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
@@ -166,8 +166,8 @@ def _init_kv_cache_quant(
|
||||
layer: The attention layer instance to initialize.
|
||||
quant_config: Optional quantization configuration.
|
||||
prefix: Layer name prefix for quantization method lookup.
|
||||
"""
|
||||
|
||||
"""
|
||||
# Note [Register q/k/v/prob scales in state dict]
|
||||
# When calling model.to(device), only parameters/buffers in state dict are
|
||||
# moved. If not registering q/k/v/prob scales in state dict, there would
|
||||
@@ -254,8 +254,7 @@ class Attention(nn.Module, AttentionLayerBase):
|
||||
head_size_v: int | None = None,
|
||||
**extra_impl_args,
|
||||
) -> None:
|
||||
"""
|
||||
The KV cache is stored inside this class and is accessed via
|
||||
"""The KV cache is stored inside this class and is accessed via
|
||||
`self.kv_cache`.
|
||||
"""
|
||||
super().__init__()
|
||||
@@ -493,8 +492,7 @@ class Attention(nn.Module, AttentionLayerBase):
|
||||
output_shape: torch.Size | None = None,
|
||||
output_dtype: torch.dtype | None = None,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
The KV cache is stored inside this class and is accessed via
|
||||
"""The KV cache is stored inside this class and is accessed via
|
||||
`self.kv_cache`.
|
||||
|
||||
Attention metadata (`attn_metadata`) is set using a context manager in
|
||||
@@ -691,6 +689,7 @@ def get_attention_context(
|
||||
|
||||
Note: attn_metadata may be None, but attn_layer and kv_cache are always
|
||||
extracted from the forward context.
|
||||
|
||||
"""
|
||||
forward_context: ForwardContext = get_forward_context()
|
||||
attn_metadata_raw = forward_context.attn_metadata
|
||||
@@ -718,8 +717,7 @@ def unified_kv_cache_update(
|
||||
value: torch.Tensor,
|
||||
layer_name: LayerNameType,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Returns a dummy that is passed to unified_attention to signal a side effect and
|
||||
"""Returns a dummy that is passed to unified_attention to signal a side effect and
|
||||
the data dependency between them to ensure torch.compile preserves ordering.
|
||||
"""
|
||||
layer_name = _resolve_layer_name(layer_name)
|
||||
|
||||
@@ -34,7 +34,6 @@ def _get_cross_slot_mapping(
|
||||
device: torch.device,
|
||||
) -> torch.Tensor:
|
||||
"""Get cross-attention slot mappings."""
|
||||
|
||||
block_size = kv_cache_spec.block_size
|
||||
slot_mappings = []
|
||||
|
||||
@@ -187,8 +186,7 @@ def create_cross_attention_backend(
|
||||
|
||||
|
||||
class CrossAttention(Attention):
|
||||
"""
|
||||
Cross-attention for encoder-decoder models.
|
||||
"""Cross-attention for encoder-decoder models.
|
||||
Handles attention between decoder queries and encoder keys/values.
|
||||
"""
|
||||
|
||||
|
||||
@@ -49,9 +49,7 @@ def create_encoder_only_attention_backend(
|
||||
|
||||
|
||||
class EncoderOnlyAttention(Attention):
|
||||
"""
|
||||
Encoder attention is a special case that doesn't need a KV Cache.
|
||||
"""
|
||||
"""Encoder attention is a special case that doesn't need a KV Cache."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
# MLA Common Components
|
||||
"""# MLA Common Components
|
||||
|
||||
This file implements common components for MLA implementations.
|
||||
|
||||
@@ -1415,8 +1414,7 @@ def unified_mla_kv_cache_update(
|
||||
kv_cache_dtype: str,
|
||||
k_scale: torch.Tensor,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Returns a dummy that is passed to unified_attention to signal a side effect and
|
||||
"""Returns a dummy that is passed to unified_attention to signal a side effect and
|
||||
the data dependency between them to ensure torch.compile preserves ordering.
|
||||
"""
|
||||
layer_name = _resolve_layer_name(layer_name)
|
||||
@@ -1539,8 +1537,7 @@ def dynamic_per_batched_tensor_quant(
|
||||
dynamic_arg_dims={"decode_ql_nope": 0, "decode_q_pe": 0},
|
||||
)
|
||||
class _DecodeConcatQuantFP8(QuantFP8):
|
||||
"""
|
||||
QuantFP8 variant that concatenates decode_ql_nope and decode_q_pe before
|
||||
"""QuantFP8 variant that concatenates decode_ql_nope and decode_q_pe before
|
||||
quantization. When disabled, forward_native is compiled via torch.compile,
|
||||
fusing cat/reshape/quant/view together.
