From e776082095afd071df69344d86000a80450bd9f3 Mon Sep 17 00:00:00 2001 From: Chen-Yo Sun Date: Tue, 24 Mar 2026 21:33:18 -0700 Subject: [PATCH] [Voxtral TTS] Use 8 step flow matching instead of 16 (#2158) Signed-off-by: Chen-Yo Sun --- .../voxtral_tts/cuda_graph_acoustic_transformer_wrapper.py | 2 +- .../models/voxtral_tts/voxtral_tts_audio_generation.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm_omni/model_executor/models/voxtral_tts/cuda_graph_acoustic_transformer_wrapper.py b/vllm_omni/model_executor/models/voxtral_tts/cuda_graph_acoustic_transformer_wrapper.py index ad7a2264b..395c0d113 100644 --- a/vllm_omni/model_executor/models/voxtral_tts/cuda_graph_acoustic_transformer_wrapper.py +++ b/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] = {} diff --git a/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts_audio_generation.py b/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts_audio_generation.py index 9b45172bd..3b8927d63 100644 --- a/vllm_omni/model_executor/models/voxtral_tts/voxtral_tts_audio_generation.py +++ b/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