whisper : add batched decoding (ggml-org#1486)

* whisper : add whisper_batch

* whisper : move kv_self to whisper_state

* whisper : full batched decoding support

* whisper : fix memory leak in whisper_batch

* whisper : fix mem leak again + remove oboslete function

* whisper : clear kv cache when using whisper_decode API

* whisper : speed-up sampling

* whisper : fix decoders initializer

* bench : add batch size 5 bench

* whisper : add comment about the KV cache size

* whisper : add check for max number of decoders

* whisper : avoid starting sampling threads with bs=1

* whisper : enable beam-search by default

* cuda : sync llama.cpp fixes

Author: ggerganov

examples/bench/bench.cpp

@@ -81,7 +81,7 @@ int whisper_bench_full(const whisper_params & params) {
     }
     // heat encoder
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
         return 4;
     }
 
@@ -90,34 +90,44 @@ int whisper_bench_full(const whisper_params & params) {
 
     // prompt heat
     if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
         return 4;
     }
 
     // text-generation heat
     if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
         return 4;
     }
 
     whisper_reset_timings(ctx);
 
     // actual run
     if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode model: %d\n", ret);
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
         return 4;
     }
 
-    for (int i = 0; i < 16; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to encode model: %d\n", ret);
+    // text-generation
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
             return 4;
         }
     }
 
-    for (int i = 0; i < 256; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to encode model: %d\n", ret);
+    // batched decoding
+    for (int i = 0; i < 64; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // prompt processing
+    for (int i = 0; i < 16; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
             return 4;
         }
     }

examples/main/main.cpp

@@ -62,8 +62,8 @@ struct whisper_params {
     int32_t progress_step =  5;
     int32_t max_context  = -1;
     int32_t max_len      =  0;
-    int32_t best_of      =  2;
-    int32_t beam_size    = -1;
+    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
 
     float word_thold    =  0.01f;
     float entropy_thold =  2.40f;
@@ -925,9 +925,9 @@ int main(int argc, char ** argv) {
             if (params.detect_language) {
                 params.language = "auto";
             }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, %d beams + best of %d, lang = %s, task = %s, %stimestamps = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
-                    params.n_threads, params.n_processors,
+                    params.n_threads, params.n_processors, params.beam_size, params.best_of,
                     params.language.c_str(),
                     params.translate ? "translate" : "transcribe",
                     params.tinydiarize ? "tdrz = 1, " : "",

extra/bench-all.sh

@@ -44,8 +44,8 @@ if [ "$encoder_only" -eq 0 ]; then
     printf "\n"
 fi
 
-printf "| %6s | %6s | %16s | %11s | %3s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "PP" "Commit"
-printf "| %6s | %6s | %16s | %11s | %3s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---"
+printf "| %6s | %6s | %16s | %11s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "Bch5" "PP" "Commit"
+printf "| %6s | %6s | %16s | %11s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"
 
 for model in "${models[@]}"; do
     # actual run
@@ -56,6 +56,7 @@ for model in "${models[@]}"; do
     # parse the output:
     encode_time=$(echo "$output" | grep "encode time" | awk '{print $11}')
     decode_time=$(echo "$output" | grep "decode time" | awk '{print $11}')
+    batchd_time=$(echo "$output" | grep "batchd time" | awk '{print $11}')
     prompt_time=$(echo "$output" | grep "prompt time" | awk '{print $11}')
     system_info=$(echo "$output" | grep "system_info")
     n_threads=$(echo "$output" | grep "system_info" | awk '{print $4}')
@@ -94,6 +95,6 @@ for model in "${models[@]}"; do
     commit=$(git rev-parse --short HEAD)
 
     if [ $ret -eq 0 ]; then
-        printf "| <todo> | <todo> | %16s | %11s | %3s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$prompt_time" "$commit"
+        printf "| <todo> | <todo> | %16s | %11s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
     fi
 done