Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -54,7 +54,8 @@ def test_audio_media_io_load_base64(dummy_audio_bytes):
|
||||
def test_audio_media_io_load_base64_rejects_malformed(dummy_audio_bytes):
|
||||
"""Malformed base64 must surface as a ValueError so the server answers 400.
|
||||
Without strict decoding the bad characters are dropped, the garbage reaches
|
||||
libsndfile, and the client gets a 500 instead."""
|
||||
libsndfile, and the client gets a 500 instead.
|
||||
"""
|
||||
encoded = base64.b64encode(dummy_audio_bytes).decode("utf-8")
|
||||
malformed = encoded[:8] + "!!!@@@###" + encoded[8:]
|
||||
|
||||
@@ -133,7 +134,8 @@ def test_small_file_passes_memory_guard():
|
||||
|
||||
def test_memory_guard_rejects_large_allocation():
|
||||
"""A file whose frames*channels*4 exceeds the byte limit must be
|
||||
rejected before allocating the buffer."""
|
||||
rejected before allocating the buffer.
|
||||
"""
|
||||
# 100_000 frames * 8 channels * 4 bytes = 3.2 MB
|
||||
payload = _make_flac_bytes(frames=100_000, channels=8, samplerate=48000)
|
||||
# Set limit to 1 MiB — should reject
|
||||
@@ -148,7 +150,8 @@ def test_memory_guard_rejects_large_allocation():
|
||||
|
||||
def test_forged_samplerate_rejected_by_memory_guard():
|
||||
"""The PoC scenario: high sample rate fools the duration guard but
|
||||
the memory guard catches the large frame*channel allocation."""
|
||||
the memory guard catches the large frame*channel allocation.
|
||||
"""
|
||||
# Forged high sample rate: 655350 Hz, 8 channels, 1M frames
|
||||
# Duration guard sees: 1_000_000 / 655_350 = 1.5s → passes
|
||||
# Memory: 1_000_000 * 8 * 4 = 32 MB
|
||||
@@ -209,7 +212,8 @@ def test_load_audio_unknown_backend_rejected(dummy_audio_bytes):
|
||||
|
||||
def test_load_audio_auto_falls_back_without_torchcodec(dummy_audio_bytes):
|
||||
"""`auto` must fall back to the soundfile → PyAV chain when torchcodec
|
||||
is not importable."""
|
||||
is not importable.
|
||||
"""
|
||||
ref_audio, ref_sr = load_audio_soundfile(BytesIO(dummy_audio_bytes), sr=None)
|
||||
with patch.object(audio_module, "load_audio_torchcodec", side_effect=ImportError):
|
||||
audio, sr = load_audio(BytesIO(dummy_audio_bytes), sr=None, backend="auto")
|
||||
@@ -218,8 +222,9 @@ def test_load_audio_auto_falls_back_without_torchcodec(dummy_audio_bytes):
|
||||
|
||||
|
||||
def test_load_audio_auto_falls_back_without_ffmpeg(dummy_audio_bytes):
|
||||
"""torchcodec installed but system ffmpeg missing (`AudioDecoder is None`)
|
||||
must surface as ImportError so `auto` falls back to soundfile → PyAV."""
|
||||
"""Torchcodec installed but system ffmpeg missing (`AudioDecoder is None`)
|
||||
must surface as ImportError so `auto` falls back to soundfile → PyAV.
|
||||
"""
|
||||
ref_audio, ref_sr = load_audio_soundfile(BytesIO(dummy_audio_bytes), sr=None)
|
||||
with patch.object(audio_module, "AudioDecoder", None):
|
||||
with pytest.raises(ImportError, match="torchcodec audio backend"):
|
||||
@@ -232,9 +237,10 @@ def test_load_audio_auto_falls_back_without_ffmpeg(dummy_audio_bytes):
|
||||
def test_load_audio_auto_falls_back_when_libtorchcodec_unloadable(
|
||||
dummy_audio_bytes,
|
||||
):
|
||||
"""torchcodec loads its ffmpeg-backed core lazily at decoder construction;
|
||||
"""Torchcodec loads its ffmpeg-backed core lazily at decoder construction;
|
||||
a "Could not load libtorchcodec" RuntimeError there (no system ffmpeg)
|
||||
must also surface as ImportError so `auto` falls back."""
