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:
Harry Mellor
2026-09-16 16:44:29 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 804e5377aa
commit 8c1557a79c
2002 changed files with 15024 additions and 14269 deletions
@@ -138,14 +138,14 @@ class ApplyRotaryEmb(CustomOp):
is_neox_style: bool = True,
enable_fp32_compute: bool = False,
) -> torch.Tensor:
"""
Args:
x: [batch_size (optional), seq_len, num_heads, head_size]
cos: [seq_len, head_size // 2]
sin: [seq_len, head_size // 2]
is_neox_style: Whether to use the Neox-style or GPT-J-style.
enable_fp32_compute: Temporarily convert x, cos, sin to FP32 dtype
for higher accuracy.
"""Args:
x: [batch_size (optional), seq_len, num_heads, head_size]
cos: [seq_len, head_size // 2]
sin: [seq_len, head_size // 2]
is_neox_style: Whether to use the Neox-style or GPT-J-style.
enable_fp32_compute: Temporarily convert x, cos, sin to FP32 dtype
for higher accuracy.
"""
origin_dtype = x.dtype
if enable_fp32_compute:
@@ -263,7 +263,6 @@ class DeepseekV4ScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
inverse: bool = False,
) -> tuple[torch.Tensor, torch.Tensor | None]:
"""PyTorch-native implementation equivalent to forward()."""
head_size = query.size(-1)
query_rot = query[..., -self.rotary_dim :]
key_rot = key[..., -self.rotary_dim :] if key is not None else None
@@ -8,7 +8,7 @@ from .mrope import MRotaryEmbedding
class Ernie4_5_VLRotaryEmbedding(MRotaryEmbedding):
"""3D rotary positional embedding. 3D is t:time h:height w:width"""
"""3D rotary positional embedding. 3D is t:time h:height w:width."""
def forward_native( # type: ignore[override]
self,
@@ -185,7 +185,7 @@ class FourierRotaryEmbedding(RotaryEmbedding):
return query, key
def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor):
"""load fope weights"""
"""Load fope weights."""
world_size = get_tensor_model_parallel_world_size()
rank = get_tensor_model_parallel_rank()
num_key_value_heads = loaded_weight.size(0)
@@ -178,8 +178,12 @@ def triton_mrope(
(T/H/W positions with multimodal inputs)
mrope_section: [t, h, w]
head_size: int
rotary_dim: Number of leading dimensions rotary is applied to.
mrope_interleaved: Whether the T/H/W sections are interleaved rather
than concatenated.
is_neox_style: Whether rotary pairs use split-half (NeoX) or
adjacent (GPT-J) layout.
"""
n_row, n_q_head_head_dim = q.shape
n_q_head = n_q_head_head_dim // head_size
@@ -320,6 +324,8 @@ class MRotaryEmbedding(RotaryEmbeddingBase):
[3, num_tokens] (T/H/W positions with multimodal inputs)
query: [num_tokens, num_heads * head_size]
key: [num_tokens, num_kv_heads * head_size]
offsets: Optional per-token position offsets added to positions.
"""
assert positions.ndim == 1 or positions.ndim == 2
assert key is not None
@@ -85,7 +85,7 @@ class MRotaryEmbeddingInterleaved(MRotaryEmbedding):
self,
positions: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Interleave the rotary embedding"""
"""Interleave the rotary embedding."""
cos_sin = self.cos_sin_cache[positions]
mrope_section_3d = [1] * len(self.mrope_dim)
mrope_dim = self.mrope_dim
@@ -139,8 +139,7 @@ class MRotaryEmbeddingInterleaved(MRotaryEmbedding):
def get_mrope_interleaved_id_list(
a: int, b: int, c: int, force_last: bool = False
) -> list[int]:
"""
Generate an interleaved list of indices for multi-modal rotary embedding.
"""Generate an interleaved list of indices for multi-modal rotary embedding.
Args:
a: Number of indices for first modality
@@ -150,6 +149,7 @@ class MRotaryEmbeddingInterleaved(MRotaryEmbedding):
Returns:
List of interleaved indices
"""
if force_last:
a -= 1
@@ -9,7 +9,8 @@ from .base import RotaryEmbedding
class NTKScalingRotaryEmbedding(RotaryEmbedding):
"""RotaryEmbedding extended with fixed and mixed NTK scaling.
https://kexue.fm/archives/9706"""
https://kexue.fm/archives/9706
"""
def __init__(
self,