whispercpp 1.3.4 → 1.3.6

data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c

@@ -5,7 +5,6 @@
 #include "ggml-backend.h"
 #include "traits.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "quants.h"
 #include "ggml-threading.h"
@@ -14,6 +13,7 @@
 #include "vec.h"
 #include "ops.h"
 #include "ggml.h"
+#include "common.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -75,12 +75,20 @@
 // precomputed f32 table for f16 (256 KB) (simd-mappings.h)
 float ggml_table_f32_f16[1 << 16];
 
+// precomputed f32 table for e8m0 half (1 KB) (simd-mappings.h)
+float ggml_table_f32_e8m0_half[1 << 8];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int sve_cnt;
 } ggml_arm_arch_features = { 0 };
 #endif
 
+#if defined(__riscv)
+struct ggml_riscv_arch_features_type {
+    int rvv_vlen;
+} ggml_riscv_arch_features = { 0 };
+#endif
 
 #if defined(_WIN32)
 
@@ -187,6 +195,9 @@ typedef void * thread_ret_t;
 
 typedef pthread_t ggml_thread_t;
 
+#define GGML_THREADPOOL_N_THREADS_MASK (0xffffU)
+#define GGML_THREADPOOL_N_THREADS_BITS (16)
+
 #if defined(__APPLE__)
 #include <unistd.h>
 #include <mach/mach.h>
@@ -259,6 +270,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
+    [GGML_TYPE_NVFP4] = {
+        .from_float               = quantize_row_nvfp4,
+        .vec_dot                  = ggml_vec_dot_nvfp4_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q2_K] = {
         .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
@@ -449,7 +466,7 @@ struct ggml_threadpool {
     struct ggml_cplan  * cplan;
 
     // synchronization primitives
-    atomic_int n_graph;       // incremented when there is work to be done (i.e each graph)
+    atomic_int n_graph;       // updated when there is work to be done (i.e each graph) holds graph and active thread counts.
     atomic_int GGML_CACHE_ALIGN n_barrier;
     atomic_int GGML_CACHE_ALIGN n_barrier_passed;
     atomic_int GGML_CACHE_ALIGN current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads.
@@ -457,12 +474,10 @@ struct ggml_threadpool {
     // these are atomic as an annotation for thread-sanitizer
     atomic_bool stop;         // Used for stopping the threadpool altogether
     atomic_bool pause;        // Used for pausing the threadpool or individual threads
-    atomic_int abort;         // Used for aborting processing of a graph
+    atomic_int  abort;        // Used for aborting processing of a graph
 
     struct ggml_compute_state * workers;   // per thread state
-    int          n_threads_max; // number of threads in the pool
-    atomic_int   n_threads_cur; // number of threads used in the current graph
-
+    int          n_threads;   // Number of threads in the pool
     int32_t      prio;        // Scheduling priority
     uint32_t     poll;        // Polling level (0 - no polling)
 
@@ -490,6 +505,15 @@ static inline void ggml_thread_cpu_relax(void) {
 static inline void ggml_thread_cpu_relax(void) {
     _mm_pause();
 }
+#elif defined(__riscv)
+static inline void ggml_thread_cpu_relax(void) {
+    #ifdef __riscv_zihintpause
+        __asm__ __volatile__ ("pause");
+    #else
+        /* Encoding of the pause instruction */
+        __asm__ __volatile__ (".4byte 0x100000F");
+    #endif
+}
 #else
 static inline void ggml_thread_cpu_relax(void) {;}
 #endif
@@ -530,7 +554,7 @@ struct ggml_state {
 static struct ggml_state g_state = {0};
 
 void ggml_barrier(struct ggml_threadpool * tp) {
-    int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
+    int n_threads = atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK;
     if (n_threads == 1) {
         return;
     }
@@ -547,7 +571,7 @@ void ggml_barrier(struct ggml_threadpool * tp) {
         // last thread
         atomic_store_explicit(&tp->n_barrier, 0, memory_order_relaxed);
 
-        // exit barrier (fill seq-cst fence)
+        // exit barrier (full seq-cst fence)
         atomic_fetch_add_explicit(&tp->n_barrier_passed, 1, memory_order_seq_cst);
         return;
     }
@@ -683,19 +707,25 @@ bool ggml_is_numa(void) {
 }
 