|
||||
must also surface as ImportError so `auto` falls back.
|
||||
"""
|
||||
|
||||
class _UnloadableDecoder:
|
||||
def __init__(self, *args, **kwargs):
|
||||
|
||||
@@ -228,7 +228,8 @@ async def test_fetch_image_local_files_with_space_in_name(image_url: str):
|
||||
@pytest.mark.asyncio
|
||||
async def test_fetch_image_data_url_with_params():
|
||||
"""RFC 2397 allows parameters between the mediatype and the base64
|
||||
marker; they must not be rejected or leak into the media type."""
|
||||
marker; they must not be rejected or leak into the media type.
|
||||
"""
|
||||
connector = MediaConnector()
|
||||
|
||||
image = Image.new("RGB", (4, 4), color=(255, 0, 0))
|
||||
|
||||
@@ -105,7 +105,6 @@ def test_image_media_io_no_mode_conversion(tmp_path):
|
||||
|
||||
def test_image_media_io_rgba_background_color_validation():
|
||||
"""Test that invalid rgba_background_color values are properly rejected."""
|
||||
|
||||
# Test invalid types
|
||||
with pytest.raises(
|
||||
ValueError, match="rgba_background_color must be a list or tuple"
|
||||
|
||||
@@ -77,8 +77,7 @@ def test_video_media_io_kwargs(monkeypatch: pytest.MonkeyPatch):
|
||||
@pytest.mark.parametrize("is_color", [True, False])
|
||||
@pytest.mark.parametrize("fourcc, ext", [("mp4v", "mp4"), ("XVID", "avi")])
|
||||
def test_opencv_video_io_colorspace(tmp_path, is_color: bool, fourcc: str, ext: str):
|
||||
"""
|
||||
Test all functions that use OpenCV for video I/O return RGB format.
|
||||
"""Test all functions that use OpenCV for video I/O return RGB format.
|
||||
Both RGB and grayscale videos are tested.
|
||||
"""
|
||||
image_path = get_vllm_public_assets(
|
||||
@@ -168,8 +167,7 @@ class TestVideoBackendOverride2(VideoLoader):
|
||||
|
||||
|
||||
def test_video_media_io_backend_kwarg_override(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test that video_backend kwarg can override the VLLM_VIDEO_LOADER_BACKEND
|
||||
"""Test that video_backend kwarg can override the VLLM_VIDEO_LOADER_BACKEND
|
||||
environment variable.
|
||||
|
||||
This allows users to dynamically select a different video backend
|
||||
@@ -201,8 +199,7 @@ def test_video_media_io_backend_kwarg_override(monkeypatch: pytest.MonkeyPatch):
|
||||
def test_video_media_io_backend_kwarg_not_passed_to_loader(
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
):
|
||||
"""
|
||||
Test that video_backend kwarg is consumed by VideoMediaIO and NOT passed
|
||||
"""Test that video_backend kwarg is consumed by VideoMediaIO and NOT passed
|
||||
through to the underlying video loader's load_bytes method.
|
||||
|
||||
This ensures the kwarg is properly popped from kwargs before forwarding.
|
||||
@@ -245,8 +242,7 @@ def test_video_media_io_backend_kwarg_not_passed_to_loader(
|
||||
|
||||
|
||||
def test_video_media_io_backend_env_var_fallback(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test that when video_backend kwarg is None or not provided,
|
||||
"""Test that when video_backend kwarg is None or not provided,
|
||||
VideoMediaIO falls back to VLLM_VIDEO_LOADER_BACKEND env var.
|
||||
"""
|
||||
with monkeypatch.context() as m:
|
||||
@@ -285,7 +281,6 @@ def test_load_base64_jpeg_returns_metadata():
|
||||
metadata, which broke downstream consumers that rely on fields like
|
||||
total_num_frames and fps. See PR #37301.
|
||||
"""
|
||||
|
||||
num_test_frames = 3
|
||||
|
||||
b64_frames = _make_jpeg_b64_frames(num_test_frames)
|
||||
@@ -403,7 +398,8 @@ def test_pynvvideocodec_unrelated_error_propagates(
|
||||
|
||||
class TestMergeKwargsGpuBackendPolicy:
|
||||
"""Verify that merge_kwargs blocks request-level GPU backend selection
|
||||
when the static (engine-level) config did not configure that backend."""
