CANN: Add RoPE contiguous check for 310I DUP device (#15735)

* ggml-cpu : optimize rvv ggml_vec_dot_f32

* ggml-cpu : optimize 128-bit rvv ggml_vec_dot_q4_K_q8_K

* ggml-cpu : fix riscv arch flags

* ggml-cpu : add more rvv ops

* ggml-cpu : optimize rvv ggml_vec_dot_q4_K_q8_K

* ggml-cpu : optimize rvv ggml_vec_dot_q6_K_q8_K

* ggml-cpu : minor rvv adjustments

* ggml-cpu : fix riscv include

.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

README.md

@@ -137,6 +137,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
+- [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
 
 #### Multimodal
 

ci/run.sh

@@ -386,10 +386,10 @@ function gg_run_open_llama_7b_v2 {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -520,8 +520,8 @@ function gg_run_pythia_1_4b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -651,10 +651,10 @@ function gg_run_pythia_2_8b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"

common/arg.cpp

@@ -1548,11 +1548,11 @@ 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()));
@@ -2962,13 +2962,6 @@ 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"),
@@ -2976,6 +2969,13 @@ 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",

common/common.h

@@ -444,7 +444,7 @@ 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;
 

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) {

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);

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",

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)
 

docs/backend/CANN.md

@@ -314,3 +314,7 @@ Controls automatic cleanup of the memory pool. This option is only effective whe
 
 Converting the matmul weight format from ND to NZ can significantly improve performance on the 310I DUO NPU.
 
+### GGML_CANN_DISABLE_ACL_GRAPH
+
+When this variable is set, ACL graph execution is disabled and operators are executed in an op-by-op (eager) mode.
+This mode is mainly intended for debugging or for cases where the overhead of graph construction and execution is not desirable.

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

examples/speculative/speculative.cpp

@@ -244,7 +244,7 @@ int main(int argc, char ** argv) {
                     // stochastic verification
                     common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
 
-                    auto & dist_tgt = *common_sampler_get_candidates(smpl);
+                    auto & dist_tgt = *common_sampler_get_candidates(smpl, true);
 
                     float p_tgt = 0.0f;
                     float p_dft = 0.0f;
@@ -493,7 +493,7 @@ int main(int argc, char ** argv) {
 
                 common_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
 
-                const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl);
+                const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl, true);
 
                 for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
                     LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",

ggml/CMakeLists.txt

@@ -129,7 +129,9 @@ endif()
 option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 option(GGML_RVV              "ggml: enable rvv"              ON)
-option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
+option(GGML_RV_ZFH           "ggml: enable riscv zfh"        ON)
+option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
+option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
 option(GGML_NNPA             "ggml: enable nnpa"             OFF)  # temp disabled by default, see: https://github.com/ggml-org/llama.cpp/issues/14877

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);

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
@@ -902,7 +902,7 @@ 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;
@@ -1687,6 +1687,8 @@ 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);

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -70,6 +70,8 @@
 #include <aclnnop/aclnn_zero.h>
 #include <aclnnop/aclnn_index_copy.h>
 #include <aclnnop/aclnn_index_select.h>
+#include <aclnnop/aclnn_clamp.h>
+#include <aclnnop/aclnn_threshold.h>
 #include <float.h>
 
 #include <cmath>
@@ -964,8 +966,8 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     }
     aclTensor* acl_gamma = get_f32_cache_acl_tensor(
         ctx,
-        &ctx.f32_one_cache,
-        ctx.f32_one_cache_element,
+        &ctx.rms_norm_one_tensor_cache.cache,
+        ctx.rms_norm_one_tensor_cache.size,
         src->ne,
         acl_gamma_nb,
         1,        // dims
@@ -980,8 +982,8 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     }
     aclTensor* acl_rstd = get_f32_cache_acl_tensor(
         ctx,
-        &ctx.f32_zero_cache,
-        ctx.f32_zero_cache_element,
+        &ctx.rms_norm_zero_tensor_cache.cache,
+        ctx.rms_norm_zero_tensor_cache.size,
         src->ne,
         acl_rstd_nb,
         GGML_MAX_DIMS,
@@ -1423,21 +1425,25 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
  * @param start         Starting exponent offset.
  * @param stop          Stopping exponent offset (exclusive).
  * @param step          Step size for the exponent increment.
+ * @param dtype         Data type for slope tensor.
  */
 static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
-    float m, int64_t size, float start, float stop, float step){
-    int64_t ne[] = {size};
-    size_t nb[] = {sizeof(uint16_t)};
+    float m, int64_t size, float start, float stop, float step, ggml_type dtype){
+    aclDataType acl_type = ggml_cann_type_mapping(dtype);
+    size_t type_size = ggml_type_size(dtype);
 
-    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(uint16_t));
+    int64_t ne[] = {size};
+    size_t nb[] = {type_size};
+
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * type_size);
     void* arange_buffer = arange_allocator.get();
 
     aclTensor* arange_tensor = ggml_cann_create_tensor(
-        arange_buffer, ACL_FLOAT16, sizeof(uint16_t), ne, nb, 1);
+        arange_buffer, acl_type, type_size, ne, nb, 1);
     aclnn_arange(ctx, arange_tensor, start, stop, step, size);
 
     aclTensor* slope_tensor = ggml_cann_create_tensor(
-        slope_buffer, ACL_FLOAT16, sizeof(uint16_t), ne, nb, 1);
+        slope_buffer, acl_type, type_size, ne, nb, 1);
 
     aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
 
@@ -1468,10 +1474,11 @@ static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_bu
  * @param n_head        Total number of attention heads.
  * @param slope_buffer  Pointer to the output buffer (float array) for storing slopes.
  * @param max_bias      Maximum bias value for slope computation.
+ * @param dtype         Data type for slope tensor.
  *
 */
 static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
-    void* slope_buffer, float max_bias) {
+    void* slope_buffer, float max_bias, ggml_type dtype) {
     const int n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
 
     float m0 = powf(2.0f, -(max_bias) / n_head_log2);
@@ -1488,7 +1495,7 @@ static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
     float step  = 1;
     float count = n_head_log2;
     // end needs to be +1 because aclnn uses a left-closed, right-open interval.
-    aclnn_get_slope_inner(ctx, slope_buffer, m0, count, start, end + 1, step);
+    aclnn_get_slope_inner(ctx, slope_buffer, m0, count, start, end + 1, step, dtype);
     if (n_head_log2 < n_head) {
         // arange2
         start = 2 * (n_head_log2 - n_head_log2) + 1;
@@ -1497,7 +1504,7 @@ static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
         count = n_head - n_head_log2;
         aclnn_get_slope_inner(
             ctx, (char *) slope_buffer + n_head_log2 * sizeof(float),
-            m1, count, start, end + 1, step);
+            m1, count, start, end + 1, step, dtype);
     }
 }
 
@@ -1534,7 +1541,7 @@ static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask,
         ggml_cann_pool_alloc bias_allocator(
                     ctx.pool(), ggml_nelements(dst) * ggml_element_size(dst));
         bias_buffer = bias_allocator.get();
-        aclnn_get_slope(ctx, n_heads, slope_buffer, max_bias);
+        aclnn_get_slope(ctx, n_heads, slope_buffer, max_bias, GGML_TYPE_F32);
     }
 
     // broadcast for mask, slop and dst;
@@ -1760,10 +1767,10 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         case GGML_TYPE_F16: {
             aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
             ggml_cann_pool_alloc src_buffer_allocator(
-                ctx.pool(), ggml_nelements(src0) * sizeof(float_t));
+                ctx.pool(), ggml_nelements(src0) * sizeof(float));
             void* src_trans_buffer = src_buffer_allocator.get();
             size_t src_trans_nb[GGML_MAX_DIMS];
-            src_trans_nb[0] = sizeof(float_t);
+            src_trans_nb[0] = sizeof(float);
             for (int i = 1; i < GGML_MAX_DIMS; i++) {
                 src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
             }
@@ -1807,14 +1814,14 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
             // [3,4,5,64] -> [3,4,5,2,32]
             dequant_ne = weight_ne;
-            dequant_nb[0] = sizeof(float_t);
+            dequant_nb[0] = sizeof(float);
             for (int i = 1; i < GGML_MAX_DIMS + 1; i++) {
                 dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1];
             }
 
             scale_offset = ggml_nelements(src0) * sizeof(int8_t);
             ggml_cann_pool_alloc dequant_buffer_allocator(
-                ctx.pool(), ggml_nelements(src0) * sizeof(float_t));
+                ctx.pool(), ggml_nelements(src0) * sizeof(float));
 
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb,
@@ -1823,11 +1830,11 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 src0->data, ACL_FLOAT16, sizeof(uint16_t), scale_ne, scale_nb,
                 GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
             aclTensor* dequant_tensor = ggml_cann_create_tensor(
-                dequant_buffer_allocator.get(), ACL_FLOAT, sizeof(float_t),
+                dequant_buffer_allocator.get(), ACL_FLOAT, sizeof(float),
                 dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
 
             aclnn_mul(ctx, acl_weight_tensor, acl_scale_tensor, dequant_tensor);
-            dequant_nb[0] = sizeof(float_t);
+            dequant_nb[0] = sizeof(float);
             dequant_ne = src0->ne;
             for (int i = 1; i < GGML_MAX_DIMS; i++) {
                 dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
@@ -2248,46 +2255,35 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
  *   5. Compute sin(θ), cos(θ) and optionally scale by attn_factor.
  *   6. Expand sin/cos values by repeat or repeat_interleave depending
  *      on whether @param is_neox is enabled.
- *   7. Store the computed values into persistent buffers
- *      (ctx.rope_sin_ptr / ctx.rope_cos_ptr).
  *
- * @param ctx         The CANN backend context, holding memory pool,
- *                    stream, and persistent buffers for rope init/cache.
- * @param dst         The destination ggml_tensor whose computation
- *                    depends on the cached RoPE values (usually Qcur/Kcur).
- * @param theta_scale Scalar exponent base for computing theta scale values.
- * @param freq_scale  Frequency scaling factor, applied to theta scale.
- * @param attn_factor Attention scaling factor, applied to sin/cos.
- * @param is_neox     Whether to use Neox-style repeat strategy
- *                    (dim expansion vs repeat_interleave).
+ * @param ctx                The CANN backend context, holding memory pool,
+ *                           stream, and persistent buffers for rope init/cache.
+ * @param dst                The destination ggml_tensor whose computation
+ *                           depends on the RoPE values (usually Qcur/Kcur).
+ * @param sin_tensor_buffer  Pre-allocated buffer for storing repeated sin values.
+ * @param cos_tensor_buffer  Pre-allocated buffer for storing repeated cos values.
+ * @param theta_scale        Scalar exponent base for computing theta scale values.
+ * @param freq_scale         Frequency scaling factor, applied to theta scale.
+ * @param attn_factor        Attention scaling factor, applied to sin/cos.
+ * @param is_neox            Whether to use Neox-style repeat strategy
+ *                           (dim expansion vs repeat_interleave).
  */
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
+                             void* sin_tensor_buffer, void* cos_tensor_buffer,
+                             float* corr_dims, float ext_factor,
                              float theta_scale, float freq_scale,
                              float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
-    bool is_q = (std::strncmp(dst->name, "Qcur-", 5) == 0);
-    bool is_k = (std::strncmp(dst->name, "Kcur-", 5) == 0);
-
-    // used for accuracy testing
-    bool is_attention = is_q || is_k;
-
-    // just compute in first layer in attention
-    bool is_fisrt_layer = (std::strncmp(dst->name, "Qcur-0", GGML_MAX_NAME) == 0);
-    if(is_attention && !is_fisrt_layer) {
-        return;
-    }
 
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
     ggml_tensor* src2 = dst->src[2];  // freq_factors
 
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    int64_t theta_scale_length = ne00 / 2;
+    int64_t theta_scale_length = src0->ne[0] / 2;
     int64_t theta_scale_ne[] = {theta_scale_length, 1, 1, 1};
-    size_t theta_scale_nb[] = {sizeof(float_t), sizeof(float_t), sizeof(float_t),
-                          theta_scale_length * sizeof(float_t)};
+    size_t theta_scale_nb[] = {sizeof(float), sizeof(float), sizeof(float),
+                          theta_scale_length * sizeof(float)};
 
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     int64_t position_length = src1->ne[0];
@@ -2297,65 +2293,115 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
 
     int64_t theta_ne[] = {theta_scale_length, 1, position_length, 1};
     size_t theta_nb[GGML_MAX_DIMS];
-    theta_nb[0] = sizeof(float_t);
+    theta_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
         theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
     }
 
-    // init theta scale, just one time
-    if(ctx.rope_init_ptr == nullptr || !is_attention) {
-        // theta_scale arange, [0,1,...,ne00/2 - 1]
-        if(ctx.rope_init_ptr != nullptr){
-            ACL_CHECK(aclrtFree(ctx.rope_init_ptr));
-        }
-        ACL_CHECK(aclrtMalloc(&ctx.rope_init_ptr, theta_scale_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
+    // theta_scale arange, [0,1,...,ne00/2 - 1]
+    aclTensor* acl_theta_scale_tensor = nullptr;
+    // cache theta scale
+    if (ctx.rope_cache.theta_scale_length != theta_scale_length ||
+        // theta_scale and freq_scale should not change during the current token inference process,
+        // so we can directly use == here instead of comparing the absolute difference.
+        ctx.rope_cache.theta_scale != theta_scale ||
+        ctx.rope_cache.freq_scale != freq_scale) {
 
-        aclTensor* acl_theta_scale_tensor =
-            ggml_cann_create_tensor(ctx.rope_init_ptr, ACL_FLOAT, sizeof(float_t),
+        ctx.rope_cache.theta_scale_length = theta_scale_length;
+        ctx.rope_cache.theta_scale = theta_scale;
+        ctx.rope_cache.freq_scale = freq_scale;
+
+        if (ctx.rope_cache.theta_scale_cache != nullptr) {
+            ACL_CHECK(aclrtFree(ctx.rope_cache.theta_scale_cache));
+        }
+        ACL_CHECK(aclrtMalloc(&ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST));
+
+        acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
                                     theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+
         float start = 0;
         float step = 1;
-        float stop = ne00 / 2;
-        float n_elements = ne00 / 2;
+        float stop = theta_scale_length;
+        float n_elements = theta_scale_length;
         aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);
 
+        ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
+        aclTensor* acl_yarn_ramp_tensor = nullptr;
+        if (ext_factor != 0) {
+            // -rope_yarn_ramp
+            // const float y = (i0 / 2 - low) / MAX(0.001f, high - low);
+            // return MIN(1, MAX(0, y)) - 1;
+            yarn_ramp_allocator.alloc(theta_scale_length * sizeof(float));
+            void* yarn_ramp_buffer = yarn_ramp_allocator.get();
+            acl_yarn_ramp_tensor = ggml_cann_create_tensor(yarn_ramp_buffer, ACL_FLOAT, sizeof(float_t),
+                                           theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+            float zero_value = 0, one_value = 1;
+            float denom_safe_value = MAX(0.001f, corr_dims[1] - corr_dims[0]);
+            aclScalar* low = aclCreateScalar(&corr_dims[0], aclDataType::ACL_FLOAT);
+            aclScalar* zero = aclCreateScalar(&zero_value, aclDataType::ACL_FLOAT);
+            aclScalar* one = aclCreateScalar(&one_value, aclDataType::ACL_FLOAT);
+            aclScalar* denom_safe = aclCreateScalar(&denom_safe_value, aclDataType::ACL_FLOAT);
+            aclScalar* ext_factor_sc = aclCreateScalar(&ext_factor, aclDataType::ACL_FLOAT);
+
+            GGML_CANN_CALL_ACLNN_OP(ctx, Subs, acl_theta_scale_tensor, low, one, acl_yarn_ramp_tensor);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceDivs, acl_yarn_ramp_tensor, denom_safe);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceThreshold, acl_yarn_ramp_tensor, zero, zero);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceClampMax, acl_yarn_ramp_tensor, one);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceSubs, acl_yarn_ramp_tensor, one, one);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor, ext_factor_sc);
+
+            // theta_interp = freq_scale * theta_extrap;
+            // theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+            // theta = freq_scale * theta_extrap * (1 - ramp_mix) + theta_extrap * ramp_mix;
+            // theta = freq_scale * theta_extrap - freq_scale * theta_extrap * ramp_mix + theta_extrap * ramp_mix;
+            // theta = theta_extrap * (freq_scale - freq_scale * ramp_mix + ramp_mix);
+            //
+            // we cache (freq_scale - freq_scale * ramp_mix + ramp_mix), Considering that the rope_yarn_ramp here is the inverse
+            // cache freq_scale + (freq_scale - 1) * ramp_mix
+            float freq_scale_1 = freq_scale - 1;
+            aclScalar* freq_scale_sc = aclCreateScalar(&freq_scale, aclDataType::ACL_FLOAT);
+            aclScalar* freq_scale_1_sc = aclCreateScalar(&freq_scale_1, aclDataType::ACL_FLOAT);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMuls, acl_yarn_ramp_tensor, freq_scale_1_sc);
+            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdds, acl_yarn_ramp_tensor, freq_scale_sc, one);
+
+            ggml_cann_release_resources(ctx, low, zero, one, denom_safe, ext_factor_sc, freq_scale_sc, freq_scale_1_sc);
+        }
+
         // power
         aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
         GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
                                 acl_theta_scale_tensor);
 
-        // freq_scale
-        if (freq_scale != 1) {
+        if (ext_factor != 0) {
+            aclnn_mul(ctx, acl_theta_scale_tensor, acl_yarn_ramp_tensor);
+        } else if (freq_scale != 1) {
             aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
         }
 