 #if defined(__ARM_ARCH)
-
-#if defined(__linux__) && defined(__aarch64__)
-#include <sys/auxv.h>
-#endif
-
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
-    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-#endif
+    ggml_arm_arch_features.sve_cnt = svcntb();
 }
-
+#else
+static void ggml_init_arm_arch_features(void) {}
+#endif
 #endif // __ARM_ARCH
 
+#if defined(__riscv) && defined(__riscv_v_intrinsic)
+#include <riscv_vector.h>
+static void ggml_init_riscv_arch_features(void) {
+    ggml_riscv_arch_features.rvv_vlen = __riscv_vlenb();
+}
+#else
+static void ggml_init_riscv_arch_features(void) {}
+#endif
+
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
 
@@ -1608,13 +1638,8 @@ static void ggml_compute_forward_mul_mat_id(
             chunk_size = 64;
         }
 
-#if defined(__aarch64__)
-        // disable for ARM
-        const bool disable_chunking = true;
-#else
         // disable for NUMA
         const bool disable_chunking = ggml_is_numa();
-#endif // defined(__aarch64__)
 
         int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
         int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
@@ -1731,6 +1756,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_sum_rows(params, tensor);
             } break;
+        case GGML_OP_CUMSUM:
+            {
+                ggml_compute_forward_cumsum(params, tensor);
+            } break;
         case GGML_OP_MEAN:
             {
                 ggml_compute_forward_mean(params, tensor);
@@ -1807,22 +1836,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_cont(params, tensor);
             } break;
-        case GGML_OP_RESHAPE:
-            {
-                ggml_compute_forward_reshape(params, tensor);
-            } break;
-        case GGML_OP_VIEW:
-            {
-                ggml_compute_forward_view(params, tensor);
-            } break;
-        case GGML_OP_PERMUTE:
-            {
-                ggml_compute_forward_permute(params, tensor);
-            } break;
-        case GGML_OP_TRANSPOSE:
-            {
-                ggml_compute_forward_transpose(params, tensor);
-            } break;
         case GGML_OP_GET_ROWS:
             {
                 ggml_compute_forward_get_rows(params, tensor);
@@ -1939,10 +1952,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_argsort(params, tensor);
             } break;
+        case GGML_OP_TOP_K:
+            {
+                ggml_compute_forward_top_k(params, tensor);
+            } break;
         case GGML_OP_LEAKY_RELU:
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
             } break;
+        case GGML_OP_TRI:
+            {
+                ggml_compute_forward_tri(params, tensor);
+            } break;
+        case GGML_OP_FILL:
+            {
+                ggml_compute_forward_fill(params, tensor);
+            } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_compute_forward_flash_attn_ext(params, tensor);
@@ -1998,6 +2023,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rwkv_wkv7(params, tensor);
             } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                ggml_compute_forward_solve_tri(params, tensor);
+            } break;
+        case GGML_OP_GATED_DELTA_NET:
+            {
+                ggml_compute_forward_gated_delta_net(params, tensor);
+            } break;
         case GGML_OP_MAP_CUSTOM1:
             {
                 ggml_compute_forward_map_custom1(params, tensor);
@@ -2042,6 +2075,22 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 // nop
             } break;
+        case GGML_OP_RESHAPE:
+            {
+                // nop
+            } break;
+        case GGML_OP_PERMUTE:
+            {
+                // nop
+            } break;
+        case GGML_OP_VIEW:
+            {
+                // nop
+            } break;
+        case GGML_OP_TRANSPOSE:
+            {
+                // nop
+            } break;
         case GGML_OP_COUNT:
             {
                 GGML_ABORT("fatal error");
@@ -2140,6 +2189,9 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ADD_ID:
         case GGML_OP_ADD1:
         case GGML_OP_ACC:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
+        case GGML_OP_FILL:
             {
                 n_tasks = n_threads;
             } break;
@@ -2157,6 +2209,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = 1;
             } break;
         case GGML_OP_COUNT_EQUAL:
+        case GGML_OP_SOLVE_TRI:
+        case GGML_OP_GATED_DELTA_NET:
             {
                 n_tasks = n_threads;
             } break;
@@ -2179,6 +2233,12 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_SOFTPLUS:
+                case GGML_UNARY_OP_EXPM1:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     {
                         n_tasks = 1;
                     } break;
@@ -2187,6 +2247,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_XIELU:
                     {
                         n_tasks = n_threads;
                     } break;
@@ -2284,6 +2345,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
+        case GGML_OP_TOP_K:
         case GGML_OP_FLASH_ATTN_EXT:
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
@@ -2426,7 +2488,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
 
