vslide1down.vx v1, v1, zero\n\t"
+                "vmv.v.x v16, zero\n\t"
                 "vslidedown.vi v2, v1, 2\n\t"
                 "vmv1r.v v3, v2\n\t"
                 "vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
-                "vsetivli zero, 2, e32, m1\n\t"
+                "vsetivli zero, 2, e32, m1, ta, ma\n\t"
                 "vmv.v.i v4, 4\n\t"
                 "vand.vx v8, v1, %[kmask1]\n\t"
                 "vslide1up.vx v5, v4, zero\n\t" // {0, 4}
                 "vsrl.vi v6, v1, 6\n\t"
                 "vsrl.vv v7, v2, v5\n\t"
+                "vsse32.v v8, (%[utmp]), %[s1]\n\t"
                 "vand.vx v0, v6, %[kmask3]\n\t"
                 "vand.vx v2, v7, %[kmask2]\n\t"
                 "vsll.vi v6, v0, 4\n\t"
-                "li %[t2], 8\n\t"
-                "addi %[t1], %[utmp], 4\n\t"
+                "addi %[s0], %[utmp], 4\n\t"
                 "vor.vv v1, v6, v2\n\t"
-                "vsse32.v v8, (%[utmp]), %[t2]\n\t"
-                "vsse32.v v1, (%[t1]), %[t2]\n\t"
-                "vsetivli zero, 8, e16, m1\n\t"
+                "vsse32.v v1, (%[s0]), %[s1]\n\t"
+                "vsetivli zero, 8, e16, m1, ta, ma\n\t"
                 "vle32.v v2, (%[bsums])\n\t"
                 "vnsrl.wi v0, v2, 0\n\t"
                 "vnsrl.wi v1, v2, 16\n\t"
@@ -1300,13 +1311,131 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 "vle8.v v3, (%[mins])\n\t"
                 "vzext.vf2 v4, v3\n\t"
                 "vwmul.vv v6, v4, v2\n\t"
+                "vsetivli zero, 4, e32, m1, ta, ma\n\t"
+                "vredsum.vs v0, v6, v16\n\t"
+                "vredsum.vs v0, v7, v0\n\t"
+                "vfcvt.f.x.v v0, v0\n\t"
+                "vfmv.f.s %[ftmp], v0\n\t"
+                "vsetivli zero, 16, e8, m1, ta, ma\n\t"
+                "vle8.v v0, (%[xs])\n\t"
+                "fnmsub.s %[sumf], %[dmin], %[ftmp], %[sumf]\n\t"
+                "addi %[q40], %[xs], 64\n\t"
+                "addi %[q41], %[xs], 16\n\t"
+                "addi %[q42], %[xs], 32\n\t"
+                "addi %[q43], %[xs], 48\n\t"
+                "addi %[q80], %[ys], 64\n\t"
+                "vle8.v v1, (%[q41])\n\t"
+                "vle8.v v2, (%[q42])\n\t"
+                "addi %[q81], %[ys], 16\n\t"
+                "addi %[q41], %[q41], 64\n\t"
+                "addi %[q82], %[ys], 32\n\t"
+                "vle8.v v3, (%[q43])\n\t"
+                "vle8.v v8, (%[ys])\n\t"
+                "addi %[q42], %[q42], 64\n\t"
+                "addi %[q83], %[ys], 48\n\t"
+                "addi %[q43], %[q43], 64\n\t"
+                "vsrl.vi v4, v0, 4\n\t"
+                "vle8.v v9, (%[q81])\n\t"
+                "vle8.v v10, (%[q82])\n\t"
+                "vand.vi v0, v0, 0xF\n\t"
+                "addi %[q81], %[q81], 64\n\t"
+                "vsrl.vi v5, v1, 4\n\t"
+                "addi %[q82], %[q82], 64\n\t"
+                "vle8.v v11, (%[q83])\n\t"
+                "vle8.v v12, (%[q80])\n\t"
+                "vand.vi v1, v1, 0xF\n\t"
+                "addi %[q83], %[q83], 64\n\t"
+                "vsrl.vi v6, v2, 4\n\t"
+                "addi %[q80], %[q80], 64\n\t"
+                "vle8.v v13, (%[q81])\n\t"
+                "vle8.v v14, (%[q82])\n\t"
+                "vand.vi v2, v2, 0xF\n\t"
+                "addi %[q81], %[q81], 64\n\t"
+                "vsrl.vi v7, v3, 4\n\t"
+                "addi %[q82], %[q82], 64\n\t"
+                "vwmul.vv v16, v0, v8\n\t"
+                "vle8.v v15, (%[q83])\n\t"
+                "vle8.v v0, (%[q40])\n\t"
+                "vand.vi v3, v3, 0xF\n\t"
+                "addi %[q83], %[q83], 64\n\t"
+                "vwmul.vv v24, v2, v12\n\t"
+                "vwmul.vv v20, v4, v10\n\t"
+                "vwmul.vv v28, v6, v14\n\t"
+                "vwmacc.vv v16, v1, v9\n\t"
+                "vle8.v v1, (%[q41])\n\t"
+                "vle8.v v2, (%[q42])\n\t"
+                "vwmacc.vv v24, v3, v13\n\t"
+                "vwmacc.vv v20, v5, v11\n\t"
+                "vwmacc.vv v28, v7, v15\n\t"
+                "addi %[q40], %[q80], 64\n\t"
+                "addi %[q41], %[q81], 64\n\t"
+                "vle8.v v3, (%[q43])\n\t"
+                "vle8.v v8, (%[q80])\n\t"
+                "addi %[q42], %[q82], 64\n\t"
+                "addi %[q43], %[q83], 64\n\t"
+                "vsrl.vi v4, v0, 4\n\t"
+                "vle8.v v9, (%[q81])\n\t"
+                "vle8.v v10, (%[q82])\n\t"
+                "vand.vi v0, v0, 0xF\n\t"
+                "vsrl.vi v5, v1, 4\n\t"
+                "vsrl.vi v7, v3, 4\n\t"
+                "vand.vi v3, v3, 0xF\n\t"
+                "vle8.v v11, (%[q83])\n\t"
+                "vle8.v v12, (%[q40])\n\t"
+                "vand.vi v1, v1, 0xF\n\t"
+                "vsrl.vi v6, v2, 4\n\t"
+                "vand.vi v2, v2, 0xF\n\t"
+                "vwmul.vv v18, v0, v8\n\t"
+                "vle8.v v13, (%[q41])\n\t"
+                "vle8.v v14, (%[q42])\n\t"
+                "vwmul.vv v26, v2, v12\n\t"
+                "vwmul.vv v22, v4, v10\n\t"
+                "vwmul.vv v30, v6, v14\n\t"
+                "vwmacc.vv v18, v1, v9\n\t"
+                "vle8.v v15, (%[q43])\n\t"
+                "vwmacc.vv v26, v3, v13\n\t"
+                "vwmacc.vv v22, v5, v11\n\t"
+                "vwmacc.vv v30, v7, v15\n\t"
                 "vmv.v.x v0, zero\n\t"
-                "vsetivli zero, 8, e32, m2\n\t"
-                "vredsum.vs v0, v6, v0\n\t"
-                "vmv.x.s %[sumi], v0"
-                : [t1] "=&r" (tmp), [t2] "=&r" (tmp2), [sumi] "=&r" (sumi)
-                : [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
-                , [s6b] "r" (x[i].scales), [kmask1] "r" (kmask1)
+                "vsetivli zero, 16, e16, m2, ta, ma\n\t"
+                "vwredsum.vs v4, v16, v0\n\t"
+                "lbu %[s0], 0(%[scale])\n\t"
+                "vwredsum.vs v5, v20, v0\n\t"
+                "lbu %[s1], 1(%[scale])\n\t"
+                "vwredsum.vs v6, v24, v0\n\t"
+                "lbu %[s2], 2(%[scale])\n\t"
+                "vwredsum.vs v7, v28, v0\n\t"
+                "lbu %[s3], 3(%[scale])\n\t"
+                "vwredsum.vs v8, v18, v0\n\t"
+                "lbu %[q40], 4(%[scale])\n\t"
+                "vwredsum.vs v9, v22, v0\n\t"
+                "lbu %[q41], 5(%[scale])\n\t"
+                "vwredsum.vs v10, v26, v0\n\t"
+                "lbu %[q42], 6(%[scale])\n\t"
+                "vwredsum.vs v11, v30, v0\n\t"
+                "lbu %[q43], 7(%[scale])\n\t"
+                "vsetivli zero, 4, e32, m1, ta, ma\n\t"
+                "vmul.vx v0, v4, %[s0]\n\t"
+                "vmul.vx v1, v8, %[q40]\n\t"
+                "vmacc.vx v0, %[s1], v5\n\t"
+                "vmacc.vx v1, %[q41], v9\n\t"
+                "vmacc.vx v0, %[s2], v6\n\t"
+                "vmacc.vx v1, %[q42], v10\n\t"
+                "vmacc.vx v0, %[s3], v7\n\t"
+                "vmacc.vx v1, %[q43], v11\n\t"
+                "vfcvt.f.x.v v0, v0\n\t"
+                "vfcvt.f.x.v v1, v1\n\t"
+                "vfmv.f.s %[ft2], v0\n\t"
+                "vfmv.f.s %[ftmp], v1\n\t"
+                "fadd.s %[ft2], %[ft2], %[ftmp]\n\t"
+                "fmadd.s %[sumf], %[d], %[ft2], %[sumf]"
+                : [ftmp] "=&f" (ftmp), [sumf] "+&f" (sumf), [ft2] "=&f" (ft2)
+                , [s0] "=&r" (s0), [s1] "=&r" (s1), [s2] "=&r" (s2), [s3] "=&r" (s3)
+                , [q40] "=&r" (q40), [q41] "=&r" (q41), [q42] "=&r" (q42), [q43] "=&r" (q43)
+                , [q80] "=&r" (q80), [q81] "=&r" (q81), [q82] "=&r" (q82), [q83] "=&r" (q83)
+                : [d] "f" (d), [ys] "r" (y[i].qs), [xs] "r" (x[i].qs), [scale] "r" (scales)
+                , [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
+                , [s6b] "r" (&x[i]), [kmask1] "r" (kmask1), [dmin] "f" (dmin)
                 , [kmask2] "r" (kmask2), [kmask3] "r" (kmask3)
                 : "memory"
                 , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
@@ -1314,59 +1443,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
                 , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
             );
-            sumf -= dmin * sumi;
-
-            const uint8_t * restrict q4 = x[i].qs;
-            const int8_t  * restrict q8 = y[i].qs;
-
-            sumi = 0;
-            const uint8_t * scale = scales;
-
-            for (int j = 0; j < QK_K/128; ++j) {
-                int vl128 = 128, vl64 = 64, vl32 = 32;
-                __asm__ __volatile__(
-                    "vsetvli zero, %[vl128], e8, m8\n\t"
-                    "vle8.v v8, (%[q8])\n\t"
-                    "vsetvli zero, %[vl64], e8, m4\n\t"
-                    "vle8.v v0, (%[q4])\n\t"
-                    "vsrl.vi v4, v0, 4\n\t"
-                    "vand.vi v0, v0, 0xF\n\t"
-                    "vsetvli zero, %[vl32], e8, m2\n\t"
-                    "vwmul.vv v28, v6, v14\n\t"
-                    "vwmul.vv v20, v4, v10\n\t"
-                    "vwmul.vv v24, v2, v12\n\t"
-                    "vwmul.vv v16, v0, v8\n\t"
-                    "vsetivli zero, 4, e32, m1\n\t"
-                    "vle8.v v2, (%[scale])\n\t"
-                    "vmv.v.x v0, zero\n\t"
-                    "vzext.vf4 v1, v2\n\t"
-                    "vsetvli zero, %[vl32], e16, m4\n\t"
-                    "vwredsum.vs v6, v24, v0\n\t"
-                    "vwredsum.vs v7, v28, v0\n\t"
-                    "vwredsum.vs v4, v16, v0\n\t"
-                    "vwredsum.vs v5, v20, v0\n\t"
-                    "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v6, v7, 1\n\t"
-                    "vslideup.vi v4, v5, 1\n\t"
-                    "vslideup.vi v4, v6, 2\n\t"
-                    "vmul.vv v8, v4, v1\n\t"
-                    "vredsum.vs v0, v8, v0\n\t"
-                    "vmv.x.s %[tmp], v0\n\t"
-                    "add %[sumi], %[sumi], %[tmp]"
-                    : [tmp] "=&r" (tmp), [sumi] "+&r" (sumi)
-                    : [vl128] "r" (vl128), [vl64] "r" (vl64), [vl32] "r" (vl32)
-                    , [q4] "r" (q4), [q8] "r" (q8), [scale] "r" (scale)
-                    : "memory"
-                    , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-                    , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
-                    , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
-                    , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
-                );
-
-                q4 += 64;    q8 += 128;    scale += 4;
-            }
-
-            sumf += d * sumi;
         }
         break;
     default:
@@ -1693,6 +1769,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     case 128:
         for (int i = 0; i < nb; ++i) {
 
+            __builtin_prefetch(&x[i + 1].d, 0, 1);
+
             const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             const uint8_t * restrict q6 = x[i].ql;
@@ -1701,23 +1779,59 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             const int8_t * restrict scale = x[i].scales;
 
-            int sum_t = 0;
-            int t0;
+            int q6h;
+            float ftmp;
 
