whisper.rn 0.3.9 → 0.4.0-rc.0

package/cpp/whisper.cpp

@@ -3,11 +3,16 @@
 #include "coreml/whisper-encoder.h"
 #endif
 
-#if WHISPER_USE_OPENVINO
+#ifdef WSP_GGML_USE_METAL
+#  include "ggml-metal.h"
+#endif
+
+#ifdef WHISPER_USE_OPENVINO
 #include "openvino/whisper-openvino-encoder.h"
 #endif
 
 #include "ggml.h"
+#include "ggml-alloc.h"
 
 #include <algorithm>
 #include <cassert>
@@ -18,11 +23,13 @@
 #include <cstring>
 #include <fstream>
 #include <map>
+#include <set>
 #include <string>
 #include <thread>
 #include <vector>
 #include <regex>
 #include <random>
+#include <functional>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -114,8 +121,66 @@ static void byteswap_tensor(wsp_ggml_tensor * tensor) {
 //#define WHISPER_USE_FLASH_FF
 #define WHISPER_MAX_DECODERS 16
 
-#define WHISPER_USE_SCRATCH
-#define WHISPER_MAX_SCRATCH_BUFFERS 16
+//
+// ggml helpers
+//
+
+static void wsp_ggml_graph_compute_helper(
+        std::vector<uint8_t> & buf,
+                 wsp_ggml_cgraph * graph,
+                         int   n_threads,
+      whisper_abort_callback   abort_callback,
+                        void * abort_callback_data) {
+    struct wsp_ggml_cplan plan = wsp_ggml_graph_plan(graph, n_threads);
+
+    plan.abort_callback = abort_callback;
+    plan.abort_callback_data = abort_callback_data;
+
+    if (plan.work_size > 0) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    wsp_ggml_graph_compute(graph, &plan);
+}
+
+// faster matrix multiplications for tensors that do not have dimension 0 divisible by "pad"
+// the idea is to represent the original matrix multiplication:
+//
+//   Z = X @ Y
+//
+// with the sum of two matrix multiplications:
+//
+//   Z = (X_0 @ Y_0) + (X_1 @ Y_1)
+//
+// here X_0 and Y_0 are views of X and Y that have dimension 0 divisible by "pad"
+// and X_1 and Y_1 are the remaining views. X_1 and Y_1 end up being small matrices that can be processed with more
+// general-purpose kernels
+//
+static struct wsp_ggml_tensor * wsp_ggml_mul_mat_pad(struct wsp_ggml_context * ctx, struct wsp_ggml_tensor * x, struct wsp_ggml_tensor * y, int pad = 32) {
+    // use padding only if dimension 0 is at least 8 times larger than the padding
+    // else we won't get much benefit from the optimization
+    const int n_pad_req = 8;
+
+    if (x->ne[0] % pad == 0 || x->ne[0] / pad < n_pad_req) {
+        return wsp_ggml_mul_mat(ctx, x, y);
+    }
+
+    struct wsp_ggml_tensor * x_0 = wsp_ggml_view_3d(ctx, x, (x->ne[0]/pad)*pad, x->ne[1], x->ne[2], x->nb[1], x->nb[2], 0);
+    struct wsp_ggml_tensor * x_1 = wsp_ggml_view_3d(ctx, x,  x->ne[0]%pad,      x->ne[1], x->ne[2], x->nb[1], x->nb[2], x_0->ne[0]*x_0->nb[0]);
+
+    struct wsp_ggml_tensor * y_0 = wsp_ggml_view_3d(ctx, y, (y->ne[0]/pad)*pad, y->ne[1], y->ne[2], y->nb[1], y->nb[2], 0);
+    struct wsp_ggml_tensor * y_1 = wsp_ggml_view_3d(ctx, y,  y->ne[0]%pad,      y->ne[1], y->ne[2], y->nb[1], y->nb[2], y_0->ne[0]*y_0->nb[0]);
+
+    return wsp_ggml_add(ctx,
+            wsp_ggml_mul_mat(ctx, x_0, y_0),
+            wsp_ggml_mul_mat(ctx, x_1, y_1));
+}
+
+// TODO: check if other platforms can benefit from this optimization
+#if defined(WSP_GGML_USE_METAL)
+#define wsp_ggml_mul_mat wsp_ggml_mul_mat_pad
+#endif
 
 // available whisper models
 enum e_model {
@@ -231,38 +296,7 @@ static const std::map<std::string, std::pair<int, std::string>> g_lang = {
 
 static const size_t MB = 1ull*1024*1024;
 
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH0 = {
-    { MODEL_TINY,     62ull*MB },
-    { MODEL_BASE,     80ull*MB },
-    { MODEL_SMALL,   120ull*MB },
-    { MODEL_MEDIUM,  158ull*MB },
-    { MODEL_LARGE,   198ull*MB },
-};
-
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH1 = {
-    { MODEL_TINY,     18ull*MB },
-    { MODEL_BASE,     24ull*MB },
-    { MODEL_SMALL,    36ull*MB },
-    { MODEL_MEDIUM,   48ull*MB },
-    { MODEL_LARGE,    60ull*MB },
-};
-
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH2 = {
-    { MODEL_TINY,      4ull*MB },
-    { MODEL_BASE,      4ull*MB },
-    { MODEL_SMALL,     6ull*MB },
-    { MODEL_MEDIUM,    7ull*MB },
-    { MODEL_LARGE,     9ull*MB },
-};
-
-static const std::map<e_model, size_t> MEM_REQ_SCRATCH3 = {
-    { MODEL_TINY,      4ull*MB },
-    { MODEL_BASE,      4ull*MB },
-    { MODEL_SMALL,     6ull*MB },
-    { MODEL_MEDIUM,    7ull*MB },
-    { MODEL_LARGE,     9ull*MB },
-};
-
+// TODO: avoid using GGUF
 static const std::map<wsp_ggml_type, std::map<e_model, size_t>> MEM_REQ_MODEL = {
     { WSP_GGML_TYPE_F32,
         {
@@ -329,38 +363,6 @@ static const std::map<wsp_ggml_type, std::map<e_model, size_t>> MEM_REQ_MODEL =
     },
 };
 
-static const std::map<e_model, size_t> MEM_REQ_KV_SELF = {
-    { MODEL_TINY,      3ull*MB },
-    { MODEL_BASE,      6ull*MB },
-    { MODEL_SMALL,    16ull*MB },
-    { MODEL_MEDIUM,   43ull*MB },
-    { MODEL_LARGE,    71ull*MB },
-};
-
-static const std::map<e_model, size_t> MEM_REQ_KV_CROSS = {
-    { MODEL_TINY,      9ull*MB },
-    { MODEL_BASE,     18ull*MB },
-    { MODEL_SMALL,    53ull*MB },
-    { MODEL_MEDIUM,  141ull*MB },
-    { MODEL_LARGE,   235ull*MB },
-};
-
-static const std::map<e_model, size_t> MEM_REQ_ENCODE = {
-    { MODEL_TINY,     30ull*MB },
-    { MODEL_BASE,     38ull*MB },
-    { MODEL_SMALL,    56ull*MB },
-    { MODEL_MEDIUM,   74ull*MB },
-    { MODEL_LARGE,    94ull*MB },
-};
-
-static const std::map<e_model, size_t> MEM_REQ_DECODE = {
-    { MODEL_TINY,      3ull*MB },
-    { MODEL_BASE,      5ull*MB },
-    { MODEL_SMALL,    10ull*MB },
-    { MODEL_MEDIUM,   18ull*MB },
-    { MODEL_LARGE,    27ull*MB },
-};
-
 struct whisper_mel {
     int n_len;
     int n_len_org;
@@ -441,6 +443,7 @@ struct whisper_hparams {
     int32_t n_text_layer  = 4;
     int32_t n_mels        = 80;
     int32_t ftype         = 1;
+    float   eps           = 1e-5f;
 };
 
 // audio encoding layer
@@ -536,6 +539,7 @@ struct whisper_kv_cache {
 
     struct wsp_ggml_context * ctx;
 
+    // buf points to the memory allocated for both wsp_ggml_tensor 'k' and 'v' (see kv_cache_init)
     std::vector<uint8_t> buf;
 
     int n; // number of tokens currently in the cache
@@ -601,7 +605,7 @@ struct whisper_sequence {
 
 // TAGS: WHISPER_DECODER_INIT
 struct whisper_decoder {
-    // each decoders keeps its own KV-cache
+    // each decoder keeps its own KV-cache
     whisper_kv_cache kv_self;
 
     // the currently generated sequence of tokens
@@ -621,15 +625,75 @@ struct whisper_decoder {
     std::vector<whisper_token> tokens_tmp; // used for whisper_decode calls
 };
 
+// replace std::pair by using customized pair struct (reason: std::pair is very slow)
+template<typename A, typename B>
+struct whisper_pair {
+    A first;
+    B second;
+
+    // Define a constructor that takes two arguments.
+    whisper_pair(const A& a, const B& b) : first(a), second(b) {}
+    // Define a constructor that takes no argument.
+    whisper_pair() : first(A()), second(B()) {}
+};
+
+// beam-search helpers
+struct kv_buf {
+    std::vector<uint8_t> k;
+    std::vector<uint8_t> v;
+};
+
+// wsp_ggml_allocr wrapper for whisper usage
+struct whisper_allocr {
+    wsp_ggml_allocr * alloc = nullptr;
+
+    std::vector<uint8_t> meta;
+    std::vector<uint8_t> data;
+};
+
+static size_t whisper_allocr_size(struct whisper_allocr & allocr) {
+    return allocr.meta.size() + allocr.data.size();
+}
+
+// measure the memory usage of a graph and prepare the allocr's internal data buffer
+static void whisper_allocr_graph_init(struct whisper_allocr & allocr, std::function<struct wsp_ggml_cgraph *()> && get_graph) {
+    const int tensor_alignment = 32;
+
+    auto & alloc = allocr.alloc;
+    auto & meta  = allocr.meta;
+    auto & data  = allocr.data;
+
+    meta.resize(wsp_ggml_tensor_overhead()*WSP_GGML_MAX_NODES + wsp_ggml_graph_overhead());
+
+    alloc = wsp_ggml_allocr_new_measure(tensor_alignment);
+
+    const size_t alloc_size = wsp_ggml_allocr_alloc_graph(alloc, get_graph()) + tensor_alignment;
+
+    wsp_ggml_allocr_free(alloc);
+
+    data.resize(alloc_size);
+
+    alloc = wsp_ggml_allocr_new(data.data(), data.size(), tensor_alignment);
+}
+
+static void whisper_allocr_free(struct whisper_allocr & allocr) {
+    if (allocr.alloc) {
+        wsp_ggml_allocr_free(allocr.alloc);
+        allocr.alloc = nullptr;
+    }
+}
+
 struct whisper_state {
     int64_t t_sample_us = 0;
     int64_t t_encode_us = 0;
     int64_t t_decode_us = 0;
+    int64_t t_prompt_us = 0;
     int64_t t_mel_us = 0;
 
     int32_t n_sample = 0; // number of tokens sampled
     int32_t n_encode = 0; // number of encoder calls
-    int32_t n_decode = 0; // number of decoder calls
+    int32_t n_decode = 0; // number of decoder calls with n_tokens == 1 (text-generation)
+    int32_t n_prompt = 0; // number of decoder calls with n_tokens >  1 (prompt encoding)
     int32_t n_fail_p = 0; // number of logprob threshold failures
     int32_t n_fail_h = 0; // number of entropy threshold failures
 
@@ -640,12 +704,23 @@ struct whisper_state {
 
     whisper_decoder decoders[WHISPER_MAX_DECODERS] = {};
 
-    // memory buffers used by encode / decode contexts
-    std::vector<uint8_t> buf_compute;
-    std::vector<uint8_t> buf_scratch[WHISPER_MAX_SCRATCH_BUFFERS];
+    // buffer for swapping KV caches between decoders during beam-search
+    std::vector<kv_buf> kv_swap_bufs;
+
+    // reusable buffer for `struct wsp_ggml_graph_plan.work_data`
+    std::vector<uint8_t> work_buffer;
 
-    int    buf_last = 0;
-    size_t buf_max_size[WHISPER_MAX_SCRATCH_BUFFERS] = { 0 };
+    // ggml-alloc:
+    // - stores meta info about the intermediate tensors into the `meta` buffers
+    // - stores the actual tensor data into the `data` buffers
+    whisper_allocr alloc_conv;
+    whisper_allocr alloc_encode;
+    whisper_allocr alloc_cross;
+    whisper_allocr alloc_decode;
+
+    // result of the encoder
+    struct wsp_ggml_tensor * embd_conv = nullptr;
+    struct wsp_ggml_tensor * embd_enc  = nullptr;
 
     // decode output (2-dimensional array: [n_tokens][n_vocab])
     std::vector<float> logits;
@@ -654,7 +729,7 @@ struct whisper_state {
     std::vector<whisper_token>   prompt_past;
 
