Add an env var to set the self attn text context factor

Add an env var to set the text context factor
WHISPER_SELFATTN_CACHE_TEXT_CTX_FACTOR
same default to 3.
Resolve:
ggerganov#2334

Author: WilliamTambellini

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;
         }
     }
@@ -3409,8 +3409,11 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
     }
 
     // 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;
+    // Note: there are cases where 3 is not enough specially when increasing beamsize
+    static const char* text_ctx_factor_evstr = getenv("WHISPER_SELFATTN_CACHE_TEXT_CTX_FACTOR");
+    static const int text_ctx_factor_ev = text_ctx_factor_evstr ? atoi(text_ctx_factor_evstr) : 0;
+    const int factor =  text_ctx_factor_ev > 0 ? text_ctx_factor_ev : 3;
+    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,