Add a new param to set the self attn text context factor ('max-decoders')

Add a param to set the text context factor.
No change of behavior: same default (3).
Resolve:
ggerganov#2334

Author: WilliamTambellini

examples/main/main.cpp

@@ -38,6 +38,7 @@ struct whisper_params {
     int32_t max_len       = 0;
     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;
+    int32_t max_decoders  = whisper_context_default_params().max_decoders;
     int32_t audio_ctx     = 0;
 
     float word_thold      =  0.01f;
@@ -131,6 +132,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
         else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
         else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
+        else if (arg == "-md"   || arg == "--max-decoders")    { params.max_decoders    = std::stoi(argv[++i]); }
         else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
         else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
         else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
@@ -198,6 +200,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
     fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
     fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -md N,     --max-decoders N    [%-7d] Max decoders, used to set the text context cache factor\n",  params.max_decoders);
     fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
@@ -981,6 +984,7 @@ int main(int argc, char ** argv) {
 
     cparams.use_gpu    = params.use_gpu;
     cparams.flash_attn = params.flash_attn;
+    cparams.max_decoders = params.max_decoders;
 
     if (!params.dtw.empty()) {
         cparams.dtw_token_timestamps = true;

include/whisper.h

@@ -126,6 +126,8 @@ extern "C" {
         struct whisper_aheads dtw_aheads;
 
         size_t dtw_mem_size; // TODO: remove
+
+        int max_decoders; // to be used to setup text context factor
     };
 
     typedef struct whisper_token_data {

src/whisper.cpp

@@ -1005,7 +1005,7 @@ static bool whisper_kv_cache_find_slot(
         }
 
         if (n_tested >= n_ctx) {
-            //WHISPER_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
+            WHISPER_LOG_ERROR("%s: failed to find a slot for %d tokens. n_tested=%d n_ctx=%d cache.head=%d\n", __func__, n_tokens, n_tested, n_ctx, cache.head);
             return false;
         }
     }
@@ -3408,9 +3408,11 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
         whisper_mel_init(state->mel, state->backends[0], n_len, n_len, n_mel);
     }
 
-    // at this point, we don't know yet how many decoders will be used, so we overallocate 3x ctx
-    // in theory, there can be a case where this is not enough, but in practice it should always be enough
-    const int factor = 3;
+    // at this point, we don't know yet how many decoders will be used, so we overallocate 3x ctx (default value)
+    // Note: there are cases where 3 is not enough specially when increasing beamsize
+    const int factor =  ctx->params.max_decoders;
+
+    WHISPER_LOG_DEBUG("%s: init self-attn cache: n_ctx: %d factor: %d\n", __func__, factor*ctx->model.hparams.n_text_ctx, factor);
 
     if (!whisper_kv_cache_init(state->kv_self, state->backends[0], ctx->itype,
                 ctx->model.hparams.n_text_state,
@@ -3635,6 +3637,7 @@ struct whisper_context_params whisper_context_default_params() {
             /*.heads            =*/ NULL,
         },
         /*.dtw_mem_size         =*/ 1024*1024*128,
+        /* max_decoders         =*/ 3
     };
     return result;
 }
@@ -3732,6 +3735,7 @@ struct whisper_context * whisper_init_with_params_no_state(struct whisper_model_
     WHISPER_LOG_INFO("%s: flash attn = %d\n", __func__, params.flash_attn);
     WHISPER_LOG_INFO("%s: gpu_device = %d\n", __func__, params.gpu_device);
     WHISPER_LOG_INFO("%s: dtw        = %d\n", __func__, params.dtw_token_timestamps);
+    WHISPER_LOG_INFO("%s: max-decoders = %d\n", __func__, params.max_decoders);
 
     whisper_context * ctx = new whisper_context;
     ctx->params = params;