|
||||
when the static (engine-level) config did not configure that backend.
|
||||
"""
|
||||
|
||||
def test_pynvvideocodec_requires_gpu(self):
|
||||
assert VIDEO_LOADER_REGISTRY.backend_requires_gpu(PYNVVIDEOCODEC_VIDEO_BACKEND)
|
||||
@@ -496,7 +492,8 @@ class TestMergeKwargsGpuBackendPolicy:
|
||||
|
||||
def test_static_pynv_with_different_runtime_gpu_backend(self):
|
||||
"""If static sets pynv via video_backend but runtime tries to set it
|
||||
via the codec-level 'backend' key (without a static match), strip it."""
|
||||
via the codec-level 'backend' key (without a static match), strip it.
|
||||
"""
|
||||
result = VideoMediaIO.merge_kwargs(
|
||||
default_kwargs={"video_backend": "pynvvideocodec"},
|
||||
runtime_kwargs={"backend": "pynvvideocodec"},
|
||||
@@ -535,7 +532,8 @@ class TestMergeKwargsGpuBackendPolicy:
|
||||
def test_strips_request_level_device(self):
|
||||
"""The decode device is a startup-only knob: a request must not move
|
||||
decoding onto the GPU when the startup config did not opt in, nor off
|
||||
it when it did."""
|
||||
it when it did.
|
||||
"""
|
||||
result = VideoMediaIO.merge_kwargs(
|
||||
default_kwargs={"backend": "torchcodec"},
|
||||
runtime_kwargs={"device": "cuda"},
|
||||
@@ -560,7 +558,8 @@ class TestMergeKwargsGpuBackendPolicy:
|
||||
def test_switching_backend_drops_stale_codec_options(default_kwargs):
|
||||
"""Codec-specific options from the static config must not leak into a
|
||||
different codec backend selected per-request, where they would fail the
|
||||
new backend's option validation."""
|
||||
new backend's option validation.
|
||||
"""
|
||||
result = VideoMediaIO.merge_kwargs(
|
||||
default_kwargs={**default_kwargs, "num_frames": 8},
|
||||
runtime_kwargs={"backend": "opencv"},
|
||||
@@ -582,7 +581,8 @@ def test_pynvvc_frames_normalized_to_nhwc(layout: str):
|
||||
"""PyNvVideoCodec frame batches are normalized to NHWC regardless of the
|
||||
per-frame layout the decoder emits (it has varied across versions), so the
|
||||
HF video processors (which materialize a PIL image per frame) receive the
|
||||
same NHWC shape as every other video backend."""
|
||||
same NHWC shape as every other video backend.
|
||||
"""
|
||||
torch = pytest.importorskip("torch")
|
||||
|
||||
n, h, w, c = 4, 5, 6, 3
|
||||
|
||||
@@ -695,7 +695,6 @@ class TestAudioChunking:
|
||||
|
||||
def test_split_audio_short_clip(self):
|
||||
"""Audio shorter than max_clip_duration_s should not be split."""
|
||||
|
||||
# 10 seconds of audio at 16kHz
|
||||
audio = np.linspace(-1.0, 1.0, 160000, dtype=np.float32)
|
||||
|
||||
@@ -712,7 +711,6 @@ class TestAudioChunking:
|
||||
|
||||
def test_split_audio_exact_length(self):
|
||||
"""Audio exactly at max_clip_duration_s should not be split."""
|
||||
|
||||
# Exactly 30 seconds at 16kHz
|
||||
audio = np.linspace(-1.0, 1.0, 480000, dtype=np.float32)
|
||||
|
||||
@@ -729,7 +727,6 @@ class TestAudioChunking:
|
||||
|
||||
def test_split_audio_long_clip(self):
|
||||
"""Long audio should be split into multiple chunks."""
|
||||
|
||||
# 65 seconds of audio at 16kHz
|
||||
audio = np.linspace(-1.0, 1.0, 1040000, dtype=np.float32)
|
||||
|
||||
@@ -749,7 +746,6 @@ class TestAudioChunking:
|
||||
|
||||
def test_split_audio_chunks_have_correct_length(self):
|
||||
"""Each chunk (except last) should be approximately max_clip_duration_s."""
|
||||
|
||||
# 65 seconds of audio at 16kHz
|
||||
audio = np.linspace(-1.0, 1.0, 1040000, dtype=np.float32)
|
||||
|
||||
@@ -841,7 +837,6 @@ class TestAudioChunking:
|
||||
|
||||
def test_split_audio_preserves_boundaries(self):
|
||||
"""Verify first and last samples are preserved when chunking."""
|
||||
|
||||
audio = np.arange(1120000, dtype=np.float32) # 70s at 16kHz
|
||||
|
||||
chunks = split_audio(
|
||||
@@ -893,7 +888,6 @@ class TestAudioChunking:
|
||||
|
||||
def test_split_audio_with_different_sample_rates(self):
|
||||
"""Test chunking works with different sample rates."""
|
||||
|
||||
# 40 seconds at 8kHz
|
||||
audio_8k = np.linspace(-1.0, 1.0, 320000, dtype=np.float32)
|
||||
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Embedding shape validation in multimodal APIs.
|
||||
"""Embedding shape validation in multimodal APIs.
|
||||
|
||||
Tests verify that embeddings with correct ndim but incorrect hidden_size
|
||||
are rejected before they can cause crashes during model inference.
|
||||
@@ -178,8 +177,7 @@ class TestShapeValidationIntegration:
|
||||
"""Integration tests verifying attack scenarios are blocked."""
|
||||
|
||||
def test_attack_scenario_multimodal_image(self):
|
||||
"""
|
||||
Simulate attack through Chat API with image embeddings.
|
||||
"""Simulate attack through Chat API with image embeddings.
|
||||
|
||||
Verifies validation occurs in multimodal parser path.
|
||||
"""
|
||||
@@ -193,8 +191,7 @@ class TestShapeValidationIntegration:
|
||||
parser.parse_mm_data({"image": attack_tensor})
|
||||
|
||||
def test_attack_scenario_multimodal_audio(self):
|
||||
"""
|
||||
Simulate attack through Chat API with audio embeddings.
|
||||
"""Simulate attack through Chat API with audio embeddings.
|
||||
|
||||
Verifies validation occurs in multimodal parser path.
|
||||
"""
|
||||
@@ -208,8 +205,7 @@ class TestShapeValidationIntegration:
|
||||
parser.parse_mm_data({"audio": attack_tensor})
|
||||
|
||||
def test_attack_scenario_multimodal_video(self):
|
||||
"""
|
||||
Simulate attack through Chat API with video embeddings.
|
||||
"""Simulate attack through Chat API with video embeddings.
|
||||
|
||||
Verifies validation occurs in multimodal parser path.
|
||||
"""
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Unit tests for embedding shape validation.
|
||||
"""Unit tests for embedding shape validation.
|
||||
|
||||
Simple, fast unit tests that can run without server fixtures.
|
||||
Run with: pytest tests/multimodal/test_embedding_shape_validation_unit.py -v
|
||||
@@ -167,16 +166,14 @@ class TestAudioEmbedBasicValidation:
|
||||
|
||||
|
||||
class TestShapeValidationDoSPrevention:
|
||||
"""
|
||||
Tests for DoS prevention through shape validation.
|
||||
"""Tests for DoS prevention through shape validation.
|
||||
|
||||
Verifies that embeddings with incorrect shapes are rejected early,
|
||||
preventing crashes during model inference.
|
||||
"""
|
||||
|
||||
def test_prevent_crash_from_wrong_shape_image_embeds(self):
|
||||
"""
|
||||
Prevent crash scenario: wrong hidden size in image embeddings.
|
||||
"""Prevent crash scenario: wrong hidden size in image embeddings.
|
||||
|
||||
Without validation, this would pass initial checks but crash later
|
||||
during model forward pass when dimensions don't match.
|
||||
@@ -198,9 +195,7 @@ class TestShapeValidationDoSPrevention:
|
||||
assert str(wrong_hidden_size) in error_msg # Received
|
||||
|
||||
def test_prevent_crash_from_wrong_shape_audio_embeds(self):
|
||||
"""
|
||||
Prevent crash scenario: wrong hidden size in audio embeddings.
|
||||
"""
|
||||
"""Prevent crash scenario: wrong hidden size in audio embeddings."""
