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
@@ -776,6 +776,7 @@ class Qwen3_VisionTransformer(nn.Module):
|
||||
instead of computing from cu_seqlens (needed for CUDA
|
||||
graph capture to cover worst-case replay scenarios).
|
||||
device: Device to place tensors on. Defaults to self.device.
|
||||
|
||||
"""
|
||||
if device is None:
|
||||
device = self.device
|
||||
@@ -1291,6 +1292,7 @@ def _replace_video_token_placeholders(
|
||||
|
||||
Returns:
|
||||
Token IDs with every placeholder replaced.
|
||||
|
||||
"""
|
||||
result: list[int] = []
|
||||
repl_idx = 0
|
||||
@@ -1319,7 +1321,8 @@ class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo])
|
||||
def _expands_only_video_token(hf_processor: ProcessorMixin) -> bool:
|
||||
"""Transformers>=5.10 processors override `replace_video_token`
|
||||
to expand only the bare video token, keeping the prompt's outer
|
||||
`<|vision_start|>`/`<|vision_end|>` markers."""
|
||||
`<|vision_start|>`/`<|vision_end|>` markers.
|
||||
"""
|
||||
mixin_impl = getattr(ProcessorMixin, "replace_video_token", None)
|
||||
proc_impl = getattr(type(hf_processor), "replace_video_token", None)
|
||||
return proc_impl is not None and proc_impl is not mixin_impl
|
||||
@@ -1647,9 +1650,11 @@ class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo])
|
||||
vision_start_token_id: Token ID for vision start marker
|
||||
vision_end_token_id: Token ID for vision end marker
|
||||
video_token_id: Token ID for video content
|
||||
select_token_id: Whether to return token ids instead of text
|
||||
|
||||
Returns:
|
||||
PromptUpdateDetails with full token sequence
|
||||
|
||||
"""
|
||||
assert len(timestamps) == len(tokens_per_frame), (
|
||||
"timestamps and tokens_per_frame must have the same length"
|
||||
@@ -2350,8 +2355,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
image_embeds_split: tuple[torch.Tensor, ...],
|
||||
image_input: Qwen2_5_VLImageInputs,
|
||||
) -> tuple[torch.Tensor, ...]:
|
||||
"""
|
||||
Append mrope positions for each for images.
|
||||
"""Append mrope positions for each for images.
|
||||
This is necessary to recover correct mrope
|
||||
positions after video pruning
|
||||
|
||||
@@ -2364,6 +2368,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
Tuple of image embeddings for each image item.
|
||||
Resulting embeddings will have extra 5 channels for
|
||||
computed mrope positions, consistent with video embeddings.
|
||||
|
||||
"""
|
||||
if self.is_multimodal_pruning_enabled:
|
||||
merge_size = self.visual.spatial_merge_size
|
||||
@@ -2393,8 +2398,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
video_embeds_split: tuple[torch.Tensor, ...],
|
||||
video_input: Qwen2_5_VLVideoInputs,
|
||||
) -> tuple[torch.Tensor, ...]:
|
||||
"""
|
||||
Prunes video embeddings via Efficient Video Sampling (EVS)
|
||||
"""Prunes video embeddings via Efficient Video Sampling (EVS)
|
||||
and then appends mrope positions for each retained embeddings
|
||||
|
||||
Args:
|
||||
@@ -2405,6 +2409,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
Tuple of video embeddings for each video item.
|
||||
Resulting embeddings will have extra 5 channels for computed mrope
|
||||
positions, and whether the index corresponds to a video embedding.
|
||||
|
||||
"""
|
||||
grid_thw = video_input["video_grid_thw"]
|
||||
assert grid_thw.ndim == 2
|
||||
@@ -2478,7 +2483,6 @@ class Qwen3VLForConditionalGeneration(
|
||||
These embeddings will replace the placeholder embeddings to create
|
||||
input_embeds for the LLM.
|
||||
"""
|
||||
|
||||
device = video_embeddings.device
|
||||
|
||||
# Generate video replacement token IDs using get_video_repl
|
||||
@@ -2664,6 +2668,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
llm_grid_h: Logical grid height (may not match actual token count with EVS).
|
||||
llm_grid_w: Logical grid width (may not match actual token count with EVS).
|
||||
actual_num_tokens: Actual number of video/image tokens in the placeholder.
|
||||
|
||||
"""
|
||||
for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset):
|
||||
offset = mm_feature.mm_position.offset
|
||||
@@ -2806,8 +2811,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
mrope_positions: torch.LongTensor,
|
||||
num_computed_tokens: int,
|
||||
) -> tuple[Sequence[torch.Tensor], torch.Tensor, int]:
|
||||
"""
|
||||
Update part of input mrope positions (starting with
|
||||
"""Update part of input mrope positions (starting with
|
||||
num_computed_tokens index). Original mrope_positions are computed
|
||||
for unpruned sequence and becomes incorrect once pruning occurs,
|
||||
so once we prune media tokens we should reflect this in the
|
||||
@@ -2825,6 +2829,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
Returns:
|
||||
Tuple of (multimodal_embeddings, mrope_positions,
|
||||
mrope_position_delta).
|
||||
|
||||
"""
|
||||
return self._recompute_mrope_positions(
|
||||
input_ids=input_ids,
|
||||
@@ -3031,8 +3036,8 @@ class Qwen3VLForConditionalGeneration(
|
||||
model. `None` if no videos are passed.
|
||||
- video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in
|
||||
LLM. `None` if no videos are passed.
|
||||
"""
|
||||
|
||||
"""
|
||||
if intermediate_tensors is not None:
|
||||
inputs_embeds = None
|
||||
|
||||
@@ -3068,9 +3073,7 @@ class Qwen3VLForConditionalGeneration(
|
||||
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
|
||||
|
||||
def get_mm_mapping(self) -> MultiModelKeys:
|
||||
"""
|
||||
Get the module prefix in multimodal models
|
||||
"""
|
||||
"""Get the module prefix in multimodal models."""
|
||||
return MultiModelKeys.from_string_field(
|
||||
language_model="language_model",
|
||||
connector=["visual.merger", "visual.deepstack_merger_list"],
|
||||
|
||||
Reference in New Issue
Block a user