     if (prio != GGML_SCHED_PRIO_LOW) {
         // Tell Windows that this thread should not be throttled (needs its own CPU core).
-        // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
+        // Newer Windows 11 versions aggressively park (offline) CPU cores and often place
         // all our threads onto the first 4 cores which results in terrible performance with
         // n_threads > 4
         #if _WIN32_WINNT >= 0x0602
@@ -2595,7 +2657,7 @@ static void ggml_thread_cpumask_next(const bool * global_mask, bool * local_mask
 void ggml_threadpool_free(struct ggml_threadpool* threadpool) {
     if (!threadpool) return;
 
-    const int n_threads = threadpool->n_threads_max;
+    const int n_threads = threadpool->n_threads;
 
 #ifndef GGML_USE_OPENMP
     struct ggml_compute_state* workers = threadpool->workers;
@@ -2671,9 +2733,14 @@ struct ggml_cplan ggml_graph_plan(
         //GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads);
     }
     if (n_threads <= 0) {
-        n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS;
+        n_threads = threadpool ? threadpool->n_threads : GGML_DEFAULT_N_THREADS;
     }
 
+#if defined(__EMSCRIPTEN__) && !defined(__EMSCRIPTEN_PTHREADS__)
+    // Emscripten without pthreads support can only use a single thread
+    n_threads = 1;
+#endif
+
     size_t work_size = 0;
 
     struct ggml_cplan cplan;
@@ -2807,12 +2874,26 @@ struct ggml_cplan ggml_graph_plan(
                         cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
                         cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
                     } break;
+                case GGML_OP_TOP_K:
+                    {
+                        cur += sizeof(int32_t)*node->src[0]->ne[0]*n_tasks;
+                    } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
-                        const int64_t ne10 = node->src[1]->ne[0]; // DK
-                        const int64_t ne20 = node->src[2]->ne[0]; // DV
+                        const int64_t neq2 = node->src[0]->ne[2]; // number of query heads
+                        const int64_t DK = node->src[1]->ne[0];
+                        const int64_t DV = node->src[2]->ne[0];
 
-                        cur = sizeof(float)*(1*ne10 + 2*ne20)*n_tasks; // 1x head size K + 2x head size V (per thread)
+                        // Tiled flash attention scratch (tile sizes defined in common.h)
+                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + K_f32 + padding
+                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV + GGML_FA_TILE_KV*DK)*n_tasks;
+
+                        // Decode path: n_kv_chunks = n_tasks (one chunk per thread)
+                        // Per-thread: VKQ accmulator (DV), partial M, partial S + intra-thread scratch for V, Q and VKQ
+                        size_t n_chunks = n_tasks;
+                        size_t decode   = sizeof(float)*(neq2*n_chunks*(2+DV) + n_tasks*(DK + 2*DV));
+
+                        cur += MAX(prefill, decode);
                     } break;
                 case GGML_OP_FLASH_ATTN_BACK:
                     {
@@ -2835,6 +2916,11 @@ struct ggml_cplan ggml_graph_plan(
                     {
                         cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
                     } break;
+                case GGML_OP_GATED_DELTA_NET:
+                    {
+                        const int64_t S_v = node->src[2]->ne[0];
+                        cur = S_v * sizeof(float) * n_tasks;
+                    } break;
                 case GGML_OP_COUNT:
                     {
                         GGML_ABORT("fatal error");
@@ -2869,16 +2955,32 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
     set_numa_thread_affinity(state->ith);
 
     struct ggml_compute_params params = {
-        /*.ith       =*/ state->ith,
-        /*.nth       =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
-        /*.wsize     =*/ cplan->work_size,
-        /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ tp,
+        /*.ith        =*/ state->ith,
+        /*.nth        =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
+        /*.wsize      =*/ cplan->work_size,
+        /*.wdata      =*/ cplan->work_data,
+        /*.threadpool =*/ tp,
+        /*.use_ref    =*/ cplan->use_ref,
     };
 
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
+
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
+        if (ggml_op_is_empty(node->op)) {
+            // skip NOPs
+            continue;
+        }
+
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&
@@ -2892,6 +2994,12 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
     }
 
+#ifdef GGML_USE_OPENMP
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p\n", state->ith, (const void *)cplan);
+#else
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d\n", state->ith, (const void *)cplan, state->last_graph);
+#endif
+
     ggml_barrier(state->threadpool);
 
     return 0;
@@ -2899,27 +3007,23 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
 #ifndef GGML_USE_OPENMP
 