|
||||
expected_hidden_size = 768
|
||||
wrong_hidden_size = 4096
|
||||
|
||||
|
||||
@@ -141,7 +141,8 @@ def test_hash_collision_video_num_frames():
|
||||
|
||||
def test_hash_video_tensor_frames():
|
||||
"""Videos holding tensor frames (e.g. NVDEC-decoded) hash like
|
||||
array-framed ones, from the original bytes without a D2H copy."""
|
||||
array-framed ones, from the original bytes without a D2H copy.
|
||||
"""
|
||||
source = b"x" * 100
|
||||
|
||||
def item_for_hash(frames):
|
||||
|
||||
@@ -80,7 +80,8 @@ def test_l_mode_no_trns_to_rgb():
|
||||
|
||||
def test_exif_transpose_normalizes_orientation():
|
||||
"""Image with EXIF orientation 3 (180-degree rotation) should be
|
||||
normalized so pixel data matches visual display."""
|
||||
normalized so pixel data matches visual display.
|
||||
"""
|
||||
# Built synthetically because the existing assets are PNGs which
|
||||
# don't carry EXIF orientation metadata; we need a JPEG with an
|
||||
# injected EXIF orientation=3 tag.
|
||||
|
||||
@@ -59,7 +59,8 @@ def test_frame_size_hwc_chw(frame):
|
||||
def test_video_with_metadata_tensor_passthrough():
|
||||
"""Tensor frames pass through unchanged regardless of device: HF video
|
||||
processors accept tensors, and device-resident frames (e.g. NVDEC-decoded)
|
||||
must not be copied back to host."""
|
||||
must not be copied back to host.
|
||||
"""
|
||||
frames = torch.zeros((4, H, W, 3), dtype=torch.uint8)
|
||||
video, metadata = MultiModalDataParser()._get_video_with_metadata(frames)
|
||||
|
||||
@@ -70,7 +71,8 @@ def test_video_with_metadata_tensor_passthrough():
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="Requires CUDA")
|
||||
def test_video_with_metadata_keeps_device_tensor():
|
||||
"""Device-resident frames (e.g. NVDEC-decoded) pass through as tensors,
|
||||
so a device-side HF processor can consume them without a D2H copy."""
|
||||
so a device-side HF processor can consume them without a D2H copy.
|
||||
"""
|
||||
frames = torch.zeros((4, H, W, 3), dtype=torch.uint8, device="cuda")
|
||||
video, metadata = MultiModalDataParser()._get_video_with_metadata(frames)
|
||||
|
||||
|
||||
@@ -968,8 +968,7 @@ def test_hf_processor_call_kwargs(
|
||||
|
||||
|
||||
def test_apply_matches_no_match_exits_quickly():
|
||||
"""
|
||||
Test that _apply_matches exits quickly when no matches are found.
|
||||
"""Test that _apply_matches exits quickly when no matches are found.
|
||||
|
||||
Previously, _apply_matches had O(n²) behavior when no match was found
|
||||
because it would increment start_idx by 1 each iteration while
|
||||
@@ -1027,8 +1026,7 @@ def test_iter_token_matches_rejects_negative_start_idx():
|
||||
|
||||
|
||||
def test_find_mm_placeholders_avoids_quadratic_false_prefixes():
|
||||
"""
|
||||
Test that placeholder scanning stays linear under adversarial candidates.
|
||||
"""Test that placeholder scanning stays linear under adversarial candidates.
|
||||
|
||||
The fast-forward scan must not rescan the prompt tail per position when
|
||||
one candidate's first token never occurs (forcing a full search) while
|
||||
@@ -1061,8 +1059,7 @@ def test_find_mm_placeholders_avoids_quadratic_false_prefixes():
|
||||
],
|
||||
)
|
||||
def test_find_mm_placeholders_stops_at_missing_item(prompt):
|
||||
"""
|
||||
Test that the scan returns no placeholders once it fails to find
|
||||
"""Test that the scan returns no placeholders once it fails to find
|
||||
an item's placeholder, leaving later items unresolved.
|
||||
"""
|
||||
result = find_mm_placeholders(
|
||||
@@ -1079,8 +1076,7 @@ def test_find_mm_placeholders_stops_at_missing_item(prompt):
|
||||
|
||||
|
||||
class _FakeTokenizer:
|
||||
"""
|
||||
Character-level tokenizer where "foo" merges into one token differently
|
||||
"""Character-level tokenizer where "foo" merges into one token differently
|
||||
depending on whether it is followed by "d", like BPE merging "foo" in
|
||||
"food" across the search-text boundary.
