pipelines: defer eager model imports in text_to_audio and token_classification (#50001)

pipelines: defer eager model imports to avoid spurious test dependencies

`token_classification.py` imported `BasicTokenizer` at module level, and
`text_to_audio.py` imported `SpeechT5HifiGan` inside the
`if is_torch_available()` block — both at import time, unconditionally.

This caused every test file that loads any pipeline (via PipelineTesterMixin
→ test_pipeline_mixin.py → pipeline source files) to record a transitive
dependency on `modeling_bert_legacy` and `modeling_speecht5`, even when
those models are completely unrelated to the test being run.

Move both imports to the call site where they are first needed (`__init__`
for `BasicTokenizer`, the spectrogram-model branch of `__init__` for
`SpeechT5HifiGan`), so they are only executed when the relevant pipeline
class is actually instantiated.

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

Author: ydshieh

src/transformers/pipelines/text_to_audio.py

@@ -27,7 +27,6 @@
     import torch
 
     from ..models.auto.modeling_auto import MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING
-    from ..models.speecht5.modeling_speecht5 import SpeechT5HifiGan
 
 DEFAULT_VOCODER_ID = "microsoft/speecht5_hifigan"
 
@@ -122,11 +121,11 @@ def __init__(self, *args, vocoder=None, sampling_rate=None, noise_scheduler=None
 
         self.vocoder = None
         if self.model.__class__ in MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING.values():
-            self.vocoder = (
-                SpeechT5HifiGan.from_pretrained(DEFAULT_VOCODER_ID).to(self.model.device)
-                if vocoder is None
-                else vocoder
-            )
+            if vocoder is None:
+                from ..models.speecht5.modeling_speecht5 import SpeechT5HifiGan
+
+                vocoder = SpeechT5HifiGan.from_pretrained(DEFAULT_VOCODER_ID).to(self.model.device)
+            self.vocoder = vocoder
 
         if self.model.config.model_type in ["musicgen", "speecht5"]:
             # MusicGen and SpeechT5 expect to use their tokenizer instead

src/transformers/pipelines/token_classification.py

@@ -4,7 +4,6 @@
 
 import numpy as np
 
-from ..models.bert.tokenization_bert_legacy import BasicTokenizer
 from ..utils import (
     ExplicitEnum,
     add_end_docstrings,
@@ -138,6 +137,8 @@ def __init__(self, args_parser=TokenClassificationArgumentHandler(), **kwargs):
 
         self.check_model_type(MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES)
 
+        from ..models.bert.tokenization_bert_legacy import BasicTokenizer
+
         self._basic_tokenizer = BasicTokenizer(do_lower_case=False)
         self._args_parser = args_parser