             for (int j = 0; j < QK_K/128; ++j) {
                 __asm__ __volatile__(
+                    "addi %[q6h], %[q6], 32\n\t"
+                    "ld t0, 0(%[scale])\n\t"
+                    "addi %[scale], %[scale], 8\n\t"
+                    "slli t6, t0, 1 * 8\n\t"
+                    "lb zero, 0(%[q6])\n\t"
+                    "slli t5, t0, 2 * 8\n\t"
+                    "slli t4, t0, 3 * 8\n\t"
+                    "lb zero, 0(%[q6h])\n\t"
+                    "slli t3, t0, 4 * 8\n\t"
+                    "slli t2, t0, 5 * 8\n\t"
+                    "lb zero, 0(%[qh])\n\t"
+                    "lb zero, 31(%[q6h])\n\t"
+                    "slli t1, t0, 6 * 8\n\t"
+                    "srai a7, t0, 56\n\t"
                     "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "vle8.v v8, (%[q6])\n\t"
+                    "srai t6, t6, 56\n\t"
+                    "srai t5, t5, 56\n\t"
+                    "srai t4, t4, 56\n\t"
+                    "srai t3, t3, 56\n\t"
+                    "vle8.v v10, (%[q6h])\n\t"
+                    "addi %[q6], %[q6], 64\n\t"
+                    "slli t0, t0, 7 * 8\n\t"
+                    "srai t2, t2, 56\n\t"
+                    "srai t1, t1, 56\n\t"
+                    "srai t0, t0, 56\n\t"
                     "vle8.v v4, (%[qh])\n\t"
+                    "vsrl.vi v12, v8, 4\n\t"
+                    "vsrl.vi v14, v10, 4\n\t"
+                    "lb zero, 0(%[q8])\n\t"
+                    "vand.vi v8, v8, 0xF\n\t"
+                    "vand.vi v10, v10, 0xF\n\t"
+                    "lb zero, 32(%[q8])\n\t"
                     "vsll.vi v0, v4, 4\n\t"
                     "vsll.vi v2, v4, 2\n\t"
+                    "lb zero, 64(%[q8])\n\t"
                     "vsrl.vi v6, v4, 2\n\t"
-                    "vsetvli zero, %[vl64], e8, m4\n\t"
-                    "vle8.v v8, (%[q6])\n\t"
-                    "vsrl.vi v12, v8, 4\n\t"
-                    "vand.vi v8, v8, 0xF\n\t"
-                    "vsetvli zero, %[vl128], e8, m8\n\t"
                     "vand.vx v0, v0, %[mask]\n\t"
+                    "lb zero, 96(%[q8])\n\t"
+                    "vand.vx v2, v2, %[mask]\n\t"
+                    "vand.vx v4, v4, %[mask]\n\t"
+                    "vand.vx v6, v6, %[mask]\n\t"
                     "vor.vv v8, v8, v0\n\t"
+                    "lb zero, 127(%[q8])\n\t"
+                    "vor.vv v10, v10, v2\n\t"
+                    "vor.vv v12, v12, v4\n\t"
+                    "vor.vv v14, v14, v6\n\t"
+                    "vsetvli zero, %[vl128], e8, m8\n\t"
                     "vle8.v v0, (%[q8])\n\t"
                     "vsub.vx v8, v8, %[vl32]\n\t"
                     "vsetvli zero, %[vl64], e8, m4\n\t"
@@ -1734,34 +1848,34 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     "vwredsum.vs v13, v28, v0\n\t"
                     "vwredsum.vs v14, v30, v0\n\t"
                     "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v10, v9, 1\n\t"
-                    "vslideup.vi v8, v7, 1\n\t"
-                    "vslideup.vi v11, v12, 1\n\t"
-                    "vslideup.vi v13, v14, 1\n\t"
-                    "vslideup.vi v10, v8, 2\n\t"
-                    "vslideup.vi v11, v13, 2\n\t"
-                    "vsetivli zero, 8, e32, m2\n\t"
-                    "vle8.v v2, (%[scale])\n\t"
-                    "vsext.vf4 v4, v2\n\t"
-                    "vmul.vv v2, v4, v10\n\t"
-                    "vredsum.vs v0, v2, v0\n\t"
-                    "vmv.x.s %[t0], v0\n\t"
-                    "add %[sumi], %[sumi], %[t0]"
-                    : [sumi] "+&r" (sum_t), [t0] "=&r" (t0)
-                    : [qh] "r" (qh), [q6] "r" (q6), [q8] "r" (q8), [scale] "r" (scale)
+                    "vmul.vx v0, v10, t0\n\t"
+                    "vmul.vx v1, v9, t1\n\t"
+                    "vmacc.vx v0, t2, v8\n\t"
+                    "vmacc.vx v1, t3, v7\n\t"
+                    "vmacc.vx v0, t4, v11\n\t"
+                    "vmacc.vx v1, t5, v12\n\t"
+                    "vmacc.vx v0, t6, v13\n\t"
+                    "vmacc.vx v1, a7, v14\n\t"
+                    "vadd.vv v0, v0, v1\n\t"
+                    "vfcvt.f.x.v v0, v0\n\t"
+                    "vfmv.f.s %[ftmp], v0\n\t"
+                    "fmadd.s %[sumf], %[d], %[ftmp], %[sumf]"
+                    : [q6] "+&r" (q6), [q6h] "=&r" (q6h)
+                    , [scale] "+&r" (scale)
+                    , [sumf] "+&f" (sumf), [ftmp] "=&f" (ftmp)
+                    : [qh] "r" (qh), [q8] "r" (q8)
                     , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
-                    , [mask] "r" (0x30)
+                    , [mask] "r" (0x30), [d] "f" (d)
                     : "memory"
                     , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
                     , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
                     , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
                     , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+                    , "t0", "t1", "t2", "t3", "t4", "t5", "t6", "a7"
+                    , "a6", "a5", "a4", "a3"
                 );
-                q6 += 64;   qh += 32;   q8 += 128;   scale += 8;
+                qh += 32;   q8 += 128;
             }
-
-            sumf += d * sum_t;
-
         }
         break;
     default:
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index c9e8ddd0d..d650b4dd7 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -4087,6 +4087,13 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
         vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
     }
+#elif defined(__riscv_zvfh)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat16m1_t vy = __riscv_vfncvt_f_f_w_f16m1(vx, vl);
+        __riscv_vse16_v_f16m1((_Float16 *)&y[i], vy, vl);
+    }
 #endif
     for (; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(x[i]);
diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h
index f71ce5807..8bd56bdac 100644
--- a/ggml/src/ggml-cpu/simd-mappings.h
+++ b/ggml/src/ggml-cpu/simd-mappings.h
@@ -215,6 +215,47 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32_VEC_MUL    GGML_F32xt_MUL
 #define GGML_F32_VEC_REDUCE GGML_F32xt_REDUCE
 
+// F16 SVE
+#define DEFAULT_PG32    svptrue_b32()
+#define DEFAULT_PG16    svptrue_b16()
+
+#define GGML_F32Cxt                         svfloat16_t
+#define GGML_F32Cxt_ZERO                    svdup_n_f16(0.0f)
+#define GGML_F32Cxt_SET1(x)                 svdup_n_f16(x)
+#define GGML_F32Cxt_LOAD(p)                 svld1_f16(DEFAULT_PG16, (const __fp16 *)(p))
+#define GGML_F32Cxt_STORE(dst_ptr, src_vec) svst1_f16(DEFAULT_PG16, (__fp16 *)(dst_ptr), (src_vec))
+
+#define GGML_F32Cxt_FMA_IMPL(pg, a, b, c)   svmad_f16_x(pg, b, c, a)
+#define GGML_F32Cxt_FMA(...)                GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_ADD_IMPL(pg, a, b)      svadd_f16_x(pg, a, b)
+#define GGML_F32Cxt_ADD(...)                GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_MUL_IMPL(pg, a, b)      svmul_f16_x(pg, a, b)
+#define GGML_F32Cxt_MUL(...)                GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, __VA_ARGS__)
+#define GGML_F32Cxt_REDUCE                  GGML_F16xt_REDUCE_MIXED
+
+#define GGML_F16x_VEC                GGML_F32Cxt
+#define GGML_F16x_VEC_ZERO           GGML_F32Cxt_ZERO
+#define GGML_F16x_VEC_SET1           GGML_F32Cxt_SET1
+#define GGML_F16x_VEC_LOAD(p, i)     GGML_F32Cxt_LOAD(p)
+#define GGML_F16x_VEC_STORE(p, r, i) GGML_F32Cxt_STORE((__fp16 *)(p), r)
+#define GGML_F16x_VEC_FMA            GGML_F32Cxt_FMA
+#define GGML_F16x_VEC_ADD            GGML_F32Cxt_ADD
+#define GGML_F16x_VEC_MUL            GGML_F32Cxt_MUL
+#define GGML_F16x_VEC_REDUCE         GGML_F32Cxt_REDUCE
+
+#define GGML_F16xt_REDUCE_ONE_IMPL(pg, a) svaddv_f16(pg, a)
+#define GGML_F16xt_REDUCE_ONE(...)        GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, __VA_ARGS__)
+
+#define GGML_F16xt_REDUCE_MIXED_IMPL(pg16, res, sum1, sum2, sum3, sum4)  \
+{                                                      \
+    sum1 = svadd_f16_x(pg16, sum1, sum2);              \
+    sum3 = svadd_f16_x(pg16, sum3, sum4);              \
+    sum1 = svadd_f16_x(pg16, sum1, sum3);              \
+    __fp16 sum_f16 = svaddv_f16(pg16, sum1);           \
+    (res) = (ggml_float) sum_f16;                      \
+}
+#define GGML_F16xt_REDUCE_MIXED(...) GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, __VA_ARGS__)
+
 // F16 NEON
 
 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
diff --git a/ggml/src/ggml-cpu/vec.cpp b/ggml/src/ggml-cpu/vec.cpp
index d8ec3b81d..437192d52 100644
--- a/ggml/src/ggml-cpu/vec.cpp
+++ b/ggml/src/ggml-cpu/vec.cpp
@@ -85,15 +85,21 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         // reduce sum1,sum2 to sum1
         GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
     #elif defined(__riscv_v_intrinsic)
-        vfloat32m1_t vsum = __riscv_vfmv_v_f_f32m1(0.0f, 1);
-        for (int i = 0, avl; i < n; i += avl) {
-            avl = __riscv_vsetvl_e32m8(n - i);
-            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
-            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
-            vfloat32m8_t prod = __riscv_vfmul_vv_f32m8(ax, ay, avl);
-            vsum = __riscv_vfredusum_vs_f32m8_f32m1(prod, vsum, avl);
+        int vl = __riscv_vsetvlmax_e32m8();
+        vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
+        vfloat32m8_t vsum;
+        vfloat32m8_t ax;
+        vfloat32m8_t ay;
+        vsum = __riscv_vfmv_v_f_f32m8_tu(vsum, 0.0f, vl);
+        for (int i = 0; i < n; i += vl) {
+            vl = __riscv_vsetvl_e32m8(n - i);
+            ax = __riscv_vle32_v_f32m8_tu(ax, &x[i], vl);
+            ay = __riscv_vle32_v_f32m8_tu(ay, &y[i], vl);
+            vsum = __riscv_vfmacc_vv_f32m8_tu(vsum, ax, ay, vl);
         }
-        sumf += __riscv_vfmv_f_s_f32m1_f32(vsum);
+        vl = __riscv_vsetvlmax_e32m8();
+        vs = __riscv_vfredusum_vs_f32m8_f32m1(vsum, vs, vl);
+        sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
     #else
         const int np = (n & ~(GGML_F32_STEP - 1));
 
@@ -207,38 +213,125 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
     ggml_float sumf = 0.0;
 
-#if defined(GGML_SIMD) && !defined(__riscv_v_intrinsic)
-    const int np = (n & ~(GGML_F16_STEP - 1));
 
-    GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
+#if defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = svcntb() * 8; //get vector length
+        const int ggml_f16_epr = sve_register_length / 16; // running when 16
+        const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
 
-    GGML_F16_VEC ax[GGML_F16_ARR];
-    GGML_F16_VEC ay[GGML_F16_ARR];
+        const int np= (n & ~(ggml_f16_step - 1));
+        svfloat16_t sum1 = svdup_n_f16(0.0f);
+        svfloat16_t sum2 = svdup_n_f16(0.0f);
+        svfloat16_t sum3 = svdup_n_f16(0.0f);
+        svfloat16_t sum4 = svdup_n_f16(0.0f);
 
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+        for (int i = 0; i < np; i += ggml_f16_step) {
+            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
+            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
+            sum1 = GGML_F16x_VEC_FMA(sum1, ax1, ay1);
 
-            sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
+            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
+            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
+            sum2 = GGML_F16x_VEC_FMA(sum2, ax2, ay2);
+
+            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
+            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+            sum3 = GGML_F16x_VEC_FMA(sum3, ax3, ay3);
+
+            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
+            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+            sum4 = GGML_F16x_VEC_FMA(sum4, ax4, ay4);
+
+            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
+            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+            sum1 = GGML_F16x_VEC_FMA(sum1, ax5, ay5);
+
+            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
+            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+            sum2 = GGML_F16x_VEC_FMA(sum2, ax6, ay6);
+
+            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
+            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+            sum3 = GGML_F16x_VEC_FMA(sum3, ax7, ay7);
+
+            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
+            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+            sum4 = GGML_F16x_VEC_FMA(sum4, ax8, ay8);
         }
-    }
 
-    // reduce sum0..sum3 to sum0
-    GGML_F16_VEC_REDUCE(sumf, sum);
+        const int np2 = (n & ~(ggml_f16_epr - 1)); // round down to multiple of 8
+        for (int k = np; k < np2; k += ggml_f16_epr) {
+            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
+            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+            sum1 = GGML_F16x_VEC_FMA(sum1, rx, ry);
+        }
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
-    }
+        if (np2 < n) {
+            svbool_t pg = svwhilelt_b16(np2, n);
+            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
+            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
 
-    // if you hit this, you are likely running outside the FP range
-    assert(!isnan(sumf) && !isinf(sumf));
+            sum1 = svmad_f16_x(pg, hx, hy, sum1);
+        }
+        GGML_F16x_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4);
+    #elif defined(__riscv_v_intrinsic)
+        #if defined(__riscv_zvfh)
+            int vl = __riscv_vsetvlmax_e32m2();
+            vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
+            vfloat32m2_t vsum;
+            vfloat16m1_t ax;
+            vfloat16m1_t ay;
+            vsum = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vmv_v_x_u32m2(0, vl));
+            for (int i = 0; i < n; i += vl) {
+                vl = __riscv_vsetvl_e16m1(n - i);
+                ax = __riscv_vle16_v_f16m1_tu(ax, (const _Float16 *)&x[i], vl);
+                ay = __riscv_vle16_v_f16m1_tu(ay, (const _Float16 *)&y[i], vl);
+                vsum = __riscv_vfwmacc_vv_f32m2_tu(vsum, ax, ay, vl);
+            }
+            vl = __riscv_vsetvlmax_e32m1();
+            vfloat32m1_t ac0 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(vsum, 0), __riscv_vget_v_f32m2_f32m1(vsum, 1), vl);
+            vs = __riscv_vfredusum_vs_f32m1_f32m1(ac0, vs, vl);
+            sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
+        #else
+            for (int i = 0; i < n; ++i) {
+                sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
+            }
+        #endif // __riscv_zvfh
+    #else
+        const int np = (n & ~(GGML_F16_STEP - 1));
+
+        GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
+
+        GGML_F16_VEC ax[GGML_F16_ARR];
+        GGML_F16_VEC ay[GGML_F16_ARR];
+
+        for (int i = 0; i < np; i += GGML_F16_STEP) {
+            for (int j = 0; j < GGML_F16_ARR; j++) {
+                ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+
+                sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
+            }
+        }
+
+        // reduce sum0..sum3 to sum0
+        GGML_F16_VEC_REDUCE(sumf, sum);
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
+        }
+        // if you hit this, you are likely running outside the FP range
+        assert(!isnan(sumf) && !isinf(sumf));
+    #endif
 #else
     for (int i = 0; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
-#endif
+#endif // GGML_SIMD
 
     *s = sumf;
 }
@@ -257,6 +350,12 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
     for (; i + 3 < n; i += 4) {
         _mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i)));
     }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    const int vlen = svcntw();
+    for (; i < n; i += vlen) {
+        const svbool_t pg = svwhilelt_b32_s32(i, n);
+        svst1_f32(pg, y + i, ggml_v_silu(pg, svld1_f32(pg, x + i)));
+    }
 #elif defined(__ARM_NEON) && defined(__aarch64__)
     for (; i + 3 < n; i += 4) {
         vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
@@ -281,10 +380,24 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
     for (; i + 3 < n; i += 4) {
         _mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
     }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    const int vlen = svcntw();
+    for (; i < n; i += vlen) {
+        const svbool_t pg = svwhilelt_b32_s32(i, n);
+        svst1_f32(pg, y + i, svmul_f32_x(pg, ggml_v_silu(pg, svld1_f32(pg, x + i)), svld1_f32(pg, g + i)));
+    }
 #elif defined(__ARM_NEON) && defined(__aarch64__)
     for (; i + 3 < n; i += 4) {
         vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
     }
+#elif defined(__riscv_v_intrinsic)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat32m2_t vg = __riscv_vle32_v_f32m2(&g[i], vl);
+        vfloat32m2_t vy = __riscv_vfmul_vv_f32m2(ggml_v_silu_m2(vx, vl), vg, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
 #endif
     for (; i < n; ++i) {
         y[i] = ggml_silu_f32(x[i]) * g[i];
@@ -328,6 +441,15 @@ ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float
 #endif
         sum += (ggml_float)_mm_cvtss_f32(val);
     }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    const int vlen = svcntw();
+    for (; i < n; i += vlen) {
+        const svbool_t pg = svwhilelt_b32_s32(i, n);
+        svfloat32_t val = ggml_v_expf(pg, svsub_f32_x(pg, svld1_f32(pg, x + i),
+                                                svdup_n_f32_x(pg, max)));
+        svst1_f32(pg, y + i, val);
+        sum += (ggml_float)svaddv_f32(pg, val);
+    }
 #elif defined(__ARM_NEON) && defined(__aarch64__)
     for (; i + 3 < n; i += 4) {
         float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i),
diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h
index 8ccf340d4..ef334d089 100644
--- a/ggml/src/ggml-cpu/vec.h
+++ b/ggml/src/ggml-cpu/vec.h
@@ -119,45 +119,149 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
     }
 
