[CI] Ignore ruff D209, rejoin the docstrings it split, and silence incompatible-rule warnings (#57212)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-16 10:57:24 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 1085b64425
commit 651a88c09a
693 changed files with 2088 additions and 4171 deletions
@@ -36,8 +36,7 @@ class BaseModelLoader(ABC):
@abstractmethod
def load_weights(self, model: nn.Module, model_config: ModelConfig) -> None:
"""Load weights into a model. This standalone API allows
inplace weights loading for an already-initialized model
"""
inplace weights loading for an already-initialized model"""
raise NotImplementedError
def create_model(
@@ -135,8 +135,7 @@ class DefaultModelLoader(BaseModelLoader):
) -> tuple[str, list[str], bool]:
"""Prepare weights for the model.
If the model is not local, it will be downloaded.
"""
If the model is not local, it will be downloaded."""
model_name_or_path = (
maybe_download_from_modelscope(model_name_or_path, revision)
or model_name_or_path
@@ -23,8 +23,7 @@ def parse_expert_id(weight_name: str) -> int | None:
Returns ``None`` for dense weights (attention, layernorm, embedding),
shared experts, and 3D fused-expert tensors where all experts are stored
in a single tensor without a numeric expert id in the name.
"""
in a single tensor without a numeric expert id in the name."""
m = _EXPERT_ID_RE.search(weight_name)
return int(m.group(1)) if m else None
@@ -71,8 +70,7 @@ def should_skip_weight(
local_expert_ids: set[int] | None,
) -> bool:
"""Return ``True`` if *weight_name* is an expert weight that does not
belong to the local rank and should be skipped during loading.
"""
belong to the local rank and should be skipped during loading."""
if local_expert_ids is None:
return False
eid = parse_expert_id(weight_name)
@@ -308,8 +308,7 @@ def _reload_attention_scales(layer: torch.nn.Module, info: LayerReloadingInfo) -
Assumes dtype/shapes of attention tensors do not change during
processing, since we use .data.copy_() to preserve kernel tensor
references.
"""
references."""
quant_method = getattr(layer, "quant_method", None)
if quant_method is not None:
# Re-create scale Parameters with sentinel values so unloaded scales
@@ -388,8 +387,7 @@ def _get_weight_loader(tensor: torch.Tensor):
def _copy_and_restore_kernel_tensors(layer: torch.nn.Module, info: LayerReloadingInfo):
"""Copy processed values into original kernel tensor storage and restore
kernel tensor references on the layer. Preserves cudagraph references.
"""
kernel tensor references on the layer. Preserves cudagraph references."""
assert info.kernel_tensors is not None
parameters, buffers = info.kernel_tensors
non_persistent = info.kernel_non_persistent_buffers
@@ -81,8 +81,7 @@ class RunaiModelStreamerLoader(BaseModelLoader):
) -> list[str]:
"""Prepare weights for the model.
If the model is not local, it will be downloaded.
"""
If the model is not local, it will be downloaded."""
is_object_storage_path = is_runai_obj_uri(model_name_or_path)
is_local = os.path.isdir(model_name_or_path)
safetensors_pattern = "*.safetensors"
@@ -497,8 +497,7 @@ def _check_tensors_on_meta_device(model: nn.Module) -> None:
def _resize_lora_embeddings(model: nn.Module):
"""Modify LoRA embedding layers to use bigger tensors
to allow for adapter added tokens.
"""
to allow for adapter added tokens."""
for child in model.modules():
if (
isinstance(child, VocabParallelEmbedding)
@@ -105,8 +105,7 @@ class TensorizerLoader(BaseModelLoader):
Expects a vLLM-tensorized model. See the
examples/features/tensorize_vllm_model.py example script
for serializing vLLM models.
"""
for serializing vLLM models."""
if is_vllm_tensorized(self.tensorizer_config):
tensorizer_config = self._patch_tensorizer_config(model_config)
deserialize_tensorizer_model(model, tensorizer_config)
@@ -1,8 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""IPC model loader: maps post-quantized weights from a local weight cache
daemon via CUDA IPC instead of loading from disk.
"""
daemon via CUDA IPC instead of loading from disk."""
import dataclasses
import socket
@@ -78,8 +78,7 @@ def maybe_retie_word_embeddings(model: nn.Module, model_config: ModelConfig) ->
Checkpoints produced by quantization or fine-tuning tooling often keep a redundant
copy of the tied `lm_head`. Sharing the storage again reclaims the memory it would
otherwise cost.
"""
otherwise cost."""
if not model_config.word_embeddings_untied_by_checkpoint:
return
if (untied := _get_untied_lm_head(model)) is None:
@@ -144,8 +144,7 @@ def atomic_writer(
def _natural_sort_key(filepath: str) -> list:
"""Natural sort key for filenames with numeric components, such as
model-00001-of-00005.safetensors -> ['model-', 1, '-of-', 5, '.safetensors']
"""
model-00001-of-00005.safetensors -> ['model-', 1, '-of-', 5, '.safetensors']"""
return [
int(s) if s.isdigit() else s
for s in re.split(r"(\d+)", os.path.basename(filepath))
@@ -162,8 +161,7 @@ def maybe_download_from_modelscope(
"""Download model from ModelScope hub if VLLM_USE_MODELSCOPE is True.
Returns the path to the downloaded model, or None if the model is not
downloaded from ModelScope.
"""
downloaded from ModelScope."""
if envs.VLLM_USE_MODELSCOPE:
# download model from ModelScope hub,
# lazy import so that modelscope is not required for normal use.
@@ -1148,8 +1146,7 @@ def instanttensor_weights_iterator(
use_tqdm_on_load: bool,
) -> Generator[tuple[str, torch.Tensor], None, None]:
"""Iterate over the weights in the model safetensor files
using instanttensor library.
"""
using instanttensor library."""
try:
import instanttensor
except ImportError as e: