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
@@ -56,10 +56,7 @@ class TestMMRSModel(torch.nn.Module):
torch.nn.init.normal_(self.gate_proj, std=0.02)
def forward(self, hidden_states):
"""
Forward pass implementing the mm + reduce scatter in the FX graph
"""
"""Forward pass implementing the mm + reduce scatter in the FX graph."""
# Reshape input
view = hidden_states.reshape(-1, self.hidden_size)
@@ -88,9 +85,7 @@ class TestAGMMModel(torch.nn.Module):
torch.nn.init.normal_(self.weight, std=0.02)
def forward(self, hidden_states):
"""
Forward pass implementing the mm + all gather in the FX graph
"""
"""Forward pass implementing the mm + all gather in the FX graph."""
# Reshape input
view = hidden_states.reshape(-1, self.hidden_size)
all_gather = tensor_model_parallel_all_gather(view, dim=0)
@@ -122,10 +117,7 @@ class _BaseScaledMMModel(torch.nn.Module):
class TestScaledMMRSModel(_BaseScaledMMModel):
def forward(self, input: torch.Tensor):
"""
Forward pass implementing the scaled_mm + reduce scatter in the FX graph
"""
"""Forward pass implementing the scaled_mm + reduce scatter in the FX graph."""
fp8_input = input.to(FP8_DTYPE)
scale_a = torch.ones(input.shape[0], 1, dtype=torch.float32)
scaled_mm = torch._scaled_mm(
@@ -147,9 +139,7 @@ class TestScaledMMRSModel(_BaseScaledMMModel):
class TestAGScaledMMModel(_BaseScaledMMModel):
def forward(self, input: torch.Tensor):
"""
Forward pass implementing the all gather + scaled_mm in the FX graph
"""
"""Forward pass implementing the all gather + scaled_mm in the FX graph."""
# Reshape input
fp8_input = input.to(FP8_DTYPE)
all_gather = tensor_model_parallel_all_gather(fp8_input, dim=0)
@@ -173,8 +163,7 @@ class TestAGScaledMMModel(_BaseScaledMMModel):
class TestCutlassScaledMMRSModel(_BaseScaledMMModel):
def forward(self, input: torch.Tensor):
"""
Forward pass implementing the cutlass_scaled_mm + reduce scatter
"""Forward pass implementing the cutlass_scaled_mm + reduce scatter
in the FX graph
"""
@@ -200,8 +189,7 @@ class TestCutlassScaledMMRSModel(_BaseScaledMMModel):
class TestAGCutlassScaledMMModel(_BaseScaledMMModel):
def forward(self, input: torch.Tensor):
"""
Forward pass implementing the all gather + cutlass_scaled_mm
"""Forward pass implementing the all gather + cutlass_scaled_mm
in the FX graph
"""
# Reshape input
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Comprehensive tests for UnsafeCloneEliminationPass.
"""Comprehensive tests for UnsafeCloneEliminationPass.
This test suite exercises all possible valid FX graph patterns involving clones:
1. Clone with no users (dead code)
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Tests for IR inplace functionalization pass integration.
"""Tests for IR inplace functionalization pass integration.
This test suite verifies that the inplace functionalization pass, lowering pass,
and clone cleanup pass work together correctly with donated buffer tracking.
@@ -108,8 +107,7 @@ class MixedModel(nn.Module):
class ModelWithTritonAfterMaybeInplace(nn.Module):
"""
Model using maybe_inplace followed by a Triton kernel.
"""Model using maybe_inplace followed by a Triton kernel.
Test clone elimination can handle Triton in the graph
"""
@@ -410,8 +408,7 @@ def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
reason="Only test on cuda and rocm platform",
)
def test_piecewise_compilation_with_donated_buffers(monkeypatch, fresh_vllm_cache):
"""
Test piecewise compilation with donated buffers across graph splits.
"""Test piecewise compilation with donated buffers across graph splits.
Utilizes a custom splitting op. Uses fresh cache to avoid compilation caching.