     // work container used to avoid memory allocations
-    std::vector<std::pair<double, whisper_vocab::id>> logits_id;
+    std::vector<whisper_pair<double, whisper_vocab::id>> logits_id;
 
     mutable std::mt19937 rng; // used for sampling at t > 0.0
 
@@ -665,6 +740,10 @@ struct whisper_state {
     whisper_coreml_context * ctx_coreml = nullptr;
 #endif
 
+#ifdef WSP_GGML_USE_METAL
+    wsp_ggml_metal_context * ctx_metal = nullptr;
+#endif
+
 #ifdef WHISPER_USE_OPENVINO
     whisper_openvino_context * ctx_openvino = nullptr;
 #endif
@@ -677,37 +756,6 @@ struct whisper_state {
 
     // [EXPERIMENTAL] speed-up techniques
     int32_t exp_n_audio_ctx = 0; // 0 - use default
-
-    void use_buf(struct wsp_ggml_context * ctx, int i) {
-#if defined(WHISPER_USE_SCRATCH)
-        size_t last_size = 0;
-
-        if (i == -1) {
-            last_size = wsp_ggml_set_scratch(ctx, { 0, 0, nullptr, });
-        } else {
-            auto & buf = buf_scratch[i];
-            last_size = wsp_ggml_set_scratch(ctx, { 0, buf.size(), buf.data(), });
-        }
-
-        if (buf_last >= 0) {
-            buf_max_size[buf_last] = std::max(buf_max_size[buf_last], last_size);
-        }
-
-        buf_last = i;
-#else
-        (void) i;
-        (void) ctx;
-#endif
-    }
-
-    size_t get_buf_max_mem(int i) const {
-#if defined(WHISPER_USE_SCRATCH)
-        return buf_max_size[i];
-#else
-        (void) i;
-        return 0;
-#endif
-    }
 };
 
 struct whisper_context {
@@ -730,6 +778,13 @@ static void whisper_default_log(const char * text) {
 
 static whisper_log_callback whisper_log = whisper_default_log;
 
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((gnu_format(printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
 static void log(const char * fmt, ...) {
     if (!whisper_log) return;
     char buf[1024];
@@ -747,10 +802,17 @@ static void read_safe(whisper_model_loader * loader, T & dest) {
 
 static bool kv_cache_init(
         const struct whisper_hparams & hparams,
-                        const size_t   mem_bytes,
              struct whisper_kv_cache & cache,
                            wsp_ggml_type   wtype,
                                  int   n_ctx) {
+    const int64_t n_text_state = hparams.n_text_state;
+    const int64_t n_text_layer = hparams.n_text_layer;
+
+    const int64_t n_mem      = n_text_layer*n_ctx;
+    const int64_t n_elements = n_text_state*n_mem;
+
+    const size_t mem_bytes = 2*(wsp_ggml_type_size(wtype)*n_elements + wsp_ggml_tensor_overhead());
+
     cache.buf.resize(mem_bytes);
 
     struct wsp_ggml_init_params params = {
@@ -766,12 +828,6 @@ static bool kv_cache_init(
         return false;
     }
 
-    const int n_text_state = hparams.n_text_state;
-    const int n_text_layer = hparams.n_text_layer;
-
-    const int n_mem      = n_text_layer*n_ctx;
-    const int n_elements = n_text_state*n_mem;
-
     cache.k = wsp_ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
     cache.v = wsp_ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
 
@@ -914,22 +970,9 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
         // print memory requirements
         {
-            // this is the total memory required to run the inference
-            const size_t mem_required =
-                     MEM_REQ_SCRATCH0.at(model.type) +
-                     MEM_REQ_SCRATCH1.at(model.type) +
-                     MEM_REQ_SCRATCH2.at(model.type) +
-                     MEM_REQ_SCRATCH3.at(model.type) +
-                scale*MEM_REQ_MODEL.at(wctx.wtype).at(model.type) +
-                scale*MEM_REQ_KV_CROSS.at(model.type) +
-                scale*std::max(MEM_REQ_ENCODE.at(model.type), MEM_REQ_DECODE.at(model.type));
-
-            // this is the memory required by one decoder
-            const size_t mem_required_decoder =
-                scale*MEM_REQ_KV_SELF.at(model.type);
-
-            log("%s: mem required  = %7.2f MB (+ %7.2f MB per decoder)\n", __func__,
-                    mem_required / 1024.0 / 1024.0, mem_required_decoder / 1024.0 / 1024.0);
+            // TODO
+            //log("%s: mem required  = %7.2f MB (+ %7.2f MB per decoder)\n", __func__,
+            //        mem_required / 1024.0 / 1024.0, mem_required_decoder / 1024.0 / 1024.0);
         }
 
         // initialize all memory buffers
@@ -1438,49 +1481,56 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
     return true;
 }
 
-// evaluate the encoder with the given state
-//
-// given audio recording (more specifically, its log mel spectrogram), runs forward pass of the encoder
-// part of the transformer model and returns the encoded features
-//
-//   - wctx:      the model
-//   - wstate:     the state of the encoder
-//   - n_threads:  number of threads to use
-//   - mel_offset: offset in the mel spectrogram (i.e. audio offset)
-//
-static bool whisper_encode_internal(
-        whisper_context & wctx,
-          whisper_state & wstate,
-              const int   mel_offset,
-              const int   n_threads){
+static bool whisper_encode_external(const whisper_state & wstate) {
+    WSP_GGML_UNUSED(wstate);
 
-    const int64_t t_start_us = wsp_ggml_time_us();
+#ifndef WHISPER_USE_COREML
+    const bool use_coreml = false;
+#else
+    const bool use_coreml = wstate.ctx_coreml != nullptr;
+#endif
+
+#ifndef WHISPER_USE_OPENVINO
+    const bool use_openvino = false;
+#else
+    const bool use_openvino = wstate.ctx_openvino != nullptr;
+#endif
+
+    return use_coreml || use_openvino;
+}
 
+static struct wsp_ggml_cgraph * whisper_build_graph_conv(
+        whisper_context & wctx,
+          whisper_state & wstate,
+              const int   mel_offset) {
     const auto & model   = wctx.model;
     const auto & mel_inp = wstate.mel;
     const auto & hparams = model.hparams;
 
     const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
-    const int n_state = hparams.n_audio_state;
-    const int n_head  = hparams.n_audio_head;
-    const int n_layer = hparams.n_audio_layer;
+    const int n_state = hparams.n_audio_state; WSP_GGML_UNUSED(n_state);
 
     const int n_mels = hparams.n_mels;
-    assert(mel_inp.n_mel == n_mels);
 
     struct wsp_ggml_init_params params = {
-        /*.mem_size   =*/ wstate.buf_compute.size(),
-        /*.mem_buffer =*/ wstate.buf_compute.data(),
-        /*.no_alloc   =*/ false,
+        /*.mem_size   =*/ wstate.alloc_conv.meta.size(),
+        /*.mem_buffer =*/ wstate.alloc_conv.meta.data(),
+        /*.no_alloc   =*/ true,
     };
 
     struct wsp_ggml_context * ctx0 = wsp_ggml_init(params);
 
-    wstate.use_buf(ctx0, 0);
+    wsp_ggml_cgraph * gf = wsp_ggml_new_graph(ctx0);
+
+    wsp_ggml_allocr * alloc = wstate.alloc_conv.alloc;
 
     struct wsp_ggml_tensor * mel = wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, 2*n_ctx, n_mels);
+    wsp_ggml_allocr_alloc(alloc, mel);
+
     assert(mel->type == WSP_GGML_TYPE_F32);
-    {
+    if (!wsp_ggml_allocr_is_measure(alloc)) {
+        assert(mel_inp.n_mel == n_mels);
+
         float * dst = (float *) mel->data;
         memset(dst, 0, wsp_ggml_nbytes(mel));
 
@@ -1494,25 +1544,11 @@ static bool whisper_encode_internal(
         }
     }
 
-    struct wsp_ggml_tensor * cur;
+    struct wsp_ggml_tensor * cur = nullptr;
 
-#ifndef WHISPER_USE_COREML
-    const bool use_coreml = false;
-#else
-    const bool use_coreml = wstate.ctx_coreml != nullptr;
-#endif
-
-#ifndef WHISPER_USE_OPENVINO
-    const bool use_openvino = false;
-#else
-    const bool use_openvino = wstate.ctx_openvino != nullptr;
-#endif
-
-    if (!use_coreml && !use_openvino) {
+    if (!whisper_encode_external(wstate)) {
         // convolution + gelu
         {
-            wstate.use_buf(ctx0, 1);
-
             cur = wsp_ggml_conv_1d_ph(ctx0, model.e_conv_1_w, mel, 1, 1);
             cur = wsp_ggml_add(ctx0,
                     wsp_ggml_repeat(ctx0,
@@ -1522,8 +1558,6 @@ static bool whisper_encode_internal(
 
             cur = wsp_ggml_gelu(ctx0, cur);
 
-            wstate.use_buf(ctx0, 0);
-
             cur = wsp_ggml_conv_1d_ph(ctx0, model.e_conv_2_w, cur, 2, 1);
             cur = wsp_ggml_add(ctx0,
                     wsp_ggml_repeat(ctx0,
@@ -1534,373 +1568,433 @@ static bool whisper_encode_internal(
             cur = wsp_ggml_gelu(ctx0, cur);
         }
 
-        wstate.use_buf(ctx0, 3);
+        wstate.embd_conv = cur;
+    } else {
+#ifdef WHISPER_USE_COREML
+        cur = wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, n_state, n_ctx);
+        wsp_ggml_allocr_alloc(alloc, cur);
 
-        // ===================================================================
-        // NOTE: experimenting with partial evaluation of the encoder (ignore)
-        //static int iter = -1;
-        //const int n_iter = 1500/n_ctx;
+        if (!wsp_ggml_allocr_is_measure(alloc)) {
+            whisper_coreml_encode(wstate.ctx_coreml, (float *) mel->data, (float *) cur->data);
+        }
+#endif
+#ifdef WHISPER_USE_OPENVINO
+        cur = wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, n_state, n_ctx);
+        wsp_ggml_allocr_alloc(alloc, cur);
 
-        //iter = (iter + 1) % n_iter;
+        if (!wsp_ggml_allocr_is_measure(alloc)) {
+            whisper_openvino_encode(wstate.ctx_openvino, mel, cur);
+        }
+#endif
 
-        //if (iter == 0) {
-        //    memset(model.memory_cross_k->data, 0, wsp_ggml_nbytes(model.memory_cross_k));
-        //    memset(model.memory_cross_v->data, 0, wsp_ggml_nbytes(model.memory_cross_v));
-        //}
+        wstate.embd_enc = cur;
+    }
 
-        static int iter = 0;
+    wsp_ggml_build_forward_expand(gf, cur);
 
-        const size_t e_pe_stride = model.e_pe->ne[0]*wsp_ggml_element_size(model.e_pe);
-        const size_t e_pe_offset = model.e_pe->ne[0]*wsp_ggml_element_size(model.e_pe)*n_ctx*iter;
+    wsp_ggml_free(ctx0);
 
-        struct wsp_ggml_tensor * e_pe = wsp_ggml_view_2d(ctx0, model.e_pe, model.e_pe->ne[0], n_ctx, e_pe_stride, e_pe_offset);
+    return gf;
+}
 
-        cur = wsp_ggml_add(ctx0, e_pe, wsp_ggml_transpose(ctx0, cur));
+static struct wsp_ggml_cgraph * whisper_build_graph_encoder(
+        whisper_context & wctx,
+          whisper_state & wstate) {
+    const auto & model   = wctx.model;
+    const auto & hparams = model.hparams;
 
-        // ===================================================================
+    const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
+    const int n_state = hparams.n_audio_state;
+    const int n_head  = hparams.n_audio_head;
+    const int n_layer = hparams.n_audio_layer;
 
-        // original:
-        //cur = wsp_ggml_add(ctx0, model.e_pe, wsp_ggml_transpose(ctx0, cur));
+    struct wsp_ggml_init_params params = {
+        /*.mem_size   =*/ wstate.alloc_encode.meta.size(),
+        /*.mem_buffer =*/ wstate.alloc_encode.meta.data(),
+        /*.no_alloc   =*/ true,
+    };
 
-        struct wsp_ggml_tensor * inpL = cur;
+    struct wsp_ggml_context * ctx0 = wsp_ggml_init(params);
 
-        for (int il = 0; il < n_layer; ++il) {
-            const auto & layer = model.layers_encoder[il];
+    wsp_ggml_cgraph * gf = wsp_ggml_new_graph(ctx0);
 
-            // norm
-            {
-                wstate.use_buf(ctx0, 0);
+    wsp_ggml_allocr * alloc = wstate.alloc_encode.alloc;
 
-                cur = wsp_ggml_norm(ctx0, inpL);
+    struct wsp_ggml_tensor * KQscale = wsp_ggml_new_tensor_1d(ctx0, WSP_GGML_TYPE_F32, 1);
+    wsp_ggml_allocr_alloc(alloc, KQscale);
 
-                // cur = ln_0_w*cur + ln_0_b
-                cur = wsp_ggml_add(ctx0,
-                        wsp_ggml_mul(ctx0,
-                            wsp_ggml_repeat(ctx0, layer.attn_ln_0_w, cur),
-                            cur),
-                        wsp_ggml_repeat(ctx0, layer.attn_ln_0_b, cur));
-            }
+    if (!wsp_ggml_allocr_is_measure(alloc)) {
+        wsp_ggml_set_f32(KQscale, 1.0f/sqrt(float(n_state)/n_head));
+    }
 
-            // self-attention
-            {
-                wstate.use_buf(ctx0, 1);
+    struct wsp_ggml_tensor * cur = wsp_ggml_view_tensor(ctx0, wstate.embd_conv);
 
-                struct wsp_ggml_tensor * Qcur = wsp_ggml_mul_mat(ctx0,
-                        layer.attn_q_w,
-                        cur);
+    // ===================================================================
+    // NOTE: experimenting with partial evaluation of the encoder (ignore)
+    //static int iter = -1;
+    //const int n_iter = 1500/n_ctx;
 
-                Qcur = wsp_ggml_add(ctx0,
-                        wsp_ggml_repeat(ctx0,
-                            layer.attn_q_b,
-                            Qcur),
-                        Qcur);
+    //iter = (iter + 1) % n_iter;
 
-                //Qcur = wsp_ggml_scale_inplace(ctx0, Qcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+    //if (iter == 0) {
+    //    memset(model.memory_cross_k->data, 0, wsp_ggml_nbytes(model.memory_cross_k));
+    //    memset(model.memory_cross_v->data, 0, wsp_ggml_nbytes(model.memory_cross_v));
+    //}
 
-                // note: no bias for Key
-                struct wsp_ggml_tensor * Kcur = wsp_ggml_mul_mat(ctx0,
-                        layer.attn_k_w,
-                        cur);
+    static int iter = 0;
 
-                //Kcur = wsp_ggml_scale_inplace(ctx0, Kcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+    const size_t e_pe_stride = model.e_pe->ne[0]*wsp_ggml_element_size(model.e_pe);
+    const size_t e_pe_offset = model.e_pe->ne[0]*wsp_ggml_element_size(model.e_pe)*n_ctx*iter;
 
-                struct wsp_ggml_tensor * Vcur = wsp_ggml_mul_mat(ctx0,
-                        layer.attn_v_w,
-                        cur);
+    struct wsp_ggml_tensor * e_pe = wsp_ggml_view_2d(ctx0, model.e_pe, model.e_pe->ne[0], n_ctx, e_pe_stride, e_pe_offset);
 
-                Vcur = wsp_ggml_add(ctx0,
-                        wsp_ggml_repeat(ctx0,
-                            layer.attn_v_b,
-                            Vcur),
-                        Vcur);
+    cur = wsp_ggml_add(ctx0, e_pe, wsp_ggml_cont(ctx0, wsp_ggml_transpose(ctx0, cur)));
+
+    // ===================================================================
+
+    // original:
+    //cur = wsp_ggml_add(ctx0, model.e_pe, wsp_ggml_transpose(ctx0, cur));
+
+    struct wsp_ggml_tensor * inpL = cur;
 
-                // ------
+    for (int il = 0; il < n_layer; ++il) {
+        const auto & layer = model.layers_encoder[il];
+
+        // norm
+        {
+            cur = wsp_ggml_norm(ctx0, inpL, hparams.eps);
+
+            // cur = ln_0_w*cur + ln_0_b
+            cur = wsp_ggml_add(ctx0,
+                    wsp_ggml_mul(ctx0, cur, layer.attn_ln_0_w),
+                    layer.attn_ln_0_b);
+        }
+
+        // self-attention
+        {
+            struct wsp_ggml_tensor * Qcur = wsp_ggml_mul_mat(ctx0,
+                    layer.attn_q_w,
+                    cur);
+
+            Qcur = wsp_ggml_add(ctx0, Qcur, layer.attn_q_b);
+
+            //Qcur = wsp_ggml_scale(ctx0, Qcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+
+            // note: no bias for Key
+            struct wsp_ggml_tensor * Kcur = wsp_ggml_mul_mat(ctx0,
+                    layer.attn_k_w,
+                    cur);
+
+            //Kcur = wsp_ggml_scale(ctx0, Kcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+
+            struct wsp_ggml_tensor * Vcur = wsp_ggml_mul_mat(ctx0,
+                    layer.attn_v_w,
+                    cur);
+
+            Vcur = wsp_ggml_add(ctx0, Vcur, layer.attn_v_b);
 
-                wstate.use_buf(ctx0, 0);
+            // ------
 
 #ifdef WHISPER_USE_FLASH_ATTN
-                struct wsp_ggml_tensor * Q =
-                    wsp_ggml_permute(ctx0,
-                            wsp_ggml_cpy(ctx0,
-                                Qcur,
-                                wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
-                            0, 2, 1, 3);
-
-                struct wsp_ggml_tensor * K =
-                    wsp_ggml_permute(ctx0,
-                            wsp_ggml_cpy(ctx0,
-                                Kcur,
-                                wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
-                            0, 2, 1, 3);
-
-                struct wsp_ggml_tensor * V =
-                    wsp_ggml_cpy(ctx0,
-                            wsp_ggml_permute(ctx0,
-                                wsp_ggml_reshape_3d(ctx0,
-                                    Vcur,
-                                    n_state/n_head, n_head, n_ctx),
-                                1, 2, 0, 3),
-                            wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head));
-
-                struct wsp_ggml_tensor * KQV = wsp_ggml_flash_attn(ctx0, Q, K, V, false);
+            struct wsp_ggml_tensor * Q =
+                wsp_ggml_permute(ctx0,
+                        wsp_ggml_cpy(ctx0,
+                            Qcur,
+                            wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
+                        0, 2, 1, 3);
+
+            struct wsp_ggml_tensor * K =
+                wsp_ggml_permute(ctx0,
+                        wsp_ggml_cpy(ctx0,
+                            Kcur,
+                            wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
+                        0, 2, 1, 3);
+
+            struct wsp_ggml_tensor * V =
+                wsp_ggml_cpy(ctx0,
+                        wsp_ggml_permute(ctx0,
+                            wsp_ggml_reshape_3d(ctx0,
+                                Vcur,
+                                n_state/n_head, n_head, n_ctx),
+                            1, 2, 0, 3),
+                        wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head));
+
+            struct wsp_ggml_tensor * KQV = wsp_ggml_flash_attn(ctx0, Q, K, V, false);
 #else
-                struct wsp_ggml_tensor * Q =
-                    wsp_ggml_permute(ctx0,
-                            wsp_ggml_cpy(ctx0,
-                                Qcur,
-                                wsp_ggml_new_tensor_3d(ctx0, WSP_GGML_TYPE_F32, n_state/n_head, n_head, n_ctx)),
-                            0, 2, 1, 3);
-
-                struct wsp_ggml_tensor * K =
-                    wsp_ggml_permute(ctx0,
-                            wsp_ggml_cpy(ctx0,
-                                Kcur,
-                                wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
-                            0, 2, 1, 3);
-
-                // K * Q
-                struct wsp_ggml_tensor * KQ = wsp_ggml_mul_mat(ctx0, K, Q);
-
-                struct wsp_ggml_tensor * KQ_scaled =
-                    wsp_ggml_scale_inplace(ctx0,
-                            KQ,
-                            wsp_ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
-                            );
-
-                struct wsp_ggml_tensor * KQ_soft_max = wsp_ggml_soft_max_inplace(ctx0, KQ_scaled);
-
-                struct wsp_ggml_tensor * V =
-                    wsp_ggml_cpy(ctx0,
-                            wsp_ggml_permute(ctx0,
-                                wsp_ggml_reshape_3d(ctx0,
-                                    Vcur,
-                                    n_state/n_head, n_head, n_ctx),
-                                1, 2, 0, 3),
-                            wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head)
-                            );
-
-                struct wsp_ggml_tensor * KQV = wsp_ggml_mul_mat(ctx0, V, KQ_soft_max);
+            struct wsp_ggml_tensor * Q =
+                wsp_ggml_permute(ctx0,
+                        wsp_ggml_cpy(ctx0,
+                            Qcur,
+                            wsp_ggml_new_tensor_3d(ctx0, WSP_GGML_TYPE_F32, n_state/n_head, n_head, n_ctx)),
+                        0, 2, 1, 3);
+
+            struct wsp_ggml_tensor * K =
+                wsp_ggml_permute(ctx0,
+                        wsp_ggml_cpy(ctx0,
+                            Kcur,
+                            wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct wsp_ggml_tensor * KQ = wsp_ggml_mul_mat(ctx0, K, Q);
+
+            struct wsp_ggml_tensor * KQ_scaled = wsp_ggml_scale(ctx0, KQ, KQscale);
+
+            struct wsp_ggml_tensor * KQ_soft_max = wsp_ggml_soft_max(ctx0, KQ_scaled);
+
+            struct wsp_ggml_tensor * V =
+                wsp_ggml_cpy(ctx0,
+                        wsp_ggml_permute(ctx0,
+                            wsp_ggml_reshape_3d(ctx0,
+                                Vcur,
+                                n_state/n_head, n_head, n_ctx),
+                            1, 2, 0, 3),
+                        wsp_ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head)
+                        );
+
+            struct wsp_ggml_tensor * KQV = wsp_ggml_mul_mat(ctx0, V, KQ_soft_max);
 #endif
-                struct wsp_ggml_tensor * KQV_merged = wsp_ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            struct wsp_ggml_tensor * KQV_merged = wsp_ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
-                wstate.use_buf(ctx0, 1);
+            cur = wsp_ggml_cpy(ctx0,
+                    KQV_merged,
+                    wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, n_state, n_ctx));
+        }
 
-                cur = wsp_ggml_cpy(ctx0,
-                        KQV_merged,
-                        wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, n_state, n_ctx));
-            }
+        // projection
+        {
+            cur = wsp_ggml_mul_mat(ctx0,
+                    layer.attn_ln_1_w,
+                    cur);
 
-            // projection
-            {
-                wstate.use_buf(ctx0, 0);
+            cur = wsp_ggml_add(ctx0, cur, layer.attn_ln_1_b);
+        }
 
-                cur = wsp_ggml_mul_mat(ctx0,
-                        layer.attn_ln_1_w,
-                        cur);
+        // add the input
+        cur = wsp_ggml_add(ctx0, cur, inpL);
 
-                wstate.use_buf(ctx0, 1);
+        struct wsp_ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = wsp_ggml_norm(ctx0, inpFF, hparams.eps);
 
+                // cur = mlp_ln_w*cur + mlp_ln_b
                 cur = wsp_ggml_add(ctx0,
-                        wsp_ggml_repeat(ctx0, layer.attn_ln_1_b, cur),
-                        cur);
+                        wsp_ggml_mul(ctx0, cur, layer.mlp_ln_w),
+                        layer.mlp_ln_b);
             }
 
-            wstate.use_buf(ctx0, 2);
+#ifdef WHISPER_USE_FLASH_FF
+            cur = wsp_ggml_flash_ff(ctx0,
+                    wsp_ggml_cpy(ctx0, cur, wsp_ggml_new_tensor_2d(ctx0, wstate.itype, n_state, n_ctx)),
+                    layer.mlp_0_w, layer.mlp_0_b, layer.mlp_1_w, layer.mlp_1_b);
+#else
+            // fully connected
+            cur = wsp_ggml_mul_mat(ctx0,
+                    layer.mlp_0_w,
+                    cur);
 
-            // add the input
-            cur = wsp_ggml_add(ctx0, cur, inpL);
+            cur = wsp_ggml_add(ctx0, cur, layer.mlp_0_b);
 
-            struct wsp_ggml_tensor * inpFF = cur;
+            // GELU activation
+            cur = wsp_ggml_gelu(ctx0, cur);
 
-            // feed-forward network
-            {
-                // norm
-                {
-                    wstate.use_buf(ctx0, 0);
+            // projection
+            cur = wsp_ggml_mul_mat(ctx0,
+                    layer.mlp_1_w,
+                    cur);
 
-                    cur = wsp_ggml_norm(ctx0, inpFF);
+            cur = wsp_ggml_add(ctx0, cur, layer.mlp_1_b);
+#endif
+        }
 
-                    wstate.use_buf(ctx0, 1);
+        inpL = wsp_ggml_add(ctx0, cur, inpFF);
+    }
 
-                    // cur = mlp_ln_w*cur + mlp_ln_b
-                    cur = wsp_ggml_add(ctx0,
-                            wsp_ggml_mul(ctx0,
-                                wsp_ggml_repeat(ctx0, layer.mlp_ln_w, cur),
-                                cur),
-                            wsp_ggml_repeat(ctx0, layer.mlp_ln_b, cur));
-                }
+    cur = inpL;
 
-#ifdef WHISPER_USE_FLASH_FF
-                wstate.use_buf(ctx0, 0);
+    // norm
+    {
+        cur = wsp_ggml_norm(ctx0, cur, hparams.eps);
 
-                cur = wsp_ggml_flash_ff(ctx0,
-                        wsp_ggml_cpy(ctx0, cur, wsp_ggml_new_tensor_2d(ctx0, wstate.itype, n_state, n_ctx)),
-                        layer.mlp_0_w, layer.mlp_0_b, layer.mlp_1_w, layer.mlp_1_b);
-#else
-                wstate.use_buf(ctx0, 0);
+        // cur = ln_f_g*cur + ln_f_b
+        cur = wsp_ggml_add(ctx0,
+                wsp_ggml_mul(ctx0, cur, model.e_ln_w),
+                model.e_ln_b);
+    }
 
-                // fully connected
-                cur = wsp_ggml_mul_mat(ctx0,
-                        layer.mlp_0_w,
-                        cur);
+    wsp_ggml_build_forward_expand(gf, cur);
 
-                wstate.use_buf(ctx0, 1);
+    wstate.embd_enc = cur;
 
-                cur = wsp_ggml_add(ctx0,
-                        wsp_ggml_repeat(ctx0, layer.mlp_0_b, cur),
-                        cur);
+    //wsp_ggml_graph_print(gf);
 
-                wstate.use_buf(ctx0, 0);
+    ////////////////////////////////////////////////////////////////////////////
 
-                // GELU activation
-                cur = wsp_ggml_gelu(ctx0, cur);
+    //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
+    //        wsp_ggml_used_mem(ctx0)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
+    //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
 
-                wstate.use_buf(ctx0, 1);
+    wsp_ggml_free(ctx0);
 
-                // projection
-                cur = wsp_ggml_mul_mat(ctx0,
-                        layer.mlp_1_w,
-                        cur);
+    return gf;
+}
 
-                wstate.use_buf(ctx0, 0);
+// pre-compute cross-attention memory
+static struct wsp_ggml_cgraph * whisper_build_graph_cross(
+        whisper_context & wctx,
+          whisper_state & wstate) {
+    const auto & model   = wctx.model;
+    const auto & hparams = model.hparams;
 
-                cur = wsp_ggml_add(ctx0,
-                        wsp_ggml_repeat(ctx0, layer.mlp_1_b, cur),
-                        cur);
-#endif
-            }
+    const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
+    const int n_state = hparams.n_audio_state;
+    const int n_head  = hparams.n_audio_head;
 
-            wstate.use_buf(ctx0, 3);
+    struct wsp_ggml_init_params params = {
+        /*.mem_size   =*/ wstate.alloc_cross.meta.size(),
+        /*.mem_buffer =*/ wstate.alloc_cross.meta.data(),
+        /*.no_alloc   =*/ true,
+    };
 
-            inpL = wsp_ggml_add(ctx0, cur, inpFF);
-        }
+    struct wsp_ggml_context * ctx0 = wsp_ggml_init(params);
 
-        cur = inpL;
+    wsp_ggml_cgraph * gf = wsp_ggml_new_graph(ctx0);
 
-        // norm
-        {
-            wstate.use_buf(ctx0, 0);
+    wsp_ggml_allocr * alloc = wstate.alloc_cross.alloc;
 
-            cur = wsp_ggml_norm(ctx0, cur);
+    struct wsp_ggml_tensor * cur = wsp_ggml_view_tensor(ctx0, wstate.embd_enc);
 
-            wstate.use_buf(ctx0, 1);
+    struct wsp_ggml_tensor * Kscale = wsp_ggml_new_tensor_1d(ctx0, WSP_GGML_TYPE_F32, 1);
+    wsp_ggml_allocr_alloc(alloc, Kscale);
 
-            // cur = ln_f_g*cur + ln_f_b
-            cur = wsp_ggml_add(ctx0,
-                    wsp_ggml_mul(ctx0,
-                        wsp_ggml_repeat(ctx0, model.e_ln_w, cur),
-                        cur),
-                    wsp_ggml_repeat(ctx0, model.e_ln_b, cur));
-        }
+    if (!wsp_ggml_allocr_is_measure(alloc)) {
+        wsp_ggml_set_f32(Kscale, pow(float(n_state) / n_head, -0.25));
+    }
 
-        wstate.use_buf(ctx0, -1);
+    for (int il = 0; il < model.hparams.n_text_layer; ++il) {
+        auto & layer = model.layers_decoder[il];
 
-        // run the computation
-        {
-            struct wsp_ggml_cgraph gf = {};
-            gf.n_threads = n_threads;
+        struct wsp_ggml_tensor* Kcross = wsp_ggml_mul_mat(ctx0,
+                layer.cross_attn_k_w,
+                cur);
 
-            wsp_ggml_build_forward_expand(&gf, cur);
-            wsp_ggml_graph_compute(ctx0, &gf);
+        Kcross = wsp_ggml_scale(ctx0, Kcross, Kscale);
 
-            //wsp_ggml_graph_print(&gf);
-        }
-    }
-#ifdef WHISPER_USE_COREML
-    else if (use_coreml) {
-        wstate.use_buf(ctx0, -1);
+        struct wsp_ggml_tensor* Vcross = wsp_ggml_mul_mat(ctx0,
+                layer.cross_attn_v_w,
+                cur);
 
-        cur = wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, n_state, n_ctx);
+        Vcross = wsp_ggml_add(ctx0,
+                    Vcross,
+                    layer.cross_attn_v_b);
 
-        whisper_coreml_encode(wstate.ctx_coreml, (float *) mel->data, (float *) cur->data);
-    }
-#endif
-#ifdef WHISPER_USE_OPENVINO
-    else if (use_openvino) {
-        wstate.use_buf(ctx0, -1);
+        Vcross = wsp_ggml_transpose(ctx0, wsp_ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));
 
-        cur = wsp_ggml_new_tensor_2d(ctx0, WSP_GGML_TYPE_F32, n_state, n_ctx);
+        struct wsp_ggml_tensor * k = wsp_ggml_view_1d(ctx0, wstate.kv_cross.k,
+                n_state*n_ctx,
+                (wsp_ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
 
-        if (!whisper_openvino_encode(wstate.ctx_openvino, mel, cur)) {
-            return false;
-        }
+        struct wsp_ggml_tensor * v = wsp_ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
+                (   n_ctx)*wsp_ggml_element_size(wstate.kv_cross.v),
+                (il*n_ctx)*wsp_ggml_element_size(wstate.kv_cross.v)*n_state);
+
+        wsp_ggml_build_forward_expand(gf, wsp_ggml_cpy(ctx0, Kcross, k));
+        wsp_ggml_build_forward_expand(gf, wsp_ggml_cpy(ctx0, Vcross, v));
     }
-#endif
 
-    // cur
-    //{
-    //    printf("ne0 = %d\n", cur->ne[0]);
-    //    printf("ne1 = %d\n", cur->ne[1]);
-    //    for (int i = 0; i < 10; ++i) {
-    //        printf("%8.4f ", ((float *)(cur->data))[i]);
-    //    }
-    //    printf("... ");
-    //    for (int i = cur->ne[0] - 10; i < cur->ne[0]; ++i) {
-    //        printf("%8.4f ", ((float *)(cur->data))[i]);
-    //    }
-    //    printf("\n");
-    //}
+    //wsp_ggml_graph_print(gf);
 
-    // pre-compute cross-attention memory
-    {
-        struct wsp_ggml_cgraph gf = {};
-        gf.n_threads = n_threads;
+    wsp_ggml_free(ctx0);
 
-        // TODO: hack to disconnect the encoded features from the previous graph
-        cur->op = WSP_GGML_OP_NONE;
-        cur->src0 = nullptr;
-        cur->src1 = nullptr;
+    return gf;
+}
 
-        for (int il = 0; il < model.hparams.n_text_layer; ++il) {
-            auto& layer = model.layers_decoder[il];
+// evaluate the encoder with the given state
+//
+// given audio recording (more specifically, its log mel spectrogram), runs forward pass of the encoder
+// part of the transformer model and returns the encoded features
+//
+//   - wctx:      the model
+//   - wstate:     the state of the encoder
+//   - n_threads:  number of threads to use
+//   - mel_offset: offset in the mel spectrogram (i.e. audio offset)
+//
+static bool whisper_encode_internal(
+        whisper_context & wctx,
+          whisper_state & wstate,
+              const int   mel_offset,
+              const int   n_threads,
+ whisper_abort_callback   abort_callback,
+                   void * abort_callback_data) {
+    const int64_t t_start_us = wsp_ggml_time_us();
 
-            wstate.use_buf(ctx0, 0);
+    // conv
+    {
+        auto & alloc = wstate.alloc_conv.alloc;
 
-            struct wsp_ggml_tensor* Kcross = wsp_ggml_mul_mat(ctx0,
-                layer.cross_attn_k_w,
-                cur);
+        wsp_ggml_allocr_reset(alloc);
 
-            Kcross = wsp_ggml_scale_inplace(ctx0, Kcross, wsp_ggml_new_f32(ctx0, pow(float(n_state) / n_head, -0.25)));
+        wsp_ggml_cgraph * gf = whisper_build_graph_conv(wctx, wstate, mel_offset);
 
-            wstate.use_buf(ctx0, 1);
+        wsp_ggml_allocr_alloc_graph(alloc, gf);
 
-            struct wsp_ggml_tensor* Vcross = wsp_ggml_mul_mat(ctx0,
-                layer.cross_attn_v_w,
-                cur);
+        if (!whisper_encode_external(wstate)) {
+            wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
+        }
+    }
 
-            Vcross = wsp_ggml_add(ctx0,
-                wsp_ggml_repeat(ctx0,
-                    layer.cross_attn_v_b,
-                    Vcross),
-                Vcross);
+    // encoder
+    if (!whisper_encode_external(wstate)) {
+        auto & alloc = wstate.alloc_encode.alloc;
 
-            wstate.use_buf(ctx0, -1);
+        wsp_ggml_allocr_reset(alloc);
 
-            Vcross = wsp_ggml_transpose(ctx0, wsp_ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));
+        wsp_ggml_cgraph * gf = whisper_build_graph_encoder(wctx, wstate);
 
-            struct wsp_ggml_tensor * k = wsp_ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx, (wsp_ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
-            struct wsp_ggml_tensor * v = wsp_ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
-                    (   n_ctx)*wsp_ggml_element_size(wstate.kv_cross.v),
-                    (il*n_ctx)*wsp_ggml_element_size(wstate.kv_cross.v)*n_state);
+        wsp_ggml_allocr_alloc_graph(alloc, gf);
 
-            wsp_ggml_build_forward_expand(&gf, wsp_ggml_cpy(ctx0, Kcross, k));
-            wsp_ggml_build_forward_expand(&gf, wsp_ggml_cpy(ctx0, Vcross, v));
+#ifdef WSP_GGML_USE_METAL
+        if (wstate.ctx_metal) {
+            wsp_ggml_metal_set_n_cb     (wstate.ctx_metal, n_threads);
+            wsp_ggml_metal_graph_compute(wstate.ctx_metal, gf);
+        } else {
+            wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
         }
-
-        wsp_ggml_graph_compute(ctx0, &gf);
-        //wsp_ggml_graph_print(&gf);
+#else
+        wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
+#endif
     }
 
-    ////////////////////////////////////////////////////////////////////////////
+    // cross
+    {
+        auto & alloc = wstate.alloc_cross.alloc;
 
-    //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
-    //        wsp_ggml_used_mem(ctx0)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(1)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(2)/1024.0/1024.0,
-    //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
+        wsp_ggml_allocr_reset(alloc);
 
-    wsp_ggml_free(ctx0);
+        wsp_ggml_cgraph * gf = whisper_build_graph_cross(wctx, wstate);
+
+        wsp_ggml_allocr_alloc_graph(alloc, gf);
+
+#ifdef WSP_GGML_USE_METAL
+        if (wstate.ctx_metal) {
+            wsp_ggml_metal_set_n_cb     (wstate.ctx_metal, n_threads);
+            wsp_ggml_metal_graph_compute(wstate.ctx_metal, gf);
+        } else {
+            wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
+        }
+#else
+        wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
+#endif
+    }
+
+    // wsp_ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
 
     wstate.t_encode_us += wsp_ggml_time_us() - t_start_us;
     wstate.n_encode++;
@@ -1908,26 +2002,13 @@ static bool whisper_encode_internal(
     return true;
 }
 
-// evaluate the decoder
-//
-// given text prompt + audio features -> computes the logits for the next token
-//
-//   - model:      the model
-//   - n_threads:  number of threads to use
-//   - tokens:     text prompt
-//   - n_tokens:   number of tokens in the prompt
-//   - n_past:     number of past tokens to prefix the prompt with
-//
-static bool whisper_decode_internal(
-        whisper_context & wctx,
-          whisper_state & wstate,
-        whisper_decoder & decoder,
-    const whisper_token * tokens,
-              const int   n_tokens,
-              const int   n_past,
-              const int   n_threads) {
-    const int64_t t_start_us = wsp_ggml_time_us();
-
+static struct wsp_ggml_cgraph * whisper_build_graph_decoder(
+         whisper_context & wctx,
+         whisper_state   & wstate,
+         whisper_decoder & decoder,
+     const whisper_token * tokens,
+                   int   n_tokens,
+                   int   n_past) {
     const auto & model   = wctx.model;
     const auto & hparams = model.hparams;
 
@@ -1935,10 +2016,6 @@ static bool whisper_decode_internal(
 
     WHISPER_ASSERT(!!kv_self.ctx);
 
-    auto & logits_out = wstate.logits;
-
-    const int n_vocab = hparams.n_vocab;
-
     const int n_ctx   = hparams.n_text_ctx;
     const int n_state = hparams.n_text_state;
     const int n_head  = hparams.n_text_head;
@@ -1950,25 +2027,39 @@ static bool whisper_decode_internal(
     //WHISPER_PRINT_DEBUG("%s: n_past = %d, N = %d, M = %d, n_ctx = %d\n", __func__, n_past, N, M, n_ctx);
 
     struct wsp_ggml_init_params params = {
-        /*.mem_size   =*/ wstate.buf_compute.size(),
-        /*.mem_buffer =*/ wstate.buf_compute.data(),
-        /*.no_alloc   =*/ false,
+        /*.mem_size   =*/ wstate.alloc_decode.meta.size(),
+        /*.mem_buffer =*/ wstate.alloc_decode.meta.data(),
+        /*.no_alloc   =*/ true,
     };
 
     struct wsp_ggml_context * ctx0 = wsp_ggml_init(params);
 
-    struct wsp_ggml_cgraph gf = {};
-    gf.n_threads = n_threads;
+    wsp_ggml_cgraph * gf = wsp_ggml_new_graph(ctx0);
+
+    wsp_ggml_allocr * alloc = wstate.alloc_decode.alloc;
 
     struct wsp_ggml_tensor * embd = wsp_ggml_new_tensor_1d(ctx0, WSP_GGML_TYPE_I32, N);
-    memcpy(embd->data, tokens, N*wsp_ggml_element_size(embd));
+    wsp_ggml_allocr_alloc(alloc, embd);
+
+    if (!wsp_ggml_allocr_is_measure(alloc)) {
+        memcpy(embd->data, tokens, N*wsp_ggml_element_size(embd));
+    }
 
     struct wsp_ggml_tensor * position = wsp_ggml_new_tensor_1d(ctx0, WSP_GGML_TYPE_I32, N);
-    for (int i = 0; i < N; ++i) {
-        ((int32_t *) position->data)[i] = n_past + i;
+    wsp_ggml_allocr_alloc(alloc, position);
+
+    if (!wsp_ggml_allocr_is_measure(alloc)) {
+        for (int i = 0; i < N; ++i) {
+            ((int32_t *) position->data)[i] = n_past + i;
+        }
     }
 
-    wstate.use_buf(ctx0, 3);
+    struct wsp_ggml_tensor * KQscale = wsp_ggml_new_tensor_1d(ctx0, WSP_GGML_TYPE_F32, 1);
+    wsp_ggml_allocr_alloc(alloc, KQscale);
+
+    if (!wsp_ggml_allocr_is_measure(alloc)) {
+        wsp_ggml_set_f32(KQscale, pow(float(n_state)/n_head, -0.25));
+    }
 
     // token encoding + position encoding
     struct wsp_ggml_tensor * cur =
@@ -1983,16 +2074,14 @@ static bool whisper_decode_internal(
 
         // norm
         {
-            wstate.use_buf(ctx0, 0);
-
-            cur = wsp_ggml_norm(ctx0, inpL);
+            cur = wsp_ggml_norm(ctx0, inpL, hparams.eps);
 
             // cur = ln_0_w*cur + ln_0_b
             cur = wsp_ggml_add(ctx0,
                     wsp_ggml_mul(ctx0,
-                        wsp_ggml_repeat(ctx0, layer.attn_ln_0_w, cur),
-                        cur),
-                    wsp_ggml_repeat(ctx0, layer.attn_ln_0_b, cur));
+                        cur,
+                        layer.attn_ln_0_w),
+                    layer.attn_ln_0_b);
         }
 
         // self-attention
@@ -2002,19 +2091,17 @@ static bool whisper_decode_internal(
                     cur);
 
             Qcur = wsp_ggml_add(ctx0,
-                    wsp_ggml_repeat(ctx0,
-                        layer.attn_q_b,
-                        Qcur),
-                    Qcur);
+                        Qcur,
+                        layer.attn_q_b);
 
-            Qcur = wsp_ggml_scale_inplace(ctx0, Qcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+            Qcur = wsp_ggml_scale(ctx0, Qcur, KQscale);
 
             // note: no bias for Key
             struct wsp_ggml_tensor * Kcur = wsp_ggml_mul_mat(ctx0,
                     layer.attn_k_w,
                     cur);
 
-            Kcur = wsp_ggml_scale_inplace(ctx0, Kcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+            Kcur = wsp_ggml_scale(ctx0, Kcur, KQscale);
 
             // store key and value to memory
             {
@@ -2023,10 +2110,8 @@ static bool whisper_decode_internal(
                         cur);
 
                 Vcur = wsp_ggml_add(ctx0,
-                        wsp_ggml_repeat(ctx0,
-                            layer.attn_v_b,
-                            Vcur),
-                        Vcur);
+                            Vcur,
+                            layer.attn_v_b);
 
                 Vcur = wsp_ggml_transpose(ctx0, wsp_ggml_reshape_2d(ctx0, Vcur, n_state, N));
 
@@ -2035,42 +2120,32 @@ static bool whisper_decode_internal(
                         (   n_ctx)*wsp_ggml_element_size(kv_self.v),
                         (il*n_ctx)*wsp_ggml_element_size(kv_self.v)*n_state + n_past*wsp_ggml_element_size(kv_self.v));
 
-                wsp_ggml_build_forward_expand(&gf, wsp_ggml_cpy(ctx0, Kcur, k));
-                wsp_ggml_build_forward_expand(&gf, wsp_ggml_cpy(ctx0, Vcur, v));
+                wsp_ggml_build_forward_expand(gf, wsp_ggml_cpy(ctx0, Kcur, k));
+                wsp_ggml_build_forward_expand(gf, wsp_ggml_cpy(ctx0, Vcur, v));
             }
 
             // ------
 
-            wstate.use_buf(ctx0, 0);
-
             struct wsp_ggml_tensor * Q =
                 wsp_ggml_permute(ctx0,
-                        wsp_ggml_cpy(ctx0,
-                            Qcur,
-                            wsp_ggml_new_tensor_3d(ctx0, WSP_GGML_TYPE_F32, n_state/n_head, n_head, N)),
+                        wsp_ggml_reshape_3d(ctx0, Qcur, n_state/n_head, n_head, N),
                         0, 2, 1, 3);
 
             struct wsp_ggml_tensor * K =
-                wsp_ggml_permute(ctx0,
-                        wsp_ggml_reshape_3d(ctx0,
-                            wsp_ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_state, il*n_ctx*wsp_ggml_element_size(kv_self.k)*n_state),
-                            n_state/n_head, n_head, n_past + N),
-                        0, 2, 1, 3);
-
-            wstate.use_buf(ctx0, 1);
+                wsp_ggml_view_3d(ctx0, kv_self.k,
+                        n_state/n_head, n_past + N, n_head,
+                        wsp_ggml_element_size(kv_self.k)*n_state,
+                        wsp_ggml_element_size(kv_self.k)*n_state/n_head,
+                        wsp_ggml_element_size(kv_self.k)*n_state*n_ctx*il);
 
             // K * Q
             struct wsp_ggml_tensor * KQ = wsp_ggml_mul_mat(ctx0, K, Q);
 
-            //struct wsp_ggml_tensor * KQ_scaled =
-            //    wsp_ggml_scale_inplace(ctx0,
-            //            KQ,
-            //            wsp_ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
-            //            );
+            //struct wsp_ggml_tensor * KQ_scaled = wsp_ggml_scale(ctx0, KQ, KQ_scale);
 
-            struct wsp_ggml_tensor * KQ_masked = wsp_ggml_diag_mask_inf_inplace(ctx0, KQ, n_past);
+            struct wsp_ggml_tensor * KQ_masked = wsp_ggml_diag_mask_inf(ctx0, KQ, n_past);
 
-            struct wsp_ggml_tensor * KQ_soft_max = wsp_ggml_soft_max_inplace(ctx0, KQ_masked);
+            struct wsp_ggml_tensor * KQ_soft_max = wsp_ggml_soft_max(ctx0, KQ_masked);
 
             struct wsp_ggml_tensor * V =
                 wsp_ggml_view_3d(ctx0, kv_self.v,
@@ -2090,36 +2165,28 @@ static bool whisper_decode_internal(
 
         // projection
         {
-            wstate.use_buf(ctx0, 0);
-
             cur = wsp_ggml_mul_mat(ctx0,
                     layer.attn_ln_1_w,
                     cur);
 
-            wstate.use_buf(ctx0, 1);
-
             cur = wsp_ggml_add(ctx0,
-                    wsp_ggml_repeat(ctx0, layer.attn_ln_1_b, cur),
-                    cur);
+                    cur,
+                    layer.attn_ln_1_b);
         }
 
-        wstate.use_buf(ctx0, 2);
-
         // add the input
         struct wsp_ggml_tensor * inpCA = wsp_ggml_add(ctx0, cur, inpL);
 
         // norm
         {
-            wstate.use_buf(ctx0, 0);
-
-            cur = wsp_ggml_norm(ctx0, inpCA); // note: we use inpCA here
+            cur = wsp_ggml_norm(ctx0, inpCA, hparams.eps); // note: we use inpCA here
 
             // cur = ln_0_w*cur + ln_0_b
             cur = wsp_ggml_add(ctx0,
                     wsp_ggml_mul(ctx0,
-                        wsp_ggml_repeat(ctx0, layer.cross_attn_ln_0_w, cur),
-                        cur),
-                    wsp_ggml_repeat(ctx0, layer.cross_attn_ln_0_b, cur));
+                        cur,
+                        layer.cross_attn_ln_0_w),
+                    layer.cross_attn_ln_0_b);
         }
 
         // cross-attention
@@ -2129,18 +2196,18 @@ static bool whisper_decode_internal(
                     cur);
 
             Qcur = wsp_ggml_add(ctx0,
-                    wsp_ggml_repeat(ctx0,
-                        layer.cross_attn_q_b,
-                        Qcur),
-                    Qcur);
+                        Qcur,
+                        layer.cross_attn_q_b);
 
-            Qcur = wsp_ggml_scale_inplace(ctx0, Qcur, wsp_ggml_new_f32(ctx0, pow(float(n_state)/n_head, -0.25)));
+            Qcur = wsp_ggml_scale(ctx0, Qcur, KQscale);
 
             // Kcross is already scaled
             struct wsp_ggml_tensor * Kcross =
-                wsp_ggml_reshape_3d(ctx0,
-                        wsp_ggml_view_1d(ctx0, wstate.kv_cross.k, M*n_state, il*M*wsp_ggml_element_size(wstate.kv_cross.k)*n_state),
-                        n_state/n_head, n_head, M);
+                wsp_ggml_view_3d(ctx0, wstate.kv_cross.k,
+                        n_state/n_head, M, n_head,
+                        wsp_ggml_element_size(wstate.kv_cross.k)*n_state,
+                        wsp_ggml_element_size(wstate.kv_cross.k)*n_state/n_head,
+                        wsp_ggml_element_size(wstate.kv_cross.k)*n_state*M*il);
 
             //struct wsp_ggml_tensor * Vcross =
             //    wsp_ggml_reshape_3d(ctx0,
@@ -2163,26 +2230,22 @@ static bool whisper_decode_internal(
 
             struct wsp_ggml_tensor * Q =
                 wsp_ggml_permute(ctx0,
-                        wsp_ggml_cpy(ctx0,
-                            Qcur,
-                            wsp_ggml_new_tensor_3d(ctx0, WSP_GGML_TYPE_F32, n_state/n_head, n_head, N)),
+                        wsp_ggml_reshape_3d(ctx0, Qcur, n_state/n_head, n_head, N),
                         0, 2, 1, 3);
 
-            struct wsp_ggml_tensor * K = wsp_ggml_permute(ctx0, Kcross, 0, 2, 1, 3);
-
             // K * Q
-            struct wsp_ggml_tensor * KQ = wsp_ggml_mul_mat(ctx0, K, Q);
+            struct wsp_ggml_tensor * KQ = wsp_ggml_mul_mat(ctx0, Kcross, Q);
 
             //struct wsp_ggml_tensor * KQ_scaled =
-            //    wsp_ggml_scale_inplace(ctx0,
+            //    wsp_ggml_scale(ctx0,
             //            KQ,
             //            wsp_ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
             //            );
 
             // no masking for cross-attention
-            //struct wsp_ggml_tensor * KQ_masked = wsp_ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+            //struct wsp_ggml_tensor * KQ_masked = wsp_ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
 