|
||||
"""
|
||||
@@ -1145,8 +1141,7 @@ def _text_fallback_processor() -> BaseMultiModalProcessor:
|
||||
|
||||
|
||||
def test_apply_prompt_updates_falls_back_to_text_matching():
|
||||
"""
|
||||
Test that the fallback in `_apply_prompt_updates` finds targets that
|
||||
"""Test that the fallback in `_apply_prompt_updates` finds targets that
|
||||
tokenize differently inside the prompt ("foo" in "food").
|
||||
"""
|
||||
processor = _text_fallback_processor()
|
||||
@@ -1166,8 +1161,7 @@ def test_apply_prompt_updates_falls_back_to_text_matching():
|
||||
|
||||
|
||||
def test_apply_prompt_updates_falls_back_with_prefix_target():
|
||||
"""
|
||||
Test that `PromptIndexTargets.prefix` targets are resolved against the
|
||||
"""Test that `PromptIndexTargets.prefix` targets are resolved against the
|
||||
decoded text in the fallback path of `_apply_prompt_updates`.
|
||||
"""
|
||||
processor = _text_fallback_processor()
|
||||
@@ -1192,8 +1186,7 @@ def test_apply_prompt_updates_falls_back_with_prefix_target():
|
||||
|
||||
|
||||
def test_apply_prompt_updates_falls_back_with_index_targets():
|
||||
"""
|
||||
Test that the text resolvers of `PromptIndexTargets.start`/`end`
|
||||
"""Test that the text resolvers of `PromptIndexTargets.start`/`end`
|
||||
match against the decoded text when another item forces the
|
||||
fallback in `_apply_prompt_updates`.
|
||||
"""
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Unit tests for MultiModalRegistry.supports_multimodal_inputs and
|
||||
"""Unit tests for MultiModalRegistry.supports_multimodal_inputs and
|
||||
Qwen2.5-VL visual component loading behavior.
|
||||
"""
|
||||
|
||||
@@ -58,7 +57,8 @@ def test_dummy_inputs_scheduler_budget(
|
||||
@pytest.mark.core_model
|
||||
def test_supports_multimodal_inputs(model_id, limit_mm_per_prompt, expected):
|
||||
"""Test supports_multimodal_inputs returns correct boolean for various
|
||||
configs."""
|
||||
configs.
|
||||
"""
|
||||
ctx = build_model_context(
|
||||
model_id,
|
||||
limit_mm_per_prompt=limit_mm_per_prompt,
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Unit tests for sparse tensor validation.
|
||||
"""Unit tests for sparse tensor validation.
|
||||
|
||||
Simple, fast unit tests that can run without server fixtures.
|
||||
Run with: pytest tests/multimodal/test_sparse_tensor_validation_unit.py -v
|
||||
@@ -56,8 +55,7 @@ class TestSparseTensorValidationContextManager:
|
||||
tensor.to_dense()
|
||||
|
||||
def test_without_context_manager_allows_invalid(self):
|
||||
"""
|
||||
WITHOUT validation, invalid tensors may not immediately error.
|
||||
"""WITHOUT validation, invalid tensors may not immediately error.
|
||||
|
||||
This demonstrates the vulnerability: PyTorch 2.8.0+ doesn't validate
|
||||
by default, which can lead to memory corruption.
|
||||
|
||||
@@ -231,7 +231,8 @@ def test_group_and_batch_mm_items_splits_shared_data_by_dtype():
|
||||
|
||||
def test_encode_image_url_uncommon_format_has_valid_mimetype():
|
||||
"""A format missing from mimetypes.types_map must still produce a
|
||||
well-formed type/subtype media type, not a bare 'image'."""
|
||||
well-formed type/subtype media type, not a bare 'image'.
|
||||
"""
|
||||
# ".tga" is not registered in mimetypes.types_map, so it exercises the
|
||||
# fallback path; Pillow can still encode it.
|
||||
url = encode_image_url(Image.new("RGB", (2, 2)), format="TGA")
|
||||
|
||||
@@ -842,8 +842,7 @@ def test_video_processor_from_model_repo(
|
||||
|
||||
|
||||
def test_video_backend_handles_broken_frames(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Regression test for handling videos with broken frames.
