[Voxtral TTS] Use 8 step flow matching instead of 16 (#2158)

Signed-off-by: Chen-Yo Sun <chenyo.sun@mistral.ai>

vllm_omni/model_executor/models/voxtral_tts/cuda_graph_acoustic_transformer_wrapper.py

@@ -48,7 +48,7 @@ class CUDAGraphAcousticTransformerWrapper:
         self.acoustic_embeddings_levels = self.acoustic_transformer.acoustic_embeddings_levels
 
         self.cfg_alpha = 1.2
-        self.n_steps = 16
+        self.n_steps = 8
 
         # Graph storage
         self.graphs: dict[int, CUDAGraph] = {}

vllm_omni/model_executor/models/voxtral_tts/voxtral_tts_audio_generation.py

@@ -437,7 +437,7 @@ class FlowMatchingAudioTransformer(nn.Module):
 
         # Flow matching constants
         # TODO(chenyo): hardcoded, need to fix
-        self._acoustic_decode_iters = 16
+        self._acoustic_decode_iters = 8
         # TODO(chenyo): hardcoded, need to fix
         self._cfg_alpha = 1.2
         self._noise_scale = 1.0