-        // freq_factors
-        if (src2) {
-            aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
-                src2->data, ggml_cann_type_mapping(src2->type),
-                ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
-            aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
-            ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
-        }
-        // release
-        ggml_cann_release_resources(ctx, acl_theta_scale_tensor,acl_theta_scale);
-    }
-
-    // init sin_repeat && cos_repeat, one token just init in 0 layer
-    if(position_length > ctx.max_prompt_length) {
-        ctx.max_prompt_length = position_length;
-        int64_t repeat_theta_length = theta_scale_length * ctx.max_prompt_length * 2;
-        if(ctx.rope_sin_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(ctx.rope_sin_ptr));
-            ACL_CHECK(aclrtFree(ctx.rope_cos_ptr));
-        }
-        ACL_CHECK(aclrtMalloc(&ctx.rope_sin_ptr, repeat_theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
-        ACL_CHECK(aclrtMalloc(&ctx.rope_cos_ptr, repeat_theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
-    }
-
-    aclTensor* acl_theta_scale_tensor =
-            ggml_cann_create_tensor(ctx.rope_init_ptr, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_release_resources(ctx, acl_yarn_ramp_tensor, acl_theta_scale);
+    } else {
+        // use cache
+        acl_theta_scale_tensor =
+            ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
                                     theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+    }
+
+    ggml_cann_pool_alloc freq_fac_res_allocator(ctx.pool());
+    // freq_factors
+    if (src2) {
+        freq_fac_res_allocator.alloc(theta_scale_length * sizeof(float));
+        void* freq_fac_res_ptr = freq_fac_res_allocator.get();
+        aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
+            src2->data, ggml_cann_type_mapping(src2->type),
+            ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+        aclTensor* acl_freq_fac_res_tensor = ggml_cann_create_tensor(
+            freq_fac_res_ptr, ACL_FLOAT, sizeof(float),
+            theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
+        aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor, acl_freq_fac_res_tensor);
+        std::swap(acl_theta_scale_tensor, acl_freq_fac_res_tensor);
+        ggml_cann_release_resources(ctx, acl_freq_factors_tensor, acl_freq_fac_res_tensor);
+    }
 
     // position
     aclTensor* acl_position_tensor = ggml_cann_create_tensor(
@@ -2365,49 +2411,53 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     // power * position
     int64_t theta_length = theta_scale_length * position_length;
     ggml_cann_pool_alloc theta_allocator(ctx.pool(),
-                                        theta_length * sizeof(float_t));
+                                        theta_length * sizeof(float));
     void* theta_buffer = theta_allocator.get();
 
     aclTensor* acl_theta_tensor =
-        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float),
                                 theta_ne, theta_nb, GGML_MAX_DIMS);
     aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
             acl_theta_tensor);
 
     // sin/cos
     ggml_cann_pool_alloc sin_allocator(ctx.pool(),
-                                    theta_length * sizeof(float_t));
+                                    theta_length * sizeof(float));
     void* sin_buffer = sin_allocator.get();
     aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
-        sin_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+        sin_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb,
         GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);
 
     ggml_cann_pool_alloc cos_allocator(ctx.pool(),
-                                    theta_length * sizeof(float_t));
+                                    theta_length * sizeof(float));
     void* cos_buffer = cos_allocator.get();
     aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
-        cos_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
+        cos_buffer, ACL_FLOAT, sizeof(float), theta_ne, theta_nb,
         GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);
 
+    if (ext_factor != 0) {
+        attn_factor *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+    }
+
     // attn_factor
     if (attn_factor != 1) {
         aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true);
         aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
     }
 
-    int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
+    int64_t sin_reshape_ne[4] = {src0->ne[0], 1, src0->ne[2], 1};
     size_t sin_reshape_nb[GGML_MAX_DIMS];
-    sin_reshape_nb[0] = sizeof(float_t);
+    sin_reshape_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
         sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
     }
     aclTensor* acl_sin_repeat_tensor =
-        ggml_cann_create_tensor(ctx.rope_sin_ptr, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(sin_tensor_buffer, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclTensor* acl_cos_repeat_tensor =
-        ggml_cann_create_tensor(ctx.rope_cos_ptr, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(cos_tensor_buffer, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
 
     // repeat
@@ -2449,6 +2499,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: use ascendc
     // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
+    ggml_tensor* src1 = dst->src[1];
 
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
@@ -2470,8 +2521,6 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: n_dims <= ne0
     GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
-    // TODO: ext_factor != 0
-    GGML_ASSERT(ext_factor == 0);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -2481,20 +2530,28 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
+    // sin/cos tensor length.
+    int64_t repeat_theta_length = src0->ne[0] * src1->ne[0];
+    ggml_cann_pool_alloc sin_tensor_allocator(ctx.pool(), repeat_theta_length * sizeof(float));
+    ggml_cann_pool_alloc cos_tensor_allocator(ctx.pool(), repeat_theta_length * sizeof(float));
+    void *sin_tensor_buffer = sin_tensor_allocator.get();
+    void *cos_tensor_buffer = cos_tensor_allocator.get();
+
     // init ctx.rope_cos/rope_sin cache
-    aclnn_cache_init(ctx, dst, theta_scale, freq_scale, attn_factor, is_neox);
+    aclnn_cache_init(ctx, dst, sin_tensor_buffer, cos_tensor_buffer, corr_dims, ext_factor,
+                    theta_scale, freq_scale, attn_factor, is_neox);
 
     int64_t sin_reshape_ne[4] = {ne00, 1, ne02, 1};
     size_t sin_reshape_nb[GGML_MAX_DIMS];
-    sin_reshape_nb[0] = sizeof(float_t);
+    sin_reshape_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
         sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
     }
     aclTensor* acl_sin_reshape_tensor =
-        ggml_cann_create_tensor(ctx.rope_sin_ptr, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(sin_tensor_buffer, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclTensor* acl_cos_reshape_tensor =
-        ggml_cann_create_tensor(ctx.rope_cos_ptr, ACL_FLOAT, sizeof(float_t),
+        ggml_cann_create_tensor(cos_tensor_buffer, ACL_FLOAT, sizeof(float),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
 
     aclTensor* acl_src = ggml_cann_create_tensor(src0);
@@ -2509,7 +2566,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     void* minus_one_scale_buffer = nullptr;
     ggml_cann_pool_alloc roll_allocator(ctx.pool(), ggml_nbytes(src0));
     ggml_cann_pool_alloc minus_one_scale_allocator(
-        ctx.pool(), sizeof(float_t) * src0->ne[0]);
+        ctx.pool(), sizeof(float) * src0->ne[0]);
     if (!is_neox) {
         // roll input: [q0,q1,q2,q3,...] -> [q1,q0,q3,q2,...]
         input_roll_buffer = roll_allocator.get();
@@ -2539,13 +2596,13 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
         int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
         size_t minus_one_nb[GGML_MAX_DIMS];
-        minus_one_nb[0] = sizeof(float_t);
+        minus_one_nb[0] = sizeof(float);
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
         }
         acl_minus_one_tensor = aclnn_values(
-            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
-            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+            ctx, minus_one_scale_buffer, sizeof(float) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float), 1);
         int64_t dim = 3;
         int64_t* index = new int64_t[src0->ne[0]];
         for (int i = 0; i < src0->ne[0]; i++) {
@@ -2573,22 +2630,22 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         minus_one_scale_buffer = minus_one_scale_allocator.get();
         int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
         size_t minus_one_nb[GGML_MAX_DIMS];
-        minus_one_nb[0] = sizeof(float_t);
+        minus_one_nb[0] = sizeof(float);
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
         }
         acl_minus_one_tensor = aclnn_values(
-            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
-            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+            ctx, minus_one_scale_buffer, sizeof(float) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float), 1);
         // -1 * first half
         int64_t first_half_ne[4] = {src0->ne[0] / 2, 1, 1, 1};
         size_t first_half_nb[GGML_MAX_DIMS];
-        first_half_nb[0] = sizeof(float_t);
+        first_half_nb[0] = sizeof(float);
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             first_half_nb[i] = first_half_nb[i - 1] * first_half_ne[i - 1];
         }
         aclTensor* acl_first_half_tensor = ggml_cann_create_tensor(
-            minus_one_scale_buffer, ACL_FLOAT, sizeof(float_t), first_half_ne,
+            minus_one_scale_buffer, ACL_FLOAT, sizeof(float), first_half_ne,
             first_half_nb, GGML_MAX_DIMS);
         bool inplace = true;
         float scale = -1;
@@ -2628,28 +2685,28 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         // TODO: ne0 != n_dims in mode2
     } else if (src0->type == GGML_TYPE_F16) {
         size_t input_fp32_nb[GGML_MAX_DIMS];
-        input_fp32_nb[0] = sizeof(float_t);
+        input_fp32_nb[0] = sizeof(float);
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             input_fp32_nb[i] = input_fp32_nb[i - 1] * dst->ne[i - 1];
         }
         ggml_cann_pool_alloc fp32_allocator1(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+            ctx.pool(), ggml_nelements(dst) * sizeof(float));
         void* input_fp32_buffer1 = fp32_allocator1.get();
         aclTensor* input_fp32_tensor1 = ggml_cann_create_tensor(
-            input_fp32_buffer1, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_buffer1, ACL_FLOAT, sizeof(float), dst->ne,
             input_fp32_nb, GGML_MAX_DIMS);
         ggml_cann_pool_alloc fp32_allocator2(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+            ctx.pool(), ggml_nelements(dst) * sizeof(float));
         void* input_fp32_buffer2 = fp32_allocator2.get();
         aclTensor* input_fp32_tensor2 = ggml_cann_create_tensor(
-            input_fp32_buffer2, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_buffer2, ACL_FLOAT, sizeof(float), dst->ne,
             input_fp32_nb, GGML_MAX_DIMS);
 
         ggml_cann_pool_alloc fp32_allocator(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+            ctx.pool(), ggml_nelements(dst) * sizeof(float));
         output_fp32_buffer = fp32_allocator.get();
         aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
-            output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
+            output_fp32_buffer, ACL_FLOAT, sizeof(float), dst->ne,
             input_fp32_nb, GGML_MAX_DIMS);
         aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor, input_fp32_tensor1);
         aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
@@ -2746,8 +2803,6 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
     aclIntArray *padding = aclCreateIntArray(paddingVal, 1);
     int64_t dilationVal[] = {1};
     aclIntArray *dilation = aclCreateIntArray(dilationVal, 1);
-    bool transposed = true;
-    int64_t groups = 1;
     int8_t cubeMathType = 0;
 
 #ifdef ASCEND_310P
@@ -2755,7 +2810,7 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
 #endif
 
     GGML_CANN_CALL_ACLNN_OP(ctx, Convolution, acl_input, acl_weight, nullptr, stride,
-        padding, dilation, transposed, padding, groups, acl_dst, cubeMathType);
+        padding, dilation, true, padding, 1, acl_dst, cubeMathType);
 
     ggml_cann_release_resources(ctx, acl_weight, acl_dst, stride, padding, dilation);
 }
@@ -2864,174 +2919,49 @@ void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst){
  */
 static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     //dst   [M, K, N, 1]
-    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
-    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
+    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]  -> [D, M, K, 1]
+    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1 -> [D, 1, K, 1]
     ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
 
-    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT(dst->ne[3] == 1);
 
-    // copy index from npu to cpu
-    int64_t n_as = ne02; // A
-    int64_t n_ids = ids->ne[0]; // K
+    int64_t batch = src1->ne[2];
+    GGML_ASSERT(batch == ids->ne[1]);
 
-    std::vector<char> ids_host(ggml_nbytes(ids));
-    ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
-        ACL_MEMCPY_DEVICE_TO_HOST);
-    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+    ggml_cann_pool_alloc export_allocator(ctx.pool(), src0->ne[0] * src0->ne[1] * ids->ne[0] * ggml_element_size(src0));
+    void* export_ptr = export_allocator.get();
+    for (int64_t i = 0; i < batch; i++) {
+        aclTensor *select_index = ggml_cann_create_tensor(ids, ids->ne, ids->nb, 1, ACL_FORMAT_ND, i * ids->nb[1]);
+        aclTensor *export_weight = ggml_cann_create_tensor(src0, src0->ne, src0->nb, 3);
 
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *)  dst->data;
-    size_t ori_src0_nb[4] = {nb00, nb01, nb02, nb03};
-
-    // src0 is F16, src1 is F32, dst is F32
-    ggml_cann_pool_alloc src0_cast_allocator;
-    if (src0->type == GGML_TYPE_F16) {
-        src0_cast_allocator.alloc(ctx.pool(), sizeof(float) * ggml_nelements(src0));
-        void* src0_cast_buf = src0_cast_allocator.get();
-
-        size_t cast_nb[GGML_MAX_DIMS];
-        cast_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            cast_nb[i] = cast_nb[i - 1] * src0->ne[i - 1];
+        int64_t select_export_ne[] = {src0->ne[0], src0->ne[1], ids->ne[0]};
+        size_t select_export_nb[3];
+        select_export_nb[0] = src0->nb[0];
+        for (int k = 1;k < 3; k++) {
+            select_export_nb[k] = select_export_nb[k-1] * select_export_ne[k-1];
         }
 
-        aclTensor* acl_src0_f16 = ggml_cann_create_tensor(src0);
-        aclTensor* acl_cast = ggml_cann_create_tensor(src0_cast_buf,
-            ACL_FLOAT, sizeof(float), src0->ne, cast_nb, 4);
-        GGML_CANN_CALL_ACLNN_OP(ctx, Cast, acl_src0_f16, ACL_FLOAT, acl_cast);
-        ggml_cann_release_resources(ctx, acl_cast, acl_src0_f16);
+        aclTensor *select_export = ggml_cann_create_tensor(export_ptr, ggml_cann_type_mapping(src0->type), ggml_element_size(src0), select_export_ne, select_export_nb, 3);
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, export_weight, 0, select_index, select_export);
 
-        src0_original = (char *) src0_cast_buf;
-        memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
+        int64_t select_transpose_ne[] = {select_export_ne[1], select_export_ne[0], select_export_ne[2]};
+        size_t select_transpose_nb[] = {select_export_nb[1], select_export_nb[0], select_export_nb[2]};
+        aclTensor *select_export_transpose = ggml_cann_create_tensor(export_ptr, ggml_cann_type_mapping(src0->type), ggml_element_size(src0), select_transpose_ne, select_transpose_nb, 3);
+
+        int64_t active_tensor_ne[] = {src1->ne[0], 1, src1->ne[1]};
+        size_t active_tensor_nb[] = {src1->nb[0], src1->nb[1], src1->nb[1]};
+        aclTensor *active_tensor = ggml_cann_create_tensor(src1, active_tensor_ne, active_tensor_nb, 3, ACL_FORMAT_ND, i * src1->nb[2]);
+
+        int64_t dst_ne[] = {dst->ne[0], 1, dst->ne[1]};
+        size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[1]};
+        aclTensor *acl_dst = ggml_cann_create_tensor(dst, dst_ne,dst_nb, 3, ACL_FORMAT_ND, i * dst->nb[2]);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, BatchMatMul, active_tensor, select_export_transpose, acl_dst, 2);
+
+        ggml_cann_release_resources(ctx, select_index, export_weight, select_export, active_tensor, acl_dst, select_export_transpose);
     }
-
-#ifdef ASCEND_310P
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row = *dst;
-
-    if (src0->type == GGML_TYPE_F16) {
-        src0_row.type = GGML_TYPE_F32;
-    }
-
-    // src0_row [D, M, 1, 1] weight without permute
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[0] = ori_src0_nb[0];
-    src0_row.nb[1] = ori_src0_nb[1];
-    src0_row.nb[2] = ori_src0_nb[1];
-    src0_row.nb[3] = ori_src0_nb[1];
-
-    // src1_row [D, 1, 1, 1] -> input
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    // dst_row [M, 1, 1, 1] -> out
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    //create weight for one row
-    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-        for (int64_t id = 0; id < n_ids; id++) {
-            // expert index
-            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-            GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-            // If B = 1 (broadcast), always use 0; otherwise, use id.
-            int64_t i11 = (ne11 == 1 ? 0 : id);
-            int64_t i12 = iid1;
-
-            int64_t i1 = id;
-            int64_t i2 = i12;
-
-            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
-            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
-            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
-
-            src0_row.data = src0_tmp_ptr;
-            src1_row.data = src1_tmp_ptr;
-            dst_row.data = dst_tmp_ptr;
-            dst_row.src[0] = &src0_row;
-            dst_row.src[1] = &src1_row;
-
-            ggml_cann_mul_mat(ctx, &dst_row);
-        }
-    }
-    return;
-#endif
-
-    std::vector<aclTensor*> src0_tensor_vec;
-    std::vector<aclTensor*> src1_tensor_vec;
-    std::vector<aclTensor*> dst_tensor_vec;
-    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-        for (int64_t id = 0; id < n_ids; id++) {
-            // src0_row [M, D] -> weight && permute
-            int64_t src0_ne[2] = {ne01, ne00};
-            size_t src0_nb[2] = {ori_src0_nb[1], ori_src0_nb[0]};
-            // src1_row [D, 1] -> input
-            int64_t src1_ne[2] = {ne10, 1};
-            size_t src1_nb[2] = {nb10, nb11};
-            // dst_row [M, 1] -> out
-            int64_t dst_ne[2] = {ne0, 1};
-            size_t dst_nb[2] = {nb0, nb1};
-
-            // expert index
-            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-            GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-            // If B = 1 (broadcast), always use 0; otherwise, use id.
-            int64_t i11 = (ne11 == 1 ? 0 : id);
-            int64_t i12 = iid1;
-
-            int64_t i1 = id;
-            int64_t i2 = i12;
-
-            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
-            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
-            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
-
-            aclTensor* acl_src0 = ggml_cann_create_tensor(src0_tmp_ptr,
-                ACL_FLOAT, sizeof(float),
-                src0_ne, src0_nb, 2);
-            aclTensor* acl_src1 = ggml_cann_create_tensor(src1_tmp_ptr,
-                ACL_FLOAT, sizeof(float),
-                src1_ne, src1_nb, 2);
-            aclTensor* acl_dst = ggml_cann_create_tensor(dst_tmp_ptr,
-                ACL_FLOAT, sizeof(float),
-                dst_ne, dst_nb, 2);
-
-            src0_tensor_vec.push_back(acl_src0);
-            src1_tensor_vec.push_back(acl_src1);
-            dst_tensor_vec.push_back(acl_dst);
-        }
-    }
-
-    size_t GROUP_SIZE = 128;
-    // GroupedMatmulV3 required tensor_list.size < 128
-    for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
-        // split and call GroupedMatmulV3
-        size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
-        std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
-        std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
-        std::vector<aclTensor*> dst_tensor_vec_split(dst_tensor_vec.begin() + i, dst_tensor_vec.begin() + end);
-
-        aclTensorList* src0_tensor_list = aclCreateTensorList(src0_tensor_vec_split.data(), src0_tensor_vec_split.size());
-        aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
-        aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
-
-        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
-            nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
-
-        ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
-    }
-    return;
 }
 
