[ROCm][CI] Gate incompatible HF references on Transformers v5 (#41532)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-06-16 20:34:11 +08:00
committed by GitHub
parent 405c7cf283
commit 7e179e4bc0
10 changed files with 217 additions and 145 deletions
@@ -42,15 +42,30 @@ class MockAudioFlamingo3Processor:
self.audio_token_id = 12345
self.max_audio_len = 60
self.feature_extractor = MockFeatureExtractor()
self.tokenizer = self._tokenize
def __call__(self, text=None, audios=None, **kwargs):
return {"input_ids": [1, 2, 3], "input_features": [np.zeros((3000, 80))]}
def __call__(self, text=None, audio=None, **kwargs):
return {
"input_ids": torch.tensor([[1, 2, 3]], dtype=torch.long),
"input_features": torch.zeros((3, 80, 3000)),
"input_features_mask": torch.ones((3, 3000), dtype=torch.long),
}
def _tokenize(self, text, **kwargs):
return {"input_ids": torch.tensor([[1, 2, 3]], dtype=torch.long)}
class MockFeatureExtractor:
def __init__(self):
self.sampling_rate = 16000
self.chunk_length = 30
self.hop_length = 160
def __call__(self, audios, **kwargs):
return {
"input_features": torch.zeros((len(audios), 80, 3000)),
"attention_mask": torch.ones((len(audios), 3000), dtype=torch.long),
}
@pytest.fixture
@@ -60,6 +75,9 @@ def mock_ctx():
ctx = MagicMock()
ctx.get_hf_config.return_value = config
ctx.get_hf_processor.return_value = MockAudioFlamingo3Processor()
ctx.call_hf_processor.side_effect = lambda processor, data, kwargs: processor(
**data, **kwargs
)
ctx.model_config.hf_config = config
return ctx
@@ -89,21 +107,14 @@ def test_audio_chunk_counting(mock_ctx):
mm_data = {"audio": [audio_1, audio_2]}
prompt = "<|user|>Listen.<|end|>"
from vllm.multimodal.processing import BaseMultiModalProcessor
processed = processor._call_hf_processor(prompt, mm_data, {}, {})
def mock_base_call(self, prompt, mm_data, mm_kwargs, tok_kwargs):
return {"input_ids": [1, 2, 3], "input_features": torch.randn(1, 80, 3000)}
chunk_counts = processed["chunk_counts"]
with pytest.MonkeyPatch.context() as mp:
mp.setattr(BaseMultiModalProcessor, "_call_hf_processor", mock_base_call)
processed = processor._call_hf_processor(prompt, mm_data, {}, {})
chunk_counts = processed["chunk_counts"]
assert chunk_counts[0].item() == 1
assert chunk_counts[1].item() == 2
assert len(chunk_counts) == 2
assert chunk_counts[0].item() == 1
assert chunk_counts[1].item() == 2
assert len(chunk_counts) == 2
assert processed["feature_attention_mask"].shape == (3, 3000)
def test_dummy_data_generation(mock_ctx):
@@ -49,6 +49,13 @@ class MockMusicFlamingoProcessor:
self.max_audio_len = 1200
self.feature_extractor = MockFeatureExtractor()
def __call__(self, text=None, audio=None, **kwargs):
return {
"input_ids": torch.tensor([[1, 2, 3]], dtype=torch.long),
"input_features": torch.zeros((3, 80, 3000)),
"input_features_mask": torch.ones((3, 3000), dtype=torch.long),
}
class MockFeatureExtractor:
def __init__(self):
@@ -63,6 +70,9 @@ def mock_ctx():
ctx = MagicMock()
ctx.get_hf_config.return_value = config
ctx.get_hf_processor.return_value = MockMusicFlamingoProcessor()
ctx.call_hf_processor.side_effect = lambda processor, data, kwargs: processor(
**data, **kwargs
)
ctx.model_config.hf_config = config
return ctx
@@ -73,7 +83,7 @@ def check_transformers_version():
model_info.check_transformers_version(on_fail="skip")
def test_musicflamingo_chunk_counting_uses_rote_timestamps(mock_ctx, monkeypatch):
def test_musicflamingo_chunk_counting_without_rote_timestamps(mock_ctx):
from vllm.model_executor.models.musicflamingo import (
MusicFlamingoDummyInputsBuilder,
MusicFlamingoMultiModalProcessor,
@@ -92,24 +102,13 @@ def test_musicflamingo_chunk_counting_uses_rote_timestamps(mock_ctx, monkeypatch
mm_data = {"audio": [audio_1, audio_2]}
prompt = "<|user|>Listen.<|end|>"
from vllm.multimodal.processing import BaseMultiModalProcessor
def mock_base_call(self, prompt, mm_data, mm_kwargs, tok_kwargs):
del self, prompt, mm_data, mm_kwargs, tok_kwargs
return {
"input_ids": [1, 2, 3],
"input_features": torch.randn(3, 80, 3000),
"rote_timestamps": torch.randn(3, 750),
}
monkeypatch.setattr(BaseMultiModalProcessor, "_call_hf_processor", mock_base_call)
processed = processor._call_hf_processor(prompt, mm_data, {}, {})
chunk_counts = processed["chunk_counts"]
assert chunk_counts.tolist() == [1, 2]
assert "rote_timestamps" in processed
assert "rote_timestamps" not in processed
assert processed["feature_attention_mask"].shape == (3, 3000)
def test_musicflamingo_dummy_text_uses_plain_audio_tokens(mock_ctx):