From 22cd9769805b79f137cc7da2664fca3dab347212 Mon Sep 17 00:00:00 2001 From: Roger Wang Date: Sat, 29 Nov 2025 16:07:26 -0800 Subject: [PATCH] [Bugfix] Fix special token `nothink` naming. (#107) Signed-off-by: Roger Wang --- pyproject.toml | 4 +++- vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py | 8 ++++---- 2 files changed, 7 insertions(+), 5 deletions(-) diff --git a/pyproject.toml b/pyproject.toml index 7f499e0a1..4480b54f6 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -164,5 +164,7 @@ extend-ignore-identifiers-re = [ "ein", ".*arange", ".*MoBA", - ".*temperal_downsample" + ".*temperal_downsample", + ".*nothink.*", + ".*NOTHINK.*" ] diff --git a/vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py b/vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py index f38523600..2709028ad 100644 --- a/vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py +++ b/vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py @@ -54,7 +54,7 @@ TTS_BOS_SINGLE_TOKEN_ID = 151674 # TALKER_CODEC_PAD_TOKEN_ID = 4196 # Padding token TALKER_CODEC_BOS_TOKEN_ID = 4197 # Beginning of speech TALKER_CODEC_EOS_TOKEN_ID = 4198 # End of speech -TALKER_CODEC_NOTHING_ID = 4203 # No-think mode +TALKER_CODEC_NOTHINK_ID = 4203 # No-think mode TALKER_CODEC_THINK_BOS_ID = 4204 # Think mode start TALKER_CODEC_THINK_EOS_ID = 4205 # Think mode end @@ -758,7 +758,7 @@ class Qwen3OmniMoeForConditionalGeneration( codec_special_tokens = torch.tensor( [ [ - talker_hf_config.codec_nothing_id, + talker_hf_config.codec_nothink_id, talker_hf_config.codec_think_bos_id, talker_hf_config.codec_think_eos_id, talker_hf_config.codec_pad_id, @@ -771,7 +771,7 @@ class Qwen3OmniMoeForConditionalGeneration( ) codec_embeds = self.talker_embedding(codec_special_tokens) # [1, 6, talker_hidden] ( - self.embed_codec_nothing_token, + self.embed_codec_nothink_token, self.embed_codec_think_bos_token, self.embed_codec_think_eos_token, self.embed_codec_pad_token, @@ -941,7 +941,7 @@ class Qwen3OmniMoeForConditionalGeneration( ) codec_special_tokens = torch.tensor( [ - self.config.talker_config.codec_nothing_id, + self.config.talker_config.codec_nothink_id, self.config.talker_config.codec_think_bos_id, self.config.talker_config.codec_think_eos_id, speaker_id,