-// check if thread is active
-static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
-    struct ggml_threadpool * threadpool = state->threadpool;
-    int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
-    return (state->ith < n_threads);
-}
-
 // check if thread is ready to proceed (exit from polling or sleeping)
+// returns true if loops should exit, sets state->pending to indicate new work
 static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
     if (state->pending || threadpool->stop || threadpool->pause) { return true; }
 
     // check for new graph/work
-    int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
-    if (new_graph != state->last_graph) {
-        state->pending    = ggml_graph_compute_thread_active(state);
-        state->last_graph = new_graph;
+    int n_graph   = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
+    int n_threads = n_graph & GGML_THREADPOOL_N_THREADS_MASK;
+    if (n_graph != state->last_graph) {
+        state->pending    = (state->ith < n_threads);
+        state->last_graph = n_graph;
+        return true;
     }
 
-    return state->pending;
+    return false;
 }
 
 // sync thread state after polling
@@ -2936,11 +3040,6 @@ static inline void ggml_graph_compute_thread_sync(struct ggml_compute_state * st
 static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
-    // Skip polling for unused threads
-    if (!ggml_graph_compute_thread_active(state)) {
-        return state->pending;
-    }
-
     // This seems to make 0 ... 100 a decent range for polling level across modern processors.
     // Perhaps, we can adjust it dynamically based on load and things.
     const uint64_t n_rounds = 1024UL * 128 * threadpool->poll;
@@ -3002,7 +3101,6 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
         ggml_graph_compute_check_for_work(state);
         if (state->pending) {
             state->pending = false;
-
             ggml_graph_compute_thread(state);
         }
     }
@@ -3017,14 +3115,15 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
 
     ggml_mutex_lock(&threadpool->mutex);
 
-    GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);
+    // Update the number of active threads and the graph count
+    int n_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed) >> GGML_THREADPOOL_N_THREADS_BITS;
+    n_graph = ((n_graph + 1) << GGML_THREADPOOL_N_THREADS_BITS) | (n_threads & GGML_THREADPOOL_N_THREADS_MASK);
 
-    // Update the number of active threads
-    atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+    GGML_PRINT_DEBUG("compute-kickoff: n_threads %d n_graph %d\n", n_threads, n_graph);
 
     // Indicate the graph is ready to be processed
     // We need the full seq-cst fence here because of the polling threads (used in thread_sync)
-    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_seq_cst);
+    atomic_store_explicit(&threadpool->n_graph, n_graph, memory_order_seq_cst);
 
     if (threadpool->pause) {
        // Update main thread prio and affinity to match the threadpool settings
@@ -3062,8 +3161,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
         threadpool->pause            = tpp->paused;
         threadpool->abort            = -1;
         threadpool->workers          = NULL;
-        threadpool->n_threads_max    = tpp->n_threads;
-        threadpool->n_threads_cur    = tpp->n_threads;
+        threadpool->n_threads        = tpp->n_threads;
         threadpool->poll             = tpp->poll;
         threadpool->prio             = tpp->prio;
         threadpool->ec               = GGML_STATUS_SUCCESS;
@@ -3158,7 +3256,7 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
             {
                 // update the number of threads from the actual number of threads that we got from OpenMP
                 n_threads = omp_get_num_threads();
-                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+                atomic_store_explicit(&threadpool->n_graph, n_threads, memory_order_relaxed);
             }
 
             // Apply thread CPU mask and priority
@@ -3171,13 +3269,13 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
             ggml_graph_compute_thread(&threadpool->workers[ith]);
         }
     } else {
-        atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
+        atomic_store_explicit(&threadpool->n_graph, 1, memory_order_relaxed);
         ggml_graph_compute_thread(&threadpool->workers[0]);
     }
 #else
-    if (n_threads > threadpool->n_threads_max) {
-        GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
-        n_threads = threadpool->n_threads_max;
+    if (n_threads > threadpool->n_threads) {
+        GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads);
+        n_threads = threadpool->n_threads;
     }
 