|
||||
"""Regression test for handling videos with broken frames.
|
||||
This test uses a pre-corrupted video file (assets/corrupted.mp4) that
|
||||
contains broken frames to verify the video loader handles
|
||||
them gracefully without crashing and returns accurate metadata.
|
||||
@@ -882,8 +881,7 @@ def test_video_backend_handles_broken_frames(monkeypatch: pytest.MonkeyPatch):
|
||||
def test_video_backend_handles_edit_list_trimmed_video(
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
):
|
||||
"""
|
||||
An mp4 edit list (e.g. from a lossless ``ffmpeg -ss ... -c copy`` cut)
|
||||
"""An mp4 edit list (e.g. from a lossless ``ffmpeg -ss ... -c copy`` cut)
|
||||
hides the decode lead-in: the header still counts every physical sample
|
||||
while sequential decode only yields the visible frames. Sampling over the
|
||||
header count used to collapse such videos to the few indices below the
|
||||
@@ -923,8 +921,7 @@ def test_video_backend_handles_edit_list_trimmed_video(
|
||||
|
||||
|
||||
def test_video_recovery_simulated_failures(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test that frame recovery correctly uses the next valid frame when
|
||||
"""Test that frame recovery correctly uses the next valid frame when
|
||||
target frames fail to load.
|
||||
|
||||
Uses corrupted.mp4 and mocks VideoCapture.grab() to fail on specific
|
||||
@@ -1019,8 +1016,7 @@ def test_video_recovery_simulated_failures(monkeypatch: pytest.MonkeyPatch):
|
||||
|
||||
|
||||
def test_video_recovery_with_corrupted_file(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test frame recovery with an actual corrupted video file using sparse sampling.
|
||||
"""Test frame recovery with an actual corrupted video file using sparse sampling.
|
||||
|
||||
This test uses corrupted.mp4 which has genuine H.264 codec errors on
|
||||
frame 17. With num_frames=8, the target frames are [0, 3, 7, 10, 14, 17, 21, 25].
|
||||
@@ -1088,8 +1084,7 @@ def test_video_recovery_with_corrupted_file(monkeypatch: pytest.MonkeyPatch):
|
||||
|
||||
|
||||
def test_video_recovery_dynamic_backend(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test that frame_recovery works with the dynamic video backend.
|
||||
"""Test that frame_recovery works with the dynamic video backend.
|
||||
|
||||
The dynamic backend samples frames based on fps/duration rather than
|
||||
loading all frames. This test verifies recovery works in that context.
|
||||
@@ -1219,7 +1214,8 @@ def test_torchcodec_backend_rejects_frame_recovery(dummy_video_path):
|
||||
@pytest.mark.skipif(not current_platform.is_cuda(), reason="Requires CUDA")
|
||||
def test_torchcodec_backend_cuda_decodes_on_gpu(dummy_video_path):
|
||||
"""With device="cuda", torchcodec decodes via NVDEC and keeps the frames
|
||||
on the GPU instead of returning a host-side numpy array."""
|
||||
on the GPU instead of returning a host-side numpy array.
|
||||
"""
|
||||
pytest.importorskip("torchcodec")
|
||||
|
||||
with open(dummy_video_path, "rb") as f:
|
||||
@@ -1512,7 +1508,8 @@ def test_glm46v_duration_estimation_from_fps():
|
||||
class TestGLMGASamplingCaps:
|
||||
"""Regression tests for GHSA-58v5-2m8f-94pr: request-controlled fps
|
||||
and max_frames must not create intermediate allocations larger than
|
||||
the actual frame count."""
|
||||
the actual frame count.
|
||||
"""
|
||||
|
||||
@staticmethod
|
||||
def _source(
|
||||
@@ -1566,7 +1563,8 @@ def test_glm5next_backend_selected_for_processor():
|
||||
"""Glm5NextVideoProcessor maps to the glm5next loader so only the
|
||||
sampled frames are decoded instead of the whole container. Both the
|
||||
borrowed-config spelling and the dedicated Glm5next class name (landing
|
||||
with the new checkpoint) must resolve."""
|
||||
with the new checkpoint) must resolve.
