[Bug] Index TTS 2 fails to load — ImportError: cannot import name 'QuantizedCacheConfig' from 'transformers.cache_utils'

[as000fm]

(With the help of Claude - Thank you for all the work on Wan2GP!)

[Bug] Index TTS 2 fails to load — ImportError: cannot import name 'QuantizedCacheConfig' from 'transformers.cache_utils'

Environment

Wan2GP version	Latest (v11.66)
GPU	NVIDIA RTX 4060 8 GB VRAM
OS	Windows 11
Python env	env_venv

Description

When selecting Index TTS 2 as the TTS engine, model loading fails immediately with an ImportError deep inside the bundled Index TTS 2 code. The error occurs because QuantizedCacheConfig was removed from transformers.cache_utils in transformers ≥ 4.57, but the file models/TTS/index_tts2/gpt/transformers_generation_utils.py still tries to import it unconditionally.

This is a known upstream issue in the index-tts repository (see index-tts/index-tts#508), but it needs to be addressed in the Wan2GP vendored copy of the file.

Steps to Reproduce

1. Launch Wan2GP (latest version).
2. Select Index TTS 2 as the TTS model.
3. Attempt to generate with any prompt.

Error Log

Loading Model 'ckpts\index_tts2_gpt_fp16.safetensors' ...
Unable to use LM Engine 'vllm' as it requires a Memory Profile such as 1,3 or 3+ that loads entirely the Main Models in VRAM. Switching to Legacy LM Engine...
Traceback (most recent call last):
  File "E:\Wan2GP\wgp.py", line 7354, in queue_worker_func
    success = generate_video(task, send_cmd, plugin_data=plugin_data,  **filtered_params)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "E:\Wan2GP\wgp.py", line 5953, in generate_video
    wan_model, offloadobj = load_models(
                            ^^^^^^^^^^^^
  File "E:\Wan2GP\wgp.py", line 3518, in load_models
    wan_model, pipe = model_type_handler.load_model(
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "E:\Wan2GP\models\TTS\index_tts2_handler.py", line 234, in load_model
    from .index_tts2.pipeline import IndexTTS2Pipeline
  File "E:\Wan2GP\models\TTS\index_tts2\pipeline.py", line 17, in <module>
    from .infer_v2 import IndexTTS2
  File "E:\Wan2GP\models\TTS\index_tts2\infer_v2.py", line 24, in <module>
    from .gpt.model_v2 import UnifiedVoice
  File "E:\Wan2GP\models\TTS\index_tts2\gpt\model_v2.py", line 18, in <module>
    from .transformers_gpt2 import GPT2PreTrainedModel, GPT2Model
  File "E:\Wan2GP\models\TTS\index_tts2\gpt\transformers_gpt2.py", line 33, in <module>
    from .transformers_generation_utils import GenerationMixin
  File "E:\Wan2GP\models\TTS\index_tts2\gpt\transformers_generation_utils.py", line 28, in <module>
    from transformers.cache_utils import (
ImportError: cannot import name 'QuantizedCacheConfig' from 'transformers.cache_utils' (E:\Wan2GP\env_venv\Lib\site-packages\transformers\cache_utils.py)
Error Queue autosaved successfully to error_queue.zip

Root Cause

QuantizedCacheConfig was removed from transformers.cache_utils in transformers 4.57+. The vendored file models/TTS/index_tts2/gpt/transformers_generation_utils.py imports it unconditionally at line 28.

Suggested Fix

Option A — Pin transformers in requirements

Pin the transformers dependency to ==4.55.0 for the Index TTS 2 code path, which is the last version known to include QuantizedCacheConfig.

Option B — Guard the import with a try/except in transformers_generation_utils.py

Replace the unconditional import at line 28 with:

try:
    from transformers.cache_utils import QuantizedCacheConfig
except ImportError:
    # QuantizedCacheConfig was removed in transformers >= 4.57
    from dataclasses import dataclass

    @dataclass
    class QuantizedCacheConfig:
        backend: str = "quanto"
        nbits: int = 4
        axis_key: int = 0
        axis_value: int = 0
        q_group_size: int = 64
        residual_length: int = 128
        compute_dtype: object = None
        device: object = None

This is a non-breaking fallback — the stub exposes the same fields that the original dataclass had, and Index TTS 2 does not actually exercise quantized KV cache at inference time.

Workaround (until fixed)

Downgrade transformers manually inside the venv:

env_venv\Scripts\pip install "transformers==4.55.0"

References

• index-tts/index-tts#508 — upstream report confirming the class was removed in 4.57.1
• billwuhao/ComfyUI_IndexTTS#48 — ComfyUI fix applying the same guard

[deepbeepmeep]

bug, should be fixed please let me know if you still have the issue