-            struct wsp_ggml_tensor * KQ_soft_max = wsp_ggml_soft_max_inplace(ctx0, KQ);
+            struct wsp_ggml_tensor * KQ_soft_max = wsp_ggml_soft_max(ctx0, KQ);
 
             struct wsp_ggml_tensor * KQV = wsp_ggml_mul_mat(ctx0, V, KQ_soft_max);
 
@@ -2196,21 +2259,15 @@ static bool whisper_decode_internal(
 
         // projection
         {
-            wstate.use_buf(ctx0, 0);
-
             cur = wsp_ggml_mul_mat(ctx0,
                     layer.cross_attn_ln_1_w,
                     cur);
 
-            wstate.use_buf(ctx0, 1);
-
             cur = wsp_ggml_add(ctx0,
-                    wsp_ggml_repeat(ctx0, layer.cross_attn_ln_1_b, cur),
-                    cur);
+                    cur,
+                    layer.cross_attn_ln_1_b);
         }
 
-        wstate.use_buf(ctx0, 2);
-
         // add the input
         cur = wsp_ggml_add(ctx0, cur, inpCA);
 
@@ -2220,54 +2277,38 @@ static bool whisper_decode_internal(
         {
             // norm
             {
-                wstate.use_buf(ctx0, 0);
-
-                cur = wsp_ggml_norm(ctx0, inpFF);
-
-                wstate.use_buf(ctx0, 1);
+                cur = wsp_ggml_norm(ctx0, inpFF, hparams.eps);
 
                 // cur = mlp_ln_w*cur + mlp_ln_b
                 cur = wsp_ggml_add(ctx0,
                         wsp_ggml_mul(ctx0,
-                            wsp_ggml_repeat(ctx0, layer.mlp_ln_w, cur),
-                            cur),
-                        wsp_ggml_repeat(ctx0, layer.mlp_ln_b, cur));
+                            cur,
+                            layer.mlp_ln_w),
+                        layer.mlp_ln_b);
             }
 
-            wstate.use_buf(ctx0, 0);
-
             // fully connected
             cur = wsp_ggml_mul_mat(ctx0,
                     layer.mlp_0_w,
                     cur);
 
-            wstate.use_buf(ctx0, 1);
-
             cur = wsp_ggml_add(ctx0,
-                    wsp_ggml_repeat(ctx0, layer.mlp_0_b, cur),
-                    cur);
-
-            wstate.use_buf(ctx0, 0);
+                    cur,
+                    layer.mlp_0_b);
 
             // GELU activation
             cur = wsp_ggml_gelu(ctx0, cur);
 
-            wstate.use_buf(ctx0, 1);
-
             // projection
             cur = wsp_ggml_mul_mat(ctx0,
                     layer.mlp_1_w,
                     cur);
 
-            wstate.use_buf(ctx0, 0);
-
             cur = wsp_ggml_add(ctx0,
-                    wsp_ggml_repeat(ctx0, layer.mlp_1_b, cur),
-                    cur);
+                    cur,
+                    layer.mlp_1_b);
         }
 
-        wstate.use_buf(ctx0, 3);
-
         inpL = wsp_ggml_add(ctx0, cur, inpFF);
     }
 
@@ -2275,21 +2316,15 @@ static bool whisper_decode_internal(
 
     // norm
     {
-        wstate.use_buf(ctx0, 0);
-
-        cur = wsp_ggml_norm(ctx0, cur);
-
-        wstate.use_buf(ctx0, 1);
+        cur = wsp_ggml_norm(ctx0, cur, hparams.eps);
 
         cur = wsp_ggml_add(ctx0,
                 wsp_ggml_mul(ctx0,
-                    wsp_ggml_repeat(ctx0, model.d_ln_w, cur),
-                    cur),
-                wsp_ggml_repeat(ctx0, model.d_ln_b, cur));
+                    cur,
+                    model.d_ln_w),
+                model.d_ln_b);
     }
 
-    wstate.use_buf(ctx0, 0);
-
     // compute logits only for the last token
     // comment this line to compute logits for all N tokens
     // might be useful in the future
@@ -2297,23 +2332,77 @@ static bool whisper_decode_internal(
 
     struct wsp_ggml_tensor * logits = wsp_ggml_mul_mat(ctx0, model.d_te, cur);
 
-    wstate.use_buf(ctx0, -1);
+    wsp_ggml_build_forward_expand(gf, logits);
+
+    wsp_ggml_free(ctx0);
+
+    return gf;
+}
+
+// evaluate the decoder
+//
+// given text prompt + audio features -> computes the logits for the next token
+//
+//   - model:      the model
+//   - n_threads:  number of threads to use
+//   - tokens:     text prompt
+//   - n_tokens:   number of tokens in the prompt
+//   - n_past:     number of past tokens to prefix the prompt with
+//
+static bool whisper_decode_internal(
+        whisper_context & wctx,
+          whisper_state & wstate,
+        whisper_decoder & decoder,
+    const whisper_token * tokens,
+              const int   n_tokens,
+              const int   n_past,
+              const int   n_threads,
+ whisper_abort_callback   abort_callback,
+                   void * abort_callback_data) {
+    const int64_t t_start_us = wsp_ggml_time_us();
+
+    const auto & model   = wctx.model;
+    const auto & hparams = model.hparams;
+
+    const int n_vocab = hparams.n_vocab;
+
+    auto & logits_out = wstate.logits;
+
+    struct wsp_ggml_tensor * logits;
 
-    // run the computation
+    // decoder
     {
-        wsp_ggml_build_forward_expand(&gf, logits);
-        wsp_ggml_graph_compute       (ctx0, &gf);
+        auto & alloc = wstate.alloc_decode.alloc;
+
+        wsp_ggml_allocr_reset(alloc);
+
+        wsp_ggml_cgraph * gf = whisper_build_graph_decoder(wctx, wstate, decoder, tokens, n_tokens, n_past);
+
+        wsp_ggml_allocr_alloc_graph(alloc, gf);
+
+        logits = gf->nodes[gf->n_nodes - 1];
+
+#ifdef WSP_GGML_USE_METAL
+        if (wstate.ctx_metal) {
+            wsp_ggml_metal_set_n_cb     (wstate.ctx_metal, n_threads);
+            wsp_ggml_metal_graph_compute(wstate.ctx_metal, gf);
+        } else {
+            wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
+        }
+#else
+        wsp_ggml_graph_compute_helper(wstate.work_buffer, gf, n_threads, abort_callback, abort_callback_data);
+#endif
     }
 
     // extract logits for all N tokens
-    //logits_out.resize(N*n_vocab);
-    //memcpy(logits_out.data(), wsp_ggml_get_data(logits), sizeof(float)*N*n_vocab);
+    //logits_out.resize(n_tokens*n_vocab);
+    //memcpy(logits_out.data(), wsp_ggml_get_data(logits), sizeof(float)*n_tokens*n_vocab);
 
     // extract logits only for the last token
     logits_out.resize(n_vocab);
     memcpy(logits_out.data(), wsp_ggml_get_data(logits), sizeof(float)*n_vocab);
 
-    if (N > 1) {
+    if (n_tokens > 1) {
         //printf("%s: used_mem = %f MB, %f MB, %f MB %f MB %f MB\n", __func__,
         //        wsp_ggml_used_mem(ctx0)/1024.0/1024.0,
         //        wstate.get_buf_max_mem(0)/1024.0/1024.0,
@@ -2322,14 +2411,18 @@ static bool whisper_decode_internal(
         //        wstate.get_buf_max_mem(3)/1024.0/1024.0);
     }
 
-    wsp_ggml_free(ctx0);
-
-    wstate.t_decode_us += wsp_ggml_time_us() - t_start_us;
-    wstate.n_decode++;
+    if (n_tokens == 1) {
+        wstate.t_decode_us += wsp_ggml_time_us() - t_start_us;
+        wstate.n_decode++;
+    } else {
+        wstate.t_prompt_us += wsp_ggml_time_us() - t_start_us;
+        wstate.n_prompt++;
+    }
 
     return true;
 }
 
+
 //  500 -> 00:05.000
 // 6000 -> 01:00.000
 static std::string to_timestamp(int64_t t, bool comma = false) {
@@ -2351,7 +2444,7 @@ static std::string to_timestamp(int64_t t, bool comma = false) {
 static float sin_vals[SIN_COS_N_COUNT];
 static float cos_vals[SIN_COS_N_COUNT];
 
-// In FFT, we frequently use sine and cosine operations with the same values. 
+// In FFT, we frequently use sine and cosine operations with the same values.
 // We can use precalculated values to speed up the process.
 static void fill_sin_cos_table() {
     static bool is_filled = false;
@@ -2446,7 +2539,7 @@ static void fft(const std::vector<float> & in, std::vector<float> & out) {
 }
 
 static bool hann_window(int length, bool periodic, std::vector<float> & output) {
-    if (output.size() < length) {
+    if (output.size() < static_cast<size_t>(length)) {
         output.resize(length);
     }
     int offset = -1;
@@ -2738,9 +2831,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
     fill_sin_cos_table();
     whisper_state * state = new whisper_state;
 
-    const size_t scale = ctx->model.hparams.ftype ? 1 : 2;
-
-    if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_SELF.at(ctx->model.type), state->decoders[0].kv_self, ctx->itype, ctx->model.hparams.n_text_ctx)) {
+    if (!kv_cache_init(ctx->model.hparams, state->decoders[0].kv_self, ctx->itype, ctx->model.hparams.n_text_ctx)) {
         log("%s: kv_cache_init() failed for self-attention cache\n", __func__);
         delete state;
         return nullptr;
@@ -2751,7 +2842,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
         log("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
     }
 
-    if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_CROSS.at(ctx->model.type), state->kv_cross, ctx->itype, ctx->model.hparams.n_audio_ctx)) {
+    if (!kv_cache_init(ctx->model.hparams, state->kv_cross, ctx->itype, ctx->model.hparams.n_audio_ctx)) {
         log("%s: kv_cache_init() failed for cross-attention cache\n", __func__);
         delete state;
         return nullptr;
@@ -2772,6 +2863,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
     if (!state->ctx_coreml) {
         log("%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
 #ifndef WHISPER_COREML_ALLOW_FALLBACK
+        delete state;
         return nullptr;
 #endif
     } else {
@@ -2786,15 +2878,111 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
     // TAGS: WHISPER_DECODER_INIT
     state->decoders[0].sequence.tokens.reserve(ctx->model.hparams.n_text_ctx);
 
-    state->decoders[0].probs.reserve(ctx->vocab.n_vocab);
-    state->decoders[0].logits.reserve(ctx->vocab.n_vocab);
+    state->decoders[0].probs.reserve   (ctx->vocab.n_vocab);
+    state->decoders[0].logits.reserve  (ctx->vocab.n_vocab);
     state->decoders[0].logprobs.reserve(ctx->vocab.n_vocab);
-    state->buf_compute.resize(scale * std::max(MEM_REQ_ENCODE.at(ctx->model.type), MEM_REQ_DECODE.at(ctx->model.type)));
 
