[BugFix] Fix Qwen3-TTS Code2Wav compatibility with transformers >= 5.9.0

[Dan250124]

Summary

• Fix compatibility between Qwen3-TTS Code2Wav decoder and transformers >= 5.9.0 by adapting to breaking API changes in create_causal_mask.

Problem

The transformers library introduced three breaking changes to masking_utils.create_causal_mask that affect vLLM-Omni's Qwen3-TTS tokenizer/decoder model:

Date	Commit	Change
2026-02-11	6c6f70c (#43916)	Renamed input_embeds to inputs_embeds
2026-03-04	421c7f6 (#44181)	Removed cache_position parameter
2026-05-11	2d6815e (#45885)	Removed backward-compatible input_embeds alias

Since vLLM's dependency constraints allow installing transformers 5.9.0, users who build vLLM-Omni from source (e.g., on CUDA 12.x where pre-built wheels require CUDA 13.0+) can hit this incompatibility. The error manifests during CUDA Graph warmup for the Code2Wav decoder:

Full error log

(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722] Failed to enable CUDA Graph for Code2Wav decoder
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722] Traceback (most recent call last):
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_code2wav.py", line 709, in load_weights
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     self.decoder.enable_cudagraph(
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py", line 881, in enable_cudagraph
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     self._cudagraph_wrapper.warmup(
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../vllm_omni/model_executor/models/qwen3_tts/cuda_graph_decoder_wrapper.py", line 334, in warmup
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     _ = self.decoder(dummy)
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]         ^^^^^^^^^^^^^^^^^^^
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../torch/nn/modules/module.py", line 1779, in _wrapped_call_impl
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     return self._call_impl(*args, **kwargs)
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../torch/nn/modules/module.py", line 1790, in _call_impl
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     return forward_call(*args, **kwargs)
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py", line 910, in forward
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     hidden = self.pre_transformer(inputs_embeds=hidden).last_hidden_state
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]   File ".../vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py", line 576, in forward
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]     "full_attention": create_causal_mask(**mask_kwargs),
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722]                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(StageEngineCoreProc pid=9773) WARNING 05-26 08:42:50 [qwen3_tts_code2wav.py:722] TypeError: create_causal_mask() got an unexpected keyword argument 'input_embeds'

This causes CUDA Graph capture to fail silently, falling back to eager mode with a performance penalty.

Fix

Use inspect.signature() to detect the parameter names accepted by create_causal_mask at runtime and pass the correct keyword arguments:

• Old transformers (< 5.9.0): passes input_embeds and cache_position
• New transformers (>= 5.9.0): passes inputs_embeds only (no cache_position)

This approach is fully backward-compatible - no minimum transformers version bump required.

Changes

• vllm_omni/model_executor/models/qwen3_tts/tokenizer_12hz/modeling_qwen3_tts_tokenizer_v2.py: Use runtime introspection of create_causal_mask signature to adapt keyword arguments across transformers versions.

Test Plan

• Verify Code2Wav decoder loads and CUDA Graph captures successfully with transformers < 5.9.0
• Verify Code2Wav decoder loads and CUDA Graph captures successfully with transformers >= 5.9.0
• Run Qwen3-TTS end-to-end TTS inference on both transformers versions

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[hsliuustc0106]

lgtm, @linyueqian @Sy0307

[Sy0307]

lgtm. Nice fix.

[linyueqian]

Looks good to me.