 /**
@@ -3342,7 +3272,7 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context& ctx, ggml_tensor* dst){
                 const int64_t n_heads = src0->ne[2];
                 ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(uint16_t));
                 void* slope_buffer = slope_allocator.get();
-                aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias);
+                aclnn_get_slope(ctx, n_heads, slope_buffer, maxBias, GGML_TYPE_F16);
 
                 int64_t slope_ne[] = {1, 1, n_heads, 1};
                 size_t slope_nb[GGML_MAX_DIMS];

ggml/src/ggml-cann/common.h

@@ -360,6 +360,30 @@ struct ggml_cann_graph {
 };
 #endif  // USE_ACL_GRAPH
 
+struct ggml_cann_rope_cache {
+    ~ggml_cann_rope_cache() {
+        if(theta_scale_cache != nullptr) {
+            ACL_CHECK(aclrtFree(theta_scale_cache));
+        }
+    }
+
+    void* theta_scale_cache = nullptr;
+    int64_t theta_scale_length = 0;
+    float theta_scale = 0.0f;
+    float freq_scale = 0.0f;
+};
+
+struct ggml_cann_tensor_cache {
+    ~ggml_cann_tensor_cache() {
+        if(cache != nullptr) {
+            ACL_CHECK(aclrtFree(cache));
+        }
+    }
+
+    void* cache = nullptr;
+    int64_t size = 0;
+};
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -371,19 +395,15 @@ struct ggml_backend_cann_context {
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
     std::unique_ptr<ggml_cann_graph> cann_graph;
+    bool acl_graph_mode = true;
 #endif
     cann_task_queue task_queue;
     bool async_mode;
     // Rope Cache
-    void* rope_init_ptr = nullptr;
-    void* rope_sin_ptr = nullptr;
-    void* rope_cos_ptr = nullptr;
-    int64_t max_prompt_length = 0;
+    ggml_cann_rope_cache rope_cache;
     // Constant Pool
-    void* f32_zero_cache = nullptr;
-    void* f32_one_cache = nullptr;
-    int64_t f32_zero_cache_element = 0;
-    int64_t f32_one_cache_element = 0;
+    ggml_cann_tensor_cache rms_norm_one_tensor_cache;
+    ggml_cann_tensor_cache rms_norm_zero_tensor_cache;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -399,6 +419,13 @@ struct ggml_backend_cann_context {
         async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
+#ifdef USE_ACL_GRAPH
+        acl_graph_mode = !(parse_bool(get_env("GGML_CANN_DISABLE_ACL_GRAPH").value_or("")));
+        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n",
+              __func__, device,
+              acl_graph_mode ? "GRAPH" : "EAGER",
+              acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
+#endif
     }
 
     /**
@@ -415,21 +442,6 @@ struct ggml_backend_cann_context {
                 ACL_CHECK(aclrtDestroyStream(streams[i]));
             }
         }
-        if(rope_init_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(rope_init_ptr));
-        }
-        if(rope_sin_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(rope_sin_ptr));
-        }
-        if(rope_cos_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(rope_cos_ptr));
-        }
-        if(f32_zero_cache != nullptr) {
-            ACL_CHECK(aclrtFree(f32_zero_cache));
-        }
-        if(f32_one_cache != nullptr) {
-            ACL_CHECK(aclrtFree(f32_one_cache));
-        }
     }
 
     /**

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

@@ -2247,10 +2247,15 @@ static enum ggml_status ggml_backend_cann_graph_compute(
         (ggml_backend_cann_context*)backend->context;
     ggml_cann_set_device(cann_ctx->device);
     release_nz_workspace();
+
 #ifdef USE_ACL_GRAPH
     bool use_cann_graph = true;
     bool cann_graph_update_required = false;
 
+    if (!cann_ctx->acl_graph_mode) {
+        use_cann_graph = false;
+    }
+
     if (use_cann_graph) {
         if (cann_ctx->cann_graph == nullptr) {
             cann_ctx->cann_graph.reset(new ggml_cann_graph());
@@ -2400,16 +2405,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_ROPE: {
             // TODO: with ops-test v == 1
-            float ext_factor = 0.0f;
-            memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float));
             // TODO: n_dims <= ne0
             if (op->src[0]->ne[0] != op->op_params[1]) {
                 return false;
             }
-            // TODO: ext_factor != 0
-            if (ext_factor != 0) {
-                return false;
-            }
 
             const int mode = ((const int32_t *) op->op_params)[2];
             if (mode & GGML_ROPE_TYPE_MROPE) {
@@ -2418,10 +2417,11 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (mode & GGML_ROPE_TYPE_VISION) {
                 return false;
             }
-
+#ifdef ASCEND_310P
             if(!ggml_is_contiguous(op->src[0])){
                 return false;
             }
+#endif
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -2483,12 +2483,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_ARGMAX:
         case GGML_OP_COS:
         case GGML_OP_SIN:
-        case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_LOG:
         case GGML_OP_MEAN:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255.
+            return (op->src[0]->ne[0] - 1) <= 255;
         case GGML_OP_SCALE:
             float bias;
             memcpy(&bias, (const float *)(op->op_params) + 1, sizeof(float));
@@ -2522,13 +2524,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // different head sizes of K and V are not supported yet
                 return false;
             }
-            if (op->src[0]->ne[0] == 192) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek MLA
-                return false;
-            }
             if (op->src[0]->ne[0] % 16 != 0) {
                 // TODO: padding to support
                 return false;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -433,15 +433,22 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ggml-cpu/arch/riscv/quants.c
             ggml-cpu/arch/riscv/repack.cpp
             )
-        if (GGML_RVV)
-            if (GGML_XTHEADVECTOR)
-                list(APPEND ARCH_FLAGS -march=rv64gc_zfhmin_xtheadvector -mabi=lp64d)
-            elseif (GGML_RV_ZFH)
-                list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -mabi=lp64d)
-            else()
-                list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+        set(MARCH_STR "rv64gc")
+        if (GGML_RV_ZFH)
+            string(APPEND MARCH_STR "_zfh")
+        endif()
+        if (GGML_XTHEADVECTOR)
+            string(APPEND MARCH_STR "_xtheadvector")
+        elseif (GGML_RVV)
+            string(APPEND MARCH_STR "_v")
+            if (GGML_RV_ZVFH)
+                string(APPEND MARCH_STR "_zvfh")
             endif()
         endif()
+        if (GGML_RV_ZICBOP)
+            string(APPEND MARCH_STR "_zicbop")
+        endif()
+        list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
         list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)

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:

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3221,6 +3221,13 @@ 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]);

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)

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),

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) {

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));
             }
         }
     }

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) {

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>;

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -2776,10 +2776,6 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
         case GGML_OP_FLASH_ATTN_EXT:
             {
-                if (op->src[4]) {
-                    return false;
-                }
-
                 const ggml_tensor * q = op->src[0];
                 const ggml_tensor * k = op->src[1];
                 const ggml_tensor * v = op->src[2];
@@ -5765,6 +5761,7 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
 static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
     const ggml_tensor * v = dst->src[2];
     const ggml_tensor * mask = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
     GGML_ASSERT(q->extra);
     GGML_ASSERT(k->extra);
     GGML_ASSERT(v->extra);
@@ -5772,6 +5769,9 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
     if (mask) {
         GGML_ASSERT(mask->extra);
     }
+    if (sinks) {
+        GGML_ASSERT(sinks->extra);
+    }
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -5813,6 +5813,7 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
     ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
     ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
     ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+    ggml_tensor_extra_cl * extra_sinks = sinks ? (ggml_tensor_extra_cl *)sinks->extra : NULL;
 
     cl_ulong offset_q = extra_q->offset + q->view_offs;
     cl_ulong offset_k = extra_k->offset + k->view_offs;
@@ -5820,6 +5821,8 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
     cl_ulong offset_o = extra_o->offset + dst->view_offs;
     cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
     cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+    cl_mem   sinks_buffer = extra_sinks ? extra_sinks->data_device : NULL;
+    cl_ulong offset_sinks = extra_sinks ? extra_sinks->offset + sinks->view_offs : 0;
 