|
||||
"""
|
||||
for name in ("Glm5NextVideoProcessor", "Glm5nextVideoProcessor"):
|
||||
assert VIDEO_LOADER_REGISTRY.get_backend_for_video_processor(name) == "glm5next"
|
||||
|
||||
@@ -1588,7 +1586,8 @@ def test_glm5next_backend_indices_match_sampler(
|
||||
total_frames, original_fps, duration, fps, max_frames
|
||||
):
|
||||
"""The loader must select exactly the frames the processor's sampler
|
||||
would, with target.fps mapping onto the raw-fps override."""
|
||||
would, with target.fps mapping onto the raw-fps override.
|
||||
"""
|
||||
from vllm.transformers_utils.processors.glm5next import (
|
||||
glm_sample_frame_indices,
|
||||
)
|
||||
@@ -1616,7 +1615,8 @@ def test_glm5next_backend_indices_match_sampler(
|
||||
|
||||
def test_glm5next_backend_metadata_contract():
|
||||
"""create_hf_metadata reports the subset so the processor skips
|
||||
re-sampling (do_sample_frames=False) and keeps the original totals."""
|
||||
re-sampling (do_sample_frames=False) and keeps the original totals.
|
||||
"""
|
||||
source = VideoSourceMetadata(total_frames_num=900, original_fps=30.0, duration=30.0)
|
||||
target = VideoTargetMetadata(num_frames=-1, fps=-1, max_duration=-1)
|
||||
indices = Glm5NextVideoBackend.compute_frames_index_to_sample(source, target)
|
||||
@@ -1645,7 +1645,8 @@ def test_glm5next_backend_metadata_contract():
|
||||
|
||||
def _write_gray_video(tmp_path, total_frames, fps, size=(32, 32)):
|
||||
"""Synthetic clip whose frame i is flat gray level i (near-lossless under
|
||||
mp4v), so a decoded frame's level maps back to its source index."""
|
||||
mp4v), so a decoded frame's level maps back to its source index.
|
||||
"""
|
||||
cv2 = pytest.importorskip("cv2")
|
||||
|
||||
path = tmp_path / f"gray_{total_frames}_{fps}.mp4"
|
||||
@@ -1686,7 +1687,8 @@ class _CountingCap:
|
||||
def test_glm5next_backend_codec_parity(tmp_path, backend):
|
||||
"""Every codec samples the same GLM indices and decodes the same
|
||||
frames; the OpenCV seek reader and torchcodec batched index-exact decode
|
||||
must agree."""
|
||||
must agree.
|
||||
"""
|
||||
if backend == "torchcodec":
|
||||
pytest.importorskip("torchcodec")
|
||||
|
||||
@@ -1716,7 +1718,8 @@ def test_glm5next_backend_codec_parity(tmp_path, backend):
|
||||
|
||||
def test_glm5next_backend_decodes_only_sampled_frames(tmp_path):
|
||||
"""End to end over a synthetic clip: load_bytes returns exactly the
|
||||
sampler's frame count, with the right frame content at each index."""
|
||||
sampler's frame count, with the right frame content at each index.
|
||||
"""
|
||||
pytest.importorskip("cv2")
|
||||
|
||||
total_frames, fps = 60, 10
|
||||
@@ -1742,7 +1745,8 @@ def test_glm5next_backend_decodes_only_sampled_frames(tmp_path):
|
||||
|
||||
def test_glm5next_read_frames_seeks_past_large_gaps(tmp_path):
|
||||
"""Sparse targets must not walk the container: the stock reader grabs
|
||||
every frame up to the last index; the GLM reader seeks instead."""
|
||||
every frame up to the last index; the GLM reader seeks instead.
|
||||
"""
|
||||
cv2 = pytest.importorskip("cv2")
|
||||
|
||||
total_frames, fps = 200, 10
|
||||
@@ -1769,7 +1773,8 @@ def test_glm5next_read_frames_seeks_past_large_gaps(tmp_path):
|
||||
|
||||
def test_glm5next_read_frames_dense_walk_matches_stock(tmp_path):
|
||||
"""Dense targets keep the sequential walk (seeking would be slower) and
|
||||
return the same frames as the stock reader."""
|
||||
return the same frames as the stock reader.
|
||||
"""
|
||||
cv2 = pytest.importorskip("cv2")
|
||||
|
||||
total_frames, fps = 120, 10
|
||||
|
||||
Reference in New Issue
Block a user