-    state->buf_scratch[0].resize(MEM_REQ_SCRATCH0.at(ctx->model.type));
-    state->buf_scratch[1].resize(MEM_REQ_SCRATCH1.at(ctx->model.type));
-    state->buf_scratch[2].resize(MEM_REQ_SCRATCH2.at(ctx->model.type));
-    state->buf_scratch[3].resize(MEM_REQ_SCRATCH3.at(ctx->model.type));
+    // conv allocator
+    {
+        whisper_allocr_graph_init(state->alloc_conv,
+                [&]() {
+                    return whisper_build_graph_conv(*ctx, *state, 0);
+                });
+
+        log("%s: compute buffer (conv)   = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_conv) / 1024.0 / 1024.0);
+    }
+
+    // encoder allocator
+    if (!whisper_encode_external(*state)) {
+        whisper_allocr_graph_init(state->alloc_encode,
+                [&]() {
+                    return whisper_build_graph_encoder(*ctx, *state);
+                });
+
+        log("%s: compute buffer (encode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_encode) / 1024.0 / 1024.0);
+    }
+
+    // cross allocator
+    {
+        whisper_allocr_graph_init(state->alloc_cross,
+                [&]() {
+                    return whisper_build_graph_cross(*ctx, *state);
+                });
+
+        log("%s: compute buffer (cross)  = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_cross) / 1024.0 / 1024.0);
+    }
+
+    // decoder allocator
+    {
+        whisper_allocr_graph_init(state->alloc_decode,
+                [&]() {
+                    const auto & hparams = ctx->model.hparams;
+
+                    // TODO: make sure this is the worst-case scenario
+                    const int n_tokens = hparams.n_text_ctx;
+                    const int n_past   = 0;
+
+                    return whisper_build_graph_decoder(*ctx, *state, state->decoders[0], nullptr, n_tokens, n_past);
+                });
+
+        log("%s: compute buffer (decode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_decode) / 1024.0 / 1024.0);
+    }
+
+#ifdef WSP_GGML_USE_METAL
+    state->ctx_metal = wsp_ggml_metal_init(1);
+    if (!state->ctx_metal) {
+        log("%s: wsp_ggml_metal_init() failed\n", __func__);
+        delete state;
+        return nullptr;
+    }
+
+    log("%s: Metal context initialized\n", __func__);
+
+    // this allocates all Metal resources and memory buffers
+
+    void * data_ptr  = NULL;
+    size_t data_size = 0;
+
+    // TODO: add mmap support
+    //if (params.use_mmap) {
+    //    data_ptr  = ctx->model.mapping->addr;
+    //    data_size = ctx->model.mapping->size;
+    //} else {
+    //    data_ptr  = wsp_ggml_get_mem_buffer(ctx->model.ctx);
+    //    data_size = wsp_ggml_get_mem_size  (ctx->model.ctx);
+    //}
+
+    data_ptr  = wsp_ggml_get_mem_buffer(ctx->model.ctx);
+    data_size = wsp_ggml_get_mem_size  (ctx->model.ctx);
+
+    const size_t max_size = wsp_ggml_get_max_tensor_size(ctx->model.ctx);
+
+    log("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+
+#define WHISPER_METAL_CHECK_BUF(result)              \
+    if (!(result)) {                                 \
+        log("%s: failed to add metal buffer\n", __func__); \
+        delete state;                                \
+        return nullptr;                              \
+    }
+
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "data", data_ptr, data_size, max_size));
+
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "meta_conv",   state->alloc_conv.meta.data(),   state->alloc_conv.meta.size(),   0));
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "meta_encode", state->alloc_encode.meta.data(), state->alloc_encode.meta.size(), 0));
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "meta_cross",  state->alloc_cross.meta.data(),  state->alloc_cross.meta.size(),  0));
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "meta_decode", state->alloc_decode.meta.data(), state->alloc_decode.meta.size(), 0));
+
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "data_conv",   state->alloc_conv.data.data(),   state->alloc_conv.data.size(),   0));
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "data_encode", state->alloc_encode.data.data(), state->alloc_encode.data.size(), 0));
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "data_cross",  state->alloc_cross.data.data(),  state->alloc_cross.data.size(),  0));
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "data_decode", state->alloc_decode.data.data(), state->alloc_decode.data.size(), 0));
+
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "kv_cross",  state->kv_cross.buf.data(), state->kv_cross.buf.size(), 0));
+
+    WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, "kv_self_0", state->decoders[0].kv_self.buf.data(), state->decoders[0].kv_self.buf.size(), 0));
+#undef WHISPER_METAL_CHECK_BUF
+#endif
 
     state->rng = std::mt19937(0);
 
@@ -2851,7 +3039,6 @@ int whisper_ctx_init_openvino_encoder(
 }
 
 struct whisper_context * whisper_init_from_file_no_state(const char * path_model) {
-
     log("%s: loading model from '%s'\n", __func__, path_model);
 
     auto fin = std::ifstream(path_model, std::ios::binary);
@@ -3004,6 +3191,13 @@ void whisper_free_state(struct whisper_state * state)
         }
 #endif
 
+#ifdef WSP_GGML_USE_METAL
+        if (state->ctx_metal) {
+            wsp_ggml_metal_free(state->ctx_metal);
+            state->ctx_metal = nullptr;
+        }
+#endif
+
 #ifdef WHISPER_USE_OPENVINO
         if (state->ctx_openvino != nullptr) {
             whisper_openvino_free(state->ctx_openvino);
@@ -3011,6 +3205,11 @@ void whisper_free_state(struct whisper_state * state)
         }
 #endif
 
+        whisper_allocr_free(state->alloc_conv);
+        whisper_allocr_free(state->alloc_decode);
+        whisper_allocr_free(state->alloc_cross);
+        whisper_allocr_free(state->alloc_encode);
+
         delete state;
     }
 }
@@ -3103,7 +3302,7 @@ int whisper_set_mel(
 }
 
 int whisper_encode_with_state(struct whisper_context * ctx, struct whisper_state * state, int offset, int n_threads) {
-    if (!whisper_encode_internal(*ctx, *state, offset, n_threads)) {
+    if (!whisper_encode_internal(*ctx, *state, offset, n_threads, nullptr, nullptr)) {
         log("%s: failed to eval\n", __func__);
         return -1;
     }
@@ -3112,7 +3311,7 @@ int whisper_encode_with_state(struct whisper_context * ctx, struct whisper_state
 }
 
 int whisper_encode(struct whisper_context * ctx, int offset, int n_threads) {
-    if (!whisper_encode_internal(*ctx, *ctx->state, offset, n_threads)) {
+    if (!whisper_encode_internal(*ctx, *ctx->state, offset, n_threads, nullptr, nullptr)) {
         log("%s: failed to eval\n", __func__);
         return -1;
     }
@@ -3123,7 +3322,7 @@ int whisper_encode(struct whisper_context * ctx, int offset, int n_threads) {
 int whisper_decode_with_state(struct whisper_context * ctx, struct whisper_state * state, const whisper_token * tokens, int n_tokens, int n_past, int n_threads) {
     const int selected_decoder_id = 0;
 
-    if (!whisper_decode_internal(*ctx, *state, state->decoders[selected_decoder_id], tokens, n_tokens, n_past, n_threads)) {
+    if (!whisper_decode_internal(*ctx, *state, state->decoders[selected_decoder_id], tokens, n_tokens, n_past, n_threads, nullptr, nullptr)) {
         log("%s: failed to eval\n", __func__);
         return 1;
     }
@@ -3140,7 +3339,7 @@ int whisper_decode(struct whisper_context * ctx, const whisper_token * tokens, i
         return false;
     }
 
-    if (!whisper_decode_internal(*ctx, *ctx->state, ctx->state->decoders[selected_decoder_id], tokens, n_tokens, n_past, n_threads)) {
+    if (!whisper_decode_internal(*ctx, *ctx->state, ctx->state->decoders[selected_decoder_id], tokens, n_tokens, n_past, n_threads, nullptr, nullptr)) {
         log("%s: failed to eval\n", __func__);
         return 1;
     }
@@ -3431,12 +3630,14 @@ void whisper_print_timings(struct whisper_context * ctx) {
         const int32_t n_sample = std::max(1, ctx->state->n_sample);
         const int32_t n_encode = std::max(1, ctx->state->n_encode);
         const int32_t n_decode = std::max(1, ctx->state->n_decode);
+        const int32_t n_prompt = std::max(1, ctx->state->n_prompt);
 
         log("%s:     fallbacks = %3d p / %3d h\n", __func__, ctx->state->n_fail_p, ctx->state->n_fail_h);
         log("%s:      mel time = %8.2f ms\n", __func__, ctx->state->t_mel_us / 1000.0f);
         log("%s:   sample time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_sample_us, n_sample, 1e-3f * ctx->state->t_sample_us / n_sample);
         log("%s:   encode time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_encode_us, n_encode, 1e-3f * ctx->state->t_encode_us / n_encode);
         log("%s:   decode time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_decode_us, n_decode, 1e-3f * ctx->state->t_decode_us / n_decode);
+        log("%s:   prompt time = %8.2f ms / %5d runs (%8.2f ms per run)\n", __func__, 1e-3f * ctx->state->t_prompt_us, n_prompt, 1e-3f * ctx->state->t_prompt_us / n_prompt);
     }
     log("%s:    total time = %8.2f ms\n", __func__, (t_end_us - ctx->t_start_us)/1000.0f);
 }
@@ -3446,6 +3647,11 @@ void whisper_reset_timings(struct whisper_context * ctx) {
         ctx->state->t_sample_us = 0;
         ctx->state->t_encode_us = 0;
         ctx->state->t_decode_us = 0;
+        ctx->state->t_prompt_us = 0;
+        ctx->state->n_sample = 0;
+        ctx->state->n_encode = 0;
+        ctx->state->n_decode = 0;
+        ctx->state->n_prompt = 0;
     }
 }
 
@@ -3475,6 +3681,7 @@ const char * whisper_print_system_info(void) {
     s += "FMA = "       + std::to_string(wsp_ggml_cpu_has_fma())       + " | ";
     s += "NEON = "      + std::to_string(wsp_ggml_cpu_has_neon())      + " | ";
     s += "ARM_FMA = "   + std::to_string(wsp_ggml_cpu_has_arm_fma())   + " | ";
+    s += "METAL = "     + std::to_string(wsp_ggml_cpu_has_metal())     + " | ";
     s += "F16C = "      + std::to_string(wsp_ggml_cpu_has_f16c())      + " | ";
     s += "FP16_VA = "   + std::to_string(wsp_ggml_cpu_has_fp16_va())   + " | ";
     s += "WASM_SIMD = " + std::to_string(wsp_ggml_cpu_has_wasm_simd()) + " | ";
@@ -3566,6 +3773,9 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
         /*.encoder_begin_callback           =*/ nullptr,
         /*.encoder_begin_callback_user_data =*/ nullptr,
 
+        /*.abort_callback           =*/ nullptr,
+        /*.abort_callback_user_data =*/ nullptr,
+
         /*.logits_filter_callback           =*/ nullptr,
         /*.logits_filter_callback_user_data =*/ nullptr,
     };
@@ -3970,17 +4180,21 @@ static std::vector<whisper_token_data> whisper_sample_token_topk(
 
     auto & logits_id = state.logits_id;
 
-    logits_id.clear();
+    logits_id.resize(n_logits);
     for (int i = 0; i < n_logits; ++i) {
-        logits_id.push_back({ logits[i], i });
+        logits_id[i].first = logits[i];
+        logits_id[i].second = i;
     }
 
-    std::partial_sort(
-            logits_id.begin(),
-            logits_id.begin() + k, logits_id.end(),
-            [](const std::pair<double, whisper_token> & a, const std::pair<double, whisper_token> & b) {
-                return a.first > b.first;
-            });
+    {
+        using pair_type = std::remove_reference<decltype(logits_id)>::type::value_type;
+        std::partial_sort(
+                logits_id.begin(),
+                logits_id.begin() + k, logits_id.end(),
+                [](const pair_type & a, const pair_type & b) {
+            return a.first > b.first;
+        });
+    }
 
     std::vector<whisper_token_data> result;
     result.reserve(k);
@@ -4075,6 +4289,115 @@ static void whisper_sequence_score(
     }
 }
 