 #if defined(GGML_SIMD)
-#if defined(__riscv_v_intrinsic)
-    // todo: RVV impl
-    for (int i = 0; i < n; ++i) {
-        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+    #if defined(__ARM_FEATURE_SVE)
+
+        const int sve_register_length = svcntb() * 8;
+        const int ggml_f16_epr = sve_register_length / 16; // running when 16
+        const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
+
+        const int np = (n & ~(ggml_f16_step - 1));
+
+        svfloat16_t sum_00 = svdup_n_f16(0.0f);
+        svfloat16_t sum_01 = svdup_n_f16(0.0f);
+        svfloat16_t sum_02 = svdup_n_f16(0.0f);
+        svfloat16_t sum_03 = svdup_n_f16(0.0f);
+
+        svfloat16_t sum_10 = svdup_n_f16(0.0f);
+        svfloat16_t sum_11 = svdup_n_f16(0.0f);
+        svfloat16_t sum_12 = svdup_n_f16(0.0f);
+        svfloat16_t sum_13 = svdup_n_f16(0.0f);
+
+        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+
+        for (int i = 0; i < np; i += ggml_f16_step) {
+            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0); // 8 elements
+
+            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elemnst
+            sum_00 = GGML_F16x_VEC_FMA(sum_00, ax1, ay1);     // sum_00 = sum_00+ax1*ay1
+            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 0*ggml_f16_epr, 0); // 8 elements
+            sum_10 = GGML_F16x_VEC_FMA(sum_10, ax1, ay1);
+
+            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1); // next 8 elements
+
+            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 ekements
+            sum_01 = GGML_F16x_VEC_FMA(sum_01, ax2, ay2);
+            ax2 = GGML_F16x_VEC_LOAD(x[1] + i + 1*ggml_f16_epr, 1);
+            sum_11 = GGML_F16x_VEC_FMA(sum_11, ax2, ay2);
+
+            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+
+            ax3 = GGML_F16x_VEC_LOAD(x[0] + i + 2*ggml_f16_epr, 2);
+            sum_02 = GGML_F16x_VEC_FMA(sum_02, ax3, ay3);
+            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
+            sum_12 = GGML_F16x_VEC_FMA(sum_12, ax3, ay3);
+
+            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+
+            ax4 = GGML_F16x_VEC_LOAD(x[0] + i + 3*ggml_f16_epr, 3);
+            sum_03 = GGML_F16x_VEC_FMA(sum_03, ax4, ay4);
+            ax4 = GGML_F16x_VEC_LOAD(x[1] + i + 3*ggml_f16_epr, 3);
+            sum_13 = GGML_F16x_VEC_FMA(sum_13, ax4, ay4);
+
+            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+
+            ax5 = GGML_F16x_VEC_LOAD(x[0] + i + 4*ggml_f16_epr, 4);
+
+            sum_00 = GGML_F16x_VEC_FMA(sum_00, ax5, ay5);
+            ax5 = GGML_F16x_VEC_LOAD(x[1] + i + 4*ggml_f16_epr, 4);
+            sum_10 = GGML_F16x_VEC_FMA(sum_10, ax5, ay5);
+
+            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+
+            ax6 = GGML_F16x_VEC_LOAD(x[0] + i + 5*ggml_f16_epr, 5);
+
+            sum_01 = GGML_F16x_VEC_FMA(sum_01, ax6, ay6);
+            ax6 = GGML_F16x_VEC_LOAD(x[1] + i + 5*ggml_f16_epr, 5);
+            sum_11 = GGML_F16x_VEC_FMA(sum_11, ax6, ay6);
+
+            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+
+            ax7 = GGML_F16x_VEC_LOAD(x[0] + i + 6*ggml_f16_epr, 6);
+
+            sum_02 = GGML_F16x_VEC_FMA(sum_02, ax7, ay7);
+            ax7 = GGML_F16x_VEC_LOAD(x[1] + i + 6*ggml_f16_epr, 6);
+            sum_12 = GGML_F16x_VEC_FMA(sum_12, ax7, ay7);
+
+            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+
+            ax8 = GGML_F16x_VEC_LOAD(x[0] + i + 7*ggml_f16_epr, 7);
+
+            sum_03 = GGML_F16x_VEC_FMA(sum_03, ax8, ay8);
+            ax8 = GGML_F16x_VEC_LOAD(x[1] + i + 7*ggml_f16_epr, 7);
+            sum_13 = GGML_F16x_VEC_FMA(sum_13, ax8, ay8);
         }
-    }
-#else
-    const int np = (n & ~(GGML_F16_STEP - 1));
 
-    GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
+        const int np2 = (n & ~(ggml_f16_epr - 1));
+        for (int k = np; k < np2; k += ggml_f16_epr) {
+            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
 
-    GGML_F16_VEC ax[GGML_F16_ARR];
-    GGML_F16_VEC ay[GGML_F16_ARR];
+            svfloat16_t rx = GGML_F16x_VEC_LOAD(x[0] + k, 0);
+            sum_00 = GGML_F16x_VEC_FMA(sum_00, rx, ry);
+            rx = GGML_F16x_VEC_LOAD(x[1] + k, 0);
+            sum_10 = GGML_F16x_VEC_FMA(sum_10, rx, ry);
+        }
 
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+        if (np2 < n) {
+            svbool_t pg = svwhilelt_b16(np2, n);
+            svfloat16_t hx_0 = svld1_f16(pg, (const __fp16 *)(x[0] + np2));
+            svfloat16_t hx_1 = svld1_f16(pg, (const __fp16 *)(x[1] + np2));
+            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
 
-            for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
-                ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
+            sum_00 = svmad_f16_x(pg, hx_0, hy, sum_00);
+            sum_10 = svmad_f16_x(pg, hx_1, hy, sum_10);
+        }
+        GGML_F16x_VEC_REDUCE(sumf[0], sum_00, sum_01, sum_02, sum_03);
+        GGML_F16x_VEC_REDUCE(sumf[1], sum_10, sum_11, sum_12, sum_13);
+    #elif defined(__riscv_v_intrinsic)
+      // todo: RVV impl
+      for (int i = 0; i < n; ++i) {
+          for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
+              sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+          }
+      }
+    #else
+        const int np = (n & ~(GGML_F16_STEP - 1));
 
-                sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
+        GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
+
+        GGML_F16_VEC ax[GGML_F16_ARR];
+        GGML_F16_VEC ay[GGML_F16_ARR];
+
+        for (int i = 0; i < np; i += GGML_F16_STEP) {
+            for (int j = 0; j < GGML_F16_ARR; j++) {
+                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+
+                for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+                    ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
+
+                    sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
+                }
             }
         }
-    }
 
-    // reduce sum0..sum3 to sum0
-    for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
-        GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
-    }
-
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+        // reduce sum0..sum3 to sum0
+        for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+            GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
         }
-    }
-#endif
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
+                sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+            }
+        }
+    #endif
 #else
     for (int i = 0; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
@@ -293,35 +397,112 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
-#if defined(__riscv_v_intrinsic)
-    // todo: RVV impl
-    // scalar
-    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-    }
-#else
-    const int np = (n & ~(GGML_F16_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = svcntb() * 8;
+        const int ggml_f16_epr = sve_register_length / 16;
+        const int ggml_f16_step = 8 * ggml_f16_epr;
 
-    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+        GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
 
-    GGML_F16_VEC ax[GGML_F16_ARR];
-    GGML_F16_VEC ay[GGML_F16_ARR];
+        const int np= (n & ~(ggml_f16_step - 1));
 
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-            ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+        svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+        svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+        for (int i = 0; i < np; i += ggml_f16_step) {
+            ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
+            ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
+            ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
 
-            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+            GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
+
+            ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
+            ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
+            ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
+
+            ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
+            ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+            ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
+
+            ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
+            ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+            ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
+
+            ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
+            ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+            ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
+
+            ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
+            ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+            ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
+
+            ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
+            ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+            ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
+
+            ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
+            ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+            ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
+
+            GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
         }
-    }
+        const int np2 = (n & ~(ggml_f16_epr - 1));
+        for (int k = np; k < np2; k += ggml_f16_epr) {
+            svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
+            svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+            ry = GGML_F16x_VEC_FMA(ry, rx, vx);
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
-    }
-#endif
+            GGML_F16x_VEC_STORE(y + k, ry, 0);
+        }
+
+        if (np2 < n) {
+            svbool_t pg = svwhilelt_b16(np2, n);
+            svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
+            svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+            hy = svmad_f16_x(pg, hx, vx, hy);
+            svst1_f16(pg, (__fp16 *)(y + np2), hy);
+        }
+
+    #elif defined(__riscv_v_intrinsic)
+        // todo: RVV impl
+        // scalar
+        for (int i = 0; i < n; ++i) {
+            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
+        }
+    #else
+        const int np = (n & ~(GGML_F16_STEP - 1));
+
+        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+        GGML_F16_VEC ax[GGML_F16_ARR];
+        GGML_F16_VEC ay[GGML_F16_ARR];
+
+        for (int i = 0; i < np; i += GGML_F16_STEP) {
+            for (int j = 0; j < GGML_F16_ARR; j++) {
+                ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+                ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+
+                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -517,33 +698,59 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 
 inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
 #if defined(GGML_SIMD)
-#if defined(__riscv_v_intrinsic)
-    // todo: RVV impl
-    // scalar
-    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-    }
-#else
-    const int np = (n & ~(GGML_F16_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = svcntb() * 8;
+        const int ggml_f16_epr = sve_register_length / 16;
+        const int ggml_f16_step = 2 * ggml_f16_epr;
 
-    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+        GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
+        const int np = (n & ~(ggml_f16_step - 1));
+        svfloat16_t ay1, ay2;
 
-    GGML_F16_VEC ay[GGML_F16_ARR];
+        for (int i = 0; i < np; i += ggml_f16_step) {
+            ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
+            ay1 = GGML_F16x_VEC_MUL(ay1, vx);
+            GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
 
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
-            ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
-
-            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+            ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
+            ay2 = GGML_F16x_VEC_MUL(ay2, vx);
+            GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
         }
-    }
+        // leftovers
+        // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
+        if (np < n) {
+            svbool_t pg = svwhilelt_b16(np, n);
+            svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
+            svfloat16_t out = svmul_f16_m(pg, hy, vx);
+            svst1_f16(pg, (__fp16 *)(y + np), out);
+        }
+    #elif defined(__riscv_v_intrinsic)
+        // todo: RVV impl
+        // scalar
+        for (int i = 0; i < n; ++i) {
+            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+        }
+    #else
+        const int np = (n & ~(GGML_F16_STEP - 1));
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-    }
-#endif
+        GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+        GGML_F16_VEC ay[GGML_F16_ARR];
+
+        for (int i = 0; i < np; i += GGML_F16_STEP) {
+            for (int j = 0; j < GGML_F16_ARR; j++) {
+                ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+                ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+                GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -795,7 +1002,39 @@ https://github.com/openvinotoolkit/openvino/blob/master/src/plugins/intel_cpu/sr
     }
 #endif
 
-#if defined(__ARM_NEON) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+
+inline static svfloat32_t ggml_v_expf(svbool_t pg, svfloat32_t x) {
+    const svfloat32_t r = svdup_n_f32_x(pg, 0x1.8p23f);
+    const svfloat32_t z = svmla_n_f32_x(pg, r, x, 0x1.715476p+0f);
+    const svfloat32_t n = svsub_f32_x(pg, z, r);
+    const svfloat32_t b = svmls_n_f32_x(pg, svmls_n_f32_x(pg, x, n, 0x1.62e4p-1f), n, 0x1.7f7d1cp-20f);
+    const svuint32_t e = svlsl_n_u32_x(pg, svreinterpret_u32_f32(z), 23);
+    const svfloat32_t k = svreinterpret_f32_u32(svadd_u32_x(pg, e, svreinterpret_u32_f32(svdup_n_f32_x(pg, 1))));
+    const svbool_t c = svacgt_n_f32(pg, n, 126);
+    const svfloat32_t u = svmul_f32_x(pg, b, b);
+    const svfloat32_t j = svmla_f32_x(pg,
+        svmul_n_f32_x(pg, b, 0x1.ffffecp-1f),
+        svmla_f32_x(pg, svmla_f32_x(pg, svdup_n_f32_x(pg, 0x1.fffdb6p-2f), svdup_n_f32_x(pg, 0x1.555e66p-3f), b),
+                        svmla_f32_x(pg, svdup_n_f32_x(pg, 0x1.573e2ep-5f), svdup_n_f32_x(pg, 0x1.0e4020p-7f), b), u), u);
+    const svuint32_t d = svdup_n_u32_z(svcmple_n_f32(pg, n, 0.0), 0x82000000);
+    const svfloat32_t s1 = svreinterpret_f32_u32(svadd_n_u32_x(pg, d, 0x7f000000));
+    const svfloat32_t s2 = svreinterpret_f32_u32(svsub_u32_x(pg, e, d));
+    return svsel_f32(svacgt_f32(pg, n, svdup_n_f32_x(pg, 192)), svmul_f32_x(pg, s1, s1),
+                     svsel_f32(c, svmul_f32_x(pg, svmla_f32_x(pg, s2, s2, j), s1), svmla_f32_x(pg, k, k, j)));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static svfloat32_t ggml_v_silu(svbool_t pg, svfloat32_t x) {
+    const svfloat32_t one = svdup_n_f32_x(pg, 1.0f);
+    const svfloat32_t zero = svdup_n_f32_x(pg, 0.0f);
+    const svfloat32_t neg_x = svsub_f32_x(pg, zero, x);
+    const svfloat32_t exp_neg_x = ggml_v_expf(pg, neg_x);
+    const svfloat32_t one_plus_exp_neg_x = svadd_f32_x(pg, one, exp_neg_x);
+    return svdiv_f32_x(pg, x, one_plus_exp_neg_x);
+}
+
+#elif defined(__ARM_NEON) && defined(__aarch64__)
 
 // adapted from arm limited optimized routine
 // the maximum error is 1.45358 plus 0.5 ulps
@@ -1030,6 +1269,14 @@ inline static vfloat32m2_t ggml_v_expf_m2(vfloat32m2_t x, int vl) {
         vl);
 }
 
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static vfloat32m2_t ggml_v_silu_m2(vfloat32m2_t x, int vl) {
+    const vfloat32m2_t neg_x = __riscv_vfneg_v_f32m2(x, vl);
+    const vfloat32m2_t exp_neg_x = ggml_v_expf_m2(neg_x, vl);
+    const vfloat32m2_t one_plus_exp_neg_x = __riscv_vfadd_vf_f32m2(exp_neg_x, 1.0f, vl);
+    return __riscv_vfdiv_vv_f32m2(x, one_plus_exp_neg_x, vl);
+}
+
 #endif // __ARM_NEON / __AVX2__ / __SSE2__ / __riscv_v_intrinsic
 
 inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
diff --git a/ggml/src/ggml-cuda/conv2d.cu b/ggml/src/ggml-cuda/conv2d.cu
index bcb70762e..142dd6690 100644
--- a/ggml/src/ggml-cuda/conv2d.cu
+++ b/ggml/src/ggml-cuda/conv2d.cu
@@ -1,4 +1,5 @@
 #include "conv2d.cuh"
+#include "convert.cuh"
 
 struct conv_params {
     const int64_t IW, IH;
@@ -94,8 +95,8 @@ static __global__ void conv2d_kernel(const float * __restrict__ input,
                 const int64_t in_x = calculate_input_coord(out_x, kx, P.ST_X, P.DL_X, P.PD_X);
 
                 const float input_val = input[Layout::input_index(n, c_in, in_y, in_x, P)];
-                const float kernel_val = kernel[Layout::kernel_index(c_out, c_in, ky, kx, P)];
-                acc += (input_val * kernel_val);
+                const T kernel_val = kernel[Layout::kernel_index(c_out, c_in, ky, kx, P)];
+                acc += (input_val * ggml_cuda_cast<float>(kernel_val));
             }
         }
     }
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 8fbc6a974..749ba09a7 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -523,13 +523,6 @@ enum ggml_metal_kernel_type {
     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_H40,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
@@ -1562,13 +1555,6 @@ 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_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_H40,      flash_attn_ext_vec_f16_h40,      has_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40,     flash_attn_ext_vec_bf16_h40,     has_simdgroup_reduction && use_bfloat);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40,     flash_attn_ext_vec_q4_0_h40,     has_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40,     flash_attn_ext_vec_q4_1_h40,     has_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40,     flash_attn_ext_vec_q5_0_h40,     has_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40,     flash_attn_ext_vec_q5_1_h40,     has_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40,     flash_attn_ext_vec_q8_0_h40,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,      flash_attn_ext_vec_f16_h64,      has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,     flash_attn_ext_vec_bf16_h64,     has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,     flash_attn_ext_vec_q4_0_h64,     has_simdgroup_reduction);
@@ -1909,9 +1895,15 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_ARANGE:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
-            if (op->src[0]->ne[0] == 32) {
-                // head size == 32 (e.g. bert-bge-small)
-                // TODO: not sure if it is worth adding kernels for this size
+            // for new head sizes, add checks here
+            if (op->src[0]->ne[0] != 40 &&
+                op->src[0]->ne[0] != 64 &&
+                op->src[0]->ne[0] != 80 &&
+                op->src[0]->ne[0] != 96 &&
+                op->src[0]->ne[0] != 112 &&
+                op->src[0]->ne[0] != 128 &&
+                op->src[0]->ne[0] != 192 &&
+                op->src[0]->ne[0] != 256) {
                 return false;
             }
             if (op->src[0]->ne[0] == 576) {
@@ -5138,10 +5130,8 @@ static int ggml_metal_encode_node(
 
                 bool use_vec_kernel = false;
 
