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
@@ -45,8 +45,7 @@ if current_platform.is_cuda_alike():
class ActivationQuantPattern(VllmPatternReplacement):
"""
Base class for Activation+Quant fusions.
"""Base class for Activation+Quant fusions.
Should not be used directly.
"""
@@ -79,9 +78,7 @@ class ActivationQuantPattern(VllmPatternReplacement):
class SiluMulFp8StaticQuantPattern(ActivationQuantPattern):
"""
Fusion for SiluMul+Fp8StaticQuant Pattern
"""
"""Fusion for SiluMul+Fp8StaticQuant Pattern."""
def __init__(self) -> None:
super().__init__(kFp8StaticTensorSym)
@@ -126,9 +123,7 @@ class SiluMulFp8StaticQuantPattern(ActivationQuantPattern):
class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
"""
Fusion for SiluMul+Nvfp4Quant Pattern
"""
"""Fusion for SiluMul+Nvfp4Quant Pattern."""
def __init__(self) -> None:
super().__init__(kNvfp4Dynamic)
@@ -182,8 +177,7 @@ class SiluMulNvfp4QuantPattern(ActivationQuantPattern):
class SiluMulBlockQuantPattern(ActivationQuantPattern):
"""
Fusion for SiluMul+BlockQuant (FP8 dynamic per-group) Pattern.
"""Fusion for SiluMul+BlockQuant (FP8 dynamic per-group) Pattern.
Supports group_size 128 and 64 via QuantKey.
Parameterized on is_scale_transposed for different scale layouts.
"""
@@ -281,8 +275,7 @@ class SiluMulBlockQuantPattern(ActivationQuantPattern):
class ActivationQuantFusionPass(VllmFusionPatternMatcherPass):
"""
This pass fuses a pre-defined set of custom ops into fused ops.
"""This pass fuses a pre-defined set of custom ops into fused ops.
It uses the torch pattern matcher to find the patterns and replace them.
Because patterns can only be registered once, the pass is a singleton.
@@ -66,7 +66,8 @@ def _view_flashinfer_nvfp4_scale_out_as_int32(
def _norm_input_weight_dtype_match(match: pm.Match) -> bool:
"""Prevent fusion when the norm input and weight dtypes differ (e.g. a Gemma
fp32 weight.float()+1 gamma), covering rms_norm and fused_add_rms_norm."""
fp32 weight.float()+1 gamma), covering rms_norm and fused_add_rms_norm.
"""
for node in match.nodes:
if node.target == _IR_RMS_NORM_OP:
x, weight = node.args[0], node.args[1]
@@ -347,8 +348,7 @@ class BasePattern:
class AllReduceRMSNormPattern(BasePattern):
"""
This pattern replaces the allreduce + rms norm (without residual)
"""This pattern replaces the allreduce + rms norm (without residual)
with fused flashinfer implementation.
Applies to allreduce + rmsnorm before attn in the first Transformer block.
"""
@@ -409,8 +409,7 @@ class AllReduceRMSNormPattern(BasePattern):
class AllReduceFusedAddRMSNormPattern(BasePattern):
"""
This pattern replaces the allreduce + rms norm (with residual)
"""This pattern replaces the allreduce + rms norm (with residual)
with fused flashinfer implementation.
Applies to o_proj + rmsnorm after attn and mlp + rmsnorm before attn.
"""
@@ -609,8 +608,7 @@ class AllReduceFusedAddGemmaRMSNormPattern(BasePattern):
class AllReduceFusedRMSNormStaticQuantFP8Pattern(BasePattern):
"""
This pattern replaces the allreduce + rms norm (without residual)
"""This pattern replaces the allreduce + rms norm (without residual)
+ static fp8 quant with fused flashinfer implementation.
Applies to allreduce + rmsnorm + quant before attn
in the first Transformer block.
@@ -684,8 +682,7 @@ class AllReduceFusedRMSNormStaticQuantFP8Pattern(BasePattern):
class AllReduceFusedAddRMSNormStaticQuantFP8Pattern(BasePattern):
"""
This pattern replaces the allreduce + rms norm (with residual)
"""This pattern replaces the allreduce + rms norm (with residual)
+ static fp8 quant with fused flashinfer implementation.