     // Kick all threads to start the new graph
@@ -3264,6 +3362,33 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
+
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfhmin)
+    // calculate step size
+    const int epr = __riscv_vsetvlmax_e16m2();
+    const int step = epr * 2;
+    const int np = (n & ~(step - 1));
+
+    // unroll by 2
+    for (; i < np; i += step) {
+        vfloat16m2_t ax0 = __riscv_vle16_v_f16m2((const _Float16*)x + i, epr);
+        vfloat32m4_t ay0 = __riscv_vfwcvt_f_f_v_f32m4(ax0, epr);
+        __riscv_vse32_v_f32m4(y + i, ay0, epr);
+
+        vfloat16m2_t ax1 = __riscv_vle16_v_f16m2((const _Float16*)x + i + epr, epr);
+        vfloat32m4_t ay1 = __riscv_vfwcvt_f_f_v_f32m4(ax1, epr);
+        __riscv_vse32_v_f32m4(y + i + epr, ay1, epr);
+    }
+
+    // leftovers
+    int vl;
+    for (i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m2(n - i);
+        vfloat16m2_t ax0 = __riscv_vle16_v_f16m2((const _Float16*)x + i, vl);
+        vfloat32m4_t ay0 = __riscv_vfwcvt_f_f_v_f32m4(ax0, vl);
+        __riscv_vse32_v_f32m4(y + i, ay0, vl);
+    }
+
 #endif
 
     for (; i < n; ++i) {
@@ -3308,6 +3433,31 @@ void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
                                         (const __m128i *)(x + i))),
                                 16)));
     }
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfmin)
+    // calculate step size
+    const int epr = __riscv_vsetvlmax_e16m2();
+    const int step = epr * 2;
+    const int np = (n & ~(step - 1));
+
+    // unroll by 2
+    for (; i < np; i += step) {
+        vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16*)x + i, epr);
+        vfloat32m4_t ay0 = __riscv_vfwcvtbf16_f_f_v_f32m4(ax0, epr);
+        __riscv_vse32_v_f32m4(y + i, ay0, epr);
+
+        vbfloat16m2_t ax1 = __riscv_vle16_v_bf16m2((const __bf16*)x + i + epr, epr);
+        vfloat32m4_t ay1 = __riscv_vfwcvtbf16_f_f_v_f32m4(ax1, epr);
+        __riscv_vse32_v_f32m4(y + i + epr, ay1, epr);
+    }
+
+    // leftovers
+    int vl;
+    for (i = np; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m2(n - i);
+        vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16*)x + i, vl);
+        vfloat32m4_t ay0 = __riscv_vfwcvtbf16_f_f_v_f32m4(ax0, vl);
+        __riscv_vse32_v_f32m4(y + i, ay0, vl);
+    }
 #endif
     for (; i < n; i++) {
         y[i] = GGML_BF16_TO_FP32(x[i]);
@@ -3410,6 +3560,14 @@ int ggml_cpu_has_riscv_v(void) {
 #endif
 }
 
+int ggml_cpu_get_rvv_vlen(void) {
+#if defined(__riscv) && defined(__riscv_v_intrinsic)
+    return ggml_riscv_arch_features.rvv_vlen;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_f16c(void) {
 #if defined(__F16C__)
     return 1;
@@ -3550,6 +3708,11 @@ void ggml_cpu_init(void) {
                 ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }
 
+            // initialize E8M0 half table (256 entries)
+            for (int i = 0; i < (1 << 8); ++i) {
+                ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
+            }
+
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
 
             GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);
@@ -3557,13 +3720,17 @@ void ggml_cpu_init(void) {
 #ifdef GGML_USE_OPENMP
             //if (!getenv("OMP_WAIT_POLICY")) {
             //    // set the wait policy to active, so that OpenMP threads don't sleep
-            //    putenv("OMP_WAIT_POLICY=active");
+            //    setenv("OMP_WAIT_POLICY", "active", 0)
             //}
 
             if (!getenv("KMP_BLOCKTIME")) {
                 // set the time to wait before sleeping a thread
                 // this is less aggressive than setting the wait policy to active, but should achieve similar results in most cases
-                putenv("KMP_BLOCKTIME=200"); // 200ms
+#ifdef _WIN32
+                _putenv_s("KMP_BLOCKTIME", "200"); // 200ms
+#else
+                setenv("KMP_BLOCKTIME", "200", 0); // 200ms
+#endif
             }
 #endif
         }
@@ -3572,6 +3739,10 @@ void ggml_cpu_init(void) {
         ggml_init_arm_arch_features();
 #endif
 
+#if defined(__riscv)
+        ggml_init_riscv_arch_features();
+#endif
+
         is_first_call = false;
     }