-                // 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 >= 20 || (ne00%128 != 0 && ne00 != 64 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
+                // use non-vec kernel if the batch size is large or if the vec-kernel is not supported for this head size
+                if (ne01 >= 20 || (ne00 == 40 || ne00 == 80 || ne00 == 112)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16:
                             {
@@ -5329,24 +5319,6 @@ static int ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
-                        case 40:
-                            {
-                                switch (src1->type) {
-                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H40].pipeline; break;
-                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H40].pipeline; break;
-                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H40].pipeline; break;
-                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H40].pipeline; break;
-                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H40].pipeline; break;
-                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H40].pipeline; break;
-                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H40].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 64:
                             {
                                 switch (src1->type) {
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 4fa16c4a5..9c5933d24 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -4803,6 +4803,9 @@ kernel void kernel_flash_attn_ext_vec(
         ushort3   ntg[[threads_per_threadgroup]],
         ushort  tiisg[[thread_index_in_simdgroup]],
         ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
+    static_assert(DK % 32 == 0, "DK must be divisible by 32");
+    static_assert(DV % 32 == 0, "DV must be divisible by 32");
+
     const short nsg = ntg.y; // number of simdgroups
     const short iwg = tgpig[2]%nwg;
 
@@ -5160,16 +5163,6 @@ 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_h40")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  40, 40, 8>;
-#if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_vec_bf16_h40")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 40, 40, 8>;
-#endif
-template [[host_name("kernel_flash_attn_ext_vec_q4_0_h40")]] 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, 40, 40, 8>;
-template [[host_name("kernel_flash_attn_ext_vec_q4_1_h40")]] 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, 40, 40, 8>;
-template [[host_name("kernel_flash_attn_ext_vec_q5_0_h40")]] 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, 40, 40, 8>;
-template [[host_name("kernel_flash_attn_ext_vec_q5_1_h40")]] 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, 40, 40, 8>;
-template [[host_name("kernel_flash_attn_ext_vec_q8_0_h40")]] 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, 40, 40, 8>;
-
 template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  64, 64, 8>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 64, 64, 8>;
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 264601cbb..1ce8fbf04 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -376,6 +376,13 @@ struct vk_fa_pipeline_state {
     }
 };
 
+enum shader_reduction_mode {
+    SHADER_REDUCTION_MODE_SHMEM,
+    SHADER_REDUCTION_MODE_HYBRID,
+    SHADER_REDUCTION_MODE_SUBGROUP,
+    SHADER_REDUCTION_MODE_COUNT,
+};
+
 static constexpr uint32_t num_argsort_pipelines = 11;
 static constexpr uint32_t max_argsort_cols = 1 << (num_argsort_pipelines-1);
 
@@ -402,15 +409,18 @@ struct vk_device_struct {
     bool uma;
     bool prefer_host_memory;
     bool float_controls_rte_fp16;
-    bool subgroup_add;
+    bool subgroup_arithmetic;
     bool subgroup_shuffle;
     bool subgroup_ballot;
+    bool subgroup_clustered;
     bool multi_add;
 
     bool add_rms_fusion;
     uint32_t partials_binding_alignment;
 
     bool integer_dot_product;
+    // 0: default, 1: force mmvq, -1: disable mmvq
+    int32_t mmvq_mode;
 
     bool subgroup_size_control;
     uint32_t subgroup_min_size;
@@ -468,12 +478,15 @@ struct vk_device_struct {
 
     vk_pipeline pipeline_matmul_split_k_reduce;
     vk_pipeline pipeline_quantize_q8_1;
+    vk_pipeline pipeline_quantize_q8_1_x4;
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
 
+    vk_pipeline pipeline_dequant_mul_mat_vec_q8_1_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio];
     vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
@@ -1373,6 +1386,7 @@ struct vk_instance_t {
     PFN_vkCmdInsertDebugUtilsLabelEXT  pfn_vkCmdInsertDebugUtilsLabelEXT  = {};
 
     std::vector<size_t> device_indices;
+    std::vector<bool>   device_supports_membudget;
     vk_device devices[GGML_VK_MAX_DEVICES];
 };
 
@@ -2780,11 +2794,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, REQSUBGROUPSIZE > 0, false, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, REQSUBGROUPSIZE > 0, false, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, REQSUBGROUPSIZE > 0, false, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
             ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
@@ -2927,60 +2941,91 @@ static void ggml_vk_load_shaders(vk_device& device) {
         rm_stdq = 2;
     uint32_t rm_iq = 2 * rm_kq;
 
-    for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
-        uint32_t wg_size_subgroup16 = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_16 : (subgroup_size_16 * 4);
-        uint32_t wg_size_subgroup   = (w == DMMV_WG_SIZE_SUBGROUP) ? device->subgroup_size : (device->subgroup_size * 4);
+    const bool use_subgroups = device->subgroup_arithmetic && device->architecture != vk_device_architecture::AMD_GCN;
+    // Ensure a subgroup size >= 16 is available
+    const bool use_subgroups16 = use_subgroups &&
+                                    (!device->subgroup_size_control && device->subgroup_size >= 16 ||
+                                    device->subgroup_size_control && device->subgroup_min_size <= 16 && device->subgroup_max_size >= 16);
 
-        const bool s = device->subgroup_add && device->architecture != vk_device_architecture::AMD_GCN;
+    const uint32_t subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control && device->subgroup_min_size <= 16 && device->subgroup_max_size >= 16) ? 16 : device->subgroup_size;
+    const uint32_t subgroup_size16 = std::max(subgroup_size, 16u);
+
+    const uint32_t force_subgroup_size = use_subgroups ? subgroup_size : 0;
+    const uint32_t force_subgroup_size16 = use_subgroups16 ? subgroup_size16 : 0;
+
+    for (uint32_t w = 0; w < DMMV_WG_SIZE_COUNT; ++w) {
+        const uint32_t wg_size_subgroup   = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size : (subgroup_size * 4);
+        const uint32_t wg_size_subgroup16 = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size16 : (subgroup_size16 * 4);
+
+        const shader_reduction_mode reduc = (use_subgroups && w == DMMV_WG_SIZE_SUBGROUP) ? SHADER_REDUCTION_MODE_SUBGROUP :
+                                            (use_subgroups && w == DMMV_WG_SIZE_LARGE) ? SHADER_REDUCTION_MODE_HYBRID :
+                                            SHADER_REDUCTION_MODE_SHMEM;
+
+        const shader_reduction_mode reduc16 = (use_subgroups16 && w == DMMV_WG_SIZE_SUBGROUP) ? SHADER_REDUCTION_MODE_SUBGROUP :
+                                              (use_subgroups16 && w == DMMV_WG_SIZE_LARGE) ? SHADER_REDUCTION_MODE_HYBRID :
+                                              SHADER_REDUCTION_MODE_SHMEM;
 
         for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32",  arr_dmmv_f32_f32_f32_len[s],  arr_dmmv_f32_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32",  arr_dmmv_f16_f32_f32_len[s],  arr_dmmv_f16_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[s], arr_dmmv_bf16_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[s], arr_dmmv_q4_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[s], arr_dmmv_q4_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[s], arr_dmmv_q5_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[s], arr_dmmv_q5_1_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[s], arr_dmmv_q8_0_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[s], arr_dmmv_q2_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[s], arr_dmmv_q3_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[s], arr_dmmv_q4_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[s], arr_dmmv_q5_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[s], arr_dmmv_q6_k_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32",   arr_dmmv_iq1_s_f32_f32_len[s],   arr_dmmv_iq1_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32",   arr_dmmv_iq1_m_f32_f32_len[s],   arr_dmmv_iq1_m_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[s], arr_dmmv_iq2_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32",  arr_dmmv_iq2_xs_f32_f32_len[s],  arr_dmmv_iq2_xs_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32",   arr_dmmv_iq2_s_f32_f32_len[s],   arr_dmmv_iq2_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[s], arr_dmmv_iq3_xxs_f32_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32",   arr_dmmv_iq3_s_f32_f32_len[s],   arr_dmmv_iq3_s_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32",  arr_dmmv_iq4_xs_f32_f32_len[s],  arr_dmmv_iq4_xs_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32",  arr_dmmv_iq4_nl_f32_f32_len[s],  arr_dmmv_iq4_nl_f32_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32",   arr_dmmv_mxfp4_f32_f32_len[s],   arr_dmmv_mxfp4_f32_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32",  arr_dmmv_f32_f32_f32_len[reduc],  arr_dmmv_f32_f32_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32",  arr_dmmv_f16_f32_f32_len[reduc],  arr_dmmv_f16_f32_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32", arr_dmmv_bf16_f32_f32_len[reduc], arr_dmmv_bf16_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32", arr_dmmv_q4_0_f32_f32_len[reduc], arr_dmmv_q4_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32", arr_dmmv_q4_1_f32_f32_len[reduc], arr_dmmv_q4_1_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32", arr_dmmv_q5_0_f32_f32_len[reduc], arr_dmmv_q5_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32", arr_dmmv_q5_1_f32_f32_len[reduc], arr_dmmv_q5_1_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32", arr_dmmv_q8_0_f32_f32_len[reduc], arr_dmmv_q8_0_f32_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32", arr_dmmv_q2_k_f32_f32_len[reduc16], arr_dmmv_q2_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32", arr_dmmv_q3_k_f32_f32_len[reduc16], arr_dmmv_q3_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32", arr_dmmv_q4_k_f32_f32_len[reduc16], arr_dmmv_q4_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32", arr_dmmv_q5_k_f32_f32_len[reduc16], arr_dmmv_q5_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32", arr_dmmv_q6_k_f32_f32_len[reduc16], arr_dmmv_q6_k_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f32_f32",   arr_dmmv_iq1_s_f32_f32_len[reduc16],   arr_dmmv_iq1_s_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f32_f32",   arr_dmmv_iq1_m_f32_f32_len[reduc16],   arr_dmmv_iq1_m_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32", arr_dmmv_iq2_xxs_f32_f32_len[reduc16], arr_dmmv_iq2_xxs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32",  arr_dmmv_iq2_xs_f32_f32_len[reduc16],  arr_dmmv_iq2_xs_f32_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32",   arr_dmmv_iq2_s_f32_f32_len[reduc16],   arr_dmmv_iq2_s_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32", arr_dmmv_iq3_xxs_f32_f32_len[reduc16], arr_dmmv_iq3_xxs_f32_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32",   arr_dmmv_iq3_s_f32_f32_len[reduc16],   arr_dmmv_iq3_s_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f32_f32",  arr_dmmv_iq4_xs_f32_f32_len[reduc16],  arr_dmmv_iq4_xs_f32_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32",  arr_dmmv_iq4_nl_f32_f32_len[reduc16],  arr_dmmv_iq4_nl_f32_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f32_f32",   arr_dmmv_mxfp4_f32_f32_len[reduc16],   arr_dmmv_mxfp4_f32_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
 
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32",  arr_dmmv_f32_f16_f32_len[s],  arr_dmmv_f32_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32",  arr_dmmv_f16_f16_f32_len[s],  arr_dmmv_f16_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[s], arr_dmmv_bf16_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[s], arr_dmmv_q4_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[s], arr_dmmv_q4_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[s], arr_dmmv_q5_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[s], arr_dmmv_q5_1_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[s], arr_dmmv_q8_0_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[s], arr_dmmv_q2_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[s], arr_dmmv_q3_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[s], arr_dmmv_q4_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[s], arr_dmmv_q5_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[s], arr_dmmv_q6_k_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32",   arr_dmmv_iq1_s_f16_f32_len[s],   arr_dmmv_iq1_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32",   arr_dmmv_iq1_m_f16_f32_len[s],   arr_dmmv_iq1_m_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[s], arr_dmmv_iq2_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32",  arr_dmmv_iq2_xs_f16_f32_len[s],  arr_dmmv_iq2_xs_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32",   arr_dmmv_iq2_s_f16_f32_len[s],   arr_dmmv_iq2_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[s], arr_dmmv_iq3_xxs_f16_f32_data[s], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32",   arr_dmmv_iq3_s_f16_f32_len[s],   arr_dmmv_iq3_s_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32",  arr_dmmv_iq4_xs_f16_f32_len[s],  arr_dmmv_iq4_xs_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32",  arr_dmmv_iq4_nl_f16_f32_len[s],  arr_dmmv_iq4_nl_f16_f32_data[s],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32",   arr_dmmv_mxfp4_f16_f32_len[s],   arr_dmmv_mxfp4_f16_f32_data[s],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32",  arr_dmmv_f32_f16_f32_len[reduc],  arr_dmmv_f32_f16_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32",  arr_dmmv_f16_f16_f32_len[reduc],  arr_dmmv_f16_f16_f32_data[reduc],  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32", arr_dmmv_bf16_f16_f32_len[reduc], arr_dmmv_bf16_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {wg_size_subgroup, 2, i+1}, 1, false, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32", arr_dmmv_q4_0_f16_f32_len[reduc], arr_dmmv_q4_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32", arr_dmmv_q4_1_f16_f32_len[reduc], arr_dmmv_q4_1_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32", arr_dmmv_q5_0_f16_f32_len[reduc], arr_dmmv_q5_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32", arr_dmmv_q5_1_f16_f32_len[reduc], arr_dmmv_q5_1_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup, 2*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32", arr_dmmv_q8_0_f16_f32_len[reduc], arr_dmmv_q8_0_f16_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup, 1*rm_stdq, i+1}, 1, true, use_subgroups, force_subgroup_size);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32", arr_dmmv_q2_k_f16_f32_len[reduc16], arr_dmmv_q2_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32", arr_dmmv_q3_k_f16_f32_len[reduc16], arr_dmmv_q3_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32", arr_dmmv_q4_k_f16_f32_len[reduc16], arr_dmmv_q4_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32", arr_dmmv_q5_k_f16_f32_len[reduc16], arr_dmmv_q5_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32", arr_dmmv_q6_k_f16_f32_len[reduc16], arr_dmmv_q6_k_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {wg_size_subgroup16, rm_kq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_S][i],   "mul_mat_vec_iq1_s_f16_f32",   arr_dmmv_iq1_s_f16_f32_len[reduc16],   arr_dmmv_iq1_s_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ1_M][i],   "mul_mat_vec_iq1_m_f16_f32",   arr_dmmv_iq1_m_f16_f32_len[reduc16],   arr_dmmv_iq1_m_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32", arr_dmmv_iq2_xxs_f16_f32_len[reduc16], arr_dmmv_iq2_xxs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32",  arr_dmmv_iq2_xs_f16_f32_len[reduc16],  arr_dmmv_iq2_xs_f16_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32",   arr_dmmv_iq2_s_f16_f32_len[reduc16],   arr_dmmv_iq2_s_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32", arr_dmmv_iq3_xxs_f16_f32_len[reduc16], arr_dmmv_iq3_xxs_f16_f32_data[reduc16], "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32",   arr_dmmv_iq3_s_f16_f32_len[reduc16],   arr_dmmv_iq3_s_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_XS][i],  "mul_mat_vec_iq4_xs_f16_f32",  arr_dmmv_iq4_xs_f16_f32_len[reduc16],  arr_dmmv_iq4_xs_f16_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32",  arr_dmmv_iq4_nl_f16_f32_len[reduc16],  arr_dmmv_iq4_nl_f16_f32_data[reduc16],  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[w][GGML_TYPE_MXFP4][i],   "mul_mat_vec_mxfp4_f16_f32",   arr_dmmv_mxfp4_f16_f32_len[reduc16],   arr_dmmv_mxfp4_f16_f32_data[reduc16],   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {wg_size_subgroup16, rm_iq, i+1}, 1, true, use_subgroups16, force_subgroup_size16);
+
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+            if (device->integer_dot_product) {
+                const uint32_t subgroup_size_int = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
+                const uint32_t wg_size_subgroup_int = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_int : (subgroup_size_int * 4);
+
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_q8_1_f32", arr_dmmv_q4_0_q8_1_f32_len[reduc], arr_dmmv_q4_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_q8_1_f32", arr_dmmv_q4_1_q8_1_f32_len[reduc], arr_dmmv_q4_1_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_q8_1_f32", arr_dmmv_q5_0_q8_1_f32_len[reduc], arr_dmmv_q5_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_q8_1_f32", arr_dmmv_q5_1_q8_1_f32_len[reduc], arr_dmmv_q5_1_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
+                ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_q8_1_f32[w][GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_q8_1_f32", arr_dmmv_q8_0_q8_1_f32_len[reduc], arr_dmmv_q8_0_q8_1_f32_data[reduc], "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq, i+1}, 1, true, use_subgroups, subgroup_size_int);
+            }
+#endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT
         }
     }
 
