Update tts.py

tts.py

@@ -161,6 +161,7 @@ def _init_vllm_engine(model_name, hf_token):
     engine_args = AsyncEngineArgs(
         model=model_name,
         dtype=dtype,
+        quantization="fp8",  # FP8 dynamic quantization — ~1.5x faster on Ada Lovelace (L4)
         gpu_memory_utilization=gpu_mem_util,
         max_model_len=4096,  # Need room for ref audio tokens + generation
         enforce_eager=True,  # Avoid CUDA graph issues on HF Spaces
@@ -433,19 +434,19 @@ def generate_audio(text: str, output_filename: str = None) -> str:
         logger.info(f"[{_ts()}] [TTS] Prompt: {len(prompt_ids)} tokens")
 
         # Scale max_tokens to text length to prevent runaway generation
-        # ~50 audio tokens per word at faster speed; model should stop naturally before cap
+        # Observed: model typically uses 35-43 tokens/word; cap at 45 for headroom
+        # This prevents runaway (50+ tok/word) while allowing natural stops
         word_count = len(text.split())
-        scaled_max = min(990, max(280, word_count * 50))
+        scaled_max = min(990, max(200, word_count * 45))
         logger.info(f"[{_ts()}] [TTS] Words: {word_count} → max_tokens: {scaled_max}")
 
         # Generation settings matched to mode
-        # Official docs: "Increasing repetition_penalty and temperature makes the model speak faster"
-        # For ~1.25x speed: bump rep_penalty from 1.1→1.3, temperature from 0.5→0.6
-        # Higher rep_penalty is the main speed lever; temperature adds variation
+        # rep_penalty=1.2 gives slight speed boost without hurting voice clone quality
+        # temperature=0.5 keeps voice stable; top_p=0.9 for natural variation
         if _use_voice_cloning:
-            gen_kwargs = dict(max_tokens=scaled_max, temperature=0.6, top_p=0.85, rep_penalty=1.3)
+            gen_kwargs = dict(max_tokens=scaled_max, temperature=0.5, top_p=0.9, rep_penalty=1.2)
         else:
-            gen_kwargs = dict(max_tokens=min(1200, word_count * 50), temperature=0.7, top_p=0.9, rep_penalty=1.3)
+            gen_kwargs = dict(max_tokens=min(1200, word_count * 35), temperature=0.7, top_p=0.9, rep_penalty=1.2)
 
         # Generate via vLLM
         t1 = time.time()