Applies to o_proj + rmsnorm after attn + quant and
mlp + rmsnorm + quant before attn.
@@ -767,8 +764,7 @@ class AllReduceFusedAddRMSNormStaticQuantFP8Pattern(BasePattern):
class AllReduceFusedRMSNormStaticQuantNVFP4Pattern(BasePattern):
"""
This pattern replaces the allreduce + rms norm (without residual)
"""This pattern replaces the allreduce + rms norm (without residual)
+ static nvfp4 quant with fused flashinfer implementation.
Applies to allreduce + rmsnorm + quant before attn
in the first Transformer block.
@@ -864,8 +860,7 @@ class AllReduceFusedRMSNormStaticQuantNVFP4Pattern(BasePattern):
class AllReduceFusedAddRMSNormStaticQuantNVFP4Pattern(BasePattern):
"""
This pattern replaces the allreduce + rms norm (with residual)
"""This pattern replaces the allreduce + rms norm (with residual)
+ static nvfp4 quant with fused flashinfer implementation.
Applies to o_proj + rmsnorm after attn + quant and
mlp + rmsnorm + quant before attn.
@@ -36,8 +36,7 @@ _FP8_QUANT_KEY = QuantKey(dtype=FP8_DTYPE, scale=kStaticTensorScale, symmetric=T
class AttnFp8StaticQuantPattern(VllmPatternReplacement[..., torch.Tensor]):
"""
Fusion for Attention+Fp8StaticQuant.
"""Fusion for Attention+Fp8StaticQuant.
Only triggers when the attention implementation returns True in
`fused_output_quant_supported()`. If the pattern is found, the
@@ -172,8 +171,7 @@ class AttnFp8StaticQuantPattern(VllmPatternReplacement[..., torch.Tensor]):
class AttnNvfp4QuantPattern(
VllmPatternReplacement[..., tuple[torch.Tensor, torch.Tensor]]
):
"""
Fusion for Attention+Nvfp4Quant.
"""Fusion for Attention+Nvfp4Quant.
Only triggers when the attention implementation returns True in
`fused_output_quant_supported()`. If the pattern is found, the
@@ -360,8 +358,7 @@ class AttnNvfp4QuantPattern(
class AttnQuantFusionPass(VllmFusionPatternMatcherPass):
"""
This pass fuses post-attention quantization onto attention if supported.
"""This pass fuses post-attention quantization onto attention if supported.
It uses the pattern matcher and matches each layer manually, as strings
cannot be wildcarded. This also lets us check support on attention layers
@@ -79,7 +79,7 @@ class MatcherCustomOp(ABC):
return torch.empty(*args, dtype=torch.float32, device=self.device, **kwargs)
def inputs(self) -> list[torch.Tensor]:
"""Utility for inputs to the pattern"""
"""Utility for inputs to the pattern."""
raise NotImplementedError
@@ -33,8 +33,7 @@ MLA_ATTN_OP = torch.ops.vllm.unified_mla_attention_with_output.default
class MLAAttnFp8StaticQuantPattern(VllmPatternReplacement[..., torch.Tensor]):
"""
Fusion for MLA Attention+Fp8StaticQuant.
"""Fusion for MLA Attention+Fp8StaticQuant.
Matches the pattern: MLA attention -> static FP8 quant, and replaces
it with MLA attention(output_scale=scale, output=fp8_buffer).
@@ -174,8 +173,7 @@ class MLAAttnFp8StaticQuantPattern(VllmPatternReplacement[..., torch.Tensor]):
class MLAAttnNvfp4QuantPattern(
VllmPatternReplacement[..., tuple[torch.Tensor, torch.Tensor]]
):
"""
Fusion for MLA Attention+Nvfp4Quant.
"""Fusion for MLA Attention+Nvfp4Quant.
Matches the pattern: MLA attention -> NVFP4 quant, and replaces
it with MLA attention(output_scale=scale, output_block_scale=block_scale,
@@ -362,8 +360,7 @@ class MLAAttnNvfp4QuantPattern(
class MLAAttnFp8GroupQuantPattern(
VllmPatternReplacement[..., tuple[torch.Tensor, torch.Tensor]]
):
"""
Fusion for MLA Attention+Fp8GroupQuant (per-group dynamic FP8).
"""Fusion for MLA Attention+Fp8GroupQuant (per-group dynamic FP8).
Matches the pattern: MLA attention -> per_token_group_fp8_quant, and
replaces it with MLA attention(output_block_scale=group_scale_buffer).
@@ -572,8 +569,7 @@ class MLAAttnFp8GroupQuantPattern(
class MLAAttnQuantFusionPass(VllmFusionPatternMatcherPass):
"""
This pass fuses post-attention quantization onto MLA attention if supported.
"""This pass fuses post-attention quantization onto MLA attention if supported.
It uses the pattern matcher and matches each MLA layer manually, as strings
cannot be wildcarded. This also lets us check support on attention layers
@@ -34,8 +34,7 @@ P = ParamSpec("P")
class QkNormRopePattern:
"""
Match the unfused sequence in attention blocks and replace with the fused op.
"""Match the unfused sequence in attention blocks and replace with the fused op.
Unfused (conceptually):
q, k, v = split(qkv, [qsz, kvsz, kvsz], -1)
@@ -106,8 +106,7 @@ direct_register_custom_op(
class QkNormRopeKvCachePattern:
"""
Match the unfused sequence:
"""Match the unfused sequence:
q, k, v = split(qkv, ...)
q = rms_norm(q.view(heads), q_weight).view(flat)
k = rms_norm(k.view(heads), k_weight).view(flat)
@@ -413,8 +412,7 @@ class QkNormRopeKvCachePattern:
class QkNormRopeKvCacheFusionPass(VllmPatternMatcherPass):
"""
Fuse QK-norm + RoPE + KV cache update into a single AITER HIP kernel.
"""Fuse QK-norm + RoPE + KV cache update into a single AITER HIP kernel.
Supersedes both QKNormRoPEFusionPass and RopeKVCacheFusionPass for
attention layers that support the combined operation, eliminating two
@@ -168,8 +168,7 @@ if current_platform.is_cuda() and hasattr(torch.ops._C, "scaled_fp4_quant"):
class FusedRMSQuantKey(NamedTuple):
"""
Named tuple for identifying the type of RMSNorm + quant fusion.
"""Named tuple for identifying the type of RMSNorm + quant fusion.
quant: type of quantization
fused_add: does the op also perform the residual add
"""
@@ -771,8 +770,7 @@ class FusedAddRMSNormNvfp4QuantPattern:
class RMSNormQuantFusionPass(VllmPatternMatcherPass):
"""
This pass fuses rms_norm & quant custom ops into a fused rms_norm_quant op.
"""This pass fuses rms_norm & quant custom ops into a fused rms_norm_quant op.
It also supports fused_add_rms_norm.
"""
@@ -180,8 +180,7 @@ class AiterFusedAddRMSNormDynamicQuantPattern(AiterRMSNormQuantPattern):
class AiterRMSFp8GroupQuantPattern(AiterRMSNormQuantPattern):
"""
This pattern fuses aiter rms_norm & group fp8 quant custom
"""This pattern fuses aiter rms_norm & group fp8 quant custom
ops into an aiter rms_norm_group_fp8_quant op.
"""
@@ -236,8 +235,7 @@ class AiterRMSFp8GroupQuantPattern(AiterRMSNormQuantPattern):
class AiterFusedAddRMSFp8GroupQuantPattern(AiterRMSNormQuantPattern):
"""
This pattern fuses aiter rms_norm_with_add & group fp8 quant custom ops
"""This pattern fuses aiter rms_norm_with_add & group fp8 quant custom ops
into a aiter rms_norm_with_add_group_fp8_quant op.
"""
@@ -298,8 +296,7 @@ class AiterFusedAddRMSFp8GroupQuantPattern(AiterRMSNormQuantPattern):
class DoubleAiterRMSFp8GroupQuantPattern(AiterRMSNormQuantPattern):
"""
Pattern matching ``rms_norm`` whose output feeds *two* distinct
"""Pattern matching ``rms_norm`` whose output feeds *two* distinct
``rocm_aiter_group_fp8_quant`` consumers, replacing it with two
independent fused ``rms_norm_group_fp8_quant`` ops.
@@ -367,8 +364,7 @@ class DoubleAiterRMSFp8GroupQuantPattern(AiterRMSNormQuantPattern):
class DoubleAiterRMSFp8GroupQuantViewPattern(AiterRMSNormQuantPattern):
"""
View-tolerant variant of ``DoubleAiterRMSFp8GroupQuantPattern``.
"""View-tolerant variant of ``DoubleAiterRMSFp8GroupQuantPattern``.
Matches the same 1-to-2 fan-out, but with a ``view``/``reshape`` between
the ``rms_norm`` output and the two ``rocm_aiter_group_fp8_quant``
@@ -453,8 +449,7 @@ class DoubleAiterRMSFp8GroupQuantViewPattern(AiterRMSNormQuantPattern):
class AiterRMSNormGatedFp8GroupQuantPattern(AiterRMSNormQuantPattern):
"""
Matches decomposed RMSNormGated + reshape + group FP8 quant and replaces
"""Matches decomposed RMSNormGated + reshape + group FP8 quant and replaces
with rocm_aiter_fused_rms_gated_fp8_group_quant.
The norm operates per-head on (N*H, D) tensors. The compiler folds the
@@ -544,8 +539,7 @@ class AiterRMSNormGatedFp8GroupQuantPattern(AiterRMSNormQuantPattern):
class RocmAiterRMSNormQuantFusionPass(VllmPatternMatcherPass):
"""
This pass fuses aiter rms_norm & vllm/aiter quant custom ops
"""This pass fuses aiter rms_norm & vllm/aiter quant custom ops
into a fused rms_norm_quant op.
It also supports fused_add_rms_norm.
"""
@@ -681,8 +675,7 @@ class RocmAiterRMSNormQuantFusionPass(VllmPatternMatcherPass):
class AiterSiluMulFp8GroupQuantPattern(VllmPatternReplacement):
"""
This pattern fuses aiter silu_and_mul & group fp8 quant custom
"""This pattern fuses aiter silu_and_mul & group fp8 quant custom
ops into an aiter silu_and_mul_group_fp8_quant op.
"""
@@ -722,8 +715,7 @@ class AiterSiluMulFp8GroupQuantPattern(VllmPatternReplacement):
class RocmAiterSiluMulFp8GroupQuantFusionPass(VllmFusionPatternMatcherPass):
"""
This pass fuses a pre-defined set of custom ops into fused ops.
"""This pass fuses a pre-defined set of custom ops into fused ops.
It uses the torch pattern matcher to find the patterns and replace them.
Because patterns can only be registered once, the pass is a singleton.
@@ -744,8 +736,7 @@ class RocmAiterSiluMulFp8GroupQuantFusionPass(VllmFusionPatternMatcherPass):
class AddAiterRMSNormPadPattern:
"""
This pattern replaces an aiter_rmsnorm_with_add & a pad op
"""This pattern replaces an aiter_rmsnorm_with_add & a pad op
with a custom triton_add_rmsnorm_pad op from AITER.
"""
@@ -820,8 +811,7 @@ class AddAiterRMSNormPadPattern:
class RocmAiterTritonAddRMSNormPadFusionPass(VllmPatternMatcherPass):
"""
This pass replaces an AITER CK RMSNorm + residual add and a pad op
"""This pass replaces an AITER CK RMSNorm + residual add and a pad op
with an triton_add_rmsnorm_pad op from AITER.
"""
@@ -854,8 +844,7 @@ class RocmAiterTritonAddRMSNormPadFusionPass(VllmPatternMatcherPass):
class MLADualRMSNormPattern(
VllmPatternReplacement[..., tuple[torch.Tensor, torch.Tensor, torch.Tensor]]
):
"""
Fuse paired q_a_layernorm + kv_a_layernorm in MLA attention into
"""Fuse paired q_a_layernorm + kv_a_layernorm in MLA attention into
AITER's ``fused_qk_rmsnorm`` HIP kernel.