|
||||
"""
|
||||
@@ -1571,7 +1568,8 @@ class MLACommonBackend(AttentionBackend):
|
||||
@classmethod
|
||||
def customize_spec(cls, spec: "AttentionSpec") -> "AttentionSpec":
|
||||
"""Per-token-head modes pack an inline fp32 scale pair after the
|
||||
latent data (single-sided: ``head_size_v == 0`` for MLA)."""
|
||||
latent data (single-sided: ``head_size_v == 0`` for MLA).
|
||||
"""
|
||||
spec = super().customize_spec(spec)
|
||||
mode = spec.kv_quant_mode
|
||||
if spec.state_content_bytes is not None or not mode.is_per_token_head:
|
||||
@@ -2030,6 +2028,7 @@ def build_mla_chunked_context_metadata(
|
||||
|
||||
Returns:
|
||||
The chunked-context metadata, or None when no prefill has any context.
|
||||
|
||||
"""
|
||||
# NOTE: it is recommended you read the `Chunked Prefill` section in the
|
||||
# comment at the top of the file before trying to understand this code.
|
||||
@@ -2230,8 +2229,7 @@ def build_mla_chunked_context_metadata(
|
||||
|
||||
|
||||
class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
|
||||
"""
|
||||
NOTE: Please read the comment at the top of the file before trying to
|
||||
"""NOTE: Please read the comment at the top of the file before trying to
|
||||
understand this class
|
||||
"""
|
||||
|
||||
@@ -2316,8 +2314,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
|
||||
vllm_config: VllmConfig,
|
||||
model_dtype: torch.dtype,
|
||||
) -> torch.dtype:
|
||||
"""
|
||||
Determine the query data type for prefill queries.
|
||||
"""Determine the query data type for prefill queries.
|
||||
Return FP8 dtype if cache is FP8 and prefill query quantization
|
||||
is enabled, else model dtype.
|
||||
"""
|
||||
@@ -2484,8 +2481,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
|
||||
def build_for_cudagraph_capture(
|
||||
self, common_attn_metadata: CommonAttentionMetadata
|
||||
) -> M:
|
||||
"""
|
||||
This method builds the metadata for full cudagraph capture.
|
||||
"""This method builds the metadata for full cudagraph capture.
|
||||
Currently, only decode is supported for full cudagraphs with MLA.
|
||||
"""
|
||||
m = common_attn_metadata
|
||||
@@ -2671,14 +2667,16 @@ def reorg_kvcache(
|
||||
max_seq_len: int,
|
||||
toks: int,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
reorg and unpad kvcache after cp local gather to tp layout for attn kernel.
|
||||
"""Reorg and unpad kvcache after cp local gather to tp layout for attn kernel.
|
||||
e.g.
|
||||
allgatered_kv_c_normed = [T0_0, T0_1, T0_2, T0_3, T1_0, T1_1, ...,
|
||||
T0_4, T0_5, pad, pad, T1_2, pad, ...]
|
||||
-> reorganized_kv_c_normed = [T0_0, T0_1, T0_2, T0_3, T0_4, T0_5,
|
||||
T1_0, T1_1, T1_2, ...]
|
||||
|
||||
Args:
|
||||
allgatered_kv_c_normed: all-gathered, padded latent KV cache.
|
||||
allgatered_k_pe: all-gathered, padded RoPE key cache.
|
||||
padded_local_chunk_seq_lens_lst: local chunk context lengths
|
||||
under current CP rank.
|
||||
local_context_lens_allranks: local context lengths on each CP rank.
|
||||
@@ -2687,6 +2685,7 @@ def reorg_kvcache(
|
||||
sum_seq_len: the sum of cp_chunk_seq_lens_lst.
|
||||
max_seq_len: the max value of cp_chunk_seq_lens_lst.
|
||||
toks: the number of tokens for local gather cache.
|
||||
|
||||
"""
|
||||
kv_c_segments = []
|
||||
k_pe_segments = []
|
||||
@@ -2789,10 +2788,16 @@ def accumulate_mla_context_chunk(
|
||||
remaining token range is initialized.
|
||||
|
||||
Args:
|
||||
chunk: The context chunk being folded in.
|
||||
attn_output: The chunk's attention output.
|
||||
attn_softmax_lse: The chunk's log-sum-exp values.
|
||||
output: Running context partial, updated in place.
|
||||
output_lse: Running log-sum-exp, updated in place.
|
||||
output_written: The chunk's attention output already landed in
|
||||
``output[chunk.token_slice]`` because the backend was handed it as
|
||||
``out``, leaving only the lse to fold. Invalid for a continuation
|
||||
chunk, whose leading tokens must be merged rather than overwritten.
|
||||
|
||||
"""
|
||||
token_start = chunk.token_slice.start
|
||||
token_end = chunk.token_slice.stop
|
||||
@@ -2819,8 +2824,7 @@ def accumulate_mla_context_chunk(
|
||||
|
||||
|
||||
class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
|
||||
"""
|
||||
Shared MLA base providing dense-MHA prefill (via the selected
|
||||
"""Shared MLA base providing dense-MHA prefill (via the selected
|
||||
MLAPrefillBackend) for both dense and sparse impls; subclasses add decode
|
||||
(``forward_mqa``).
|
||||
"""
|
||||
@@ -2856,8 +2860,7 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
|
||||
def _concat_k_nope_k_pe(
|
||||
self, k_nope: torch.Tensor, k_pe: torch.Tensor
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Efficiently concatenate k_nope and k_pe tensors along the last dimension.
|
||||
"""Efficiently concatenate k_nope and k_pe tensors along the last dimension.
|
||||
|
||||
This function avoids the performance penalty of torch.cat with expanded
|
||||
non-contiguous tensors by pre-allocating the output and using direct copies.
|
||||
@@ -2869,6 +2872,7 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
|
||||
|
||||
Returns:
|
||||
Tensor of shape [..., nope_dim + pe_dim]
|
||||
|
||||
"""
|
||||
if k_pe.shape[-1] == 0:
|
||||
# NoPE MLA: nothing to append, so no copy either.
|
||||
@@ -3224,8 +3228,7 @@ class MLACommonBaseImpl(MLAAttentionImpl[A], Generic[A]):
|
||||
|
||||
|
||||
class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
|
||||
"""
|
||||
NOTE: Please read the comment at the top of the file before trying to
|
||||
"""NOTE: Please read the comment at the top of the file before trying to
|
||||
understand this class
|
||||
"""
|
||||
|
||||
|
||||
@@ -322,14 +322,14 @@ class MMEncoderAttention(CustomOp):
|
||||
num_kv_heads: int | None = None,
|
||||
prefix: str = "",
|
||||
) -> None:
|
||||
"""
|
||||
Args:
|
||||
num_heads: number of attention heads per partition.
|
||||
head_size: hidden_size per attention head.
|
||||
scale: scale factor.
|
||||
num_kv_heads: number of kv heads.
|
||||
prefix: This has no effect, it is only here to make it easier to
|
||||
swap between Attention and MultiHeadAttention
|
||||
"""Args:
|
||||
num_heads: number of attention heads per partition.
|
||||
head_size: hidden_size per attention head.
|
||||
scale: scale factor.
|
||||
num_kv_heads: number of kv heads.
|
||||
prefix: This has no effect, it is only here to make it easier to
|
||||
swap between Attention and MultiHeadAttention
|
||||
|
||||
"""
|
||||
super().__init__()
|
||||
|
||||
@@ -517,8 +517,7 @@ class MMEncoderAttention(CustomOp):
|
||||
q_len: int,
|
||||
kv_len: int,
|
||||
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Reshape query, key, value to 4D tensors:
|
||||
"""Reshape query, key, value to 4D tensors:
|
||||
(batch_size, seq_len, num_heads, head_size)
|
||||
"""
|
||||
query = query.view(bsz, q_len, self.num_heads, self.head_size)
|
||||
|
||||
@@ -114,9 +114,7 @@ def create_static_sink_attention_backend(
|
||||
|
||||
@CustomOp.register("static_sink_attention")
|
||||
class StaticSinkAttention(Attention, CustomOp):
|
||||
"""
|
||||
Attention with static sink tokens
|
||||
"""
|
||||
"""Attention with static sink tokens."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user