@@ -3072,10 +3117,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
+
+    if (device->subgroup_clustered && device->subgroup_require_full_support) {
+        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_subgroup_len, quantize_q8_1_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
+        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_len, quantize_q8_1_x4_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
+    }
 
     for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
-        if (device->subgroup_add && device->subgroup_require_full_support) {
+        if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true);
         } else {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
@@ -3604,11 +3656,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
         }
         device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
 
-        device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
-                               (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
-
+        device->subgroup_arithmetic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
+                                      (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
         device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                    (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
+        device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
+                                     (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eClustered);
 
         device->subgroup_ballot = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                   (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eBallot);
@@ -4068,11 +4121,18 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;
 
         device->add_rms_fusion = !device->disable_fusion &&
-                                 device->subgroup_add &&
+                                 device->subgroup_arithmetic &&
                                  device->vendor_id != VK_VENDOR_ID_INTEL;
         device->partials_binding_alignment =
             std::max(4u, (uint32_t)device->properties.limits.minStorageBufferOffsetAlignment);
 
+        device->mmvq_mode = 0;
+        if (getenv("GGML_VK_DISABLE_MMVQ")) {
+            device->mmvq_mode = -1;
+        } else if (getenv("GGML_VK_FORCE_MMVQ")) {
+            device->mmvq_mode = 1;
+        }
+
         return device;
     }
 
@@ -4311,15 +4371,16 @@ static void ggml_vk_instance_init() {
         vk_instance.pfn_vkCmdBeginDebugUtilsLabelEXT = (PFN_vkCmdBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdBeginDebugUtilsLabelEXT");
         vk_instance.pfn_vkCmdEndDebugUtilsLabelEXT =   (PFN_vkCmdEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdEndDebugUtilsLabelEXT");
         vk_instance.pfn_vkCmdInsertDebugUtilsLabelEXT = (PFN_vkCmdInsertDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdInsertDebugUtilsLabelEXT");
-
     }
 
     vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
 
+    std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
+
     // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
     char * devices_env = getenv("GGML_VK_VISIBLE_DEVICES");
     if (devices_env != nullptr) {
-        size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
+        size_t num_available_devices = devices.size();
 
         std::string devices(devices_env);
         std::replace(devices.begin(), devices.end(), ',', ' ');
@@ -4334,8 +4395,6 @@ static void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(tmp);
         }
     } else {
-        std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
-
         // If no vulkan devices are found, return early
         if (devices.empty()) {
             GGML_LOG_INFO("ggml_vulkan: No devices found.\n");
@@ -4440,6 +4499,19 @@ static void ggml_vk_instance_init() {
     GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size());
 
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
+        vk::PhysicalDevice vkdev = devices[vk_instance.device_indices[i]];
+        std::vector<vk::ExtensionProperties> extensionprops = vkdev.enumerateDeviceExtensionProperties();
+
+        bool membudget_supported = false;
+        for (const auto & ext : extensionprops) {
+            if (strcmp(VK_EXT_MEMORY_BUDGET_EXTENSION_NAME, ext.extensionName) == 0) {
+                membudget_supported = true;
+                break;
+            }
+        }
+
+        vk_instance.device_supports_membudget.push_back(membudget_supported);
+
         ggml_vk_print_gpu_info(i);
     }
 }
@@ -4586,9 +4658,22 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols, uint32_t m, uint32_t k) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
-    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16 || b_type == GGML_TYPE_Q8_1);
     GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
 
+    if (b_type == GGML_TYPE_Q8_1) {
+        switch (a_type) {
+            case GGML_TYPE_Q4_0:
+            case GGML_TYPE_Q4_1:
+            case GGML_TYPE_Q5_0:
+            case GGML_TYPE_Q5_1:
+            case GGML_TYPE_Q8_0:
+                break;
+            default:
+                return nullptr;
+        }
+    }
+
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
@@ -4620,7 +4705,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
 
     // heuristic to choose workgroup size
     uint32_t dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
-    if (ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
+    if ((ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA && ctx->device->architecture != vk_device_architecture::NVIDIA_PRE_TURING) || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
         // Prefer larger workgroups when M is small, to spread the work out more
         // and keep more SMs busy.
         // q6_k seems to prefer small workgroup size even for "medium" values of M.
@@ -4635,6 +4720,13 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
         }
     }
 
+    if (b_type == GGML_TYPE_Q8_1) {
+        if (ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
+            dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
+        }
+        return ctx->device->pipeline_dequant_mul_mat_vec_q8_1_f32[dmmv_wg][a_type][num_cols-1];
+    }
+
     return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[dmmv_wg][a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[dmmv_wg][a_type][num_cols-1];
 }
 
@@ -4704,7 +4796,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
-    VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
+    VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec_id()");
     GGML_ASSERT(b_type == GGML_TYPE_F32);
 
     switch (a_type) {
@@ -5617,20 +5709,20 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
     ggml_vk_sync_buffers(ctx, subctx);
 }
 
-static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
+static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type, bool use_x4_blocks) {
     switch(type) {
         case GGML_TYPE_Q8_1:
-            return ctx->device->pipeline_quantize_q8_1;
+            return use_x4_blocks ? ctx->device->pipeline_quantize_q8_1_x4 : ctx->device->pipeline_quantize_q8_1;
         default:
             std::cerr << "Missing quantize pipeline for type: " << ggml_type_name(type) << std::endl;
             GGML_ABORT("fatal error");
     }
 }
 
-static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne) {
+static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne, bool use_x4_blocks = false) {
     VK_LOG_DEBUG("ggml_vk_quantize_q8_1(" << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ", " << ne << ")");
 
-    vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
+    vk_pipeline pipeline = use_x4_blocks ? ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true) : ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, false);
 
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
     ggml_vk_sync_buffers(ctx, subctx);
@@ -5750,12 +5842,15 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
 
     if (quantize_y) {
-        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
     }
 
     if (dryrun) {
         const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        if (quantize_y) {
+            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
+        }
         const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0;
         if (
                 (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
@@ -5821,7 +5916,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13);
     } else if (quantize_y) {
         d_Y = ctx->prealloc_y;
-        GGML_ASSERT(d_Y->size >= y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1));
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
     } else {
         d_Y = d_Qy;
         y_buf_offset = qy_buf_offset;
@@ -5858,7 +5953,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
             if (ctx->prealloc_y_need_sync) {
                 ggml_vk_sync_buffers(ctx, subctx);
             }
-            ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13);
+            ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13, true);
             ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
             ctx->prealloc_y_last_tensor_used = src1;
         }
@@ -5875,10 +5970,15 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
     }
 
+    uint32_t y_sz_total = y_sz * ne12 * ne13;
+    if (quantize_y) {
+        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
+    }
+
     // compute
     ggml_vk_matmul(
         ctx, subctx, pipeline,
-        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
+        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz_total },
         { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
         ne01, ne11, ne10,
         ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21,
@@ -5893,6 +5993,51 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     }
 }
 
+// Device tuning
+static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_t n, uint32_t k, ggml_type src0_type) {
+    if (device->mmvq_mode == 1) {
+        return true;
+    } else if (device->mmvq_mode == -1) {
+        return false;
+    }
+
+    // MMVQ is generally good for batches
+    if (n > 1) {
+        return true;
+    }
+
+    switch (device->vendor_id) {
+    case VK_VENDOR_ID_NVIDIA:
+        switch (src0_type) {
+        case GGML_TYPE_Q8_0:
+            return device->architecture == vk_device_architecture::NVIDIA_PRE_TURING;
+        default:
+            return true;
+        }
+    case VK_VENDOR_ID_AMD:
+        switch (src0_type) {
+        case GGML_TYPE_Q8_0:
+            return device->architecture == vk_device_architecture::AMD_GCN;
+        default:
+            return true;
+        }
+    case VK_VENDOR_ID_INTEL:
+        switch (src0_type) {
+        // From tests on A770 Linux, may need more tuning
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q5_1:
+            return false;
+        default:
+            return true;
+        }
+    default:
+        return true;
+    }
+
+    GGML_UNUSED(m);
+    GGML_UNUSED(k);
+}
+
 static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -5947,22 +6092,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
 
     const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
-
-    const bool qx_needs_dequant = x_non_contig;
-    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
-
-    // Not implemented
-    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
-
-    const uint64_t x_ne = ne01 * ne00;
-    const uint64_t y_ne = ne11 * ne10;
-    const uint64_t d_ne = ne11 * ne01;
-
-    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
-    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
-    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
-    const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
-    const uint64_t d_sz = sizeof(float) * d_ne;
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne11, ne10, src0->type);
 
     vk_pipeline to_fp16_vk_0 = nullptr;
     vk_pipeline to_fp16_vk_1 = nullptr;
@@ -5974,14 +6104,47 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11, ne20, ne00);
+
+    // Check for mmq first
+    vk_pipeline dmmv = quantize_y ? ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, GGML_TYPE_Q8_1, ne11, ne20, ne00) : nullptr;
+    vk_pipeline to_q8_1 = nullptr;
+
+    if (dmmv == nullptr) {
+        // Fall back to f16 dequant mul mat
+        dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11, ne20, ne00);
+        quantize_y = false;
+    }
+
+    if (quantize_y) {
+        to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1, true);
+    }
+
+    const bool qx_needs_dequant = x_non_contig;
+    const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
 
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = ne11 * ne10;
+    const uint64_t d_ne = ne11 * ne01;
+
+    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
+    const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne);
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
     if (dryrun) {
         const uint64_t x_sz_upd = x_sz * ne02 * ne03;
-        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        if (quantize_y) {
+            y_sz_upd = CEIL_DIV(y_sz_upd, 144) * 144;
+        }
         if (
                 (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
                 (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) {
@@ -5990,7 +6153,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
             ctx->prealloc_size_x = x_sz_upd;
         }
-        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
+        if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) {
             ctx->prealloc_size_y = y_sz_upd;
         }
 
@@ -6001,6 +6164,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         if (qy_needs_dequant) {
             ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
+        if (quantize_y) {
+            ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
+        }
         ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
         return;
     }
@@ -6031,6 +6197,9 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     }
     if (qy_needs_dequant) {
         d_Y = ctx->prealloc_y;
+    } else if (quantize_y) {
+        d_Y = ctx->prealloc_y;
+        GGML_ASSERT(d_Y->size >= CEIL_DIV(y_sz * ne12 * ne13, 144) * 144);
     } else {
         d_Y = d_Qy;
         y_buf_offset = qy_buf_offset;
@@ -6041,9 +6210,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         if (ctx->prealloc_x_need_sync) {
             ggml_vk_sync_buffers(ctx, subctx);
         }
-    }
 
-    if (x_non_contig) {
         GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
     }
@@ -6059,6 +6226,17 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
             ctx->prealloc_y_last_tensor_used = src1;
         }
     }
+    if (quantize_y) {
+        if (ctx->prealloc_y_last_pipeline_used != to_q8_1.get() ||
+            ctx->prealloc_y_last_tensor_used != src1) {
+            if (ctx->prealloc_y_need_sync) {
+                ggml_vk_sync_buffers(ctx, subctx);
+            }
+            ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13, true);
+            ctx->prealloc_y_last_pipeline_used = to_q8_1.get();
+            ctx->prealloc_y_last_tensor_used = src1;
+        }
+    }
 
     // For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
     uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01;
@@ -6083,6 +6261,12 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
+    // TODO: Clean up this whole sz * ne_2 * ne_3 thing, it hasn't been necessary for a long time
+    uint32_t y_sz_total = y_sz * ne12 * ne13;
+    if (quantize_y) {
+        y_sz_total = CEIL_DIV(y_sz_total, 144) * 144;
+    }
+
     // compute
     const vk_mat_vec_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
@@ -6090,13 +6274,13 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-                              { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
+                              { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz_total }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
                               pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
 
     if (x_non_contig) {
         ctx->prealloc_x_need_sync = true;
     }
-    if (y_non_contig) {
+    if (y_non_contig || quantize_y) {
         ctx->prealloc_y_need_sync = true;
     }
 }
@@ -11513,15 +11697,29 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
 
 void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
     GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+    GGML_ASSERT(device < (int) vk_instance.device_supports_membudget.size());
 
     vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
+    vk::PhysicalDeviceMemoryBudgetPropertiesEXT budgetprops;
+    vk::PhysicalDeviceMemoryProperties2 memprops = {};
+    bool membudget_supported = vk_instance.device_supports_membudget[device];
 
-    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+    if (membudget_supported) {
+        memprops.pNext = &budgetprops;
+    }
+    vkdev.getMemoryProperties2(&memprops);
+
+    for (uint32_t i = 0; i < memprops.memoryProperties.memoryHeapCount; ++i) {
+        const vk::MemoryHeap & heap = memprops.memoryProperties.memoryHeaps[i];
 
-    for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
         if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
             *total = heap.size;
-            *free = heap.size;
+
+            if (membudget_supported && i < budgetprops.heapUsage.size()) {
+                *free = budgetprops.heapBudget[i] - budgetprops.heapUsage[i];
+            } else {
+                *free = heap.size;
+            }
             break;
         }
     }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
index b93e9948f..f761391ea 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
@@ -1,7 +1,8 @@
 #extension GL_EXT_control_flow_attributes : enable
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
-#if USE_SUBGROUP_ADD
+
+#if USE_SUBGROUP_ADD || USE_SUBGROUP_ADD_NO_SHMEM
 #extension GL_KHR_shader_subgroup_basic : require
 #extension GL_KHR_shader_subgroup_arithmetic : require
 #endif
@@ -12,10 +13,19 @@
 
 #include "types.comp"
 
+#ifndef MMQ
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#else
+layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
+#endif
+
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+#ifdef B_TYPE_VEC2
 layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+#endif
+#ifdef B_TYPE_VEC4
 layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#endif
 
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
@@ -92,6 +102,23 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 layout (constant_id = 1) const uint NUM_ROWS = 1;
 layout (constant_id = 2) const uint NUM_COLS = 1;
 