+static bool whisper_kv_swap_fast(
+                   std::vector<int> & view,
+                    whisper_decoder   src[],
+                std::vector<kv_buf> & kv_swap_bufs,
+                          const int & n_decoders) {
+    WHISPER_PRINT_DEBUG("%s: n_decoders %d\n", __func__, n_decoders);
+
+    // (decoder->buffer->decoder or decoder->buffer + decoder->decoder)
+    std::set<int> two_copy; // decoder indices require two copies to safely modify KV caches
+
+    // (buffer->decoder or decoder->decoder)
+    std::set<int> one_copy; // decoder indices require one copy to safely modify KV caches
+
+    // (decoder<->decoder)
+    std::set<int> p_swap_set; // decoder indices able to swap KV-cache pointers
+    std::vector<whisper_pair<int, int>> p_swap_vec;
+    p_swap_vec.reserve(n_decoders);
+
+    // see https://github.com/ggerganov/whisper.cpp/wiki
+    for (int i = 0; i < n_decoders; i++) {
+        // zero-copy (no modification)
+        if (i == view[i] || view[i] < 0) {
+            continue;
+        }
+
+        bool is_one_copy = true;
+        // since we modify data sequentially, we only consider decoder indices after current index
+        for (int j = i + 1; j < n_decoders; j++) {
+            if (i == view[j]) {
+                // detect symmetric diagram
+                if (j == view[i]) {
+                    p_swap_set.insert(i);
+                    p_swap_set.insert(j);
+                    p_swap_vec.emplace_back(i, j);
+                } else {
+                    two_copy.insert(i);
+                    is_one_copy = false;
+                }
+                break;
+            }
+        }
+        if (is_one_copy) {
+            one_copy.insert(i);
+        }
+    }
+
+    kv_swap_bufs.resize(n_decoders);
+
+    for (int i = 0; i < n_decoders; i++) {
+        kv_swap_bufs[i].k.resize(wsp_ggml_nbytes(src[i].kv_self.k));
+        kv_swap_bufs[i].v.resize(wsp_ggml_nbytes(src[i].kv_self.v));
+    }
+
+    for (auto & i : two_copy) {
+        // make a copy of KV caches
+        WHISPER_PRINT_DEBUG("%s: store KV cache into swap: idx %d\n", __func__, i);
+        memcpy(kv_swap_bufs[i].k.data(), src[i].kv_self.k->data, kv_swap_bufs[i].k.size());
+        memcpy(kv_swap_bufs[i].v.data(), src[i].kv_self.v->data, kv_swap_bufs[i].v.size());
+    }
+
+    // since two-copy decoder KV caches are protected by kv_swap_bufs, modify them first
+    for (auto & i : two_copy) {
+        // skip the decoder indices that require pointer swapping
+        if (p_swap_set.find(i) != p_swap_set.end()) {
+            continue;
+        }
+
+        if (two_copy.find(view[i]) != two_copy.end()) {
+            // modify KV caches of decoder using data from kv_swap_bufs
+            WHISPER_PRINT_DEBUG("%s: two-copy decoder using   swap buffers: swap[%d] -> %d\n", __func__, view[i], i);
+            memcpy(src[i].kv_self.k->data, kv_swap_bufs[view[i]].k.data(), kv_swap_bufs[view[i]].k.size());
+            memcpy(src[i].kv_self.v->data, kv_swap_bufs[view[i]].v.data(), kv_swap_bufs[view[i]].v.size());
+        } else {
+            // modify KV caches of decoder using data from correspond decoder KV caches directly
+            WHISPER_PRINT_DEBUG("%s: two-copy decoder without swap buffers:      %d  -> %d\n", __func__, view[i], i);
+            memcpy(src[i].kv_self.k->data, src[view[i]].kv_self.k->data, wsp_ggml_nbytes(src[view[i]].kv_self.k));
+            memcpy(src[i].kv_self.v->data, src[view[i]].kv_self.v->data, wsp_ggml_nbytes(src[view[i]].kv_self.v));
+        }
+    }
+
+    // then modify one-copy decoder KV caches
+    for (auto & i : one_copy) {
+        // skip the decoder indices that require pointer swapping
+        if (p_swap_set.find(i) != p_swap_set.end()) {
+            continue;
+        }
+
+        if (two_copy.find(view[i]) != two_copy.end()) {
+            // modify KV caches of decoder using data from kv_swap_bufs
+            WHISPER_PRINT_DEBUG("%s: one-copy decoder using   swap buffers: swap[%d] -> %d\n", __func__, view[i], i);
+            memcpy(src[i].kv_self.k->data, kv_swap_bufs[view[i]].k.data(), kv_swap_bufs[view[i]].k.size());
+            memcpy(src[i].kv_self.v->data, kv_swap_bufs[view[i]].v.data(), kv_swap_bufs[view[i]].v.size());
+        } else {
+            // modify KV caches of decoder using data from correspond decoder KV caches directly
+            WHISPER_PRINT_DEBUG("%s: one-copy decoder without swap buffers:      %d  -> %d\n", __func__, view[i], i);
+            memcpy(src[i].kv_self.k->data, src[view[i]].kv_self.k->data, wsp_ggml_nbytes(src[view[i]].kv_self.k));
+            memcpy(src[i].kv_self.v->data, src[view[i]].kv_self.v->data, wsp_ggml_nbytes(src[view[i]].kv_self.v));
+        }
+    }
+
+    // swap the pointers
+    for (auto & i : p_swap_vec) {
+        WHISPER_PRINT_DEBUG("%s: swap pointers: %d <-> %d\n", __func__, i.first, i.second);
+        std::swap(src[i.first].kv_self, src[i.second].kv_self);
+    }
+
+    return true;
+}
+
 int whisper_full_with_state(
         struct whisper_context * ctx,
           struct whisper_state * state,
@@ -4182,6 +4505,21 @@ int whisper_full_with_state(
             decoder.probs.resize   (ctx->vocab.n_vocab);
             decoder.logits.resize  (ctx->vocab.n_vocab);
             decoder.logprobs.resize(ctx->vocab.n_vocab);
+
+            // TODO: not very clean - look for a better way and potentially merging with the init of decoder 0
+#ifdef WSP_GGML_USE_METAL
+#define WHISPER_METAL_CHECK_BUF(result)              \
+            if (!(result)) {                                 \
+                log("%s: failed to add metal buffer\n", __func__); \
+                return 0;                              \
+            }
+
+            const std::string kv_name = "kv_self_" + std::to_string(j);
+            auto & kv_self = decoder.kv_self;
+
+            WHISPER_METAL_CHECK_BUF(wsp_ggml_metal_add_buffer(state->ctx_metal, kv_name.c_str(), kv_self.buf.data(), kv_self.buf.size(), 0));
+#undef WHISPER_METAL_CHECK_BUF
+#endif
         }
     }
 
@@ -4197,7 +4535,7 @@ int whisper_full_with_state(
 
         // initial prompt
         if (!params.prompt_tokens && params.initial_prompt) {
-            prompt_tokens.resize(1024);
+            prompt_tokens.resize(2048);
             prompt_tokens.resize(whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size()));
             params.prompt_tokens   = prompt_tokens.data();
             params.prompt_n_tokens = prompt_tokens.size();
@@ -4238,14 +4576,6 @@ int whisper_full_with_state(
     std::vector<whisper_token> prompt;
     prompt.reserve(whisper_n_text_ctx(ctx));
 
-    // beam-search helpers
-    struct kv_buf {
-        std::vector<uint8_t> k;
-        std::vector<uint8_t> v;
-    };
-
-    std::vector<kv_buf> kv_bufs;
-
     struct beam_candidate {
         int decoder_idx;
         int seek_delta;
@@ -4279,7 +4609,7 @@ int whisper_full_with_state(
         }
 
         // encode audio features starting at offset seek
-        if (!whisper_encode_internal(*ctx, *state, seek, params.n_threads)) {
+        if (!whisper_encode_internal(*ctx, *state, seek, params.n_threads, params.abort_callback, params.abort_callback_user_data)) {
             log("%s: failed to encode\n", __func__);
             return -6;
         }
@@ -4362,7 +4692,7 @@ int whisper_full_with_state(
                 }
                 WHISPER_PRINT_DEBUG("\n\n");
 
-                if (!whisper_decode_internal(*ctx, *state, state->decoders[0], prompt.data(), prompt.size(), 0, params.n_threads)) {
+                if (!whisper_decode_internal(*ctx, *state, state->decoders[0], prompt.data(), prompt.size(), 0, params.n_threads, params.abort_callback, params.abort_callback_user_data)) {
                     log("%s: failed to decode\n", __func__);
                     return -7;
                 }
@@ -4382,8 +4712,8 @@ int whisper_full_with_state(
 
                         decoder.kv_self.n += prompt.size();
 
-                        memcpy(decoder.probs.data(), state->decoders[0].probs.data(),    decoder.probs.size()*sizeof(decoder.probs[0]));
-                        memcpy(decoder.logits.data(), state->decoders[0].logits.data(),   decoder.logits.size()*sizeof(decoder.logits[0]));
+                        memcpy(decoder.probs.data(),    state->decoders[0].probs.data(),    decoder.probs.size()*sizeof(decoder.probs[0]));
+                        memcpy(decoder.logits.data(),   state->decoders[0].logits.data(),   decoder.logits.size()*sizeof(decoder.logits[0]));
                         memcpy(decoder.logprobs.data(), state->decoders[0].logprobs.data(), decoder.logprobs.size()*sizeof(decoder.logprobs[0]));
                     }
 
@@ -4394,23 +4724,7 @@ int whisper_full_with_state(
             for (int i = 0, n_max = whisper_n_text_ctx(ctx)/2 - 4; i < n_max; ++i) {
                 const int64_t t_start_sample_us = wsp_ggml_time_us();
 
-                // store the KV caches of all decoders when doing beam-search
                 if (params.strategy == whisper_sampling_strategy::WHISPER_SAMPLING_BEAM_SEARCH) {
-                    kv_bufs.resize(n_decoders_cur);
-                    for (int j = 0; j < n_decoders_cur; ++j) {
-                        auto & decoder = state->decoders[j];
-
-                        if (decoder.completed || decoder.failed) {
-                            continue;
-                        }
-
-                        kv_bufs[j].k.resize(wsp_ggml_nbytes(decoder.kv_self.k));
-                        kv_bufs[j].v.resize(wsp_ggml_nbytes(decoder.kv_self.v));
-
-                        memcpy(kv_bufs[j].k.data(), decoder.kv_self.k->data, kv_bufs[j].k.size());
-                        memcpy(kv_bufs[j].v.data(), decoder.kv_self.v->data, kv_bufs[j].v.size());
-                    }
-
                     beam_candidates.clear();
                 }
 
@@ -4458,6 +4772,7 @@ int whisper_full_with_state(
                     });
 
                     uint32_t cur_c = 0;
+                    std::vector<int> decoder_idx(n_decoders_cur, -1);
 
                     for (int j = 0; j < n_decoders_cur; ++j) {
                         auto & decoder = state->decoders[j];
@@ -4476,12 +4791,13 @@ int whisper_full_with_state(
                         decoder.seek_delta = cur.seek_delta;
                         decoder.has_ts     = cur.has_ts;
 
-                        memcpy(decoder.kv_self.k->data, kv_bufs[cur.decoder_idx].k.data(), kv_bufs[cur.decoder_idx].k.size());
-                        memcpy(decoder.kv_self.v->data, kv_bufs[cur.decoder_idx].v.data(), kv_bufs[cur.decoder_idx].v.size());
-
+                        decoder_idx[j] = cur.decoder_idx;
                         WHISPER_PRINT_DEBUG("%s: beam search: decoder %d: from decoder %d: token = %10s, plog = %8.5f, sum_logprobs = %8.5f\n",
                                 __func__, j, cur.decoder_idx, ctx->vocab.id_to_token.at(decoder.sequence.tokens.back().id).c_str(), decoder.sequence.tokens.back().plog, decoder.sequence.sum_logprobs_all);
                     }
+
+                    // update KV caches
+                    whisper_kv_swap_fast(decoder_idx, state->decoders, state->kv_swap_bufs, n_decoders_cur);
                 }
 
                 // update the decoder state
@@ -4600,7 +4916,7 @@ int whisper_full_with_state(
 
                     //WHISPER_PRINT_DEBUG("%s: decoder %d: token %d, kv_self.n %d, seek_delta %d\n", __func__, j, decoder.tokens_tmp[0], decoder.kv_self.n, decoder.seek_delta);
 
-                    if (!whisper_decode_internal(*ctx, *state, decoder, decoder.tokens_tmp.data(), decoder.tokens_tmp.size(), decoder.kv_self.n, params.n_threads)) {
+                    if (!whisper_decode_internal(*ctx, *state, decoder, decoder.tokens_tmp.data(), decoder.tokens_tmp.size(), decoder.kv_self.n, params.n_threads, params.abort_callback, params.abort_callback_user_data)) {
                         log("%s: failed to decode\n", __func__);
                         return -8;
                     }
@@ -4910,6 +5226,12 @@ int whisper_full_parallel(
         ctx->state->t_sample_us += states[i]->t_sample_us;
         ctx->state->t_encode_us += states[i]->t_encode_us;
         ctx->state->t_decode_us += states[i]->t_decode_us;
+        ctx->state->t_prompt_us += states[i]->t_prompt_us;
+
+        ctx->state->n_sample += states[i]->n_sample;
+        ctx->state->n_encode += states[i]->n_encode;
+        ctx->state->n_decode += states[i]->n_decode;
+        ctx->state->n_prompt += states[i]->n_prompt;
 
         whisper_free_state(states[i]);
     }
@@ -4963,6 +5285,10 @@ int64_t whisper_full_get_segment_t1(struct whisper_context * ctx, int i_segment)
     return ctx->state->result_all[i_segment].t1;
 }
 
+bool whisper_full_get_segment_speaker_turn_next_from_state(struct whisper_state * state, int i_segment) {
+    return state->result_all[i_segment].speaker_turn_next;
+}
+
 bool whisper_full_get_segment_speaker_turn_next(struct whisper_context * ctx, int i_segment) {
     return ctx->state->result_all[i_segment].speaker_turn_next;
 }
@@ -5106,7 +5432,8 @@ WHISPER_API const char * whisper_bench_wsp_ggml_mul_mat_str(int n_threads) {
     // b: N*N*sizeof(float)
     // c: N*N*sizeof(float)
     // when F16 is used, there is an extra work buffer of size N*N*sizeof(float)
-    std::vector<char> buf(4llu*N_max*N_max*sizeof(float) + 4*512);
+    std::vector<uint8_t> buf(3llu*N_max*N_max*sizeof(float) + 3*wsp_ggml_tensor_overhead());
+    std::vector<uint8_t> work;
 
     // put a bunch of random data in the buffer
     for (size_t i = 0; i < buf.size(); i++) buf[i] = i;
@@ -5158,17 +5485,15 @@ WHISPER_API const char * whisper_bench_wsp_ggml_mul_mat_str(int n_threads) {
 
             struct wsp_ggml_cgraph gf = wsp_ggml_build_forward(c);
 
-            gf.n_threads = n_threads;
-
             double tsum = 0.0;
 
             // heat-up
-            wsp_ggml_graph_compute(ctx0, &gf);
+            wsp_ggml_graph_compute_helper(work, &gf, n_threads, nullptr , nullptr);
 
             for (int i = 0; i < n_max; ++i) {
                 const int64_t t0 = wsp_ggml_time_us();
 
-                wsp_ggml_graph_compute(ctx0, &gf);
+                wsp_ggml_graph_compute_helper(work, &gf, n_threads, nullptr, nullptr);
 
                 const int64_t t1 = wsp_ggml_time_us();