Target FX-graph pattern (unfused, ``vllm_ir`` stage)::
@@ -946,8 +935,7 @@ class MLADualRMSPerTokenQuantPattern(
],
]
):
"""
Fuse the MLA FP8 attention path -- q-latent RMSNorm + FP8 *per-token* quant
"""Fuse the MLA FP8 attention path -- q-latent RMSNorm + FP8 *per-token* quant
plus kv-latent RMSNorm -- into AITER's ``fused_qk_rmsnorm_per_token_quant``.
With a per-token FP8 ``q_b_proj`` (Quark / ModelOpt), the earlier
@@ -1064,8 +1052,7 @@ class MLADualRMSPerTokenQuantPattern(
class MLADualRMSNormFusionPass(VllmFusionPatternMatcherPass):
"""
Post-grad PatternMatcher pass that fuses paired q / kv RMS norms in
"""Post-grad PatternMatcher pass that fuses paired q / kv RMS norms in
MLA attention into ``fused_mla_dual_rms_norm`` backed by aiter's
``fused_qk_rmsnorm`` HIP kernel.
@@ -52,8 +52,7 @@ def fused_rope_and_unified_kv_cache_update_impl(
is_neox: bool,
layer_name: LayerNameType,
) -> torch.Tensor:
"""
This impl fetches the KV cache and slot mapping from the forward context,
"""This impl fetches the KV cache and slot mapping from the forward context,
then calls the layer impl's `AttentionImpl.do_rope_and_kv_cache_update` method.
It also returns a dummy tensor, similar to `Attention.unified_kv_cache_update`,
that is passed to unified_attention to signal a side effect and
@@ -98,8 +97,7 @@ direct_register_custom_op(
class RopeStaticQQuantKVCachePattern:
"""
Fuse rope + static Q fp8 quant while preserving explicit KV-cache update
"""Fuse rope + static Q fp8 quant while preserving explicit KV-cache update
dependency ordering.
"""
@@ -277,8 +275,7 @@ class RopeStaticQQuantKVCachePattern:
class RopeReshapeKVCachePattern:
"""
This pattern matches the following unfused inplace ops:
"""This pattern matches the following unfused inplace ops:
q, k = rotary_embedding(positions, q, k, head_size, cos_sin_cache, is_neox)
kv_cache_dummy = unified_kv_cache_update(k, v, layer_name)
@@ -410,8 +407,7 @@ class RopeReshapeKVCachePattern:
class RopeKVCacheFusionPass(VllmPatternMatcherPass):
"""
This pass fuses the rotary embedding and KV cache update operations
"""This pass fuses the rotary embedding and KV cache update operations
into a single fused kernel if available.
It uses the pattern matcher and matches each layer manually, as strings
@@ -57,8 +57,7 @@ def get_sequence_parallelism_threshold(
tp_size: int,
element_size: int,
) -> int | None:
"""
Calculate the minimum token threshold for applying sequence parallelism.
"""Calculate the minimum token threshold for applying sequence parallelism.
Returns None if sequence parallelism should not be applied based on model size.
@@ -496,8 +495,7 @@ class MiddleAllReduceRMSNormStaticNVFP4Pattern(_SequenceParallelPatternHelper):
class SequenceParallelismPass(VllmPatternMatcherPass):
"""
This pass enables sequence parallelism for models.
"""This pass enables sequence parallelism for models.
It identifies patterns where an AllReduce operation is followed by
an RMSNorm (or RMSNorm and then Quantization) operation.
These patterns are replaced with a ReduceScatter operation, followed by
@@ -590,8 +588,7 @@ class SequenceParallelismPass(VllmPatternMatcherPass):
self.dump_patterns(config, self.patterns)
def is_applicable_for_range(self, compile_range: Range) -> bool:
"""
Determines if sequence parallelism should be applied for the given
"""Determines if sequence parallelism should be applied for the given
compile range.
SP is only beneficial for larger batch sizes where the communication