+#ifdef USE_SUBGROUP_ADD_NO_SHMEM
+void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = subgroupAdd(temp[j][n]);
+        }
+    }
+
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
+            }
+        }
+    }
+}
+#else
 shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
 
 void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
@@ -152,3 +179,4 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
     }
 #endif
 }
+#endif
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp
new file mode 100644
index 000000000..8fb314fa0
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp
@@ -0,0 +1,140 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_integer_dot_product : require
+
+#define MMQ
+#define B_TYPE block_q8_1_x4
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+#define K_PER_ITER 8
+
+#include "mul_mmq_funcs.comp"
+
+uint a_offset, b_offset, d_offset;
+
+int32_t cache_b_qs[2];
+vec2 cache_b_ds;
+
+void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
+
+        // Preload data_b block
+        const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
+        const uint b_qs_idx = tid % 4;
+        const uint b_block_idx_outer = b_block_idx / 4;
+        const uint b_block_idx_inner = b_block_idx % 4;
+        cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
+
+#if QUANT_R == 2
+        cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx];
+        cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx + 4];
+#else
+        cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2];
+        cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2 + 1];
+#endif
+
+        uint ibi = first_row*p.ncols;
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint a_block_idx = (ibi + col)/QUANT_K + a_offset;
+            ibi += p.ncols;
+
+            int32_t q_sum = 0;
+#if QUANT_R == 2
+            const i32vec2 data_a_qs = repack(a_block_idx, b_qs_idx);
+            q_sum += dotPacked4x8EXT(data_a_qs.x,
+                                     cache_b_qs[0]);
+            q_sum += dotPacked4x8EXT(data_a_qs.y,
+                                     cache_b_qs[1]);
+#else
+            int32_t data_a_qs = repack(a_block_idx, b_qs_idx * 2);
+            q_sum += dotPacked4x8EXT(data_a_qs,
+                                     cache_b_qs[0]);
+            data_a_qs = repack(a_block_idx, b_qs_idx * 2 + 1);
+            q_sum += dotPacked4x8EXT(data_a_qs,
+                                     cache_b_qs[1]);
+#endif
+
+#if QUANT_AUXF == 1
+            temp[j][n] += mul_q8_1(q_sum,  get_d(a_block_idx), cache_b_ds, 4);
+#else
+            temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds, 4);
+#endif
+        }
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    const uint tid = gl_LocalInvocationID.x;
+
+    get_offsets(a_offset, b_offset, d_offset);
+    a_offset /= QUANT_K;
+    b_offset /= QUANT_K_Q8_1;
+
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = FLOAT_TYPE(0.0f);
+        }
+    }
+
+    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
+    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+        num_iters++;
+    }
+    int unroll_count = 4;
+    uint unrolled_iters = num_iters & ~(unroll_count - 1);
+
+    uint i = 0;
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
+            i++;
+        }
+    }
+
+    unroll_count = 2;
+    unrolled_iters = num_iters & ~(unroll_count - 1);
+
+#if K_PER_ITER == 2
+    if ((p.ncols & 1) != 0 &&
+        unrolled_iters == num_iters &&
+        unrolled_iters > 0) {
+        unrolled_iters -= unroll_count;
+    }
+#endif
+
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
+            i++;
+        }
+    }
+    while (i < num_iters) {
+        iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
+        i++;
+    }
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
index 654105a49..69ac38fd4 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
@@ -494,6 +494,9 @@ void main() {
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
             }
+#if defined(ACC_TYPE_MAX)
+            [[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
+#endif
 
             // Convert from ACC_TYPE to D_TYPE
             coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover4, gl_MatrixUseAccumulator> mat_d;
@@ -535,6 +538,9 @@ void main() {
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
             }
+#if defined(ACC_TYPE_MAX)
+            [[unroll]] for (uint i = 0; i < sum.length(); ++i) { sum[i] = clamp(sum[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
+#endif
 
             // Convert from ACC_TYPE to D_TYPE
             coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> mat_d;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
index 83de90eb7..f36add62a 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp
@@ -28,7 +28,7 @@ layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
 #if defined(A_TYPE_PACKED32)
 layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
 #endif
-layout (binding = 1) readonly buffer B {block_q8_1_packed32 data_b[];};
+layout (binding = 1) readonly buffer B {block_q8_1_x4_packed128 data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
@@ -98,7 +98,7 @@ shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
 #endif
 
 #define LOAD_VEC_A (4 * QUANT_R)
-#define LOAD_VEC_B 4
+#define LOAD_VEC_B 16
 
 #ifdef MUL_MAT_ID
 shared u16vec2 row_ids[4096];
@@ -270,15 +270,22 @@ void main() {
             const uint iqs = idx & 0x7;
 #else
             const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK;
+            const uint ib_outer = ib / 4;
+            const uint ib_inner = ib % 4;
+
             const uint iqs = loadr_b;
 #endif
 
             const uint buf_ib = loadc_b + l;
 
             if (iqs == 0) {
-                buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib].ds);
+                buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib_outer].ds[ib_inner]);
             }
-            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs] = data_b[ib].qs[iqs];
+            const ivec4 values = data_b[ib_outer].qs[ib_inner * 2 + iqs];
+            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4    ] = values.x;
+            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 1] = values.y;
+            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 2] = values.z;
+            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs * 4 + 3] = values.w;
         }
 
         barrier();
@@ -349,7 +356,7 @@ void main() {
                                                      cache_b_qs[cc * (BK / 4) + idx_k]);
                         }
 
-                        sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]);
+                        sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc], 1);
                     }
                 }
             }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
index 34e8db977..cdfb230f4 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp
@@ -16,8 +16,8 @@ i32vec2 repack(uint ib, uint iqs) {
                    (vui >> 4) & 0x0F0F0F0F);
 }
 
-ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y));
+ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return ACC_TYPE(da * (float(q_sum) * dsb.x - (8 / sum_divisor) * dsb.y));
 }
 #endif
 
@@ -29,8 +29,8 @@ i32vec2 repack(uint ib, uint iqs) {
                    (vui >> 4) & 0x0F0F0F0F);
 }
 
-ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
+ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
+    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
 }
 #endif
 
@@ -50,8 +50,8 @@ i32vec2 repack(uint ib, uint iqs) {
     return i32vec2(v0, v1);
 }
 
-ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
-    return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y));
+ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
+    return ACC_TYPE(da * (float(q_sum) * dsb.x - (16 / sum_divisor) * dsb.y));
 }
 #endif
 
@@ -69,8 +69,8 @@ i32vec2 repack(uint ib, uint iqs) {
     return i32vec2(v0, v1);
 }
 
-ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
-    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
+ACC_TYPE mul_q8_1(const int32_t q_sum, const vec2 dma, const vec2 dsb, const int32_t sum_divisor) {
+    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y / sum_divisor);
 }
 #endif
 
@@ -81,7 +81,7 @@ int32_t repack(uint ib, uint iqs) {
                           data_a[ib].qs[iqs * 2 + 1]));
 }
 
-ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
+ACC_TYPE mul_q8_1(const int32_t q_sum, const float da, const vec2 dsb, const int32_t sum_divisor) {
     return ACC_TYPE(float(q_sum) * da * dsb.x);
 }
 #endif
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
index e2e020fec..145c9fbdc 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp
@@ -3,6 +3,15 @@
 #extension GL_EXT_control_flow_attributes : require
 #extension GL_EXT_shader_16bit_storage : require
 
+#ifdef USE_SUBGROUPS
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_clustered : require
+
+#define INVOCATION_ID gl_SubgroupInvocationID.x
+#else
+#define INVOCATION_ID gl_LocalInvocationID.x
+#endif
+
 layout (push_constant) uniform parameter
 {
     uint ne;
@@ -14,13 +23,19 @@ layout(constant_id = 0) const uint GROUP_SIZE = 32;
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {vec4 data_a[];};
+#ifndef QBLOCK_X4
 layout (binding = 1) writeonly buffer D {block_q8_1_packed32 data_b[];};
+#else
+layout (binding = 1) writeonly buffer D {block_q8_1_x4 data_b[];};
+#endif
 
+#ifndef USE_SUBGROUPS
 shared float shmem[GROUP_SIZE];
+#endif
 
 void quantize() {
     const uint wgid = gl_WorkGroupID.x;
-    const uint tid = gl_LocalInvocationID.x;
+    const uint tid = INVOCATION_ID;
 
     // Each thread handles a vec4, so 8 threads handle a block
     const uint blocks_per_group = GROUP_SIZE / 8;
@@ -30,9 +45,19 @@ void quantize() {
     const uint ib = wgid * blocks_per_group + block_in_wg;
     const uint iqs = tid % 8;
 
+#ifndef QBLOCK_X4
     if (ib >= gl_NumWorkGroups.x * blocks_per_group) {
         return;
     }
+#else
+    const uint ibx4_outer = ib / 4;
+    const uint ibx4_inner = ib % 4;
+
+    const uint required_x4_blocks = (p.ne + 127) / 128;
+    if (ibx4_outer >= required_x4_blocks) {
+        return;
+    }
+#endif
 
     const uint a_idx = ib * 8 + iqs;
 
@@ -40,7 +65,9 @@ void quantize() {
     const vec4 abs_vals = abs(vals);
 
     // Find absolute max for each block
-    shmem[tid] = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
+    const float thread_max = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
+#ifndef USE_SUBGROUPS
+    shmem[tid] = thread_max;
     barrier();
     [[unroll]] for (uint s = 4; s > 0; s >>= 1) {
         if (iqs < s) {
@@ -50,14 +77,28 @@ void quantize() {
     }
 
     const float amax = shmem[block_in_wg * 8];
+#else
+    const float amax = subgroupClusteredMax(thread_max, 8);
+#endif
+
     const float d = amax / 127.0;
     const float d_inv = d != 0.0 ? 1.0 / d : 0.0;
     vals = round(vals * d_inv);
+
+#ifndef QBLOCK_X4
     data_b[ib].qs[iqs] = pack32(i8vec4(round(vals)));
+#else
+    data_b[ibx4_outer].qs[ibx4_inner * 8 + iqs] = pack32(i8vec4(round(vals)));
+#endif
+
+#ifndef USE_SUBGROUPS
     barrier();
+#endif
 
     // Calculate the sum for each block
-    shmem[tid] = vals.x + vals.y + vals.z + vals.w;
+    const float thread_sum = vals.x + vals.y + vals.z + vals.w;
+#ifndef USE_SUBGROUPS
+    shmem[tid] = thread_sum;
     barrier();
     [[unroll]] for (uint s = 4; s > 0; s >>= 1) {
         if (iqs < s) {
@@ -65,10 +106,19 @@ void quantize() {
         }
         barrier();
     }
+#else
+    const float sum = subgroupClusteredAdd(thread_sum, 8);
+#endif
     if (iqs == 0) {
+#ifndef USE_SUBGROUPS
         const float sum = shmem[tid];
+#endif
 
+#ifndef QBLOCK_X4
         data_b[ib].ds = f16vec2(vec2(d, sum * d));
+#else
+        data_b[ibx4_outer].ds[ibx4_inner] = f16vec2(vec2(d, sum * d));
+#endif
     }
 }
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
index a36c33e26..408722c87 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -207,6 +207,18 @@ struct block_q8_1_packed32
     int32_t qs[8];
 };
 
+// 4 blocks in one to allow 16-byte/128-bit alignment and loads
+struct block_q8_1_x4
+{
+    f16vec2 ds[4];
+    int32_t qs[32];
+};
+struct block_q8_1_x4_packed128
+{
+    f16vec2 ds[4];
+    ivec4 qs[8];
+};
+
 // K-quants
 #define QUANT_K_Q2_K 256
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index d7c8a83da..443c2066a 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -218,6 +218,22 @@ bool string_ends_with(const std::string& str, const std::string& suffix) {
     return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
 }
 
+bool is_quantized_type(const std::string& type_name) {
+    return type_name != "f32" && type_name != "f16" && type_name != "bf16";
+}
+
+bool is_legacy_quant(const std::string& type_name) {
+    return type_name == "q4_0" || type_name == "q4_1" || type_name == "q5_0" || type_name == "q5_1" || type_name == "q8_0";
+}
+
+bool is_k_quant(const std::string& type_name) {
+    return string_ends_with(type_name, "_k");
+}
+
+bool is_iq_quant(const std::string& type_name) {
+    return string_starts_with(type_name, "iq");
+}
+
 static const char path_separator = '/';
 
 std::string join_paths(const std::string& path1, const std::string& path2) {
@@ -416,7 +432,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-        if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
+        if (!coopmat && !coopmat2 && matmul_id_type == MatMulIdType::NONE && is_legacy_quant(tname)) {
             string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
@@ -509,8 +525,20 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
 
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
+
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
+        // mul mat vec with integer dot product
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+        if (is_legacy_quant(tname)) {
+            string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
+            string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
+            string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
+        }
+#endif
+
         // Dequant shaders
         if (tname != "f16" && tname != "bf16") {
             string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
@@ -593,7 +621,12 @@ void process_shaders() {
 
     string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
     string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {});
+
     string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {});
+    string_to_spv("quantize_q8_1_subgroup", "quantize_q8_1.comp", {{"USE_SUBGROUPS", "1"}});
+
+    string_to_spv("quantize_q8_1_x4", "quantize_q8_1.comp", {{"QBLOCK_X4", "1"}});
+    string_to_spv("quantize_q8_1_x4_subgroup", "quantize_q8_1.comp", {{"QBLOCK_X4", "1"}, {"USE_SUBGROUPS", "1"}});
 
     string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
@@ -835,12 +868,21 @@ void write_output_files() {
         fputs(len.c_str(), src);
     }
 
-    for (const std::string& btype : {"f16", "f32"}) {
+    std::vector<std::string> btypes = {"f16", "f32"};
+
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+    btypes.push_back("q8_1");
+#endif
+
+    for (const std::string& btype : btypes) {
     for (const auto& tname : type_names) {
-        fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[2];\n", tname.c_str(), btype.c_str());
-        fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[2];\n", tname.c_str(), btype.c_str());
-        std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[2] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data};\n";
-        std::string len =  "uint64_t arr_dmmv_"       + tname + "_" + btype + "_f32_len[2] =  {mul_mat_vec_" + tname + "_" + btype + "_f32_len,  mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len};\n";
+        if (btype == "q8_1" && !is_legacy_quant(tname)) {
+            continue;
+        }
+        fprintf(hdr, "extern unsigned char *arr_dmmv_%s_%s_f32_data[3];\n", tname.c_str(), btype.c_str());
+        fprintf(hdr, "extern uint64_t arr_dmmv_%s_%s_f32_len[3];\n", tname.c_str(), btype.c_str());
+        std::string data = "unsigned char *arr_dmmv_" + tname + "_" + btype + "_f32_data[3] = {mul_mat_vec_" + tname + "_" + btype + "_f32_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_data, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_no_shmem_data};\n";
+        std::string len =  "uint64_t arr_dmmv_"       + tname + "_" + btype + "_f32_len[3] =  {mul_mat_vec_" + tname + "_" + btype + "_f32_len,  mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_len, mul_mat_vec_" + tname + "_" + btype + "_f32_subgroup_no_shmem_len};\n";
         fputs(data.c_str(), src);
         fputs(len.c_str(), src);
     }
diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp
index 6f7d89292..c5e98789a 100644
--- a/ggml/src/gguf.cpp
+++ b/ggml/src/gguf.cpp
@@ -274,7 +274,7 @@ struct gguf_reader {
     }
 
     bool read(std::string & dst) const {
-        uint64_t size = -1;
+        uint64_t size = 0;
         if(isggufv1)
         {
             uint32_t size_32 = -1;
@@ -560,7 +560,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
         // tensor shape
         {
-            uint32_t n_dims = -1;
+            uint32_t n_dims = 0;
             ok = ok && gr.read(n_dims);
             if (n_dims > GGML_MAX_DIMS) {
                 GGML_LOG_ERROR("%s: tensor '%s' has invalid number of dimensions: %" PRIu32 " > %" PRIu32 "\n",
diff --git a/include/llama.h b/include/llama.h
index 35c687070..7b0811c6c 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -209,7 +209,7 @@ extern "C" {
         llama_token_data * data;
         size_t size;
         int64_t selected; // this is the index in the data array (i.e. not the token id)
-        bool sorted;
+        bool sorted;      // note: do not assume the data is sorted - always check this flag
     } llama_token_data_array;
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
@@ -1159,11 +1159,6 @@ extern "C" {
     LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
     LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
 
-    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
-    /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
-    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
-        "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
-
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 1dda251ec..bc42e7a21 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -270,19 +270,7 @@ llama_context::llama_context(
         }
     }
 
-    // resolve automatic Flash Attention use and reserve worst-case graph
     if (!hparams.vocab_only) {
-        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
-        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-
-        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
-
-        int n_splits_pp = -1;
-        int n_nodes_pp  = -1;
-
-        int n_splits_tg = -1;
-        int n_nodes_tg  = -1;
-
         llama_memory_context_ptr mctx;
         if (memory) {
             LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__);
@@ -294,6 +282,60 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+
+        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
+
+        // resolve automatic Flash Attention use
+        if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
+            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
+            if (!gf) {
+                throw std::runtime_error("failed to split graph for Flash Attention check");
+            }
+
+            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
+            bool fa_device_mismatch = false;
+            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+                ggml_tensor * n = ggml_graph_node(gf, i);
+                if (n->op != GGML_OP_FLASH_ATTN_EXT) {
+                    continue;
+                }
+                ggml_backend_dev_t device_fa = ggml_backend_get_device(
+                    ggml_backend_sched_get_tensor_backend(sched.get(), n));
+
+                // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
+                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
+                const int il = std::stoi(n->name + prefix_len);
+                ggml_backend_dev_t device_kv = model.dev_layer(il);
+                if (device_fa != device_kv) {
+                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention tensor "
+                        "is assigned to device %s (usually due to missing support)\n",
+                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_fa));
+                    // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
+                    fa_device_mismatch = true;
+                    break;
+                }
+            }
+            if (fa_device_mismatch) {
+                cparams.flash_attn = false;
+                LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
+                if (ggml_is_quantized(params.type_v)) {
+                    throw std::runtime_error("quantized V cache was requested, but this requires Flash Attention");
+                }
+            } else {
+                cparams.flash_attn = true;
+                LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
+            }
+        }
+
+        // reserve worst-case graph
+        int n_splits_pp = -1;
+        int n_nodes_pp  = -1;
+
+        int n_splits_tg = -1;
+        int n_nodes_tg  = -1;
+
         // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -301,48 +343,6 @@ llama_context::llama_context(
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
 
-            if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO) {
-                ggml_backend_sched_alloc_graph(sched.get(), gf);
-
-                const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FATTN) + 1;
-                bool fa_device_mismatch = false;
-                for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-                    ggml_tensor * n = ggml_graph_node(gf, i);
-                    if (n->op != GGML_OP_FLASH_ATTN_EXT) {
-                        continue;
-                    }
-                    ggml_backend_dev_t device_fa = ggml_backend_get_device(
-                        ggml_backend_sched_get_tensor_backend(sched.get(), n));
-
-                    // TODO: instead of the tensor names, use a map to keep track of which (FA) tensors belong to which layer
-                    GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FATTN "-", prefix_len) == 0);
-                    const int il = std::stoi(n->name + prefix_len);
-                    ggml_backend_dev_t device_kv = model.dev_layer(il);
-                    if (device_fa != device_kv) {
-                        LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the Flash Attention tensor "
-                            "is assigned to device %s (usually due to missing support)\n",
-                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_fa));
-                        // FIXME: fa_device_mismatch logic is wrong for --no-kv-offload, but this is broken anyways
-                        fa_device_mismatch = true;
-                        break;
-                    }
-                }
-                if (fa_device_mismatch) {
-                    cparams.flash_attn = false;
-                    LLAMA_LOG_WARN("%s: Flash Attention was auto, set to disabled\n", __func__);
-                    if (ggml_is_quantized(params.type_v)) {
-                        throw std::runtime_error("quantized V cache was requested, but this requires Flash Attention");
-                    }
-                    auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
-                    if (!gf) {
-                        throw std::runtime_error("failed to allocate compute pp buffers");
-                    }
-                } else {
-                    cparams.flash_attn = true;
-                    LLAMA_LOG_INFO("%s: Flash Attention was auto, set to enabled\n", __func__);
-                }
-            }
-
             n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
             n_nodes_pp  = ggml_graph_n_nodes(gf);
         }
@@ -1366,8 +1366,8 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
-   // LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
+ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
+    // LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
 
     if (n_tokens % n_seqs != 0) {
         n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
@@ -1401,7 +1401,9 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     this->n_outputs = save_n_outputs;
 
     // initialize scheduler with the specified graph
-    if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+    if (split_only) {
+        ggml_backend_sched_split_graph(sched.get(), gf);
+    } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
diff --git a/src/llama-context.h b/src/llama-context.h
index a372bcfbe..f23aa8ee1 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -196,7 +196,7 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
 
 private:
     llm_graph_params graph_params(
diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
index bfbf5fa23..e8c0fc341 100644
--- a/src/llama-sampling.cpp
+++ b/src/llama-sampling.cpp
@@ -128,6 +128,89 @@ struct ring_buffer {
     std::vector<T> data;
 };
 
+// writes result in res, does not mutate cur
+static void llama_token_data_array_partial_sort(const llama_token_data_array & cur, int npartial, std::vector<llama_token_data> & res) {
+    static const auto comp = [](const llama_token_data & a, const llama_token_data & b) {
+        return a.logit > b.logit;
+    };
+
+    constexpr int   nbuckets     = 128;
+    constexpr float bucket_low   = -10.0f;
+    constexpr float bucket_high  =  10.0f;
+    constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
+    constexpr float bucket_inter = -bucket_low * bucket_scale;
+
+    std::vector<int> bucket_idx;
+    std::vector<int> histo(nbuckets, 0);
+
+    std::vector<llama_token_data*> bucket_ptrs;
+
+    bucket_idx.reserve(cur.size);
+
+    for (int i = 0; i < (int)cur.size; ++i) {
+        const float val = cur.data[i].logit;
+        int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
+        ib = std::max(0, std::min(nbuckets - 1, ib));
+        bucket_idx.push_back(ib);
+        ++histo[ib];
+    }
+    int nhave = 0;
+    int ib = nbuckets - 1;
+    for ( ; ib >= 0; --ib) {
+        nhave += histo[ib];
+        if (nhave >= npartial) {
+            break;
+        }
+    }
+    res.resize(nhave);
+    auto * ptr = res.data();
+    bucket_ptrs.reserve(nbuckets - ib);
+    for (int j = nbuckets - 1; j >= ib; --j) {
+        bucket_ptrs.push_back(ptr);
+        ptr += histo[j];
+    }
+    for (int i = 0; i < (int)cur.size; ++i) {
+        int j = bucket_idx[i];
+        if (j >= ib) {
+            *bucket_ptrs[nbuckets - 1 - j]++ = cur.data[i];
+        }
+    }
+
+    ptr = res.data();
+    int ndone = 0;
+    for (int j = nbuckets - 1; j > ib; --j) {
+        std::sort(ptr, ptr + histo[j], comp);
+        ptr += histo[j];
+        ndone += histo[j];
+    }
+    std::partial_sort(ptr, ptr + npartial - ndone, ptr + histo[ib], comp);
+}
+
+// reduces the size of cur_p to npartial, keeping only the top npartial elements
+static void llama_token_data_array_partial_sort_inplace(llama_token_data_array * cur_p, int npartial) {
+    static const auto comp = [](const llama_token_data & a, const llama_token_data & b) {
+        return a.logit > b.logit;
+    };
+
+    if (npartial <= 128) {
+        std::partial_sort(cur_p->data, cur_p->data + npartial, cur_p->data + cur_p->size, comp);
+
+        cur_p->size = npartial;
+        cur_p->sorted = true;
+
+        return;
+    }
+
+    std::vector<llama_token_data> tmp;
+
+    llama_token_data_array_partial_sort(*cur_p, npartial, tmp);
+
+    std::copy(tmp.data(), tmp.data() + npartial, cur_p->data);
+
+    cur_p->size = npartial;
+    cur_p->sorted = true;
+}
+
 static int llama_sample_dist(llama_token_data_array * cur_p, std::mt19937 & rng) {
     // iterator for the probabilities
 #ifdef __GNUC__
@@ -200,18 +283,21 @@ static void llama_sampler_temp_impl(llama_token_data_array * cur_p, float temp)
     }
 }
 
-static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
+static void llama_sampler_softmax_impl(llama_token_data_array * cur_p, bool do_sort) {
     GGML_ASSERT(cur_p->size > 0);
 
-    // Sort the logits in descending order
-    if (!cur_p->sorted) {
-        std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
-            return a.logit > b.logit;
-        });
-        cur_p->sorted = true;
+    // Sort the logits in descending order if requested
+    if (do_sort && !cur_p->sorted) {
+        llama_token_data_array_partial_sort_inplace(cur_p, cur_p->size);
     }
 
     float max_l = cur_p->data[0].logit;
+    if (!cur_p->sorted) {
+        for (size_t i = 1; i < cur_p->size; ++i) {
+            max_l = std::max(max_l, cur_p->data[i].logit);
+        }
+    }
+
     float cum_sum = 0.0f;
 
     for (size_t i = 0; i < cur_p->size; ++i) {
@@ -226,7 +312,6 @@ static void llama_sampler_softmax_impl(llama_token_data_array * cur_p) {
 }
 
 static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) {
-    // TODO: move bucket sort to separate function so that top_p/typical/softmax first is equally fast
     // if (k >= (int32_t)cur_p->size) {
     //     return;
     // }
@@ -239,64 +324,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k)
 
     // Sort scores in descending order
     if (!cur_p->sorted) {
-        auto comp = [](const llama_token_data & a, const llama_token_data & b) {
-            return a.logit > b.logit;
-        };
-        if (k <= 128) {
-            std::partial_sort(cur_p->data, cur_p->data + k, cur_p->data + cur_p->size, comp);
-        } else {
-            constexpr int   nbuckets     = 128;
-            constexpr float bucket_low   = -10.0f;
-            constexpr float bucket_high  =  10.0f;
-            constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low);
-            constexpr float bucket_inter = -bucket_low * bucket_scale;
-
-            std::vector<int> bucket_idx(cur_p->size);
-            std::vector<int> histo(nbuckets, 0);
-
-            for (int i = 0; i < (int)cur_p->size; ++i) {
-                const float val = cur_p->data[i].logit;
-                int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low);
-                ib = std::max(0, std::min(nbuckets - 1, ib));
-                bucket_idx[i] = ib;
-                ++histo[ib];
-            }
-            int nhave = 0;
-            int ib = nbuckets - 1;
-            for ( ; ib >= 0; --ib) {
-                nhave += histo[ib];
-                if (nhave >= k) {
-                    break;
-                }
-            }
-            std::vector<llama_token_data> tmp_tokens(nhave);
-            auto * ptr = tmp_tokens.data();
-            std::vector<llama_token_data*> bucket_ptrs;
-            bucket_ptrs.reserve(nbuckets - ib);
-            for (int j = nbuckets - 1; j >= ib; --j) {
-                bucket_ptrs.push_back(ptr);
-                ptr += histo[j];
-            }
-            for (int i = 0; i < (int)cur_p->size; ++i) {
-                int j = bucket_idx[i];
-                if (j >= ib) {
-                    *bucket_ptrs[nbuckets - 1 - j]++ = cur_p->data[i];
-                }
-            }
-
-            ptr = tmp_tokens.data();
-            int ndone = 0;
-            for (int j = nbuckets - 1; j > ib; --j) {
-                std::sort(ptr, ptr + histo[j], comp);
-                ptr += histo[j];
-                ndone += histo[j];
-            }
-            std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp);
-
-            std::memcpy(cur_p->data, tmp_tokens.data(), k*sizeof(llama_token_data));
-
-        }
-        cur_p->sorted = true;
+        llama_token_data_array_partial_sort_inplace(cur_p, k);
     }
 
     cur_p->size = k;
@@ -576,7 +604,8 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
 static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_dist *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    // sorting is not necessary here
+    llama_sampler_softmax_impl(cur_p, false);
 
     cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
 }
@@ -626,32 +655,6 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
     );
 }
 
-// softmax
-
-static const char * llama_sampler_softmax_name(const struct llama_sampler * /*smpl*/) {
-    return "softmax";
-}
-
-static void llama_sampler_softmax_apply(struct llama_sampler * /*smpl*/, llama_token_data_array * cur_p) {
-    llama_sampler_softmax_impl(cur_p);
-}
-
-static struct llama_sampler_i llama_sampler_softmax_i = {
-    /* .name   = */ llama_sampler_softmax_name,
-    /* .accept = */ nullptr,
-    /* .apply  = */ llama_sampler_softmax_apply,
-    /* .reset  = */ nullptr,
-    /* .clone  = */ nullptr,
-    /* .free   = */ nullptr,
-};
-
-struct llama_sampler * llama_sampler_init_softmax() {
-    return llama_sampler_init(
-        /* .iface = */ &llama_sampler_softmax_i,
-        /* .ctx   = */ nullptr
-    );
-}
-
 // top-k
 
 struct llama_sampler_top_k {
@@ -663,7 +666,7 @@ static const char * llama_sampler_top_k_name(const struct llama_sampler * /*smpl
 }
 
 static void llama_sampler_top_k_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_top_k *) smpl->ctx;
+    auto * ctx = (llama_sampler_top_k *) smpl->ctx;
     llama_sampler_top_k_impl(cur_p, ctx->k);
 }
 
@@ -699,6 +702,8 @@ struct llama_sampler * llama_sampler_init_top_k(int32_t k) {
 struct llama_sampler_top_p {
     const float  p;
     const size_t min_keep;
+
+    std::vector<llama_token_data> buf_sort;
 };
 
 static const char * llama_sampler_top_p_name(const struct llama_sampler * /*smpl*/) {
@@ -706,20 +711,35 @@ static const char * llama_sampler_top_p_name(const struct llama_sampler * /*smpl
 }
 
 static void llama_sampler_top_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_top_p *) smpl->ctx;
+    auto * ctx = (llama_sampler_top_p *) smpl->ctx;
 
     if (ctx->p >= 1.0f) {
         return;
     }
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, false);
+
+    size_t k = cur_p->size;
+    auto * pdata = cur_p->data;
+
+    auto & buf_sort = ctx->buf_sort;
+
+    // if not sorted, try adaptive top-k sorting
+    if (!cur_p->sorted && cur_p->size > 1024) {
+        k = std::min<size_t>(256, cur_p->size);
+        llama_token_data_array_partial_sort(*cur_p, k, buf_sort);
+        pdata = buf_sort.data();
+    } else if (!cur_p->sorted) {
+        // small candidates -> sort inplace
+        llama_token_data_array_partial_sort_inplace(cur_p, k);
+    }
 
     // Compute the cumulative probabilities
     float cum_sum = 0.0f;
     size_t last_idx = cur_p->size;
 
     for (size_t i = 0; i < cur_p->size; ++i) {
-        cum_sum += cur_p->data[i].p;
+        cum_sum += pdata[i].p;
 
         // Check if the running sum is at least p or if we have kept at least min_keep tokens
         // we set the last index to i+1 to indicate that the current iterate should be included in the set
@@ -727,9 +747,21 @@ static void llama_sampler_top_p_apply(struct llama_sampler * smpl, llama_token_d
             last_idx = i + 1;
             break;
         }
+
+        // we exceeded the current top-k heuristic -> increase k and continue
+        if (!cur_p->sorted && i == k - 1) {
+            k = cur_p->size;
+            llama_token_data_array_partial_sort(*cur_p, k, buf_sort);
+            pdata = buf_sort.data();
+        }
     }
 