     const cl_ulong q_nb1 = q->nb[1], q_nb2 = q->nb[2], q_nb3 = q->nb[3];
     const cl_ulong k_nb1 = k->nb[1], k_nb2 = k->nb[2], k_nb3 = k->nb[3];
@@ -5874,6 +5877,8 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
     CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &mask_nb3));
     CL_CHECK(clSetKernelArg(kernel, 36, sizeof(int),      &mask_ne2));
     CL_CHECK(clSetKernelArg(kernel, 37, sizeof(int),      &mask_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 38, sizeof(cl_mem),   &sinks_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 39, sizeof(cl_ulong), &offset_sinks));
 
     if (n_q == 1) {
         const size_t wg_size = 64;

ggml/src/ggml-opencl/kernels/flash_attn_f16.cl

@@ -49,7 +49,9 @@ __kernel void flash_attn_f16(
     const ulong mask_nb2,
     const ulong mask_nb3,
     const int mask_ne2,
-    const int mask_ne3
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
 ) {
     const int tid = get_local_id(0);
     const int block_q_idx = get_group_id(0);
@@ -171,6 +173,20 @@ __kernel void flash_attn_f16(
     }
 
     if (my_query_row < n_q) {
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+
+            const ACC_TYPE scale_o = exp(m_i - m_final);
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] *= scale_o;
+            }
+
+            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
+        }
+
         const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
         global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
         if (l_i > 0.0f) {
@@ -214,7 +230,9 @@ __kernel void flash_attn_f16_q1(
     const ulong mask_nb2,
     const ulong mask_nb3,
     const int mask_ne2,
-    const int mask_ne3
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
 ) {
     const int tid = get_local_id(0);
     const int head_batch_idx = get_global_id(1);
@@ -247,7 +265,12 @@ __kernel void flash_attn_f16_q1(
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
 
-    ACC_TYPE m_i = -INFINITY;
+    const global ACC_TYPE* sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
     for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
         const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
         const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
@@ -320,7 +343,11 @@ __kernel void flash_attn_f16_q1(
 
     const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
     global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
-    const ACC_TYPE l_final = local_l[0];
+    ACC_TYPE l_final = local_l[0];
+
+    if (sinks_ptr != NULL) {
+        l_final += exp(sinks_ptr[head_idx] - m_final);
+    }
 
     if (l_final > 0.0f) {
         const ACC_TYPE l_inv = 1.0f / l_final;

ggml/src/ggml-opencl/kernels/flash_attn_f32.cl

@@ -49,7 +49,9 @@ __kernel void flash_attn_f32(
     const ulong mask_nb2,
     const ulong mask_nb3,
     const int mask_ne2,
-    const int mask_ne3
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
 ) {
     const int tid = get_local_id(0);
     const int block_q_idx = get_group_id(0);
@@ -171,6 +173,20 @@ __kernel void flash_attn_f32(
     }
 
     if (my_query_row < n_q) {
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+
+            const ACC_TYPE scale_o = exp(m_i - m_final);
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] *= scale_o;
+            }
+
+            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
+        }
+
         const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
         global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
         if (l_i > 0.0f) {
@@ -214,7 +230,9 @@ __kernel void flash_attn_f32_q1(
     const ulong mask_nb2,
     const ulong mask_nb3,
     const int mask_ne2,
-    const int mask_ne3
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
 ) {
     const int tid = get_local_id(0);
     const int head_batch_idx = get_global_id(1);
@@ -247,7 +265,12 @@ __kernel void flash_attn_f32_q1(
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
 
-    ACC_TYPE m_i = -INFINITY;
+    const global ACC_TYPE* sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
     for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
         const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
         const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
@@ -320,7 +343,11 @@ __kernel void flash_attn_f32_q1(
 
     const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
     global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
-    const ACC_TYPE l_final = local_l[0];
+    ACC_TYPE l_final = local_l[0];
+
+    if (sinks_ptr != NULL) {
+        l_final += exp(sinks_ptr[head_idx] - m_final);
+    }
 
     if (l_final > 0.0f) {
         const ACC_TYPE l_inv = 1.0f / l_final;

ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl

@@ -52,7 +52,9 @@ __kernel void flash_attn_f32_f16(
     const ulong mask_nb2,
     const ulong mask_nb3,
     const int mask_ne2,
-    const int mask_ne3
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
 ) {
     const int tid = get_local_id(0);
     const int block_q_idx = get_group_id(0);
@@ -174,6 +176,20 @@ __kernel void flash_attn_f32_f16(
     }
 
     if (my_query_row < n_q) {
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+
+            const ACC_TYPE scale_o = exp(m_i - m_final);
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] *= scale_o;
+            }
+
+            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
+        }
+
         const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
         global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
         if (l_i > 0.0f) {
@@ -217,7 +233,9 @@ __kernel void flash_attn_f32_f16_q1(
     const ulong mask_nb2,
     const ulong mask_nb3,
     const int mask_ne2,
-    const int mask_ne3
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
 ) {
     const int tid = get_local_id(0);
     const int head_batch_idx = get_global_id(1);
@@ -250,7 +268,12 @@ __kernel void flash_attn_f32_f16_q1(
 
     float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
 
-    ACC_TYPE m_i = -INFINITY;
+    const global ACC_TYPE* sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
     for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
         const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
         const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
@@ -323,7 +346,11 @@ __kernel void flash_attn_f32_f16_q1(
 
     const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
     global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
-    const ACC_TYPE l_final = local_l[0];
+    ACC_TYPE l_final = local_l[0];
+
+    if (sinks_ptr != NULL) {
+        l_final += exp(sinks_ptr[head_idx] - m_final);
+    }
 
     if (l_final > 0.0f) {
         const ACC_TYPE l_inv = 1.0f / l_final;

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

@@ -360,6 +360,13 @@ 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);
 
@@ -386,15 +393,18 @@ 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;
@@ -452,12 +462,15 @@ 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];
@@ -1357,6 +1370,7 @@ 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];
 };
 
@@ -2764,11 +2778,11 @@ 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]) \
@@ -2911,60 +2925,91 @@ 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
         }
     }
 
@@ -3056,10 +3101,17 @@ 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);
@@ -3578,11 +3630,12 @@ 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);
@@ -4038,11 +4091,18 @@ 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;
     }
 
@@ -4281,15 +4341,16 @@ 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(), ',', ' ');
@@ -4304,8 +4365,6 @@ 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");
@@ -4410,6 +4469,19 @@ 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);
     }
 }
@@ -4556,9 +4628,22 @@ 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:
@@ -4590,7 +4675,7 @@ 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.
@@ -4605,6 +4690,13 @@ 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];
 }
 
@@ -4674,7 +4766,7 @@ 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) {
@@ -5587,20 +5679,20 @@ 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);
@@ -5720,12 +5812,15 @@ 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) ||
@@ -5791,7 +5886,7 @@ 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;
@@ -5828,7 +5923,7 @@ 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;
         }
@@ -5845,10 +5940,15 @@ 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,
@@ -5863,6 +5963,51 @@ 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];
@@ -5917,22 +6062,7 @@ 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;
@@ -5944,14 +6074,47 @@ 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)) {
@@ -5960,7 +6123,7 @@ 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;
         }
 
@@ -5971,6 +6134,9 @@ 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;
     }
@@ -6001,6 +6167,9 @@ 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;
@@ -6011,9 +6180,7 @@ 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 });
     }
@@ -6029,6 +6196,17 @@ 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;
@@ -6053,6 +6231,12 @@ 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,
@@ -6060,13 +6244,13 @@ 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;
     }
 }
@@ -11483,15 +11667,29 @@ 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;
         }
     }

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

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);
+    }
+}

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;

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);
                     }
                 }
             }

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

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
     }
 }
 

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
 

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -206,6 +206,22 @@ 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) {
@@ -402,7 +418,7 @@ 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
@@ -495,8 +511,20 @@ 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"}}));
@@ -579,7 +607,12 @@ 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"}});
 
@@ -821,12 +854,21 @@ 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);
     }

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -611,6 +611,8 @@ static bool ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node) {
         case GGML_OP_NONE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_RESHAPE:
             return false;
         case GGML_OP_CPY:
             {
@@ -1062,6 +1064,8 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_NONE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_RESHAPE:
             return true;
         case GGML_OP_CPY:
         case GGML_OP_SET_ROWS:

ggml/src/gguf.cpp

@@ -273,7 +273,7 @@ struct gguf_reader {
     }
 
     bool read(std::string & dst) const {
-        uint64_t size = -1;
+        uint64_t size = 0;
         if (!read(size)) {
             return false;
         }
@@ -523,7 +523,7 @@ 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",

include/llama.h

@@ -206,7 +206,7 @@ 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);
@@ -1156,11 +1156,6 @@ 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);

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);
         }
@@ -1366,7 +1366,7 @@ 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) {
+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) {
@@ -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;
     }

