Convert formatting to use ruff instead of yapf + isort (#26247)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 07:06:22 -07:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115205 additions and 94107 deletions
@@ -20,8 +20,8 @@ logger = init_logger(__name__)
# common functions
def rotate_neox(x: torch.Tensor) -> torch.Tensor:
x1 = x[..., :x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2:]
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
@@ -53,9 +53,9 @@ def apply_rotary_emb_torch(
return torch.stack((o1, o2), dim=-1).flatten(-2)
def apply_rotary_emb_dispatch(x: torch.Tensor, cos: torch.Tensor,
sin: torch.Tensor,
is_neox_style: bool) -> torch.Tensor:
def apply_rotary_emb_dispatch(
x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, is_neox_style: bool
) -> torch.Tensor:
"""
Args:
x: [num_tokens, num_heads, head_size]
@@ -65,15 +65,14 @@ def apply_rotary_emb_dispatch(x: torch.Tensor, cos: torch.Tensor,
positional embeddings.
"""
if current_platform.is_cuda():
return apply_rotary_emb(x.unsqueeze(0), cos, sin,
not is_neox_style).squeeze(0)
return apply_rotary_emb(x.unsqueeze(0), cos, sin, not is_neox_style).squeeze(0)
else:
return apply_rotary_emb_torch(x, cos, sin, is_neox_style)
@cache
def dispatch_rotary_emb_function(
default: Optional[Callable[..., torch.Tensor]] = None
default: Optional[Callable[..., torch.Tensor]] = None,
) -> Callable[..., torch.Tensor]:
if current_platform.is_cuda():
return apply_rotary_emb
@@ -81,11 +80,13 @@ def dispatch_rotary_emb_function(
if current_platform.is_rocm():
if find_spec("flash_attn") is not None:
from flash_attn.ops.triton.rotary import apply_rotary
return apply_rotary
else:
logger.warning(
"flash_attn is not installed. Falling back to PyTorch "
"implementation for rotary embeddings.")
"implementation for rotary embeddings."
)
if default is not None:
return default
@@ -95,31 +96,37 @@ def dispatch_rotary_emb_function(
# yarn functions
# Inverse dim formula to find dim based on number of rotations
def yarn_find_correction_dim(num_rotations: int,
dim: int,
base: float = 10000,
max_position_embeddings: int = 2048) -> float:
return (dim * math.log(max_position_embeddings /
(num_rotations * 2 * math.pi))) / (2 *
math.log(base))
def yarn_find_correction_dim(
num_rotations: int,
dim: int,
base: float = 10000,
max_position_embeddings: int = 2048,
) -> float:
return (dim * math.log(max_position_embeddings / (num_rotations * 2 * math.pi))) / (
2 * math.log(base)
)
# Find dim range bounds based on rotations
def yarn_find_correction_range(
low_rot: int,
high_rot: int,
dim: int,
base: float = 10000,
max_position_embeddings: int = 2048) -> tuple[int, int]:
low_rot: int,
high_rot: int,
dim: int,
base: float = 10000,
max_position_embeddings: int = 2048,
) -> tuple[int, int]:
low = math.floor(
yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings))
yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings)
)
high = math.ceil(
yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings))
yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings)
)
return max(low, 0), min(high, dim - 1) # Clamp values just in case
def yarn_linear_ramp_mask(low: float, high: float, dim: int,
dtype: torch.dtype) -> torch.Tensor:
def yarn_linear_ramp_mask(
low: float, high: float, dim: int, dtype: torch.dtype
) -> torch.Tensor:
if low == high:
high += 0.001 # Prevent singularity
@@ -143,7 +150,7 @@ def _flashinfer_rotary_embedding(
is_neox: bool,
) -> None:
"""Custom op wrapper for flashinfer's rotary embedding.
This is an in-place operation that modifies query and key tensors directly.
"""
from flashinfer.rope import apply_rope_with_cos_sin_cache_inplace