     // Resize the output vector to keep only the top-p tokens
+    if (!cur_p->sorted) {
+        std::copy(buf_sort.data(), buf_sort.data() + last_idx, cur_p->data);
+        cur_p->sorted = true;
+    }
+
     cur_p->size = last_idx;
 }
 
@@ -757,6 +789,7 @@ struct llama_sampler * llama_sampler_init_top_p(float p, size_t min_keep) {
         /* .ctx   = */ new llama_sampler_top_p {
             /* .p        = */ p,
             /* .min_keep = */ min_keep,
+            /* .buf_sort = */ {},
         }
     );
 }
@@ -773,7 +806,7 @@ static const char * llama_sampler_min_p_name(const struct llama_sampler * /*smpl
 }
 
 static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_min_p *) smpl->ctx;
+    auto * ctx = (llama_sampler_min_p *) smpl->ctx;
 
     if (ctx->p <= 0.0f || !cur_p->size) {
         return;
@@ -799,7 +832,7 @@ static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_d
 
         // if we have enough values the operation was a success
         if (!filtered_tokens.empty() && filtered_tokens.size() >= ctx->min_keep) {
-            memcpy(cur_p->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data));
+            std::copy(filtered_tokens.begin(), filtered_tokens.end(), cur_p->data);
             cur_p->size = filtered_tokens.size();
             min_p_applied = true;
         }
@@ -809,10 +842,7 @@ static void llama_sampler_min_p_apply(struct llama_sampler * smpl, llama_token_d
     if (!min_p_applied) {
         // Sort the logits in descending order
         if (!cur_p->sorted) {
-            std::sort(cur_p->data, cur_p->data + cur_p->size, [](const llama_token_data & a, const llama_token_data & b) {
-                return a.logit > b.logit;
-            });
-            cur_p->sorted = true;
+            llama_token_data_array_partial_sort_inplace(cur_p, cur_p->size);
         }
 
         const float min_logit = cur_p->data[0].logit + logf(ctx->p); // min logit for p_i >= p * p_max
@@ -869,7 +899,7 @@ static const char * llama_sampler_typical_name(const struct llama_sampler * /*sm
 }
 
 static void llama_sampler_typical_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_typical *) smpl->ctx;
+    auto * ctx = (llama_sampler_typical *) smpl->ctx;
 
     // Reference implementation:
     // https://github.com/huggingface/transformers/compare/main...cimeister:typical-sampling:typical-pr
@@ -878,7 +908,7 @@ static void llama_sampler_typical_apply(struct llama_sampler * smpl, llama_token
     }
 
     // Compute the softmax of logits and calculate entropy
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     float entropy = 0.0f;
     for (size_t i = 0; i < cur_p->size; ++i) {
@@ -1012,7 +1042,7 @@ static const char * llama_sampler_temp_ext_name(const struct llama_sampler * /*s
 }
 
 static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_temp_ext *) smpl->ctx;
+    auto * ctx = (llama_sampler_temp_ext *) smpl->ctx;
     if (ctx->delta > 0) {
         const float min_temp = std::max(0.0f, ctx->temp - ctx->delta);
         const float max_temp = ctx->temp + ctx->delta;
@@ -1027,7 +1057,7 @@ static void llama_sampler_temp_ext_apply(struct llama_sampler * smpl, llama_toke
         // Calculate maximum possible entropy
         float max_entropy = -logf(1.0f / cur_p->size);
 
-        llama_sampler_softmax_impl(cur_p);
+        llama_sampler_softmax_impl(cur_p, true);
 
         // Calculate entropy of the softmax probabilities
         float entropy = 0.0f;
@@ -1121,7 +1151,7 @@ struct llama_sampler_xtc {
     const uint32_t seed;
     uint32_t       seed_cur;
 
-    std::mt19937   rng;
+    std::mt19937    rng;
 };
 
 static const char * llama_sampler_xtc_name(const struct llama_sampler * /*smpl*/) {
@@ -1139,17 +1169,20 @@ static void llama_sample_xtc_apply(struct llama_sampler * smpl, llama_token_data
 
     std::uniform_real_distribution<float> distribution(0.0f, 1.0f);
     float chance = distribution(ctx->rng);
-    if (chance > ctx->probability) return;
+    if (chance > ctx->probability) {
+        return;
+    }
 
-    // in case it's not sorted/recalculated yet
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     int pos_last = 0;
 
     for (size_t i = 0; i < cur_p->size; ++i) {
         if (cur_p->data[i].p >= ctx->threshold) {
             pos_last = i;
-        } else break;
+        } else {
+            break;
+        }
     }
 
     if (cur_p->size - pos_last >= ctx->min_keep && pos_last > 0) {
@@ -1221,7 +1254,7 @@ struct llama_sampler_mirostat {
 
     float mu;
 
-    std::mt19937 rng;
+    std::mt19937    rng;
 };
 
 static const char * llama_sampler_mirostat_name(const struct llama_sampler * /*smpl*/) {
@@ -1231,7 +1264,7 @@ static const char * llama_sampler_mirostat_name(const struct llama_sampler * /*s
 static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_mirostat *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     // Estimate s_hat using the most probable m tokens
     float s_hat = 0.0;
@@ -1250,7 +1283,8 @@ static void llama_sampler_mirostat_apply(struct llama_sampler * smpl, llama_toke
     float k = powf((epsilon_hat * powf(2, ctx->mu)) / (1 - powf(ctx->n_vocab, -epsilon_hat)), 1 / s_hat);
 
     llama_sampler_top_k_impl(cur_p, std::max(int(k), 1));
-    llama_sampler_softmax_impl(cur_p);
+
+    llama_sampler_softmax_impl(cur_p, true);
 
     const int idx = llama_sample_dist(cur_p, ctx->rng);
 
@@ -1336,7 +1370,7 @@ static const char * llama_sampler_mirostat_v2_name(const struct llama_sampler *
 static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_mirostat_v2 *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     // Truncate the words with surprise values greater than mu
     cur_p->size = std::distance(cur_p->data, std::find_if(cur_p->data, cur_p->data + cur_p->size, [&](const llama_token_data & candidate) {
@@ -1348,7 +1382,7 @@ static void llama_sampler_mirostat_v2_apply(struct llama_sampler * smpl, llama_t
     }
 
     // Normalize the probabilities of the remaining words
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
     const int idx = llama_sample_dist(cur_p, ctx->rng);
 
@@ -1540,7 +1574,7 @@ static struct llama_sampler * llama_sampler_init_grammar_impl(
                 trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0");
             }
             trigger_pattern += ")[\\s\\S]*";
-            auto trigger_pattern_c = trigger_pattern.c_str();
+            const auto * trigger_pattern_c = trigger_pattern.c_str();
             trigger_patterns = &trigger_pattern_c;
             num_trigger_patterns = 1;
         }
@@ -1748,7 +1782,7 @@ static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler *
 }
 
 static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
+    auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx;
 
     if (ctx->n <= 0.0f || cur_p->size <= 1) {
         return;
@@ -1780,13 +1814,14 @@ static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_t
     }
     float std = valid_count > 0 ? sqrt(acc/valid_count) : 0;
 
-    //apply mask
+    // apply mask
     for (size_t i = 0; i < cur_p->size; ++i) {
         if (cur_p->data[i].logit < max - (ctx->n * std)) {
             cur_p->data[i].logit = -INFINITY;
         }
     }
-    llama_sampler_softmax_impl(cur_p);
+
+    llama_sampler_softmax_impl(cur_p, true);
 }
 
 static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) {
@@ -1991,7 +2026,9 @@ static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_dat
 
     {
         const int last = last_n_repeat - 1;
-        int rt = 0, lt = 0;
+
+        int rt = 0;
+        int lt = 0;
 
         for (int k = 1; k < last_n_repeat; ++k) {
             if (k > rt) {
@@ -2135,8 +2172,8 @@ static struct llama_sampler_i llama_sampler_dry_i = {
     /* .free   = */ llama_sampler_dry_free,
 };
 
-struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab, int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
-    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? context_size : std::max(dry_penalty_last_n, 0);
+struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab, int32_t n_ctx_train, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? n_ctx_train : std::max(dry_penalty_last_n, 0);
     std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
     const int MAX_CHAR_LEN = 40;
     const int MAX_SEQ_LEN = 20;
@@ -2169,7 +2206,7 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_dry_i,
         /* .ctx   = */ new llama_sampler_dry {
-            /* .total_context_size     = */ context_size,
+            /* .total_context_size     = */ n_ctx_train,
             /* .dry_multiplier         = */ dry_multiplier,
             /* .dry_base               = */ dry_base,
             /* .dry_allowed_length     = */ dry_allowed_length,
@@ -2308,7 +2345,7 @@ static const char * llama_sampler_infill_name(const struct llama_sampler * /*smp
 static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_infill *) smpl->ctx;
 
-    llama_sampler_softmax_impl(cur_p);
+    llama_sampler_softmax_impl(cur_p, true);
 
 #if defined(GGML_DEBUG_SAMPLER_INFILL)
 #define LOG_DBG_CUR LLAMA_LOG_DEBUG
diff --git a/tools/server/server.cpp b/tools/server/server.cpp
index 6aa319d2f..e0302e2f2 100644
--- a/tools/server/server.cpp
+++ b/tools/server/server.cpp
@@ -141,7 +141,7 @@ struct slot_params {
     // Embeddings
     int32_t embd_normalize = 2; // (-1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm)
 
-    json to_json() const {
+    json to_json(bool only_metrics = false) const {
         std::vector<std::string> samplers;
         samplers.reserve(sampling.samplers.size());
         for (const auto & sampler : sampling.samplers) {
@@ -153,9 +153,55 @@ struct slot_params {
             lora.push_back({{"id", i}, {"scale", this->lora[i].scale}});
         }
 
+        if (only_metrics) {
+            return json {
+                {"n_predict",                 n_predict},     // Server configured n_predict
+                {"seed",                      sampling.seed},
+                {"temperature",               sampling.temp},
+                {"dynatemp_range",            sampling.dynatemp_range},
+                {"dynatemp_exponent",         sampling.dynatemp_exponent},
+                {"top_k",                     sampling.top_k},
+                {"top_p",                     sampling.top_p},
+                {"min_p",                     sampling.min_p},
+                {"top_n_sigma",               sampling.top_n_sigma},
+                {"xtc_probability",           sampling.xtc_probability},
+                {"xtc_threshold",             sampling.xtc_threshold},
+                {"typical_p",                 sampling.typ_p},
+                {"repeat_last_n",             sampling.penalty_last_n},
+                {"repeat_penalty",            sampling.penalty_repeat},
+                {"presence_penalty",          sampling.penalty_present},
+                {"frequency_penalty",         sampling.penalty_freq},
+                {"dry_multiplier",            sampling.dry_multiplier},
+                {"dry_base",                  sampling.dry_base},
+                {"dry_allowed_length",        sampling.dry_allowed_length},
+                {"dry_penalty_last_n",        sampling.dry_penalty_last_n},
+                {"mirostat",                  sampling.mirostat},
+                {"mirostat_tau",              sampling.mirostat_tau},
+                {"mirostat_eta",              sampling.mirostat_eta},
+                {"max_tokens",                n_predict}, // User configured n_predict
+                {"n_keep",                    n_keep},
+                {"n_discard",                 n_discard},
+                {"ignore_eos",                sampling.ignore_eos},
+                {"stream",                    stream},
+                {"n_probs",                   sampling.n_probs},
+                {"min_keep",                  sampling.min_keep},
+                {"chat_format",               common_chat_format_name(oaicompat_chat_syntax.format)},
+                {"reasoning_format",          common_reasoning_format_name(oaicompat_chat_syntax.reasoning_format)},
+                {"reasoning_in_content",      oaicompat_chat_syntax.reasoning_in_content},
+                {"thinking_forced_open",      oaicompat_chat_syntax.thinking_forced_open},
+                {"samplers",                  samplers},
+                {"speculative.n_max",         speculative.n_max},
+                {"speculative.n_min",         speculative.n_min},
+                {"speculative.p_min",         speculative.p_min},
+                {"timings_per_token",         timings_per_token},
+                {"post_sampling_probs",       post_sampling_probs},
+                {"lora",                      lora},
+            };
+        }
+
         auto grammar_triggers = json::array();
         for (const auto & trigger : sampling.grammar_triggers) {
-            server_grammar_trigger ct(std::move(trigger));
+            server_grammar_trigger ct(trigger);
             grammar_triggers.push_back(ct.to_json());
         }
 
@@ -1572,7 +1618,26 @@ struct server_slot {
         }
     }
 
-    json to_json() const {
+    json to_json(bool only_metrics = false) const {
+        if (only_metrics) {
+            return json {
+                {"id",            id},
+                {"id_task",       id_task},
+                {"n_ctx",         n_ctx},
+                {"speculative",   can_speculate()},
+                {"is_processing", is_processing()},
+                {"params",        params.to_json(true)},
+                {"next_token",
+                    {
+                        {"has_next_token", has_next_token},
+                        {"has_new_line",   has_new_line},
+                        {"n_remain",       n_remaining},
+                        {"n_decoded",      n_decoded},
+                    }
+                },
+            };
+        }
+
         return json {
             {"id",            id},
             {"id_task",       id_task},
@@ -2485,11 +2550,12 @@ struct server_context {
         return slot.has_next_token; // continue
     }
 
-    void populate_token_probs(const server_slot & slot, completion_token_output & result, bool post_sampling, bool special, int idx) {
+    void populate_token_probs(const server_slot & slot, completion_token_output & result, bool post_sampling, bool special, int idx) const {
         size_t n_probs = slot.params.sampling.n_probs;
         size_t n_vocab = llama_vocab_n_tokens(vocab);
+
         if (post_sampling) {
-            const auto * cur_p = common_sampler_get_candidates(slot.smpl);
+            const auto * cur_p = common_sampler_get_candidates(slot.smpl, true);
             const size_t max_probs = cur_p->size;
 
             // set probability for sampled token
@@ -2874,7 +2940,7 @@ struct server_context {
                     int n_processing_slots = 0;
 
                     for (server_slot & slot : slots) {
-                        json slot_data = slot.to_json();
+                        json slot_data = slot.to_json(true);
 
                         if (slot.is_processing()) {
                             n_processing_slots++;
@@ -4271,16 +4337,20 @@ int main(int argc, char ** argv) {
         }
     };
 
-    const auto handle_props = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+    const auto handle_props = [&params, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
         // this endpoint is publicly available, please only return what is safe to be exposed
         json data = {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
             { "model_path",                  ctx_server.params_base.model.path },
-            { "modalities",                  json{
+            { "modalities",                  json {
                 {"vision", ctx_server.oai_parser_opt.allow_image},
                 {"audio",  ctx_server.oai_parser_opt.allow_audio},
             } },
+            { "endpoint_slots",              params.endpoint_slots },
+            { "endpoint_props",              params.endpoint_props },
+            { "endpoint_metrics",            params.endpoint_metrics },
+            { "webui",                       params.webui },
             { "chat_template",               common_chat_templates_source(ctx_server.chat_templates.get()) },
             { "bos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
             { "eos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py
index 82f7215d5..cda7434d7 100644
--- a/tools/server/tests/utils.py
+++ b/tools/server/tests/utils.py
@@ -148,6 +148,8 @@ class ServerProcess:
             server_args.append("--metrics")
         if self.server_slots:
             server_args.append("--slots")
+        else:
+            server_args.append("--no-slots")
         if self.pooling:
             server_args.extend(["--pooling", self.pooling])
         if self.model_alias:
diff --git a/tools/tts/tts.cpp b/tools/tts/tts.cpp
index 18f01a994..eaf56591d 100644
--- a/tools/tts/tts.cpp
+++ b/tools/tts/tts.cpp
@@ -895,7 +895,7 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14
 
                 codes.push_back(new_token_id);
 
-                const auto * cands = common_sampler_get_candidates(smpl[i]);
+                const auto * cands = common_sampler_get_candidates(smpl[i], false);
 
                 // is it an end of generation? -> mark the stream as finished
                 if (llama_vocab_is_eog(vocab, new_token_id) || n_decode == n_predict) {