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(

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

tests/test-sampling.cpp

@@ -197,10 +197,10 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
     sampler_tester tester(n_vocab);
 
           llama_token min_token_id = 0;
-    const llama_token max_token_id = n_vocab-1;
+    const llama_token max_token_id = n_vocab - 1;
 
     for (auto s : samplers_sequence) {
-        switch (s){
+        switch (s) {
             case 'k': tester.apply(llama_sampler_init_top_k(top_k)); break;
             case 'y': GGML_ABORT("typical test not implemented");
             case 'p': tester.apply(llama_sampler_init_top_p(top_p, 1)); break;
@@ -243,10 +243,10 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
             }
 
             GGML_ASSERT(size == expected_size);
-            GGML_ASSERT(cur_p.data[0].id == max_token_id);
-            GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
+            GGML_ASSERT(!cur_p.sorted || cur_p.data[0].id == max_token_id);
+            GGML_ASSERT(!cur_p.sorted || cur_p.data[expected_size-1].id == min_token_id);
         } else if (s == 'm') {
-            int expected_size = ceilf((1.0f-min_p) * n_vocab);
+            int expected_size = ceilf((1.0f - min_p) * n_vocab);
             expected_size = std::max(expected_size, 1);
             expected_size = std::min(expected_size, size);
 
@@ -256,14 +256,14 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
             min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1));
 
             GGML_ASSERT(size == expected_size);
-            GGML_ASSERT(cur_p.data[0].id == max_token_id);
-            GGML_ASSERT(cur_p.data[expected_size-1].id == min_token_id);
+            GGML_ASSERT(!cur_p.sorted || cur_p.data[0].id == max_token_id);
+            GGML_ASSERT(!cur_p.sorted || cur_p.data[expected_size-1].id == min_token_id);
         } else {
             GGML_ABORT("fatal error");
         }
     }
 
-    printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%05d top_p=%f min_p=%f\n",
+    printf("Sampler queue %3s OK with n_vocab=%05zu top_k=%5d top_p=%f min_p=%f\n",
            samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
 }
 
@@ -308,28 +308,28 @@ static void test_perf() {
 int main(void) {
     ggml_time_init();
 
-    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
-    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f);
+    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f);
+    test_temp({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.0f, 1.0f}, 0.0f);
 
-    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f, 0.0f, 1.0f);
-    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.0f, 0.0f, 1.0f);
+    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f, 0.0f, 1.0f);
+    test_temp_ext({0.1f, 0.2f, 0.3f, 0.4f}, {0.0f, 0.0f, 0.0f, 1.0f}, 0.0f, 0.0f, 1.0f);
 
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 1);
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 3);
     test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 4);
-    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0);
+    test_top_k({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0);
 
     test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f}, 0);
     test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f}, 0.7f);
     test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.44444f, 0.33333f, 0.22222f}, 0.8f);
-    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1.0f);
+    test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 1.0f);
 
-    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f);
-    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f);
-    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f},            0.26f);
-    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f},            0.49f);
-    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f},                       0.51f);
-    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f},                       0.74f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f/1.0f, 0.2f/1.0f, 0.3f/1.0f, 0.4f/1.0f}, 0.00f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f/1.0f, 0.2f/1.0f, 0.3f/1.0f, 0.4f/1.0f}, 0.24f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.2f/0.9f, 0.3f/0.9f, 0.4f/0.9f},            0.26f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.2f/0.9f, 0.3f/0.9f, 0.4f/0.9f},            0.49f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.3f/0.7f, 0.4f/0.7f},                       0.51f);
+    test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.3f/0.7f, 0.4f/0.7f},                       0.74f);
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  0.76f);
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.00f);
     test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f},                                  1.05f);
@@ -345,23 +345,23 @@ int main(void) {
     test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f},            0.5f);
     test_typical({0.4f, 0.2f, 0.2f, 0.2f},     {0.2f, 0.2f, 0.2f}, 0.5f);
 
-    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0},   50.0f, 0.0f, 0.0f);
-    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0},       50.0f, 0.0f, 0.0f);
-    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0, 0.25f, 0.25f, 0.25f, 0.25f},   50.0f, 0.0f, 0.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0, 0, 0, 0.5f, 0.5f},       50.0f, 0.0f, 0.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0, 0, 0, 0.5f, 0.5f}, 50.0f, 0.0f, 0.0f);
 
-    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 1.0f, 5.0f, 5.0f);
-    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
-    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0},             {0.000011f, 0.249997f, 0.249997f, 0.249997f, 0.249997f}, 1.0f, 5.0f, 5.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2},       {0.000023f, 0.000023f, 0.000023f, 0.499966f, 0.499966f}, 1.0f, 5.0f, 5.0f);
+    test_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.000000f, 0.000023f, 0.000023f, 0.499977f, 0.499977f}, 1.0f, 5.0f, 5.0f);
 
 
     test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1}, {0.25f, 0.25f, 0.25f, 0.25f}, 1.0f, 1.1f, 2, 4, {});
-    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.296923f, 0.109232f}, 1.0f, 1.1f, 2, 5, {});
+    test_dry({0.25f, 0.25f, 0.25f, 0.25f}, {0, 1, 2, 0, 1}, {0.296923f, 0.296923f, 0.109232f, 0.296923f}, 1.0f, 1.1f, 2, 5, {});
     test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 2, 6, {{3}});
-    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.241818f, 0.241818f, 0.032727f}, 2.0f, 1.1f, 2, 5, {});
+    test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.032727f, 0.241818f, 0.241818f}, 2.0f, 1.1f, 2, 5, {});
     test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
 
     test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f, 0.0f, 0.0f}, 1.00f);
-    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345
+    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.1f, 0.2f, 0.3f, 0.4f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345
     test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 3.00f);
 
     test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
@@ -372,7 +372,7 @@ int main(void) {
     test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12);
 
     test_sampler_queue(10000, "k",   100, 1.0000f, 1.0f);
-    test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f);
+    test_sampler_queue(10000, "p", 10000, 0.0003f, 1.0f);
     test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f);
     test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f);
     test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f);

tools/server/README.md

@@ -37,7 +37,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
 | `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
 | `--cpu-strict <0\|1>` | use strict CPU placement (default: 0)<br/> |
-| `--prio N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
+| `--prio N` | set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: 0)<br/> |
 | `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50)<br/> |
 | `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
 | `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
@@ -49,6 +49,8 @@ The project is under active development, and we are [looking for feedback and co
 | `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
 | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
 | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
+| `--swa-full` | use full-size SWA cache (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)<br/>(env: LLAMA_ARG_SWA_FULL) |
+| `--kv-unified, -kvu` | use single unified KV buffer for the KV cache of all sequences (default: false)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)<br/>(env: LLAMA_ARG_KV_SPLIT) |
 | `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `--no-perf` | disable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_NO_PERF) |
 | `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
@@ -63,6 +65,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
 | `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
+| `-nr, --no-repack` | disable weight repacking<br/>(env: LLAMA_ARG_NO_REPACK) |
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
 | `-ctv, --cache-type-v TYPE` | KV cache data type for V<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
 | `-dt, --defrag-thold N` | KV cache defragmentation threshold (DEPRECATED)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
@@ -73,12 +76,15 @@ The project is under active development, and we are [looking for feedback and co
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
 | `--override-tensor, -ot <tensor name pattern>=<buffer type>,...` | override tensor buffer type |
+| `--cpu-moe, -cmoe` | keep all Mixture of Experts (MoE) weights in the CPU<br/>(env: LLAMA_ARG_CPU_MOE) |
+| `--n-cpu-moe, -ncmoe N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU<br/>(env: LLAMA_ARG_N_CPU_MOE) |
 | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
 | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs<br/>(env: LLAMA_ARG_SPLIT_MODE) |
 | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1<br/>(env: LLAMA_ARG_TENSOR_SPLIT) |
 | `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)<br/>(env: LLAMA_ARG_MAIN_GPU) |
 | `--check-tensors` | check model tensor data for invalid values (default: false) |
 | `--override-kv KEY=TYPE:VALUE` | advanced option to override model metadata by key. may be specified multiple times.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false |
+| `--no-op-offload` | disable offloading host tensor operations to device (default: false) |
 | `--lora FNAME` | path to LoRA adapter (can be repeated to use multiple adapters) |
 | `--lora-scaled FNAME SCALE` | path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters) |
 | `--control-vector FNAME` | add a control vector<br/>note: this argument can be repeated to add multiple control vectors |
@@ -96,9 +102,12 @@ The project is under active development, and we are [looking for feedback and co
 | `--log-file FNAME` | Log to file |
 | `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |
 | `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) |
+| `--offline` | Offline mode: forces use of cache, prevents network access<br/>(env: LLAMA_OFFLINE) |
 | `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.<br/>(env: LLAMA_LOG_VERBOSITY) |
 | `--log-prefix` | Enable prefix in log messages<br/>(env: LLAMA_LOG_PREFIX) |
 | `--log-timestamps` | Enable timestamps in log messages<br/>(env: LLAMA_LOG_TIMESTAMPS) |
+| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) |
+| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) |
 
 
 **Sampling params**
@@ -113,6 +122,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
 | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
 | `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
 | `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
 | `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
 | `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
@@ -141,7 +151,10 @@ The project is under active development, and we are [looking for feedback and co
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--no-context-shift` | disables context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
+| `--swa-checkpoints N` | max number of SWA checkpoints per slot to create (default: 3)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)<br/>(env: LLAMA_ARG_SWA_CHECKPOINTS) |
+| `--no-context-shift` | disables context shift on infinite text generation (default: enabled)<br/>(env: LLAMA_ARG_NO_CONTEXT_SHIFT) |
+| `--context-shift` | enables context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
+| `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode<br/> |
 | `-sp, --special` | special tokens output enabled (default: false) |
 | `--no-warmup` | skip warming up the model with an empty run |
 | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
@@ -152,10 +165,14 @@ The project is under active development, and we are [looking for feedback and co
 | `--mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md<br/>(env: LLAMA_ARG_MMPROJ_URL) |
 | `--no-mmproj` | explicitly disable multimodal projector, useful when using -hf<br/>(env: LLAMA_ARG_NO_MMPROJ) |
 | `--no-mmproj-offload` | do not offload multimodal projector to GPU<br/>(env: LLAMA_ARG_NO_MMPROJ_OFFLOAD) |
+| `--override-tensor-draft, -otd <tensor name pattern>=<buffer type>,...` | override tensor buffer type for draft model |
+| `--cpu-moe-draft, -cmoed` | keep all Mixture of Experts (MoE) weights in the CPU for the draft model<br/>(env: LLAMA_ARG_CPU_MOE_DRAFT) |
+| `--n-cpu-moe-draft, -ncmoed N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model<br/>(env: LLAMA_ARG_N_CPU_MOE_DRAFT) |
 | `-a, --alias STRING` | set alias for model name (to be used by REST API)<br/>(env: LLAMA_ARG_ALIAS) |
 | `--host HOST` | ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
 | `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
 | `--path PATH` | path to serve static files from (default: )<br/>(env: LLAMA_ARG_STATIC_PATH) |
+| `--api-prefix PREFIX` | prefix path the server serves from, without the trailing slash (default: )<br/>(env: LLAMA_ARG_API_PREFIX) |
 | `--no-webui` | Disable the Web UI (default: enabled)<br/>(env: LLAMA_ARG_NO_WEBUI) |
 | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
 | `--reranking, --rerank` | enable reranking endpoint on server (default: disabled)<br/>(env: LLAMA_ARG_RERANKING) |
@@ -163,23 +180,25 @@ The project is under active development, and we are [looking for feedback and co
 | `--api-key-file FNAME` | path to file containing API keys (default: none) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
-| `--chat-template-kwargs STRING` | JSON object containing additional params for the json template parser. Example: `--chat_template_kwargs "{\"enable_thinking\":false}`"<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
+| `--chat-template-kwargs STRING` | sets additional params for the json template parser<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
 | `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
-| `--slots` | enable slots monitoring endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
+| `--slots` | enable slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
 | `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
 | `--jinja` | use jinja template for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
-| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)<br/>(default: deepseek)<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
-| `--no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/>(env: LLAMA_ARG_NO_PREFILL_ASSISTANT) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/><br/>(env: LLAMA_ARG_NO_PREFILL_ASSISTANT) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
+| `-td, --threads-draft N` | number of threads to use during generation (default: same as --threads) |
+| `-tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |
 | `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
 | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
 | `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
@@ -187,8 +206,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
 | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
-| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for speculative decoding model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) |
-| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for speculative decoding model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) |
+| `--spec-replace TARGET DRAFT` | translate the string in TARGET into DRAFT if the draft model and main model are not compatible |
 | `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) |
 | `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall |
 | `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) |
@@ -199,6 +217,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--fim-qwen-7b-default` | use default Qwen 2.5 Coder 7B (note: can download weights from the internet) |
 | `--fim-qwen-7b-spec` | use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet) |
 | `--fim-qwen-14b-spec` | use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet) |
+| `--fim-qwen-30b-default` | use default Qwen 3 Coder 30B A3B Instruct (note: can download weights from the internet) |
 
 
 Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
@@ -865,25 +884,23 @@ Same as the `/v1/embeddings` endpoint.
 
 ### GET `/slots`: Returns the current slots processing state
 
-> [!WARNING]
-> This endpoint is intended for debugging and may be modified in future versions. For security reasons, we strongly advise against enabling it in production environments.
-
-This endpoint is disabled by default and can be enabled with `--slots`
+This endpoint is enabled by default and can be disabled with `--no-slots`. It can be used to query various per-slot metrics, such as speed, processed tokens, sampling parameters, etc.
 
 If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots.
 
 **Response format**
 
-Example:
+<details>
+<summary>Example with 2 slots</summary>
 
 ```json
 [
   {
     "id": 0,
-    "id_task": -1,
-    "n_ctx": 1024,
+    "id_task": 135,
+    "n_ctx": 65536,
     "speculative": false,
-    "is_processing": false,
+    "is_processing": true,
     "params": {
       "n_predict": -1,
       "seed": 4294967295,
@@ -893,6 +910,7 @@ Example:
       "top_k": 40,
       "top_p": 0.949999988079071,
       "min_p": 0.05000000074505806,
+      "top_n_sigma": -1.0,
       "xtc_probability": 0.0,
       "xtc_threshold": 0.10000000149011612,
       "typical_p": 1.0,
@@ -903,17 +921,10 @@ Example:
       "dry_multiplier": 0.0,
       "dry_base": 1.75,
       "dry_allowed_length": 2,
-      "dry_penalty_last_n": -1,
-      "dry_sequence_breakers": [
-        "\n",
-        ":",
-        "\"",
-        "*"
-      ],
+      "dry_penalty_last_n": 131072,
       "mirostat": 0,
       "mirostat_tau": 5.0,
       "mirostat_eta": 0.10000000149011612,
-      "stop": [],
       "max_tokens": -1,
       "n_keep": 0,
       "n_discard": 0,
@@ -921,8 +932,12 @@ Example:
       "stream": true,
       "n_probs": 0,
       "min_keep": 0,
-      "grammar": "",
+      "chat_format": "GPT-OSS",
+      "reasoning_format": "none",
+      "reasoning_in_content": false,
+      "thinking_forced_open": false,
       "samplers": [
+        "penalties",
         "dry",
         "top_k",
         "typ_p",
@@ -932,22 +947,89 @@ Example:
         "temperature"
       ],
       "speculative.n_max": 16,
-      "speculative.n_min": 5,
-      "speculative.p_min": 0.8999999761581421,
-      "timings_per_token": false
+      "speculative.n_min": 0,
+      "speculative.p_min": 0.75,
+      "timings_per_token": false,
+      "post_sampling_probs": false,
+      "lora": []
     },
-    "prompt": "",
     "next_token": {
       "has_next_token": true,
       "has_new_line": false,
       "n_remain": -1,
-      "n_decoded": 0,
-      "stopping_word": ""
+      "n_decoded": 0
+    }
+  },
+  {
+    "id": 1,
+    "id_task": 0,
+    "n_ctx": 65536,
+    "speculative": false,
+    "is_processing": true,
+    "params": {
+      "n_predict": -1,
+      "seed": 4294967295,
+      "temperature": 0.800000011920929,
+      "dynatemp_range": 0.0,
+      "dynatemp_exponent": 1.0,
+      "top_k": 40,
+      "top_p": 0.949999988079071,
+      "min_p": 0.05000000074505806,
+      "top_n_sigma": -1.0,
+      "xtc_probability": 0.0,
+      "xtc_threshold": 0.10000000149011612,
+      "typical_p": 1.0,
+      "repeat_last_n": 64,
+      "repeat_penalty": 1.0,
+      "presence_penalty": 0.0,
+      "frequency_penalty": 0.0,
+      "dry_multiplier": 0.0,
+      "dry_base": 1.75,
+      "dry_allowed_length": 2,
+      "dry_penalty_last_n": 131072,
+      "mirostat": 0,
+      "mirostat_tau": 5.0,
+      "mirostat_eta": 0.10000000149011612,
+      "max_tokens": -1,
+      "n_keep": 0,
+      "n_discard": 0,
+      "ignore_eos": false,
+      "stream": true,
+      "n_probs": 0,
+      "min_keep": 0,
+      "chat_format": "GPT-OSS",
+      "reasoning_format": "none",
+      "reasoning_in_content": false,
+      "thinking_forced_open": false,
+      "samplers": [
+        "penalties",
+        "dry",
+        "top_k",
+        "typ_p",
+        "top_p",
+        "min_p",
+        "xtc",
+        "temperature"
+      ],
+      "speculative.n_max": 16,
+      "speculative.n_min": 0,
+      "speculative.p_min": 0.75,
+      "timings_per_token": false,
+      "post_sampling_probs": false,
+      "lora": []
+    },
+    "next_token": {
+      "has_next_token": true,
+      "has_new_line": true,
+      "n_remain": -1,
+      "n_decoded": 136
     }
   }
 ]
 ```
 
+</details>
+
 ### GET `/metrics`: Prometheus compatible metrics exporter
 
 This endpoint is only accessible if `--metrics` is set.

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)},

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:

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) {