"""
torch.set_default_device(current_platform.device_type)
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Unit tests for the DoubleQuant fan-out variants registered by
"""Unit tests for the DoubleQuant fan-out variants registered by
``RocmAiterRMSNormQuantFusionPass``.
Both variants target a 1-to-2 fan-out where one ``rms_norm`` output feeds
@@ -90,8 +89,7 @@ def test_double_aiter_rms_fp8_group_quant_fusion(
model_cls: type[torch.nn.Module],
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""
Both fan-out shapes (with and without an intermediate view) must fuse
"""Both fan-out shapes (with and without an intermediate view) must fuse
into ``rocm_aiter_rmsnorm_fp8_group_quant``: the no-view shape via
``DoubleAiterRMSFp8GroupQuantPattern`` and the viewed shape via the
new ``DoubleAiterRMSFp8GroupQuantViewPattern`` sibling.
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Unit test for the MLADualRMSNormFusionPass.
"""Unit test for the MLADualRMSNormFusionPass.
The pass fuses paired q/kv RMS norms in MLA attention into a single
fused_mla_dual_rms_norm op backed by AITER's fused_qk_rmsnorm kernel.
@@ -38,13 +37,12 @@ FP8_DTYPE = current_platform.fp8_dtype()
class MLADualRMSNormTestModel(torch.nn.Module):
"""
Minimal model reproducing the MLA dual RMS norm pattern:
linear -> split([q_dim, kv_dim])
+-- q_c (getitem 0) -> rms_norm(q_w, eps) -> linear
+-- kv_lora (getitem 1) -> split([kv_c_dim, k_pe_dim])
+-- kv_c (getitem 0) -> rms_norm(kv_w, eps)
+-- k_pe
"""Minimal model reproducing the MLA dual RMS norm pattern:
linear -> split([q_dim, kv_dim])
+-- q_c (getitem 0) -> rms_norm(q_w, eps) -> linear
+-- kv_lora (getitem 1) -> split([kv_c_dim, k_pe_dim])
+-- kv_c (getitem 0) -> rms_norm(kv_w, eps)
+-- k_pe
"""
def __init__(
@@ -155,13 +153,12 @@ def test_fuse_mla_dual_rms_norm(
class MLADualRMSNormFp8PerTokenTestModel(torch.nn.Module):
"""
Minimal model reproducing the FP8 MLA attention path with *per-token* quant:
linear -> split([q_dim, kv_dim])
+-- q_c (getitem 0) -> rocm_aiter_rmsnorm_fused_dynamic_quant -> dequant
+-- kv_lora (getitem 1) -> split([kv_c_dim, k_pe_dim])
+-- kv_c (getitem 0) -> rms_norm (bf16)
+-- k_pe
"""Minimal model reproducing the FP8 MLA attention path with *per-token* quant:
linear -> split([q_dim, kv_dim])
+-- q_c (getitem 0) -> rocm_aiter_rmsnorm_fused_dynamic_quant -> dequant
+-- kv_lora (getitem 1) -> split([kv_c_dim, k_pe_dim])
+-- kv_c (getitem 0) -> rms_norm (bf16)
+-- k_pe
"""
def __init__(
+2 -1
View File
@@ -732,7 +732,8 @@ def test_aiter_fusion_rmsnorm_gated_quant_no_gdn_layers(
monkeypatch: pytest.MonkeyPatch,
):
"""Verify that without GDN layers in static_forward_context,
the gated pattern is not registered and no matches occur."""
the gated pattern is not registered and no matches occur.
"""
group_shape = GroupShape(1, 128)
vllm_config = VllmConfig(
model_config=ModelConfig(dtype=dtype),
+1 -2
View File
@@ -107,7 +107,6 @@ class AttentionQuantPatternModel(torch.nn.Module):
def build_attn_metadata(self, batch_size: int) -> AttentionMetadata:
"""Initialize attention metadata."""
# TODO (Rohan138) reuse utils from vllm/v1/worker/gpu/attn_utils.py
# Create common attn metadata
@@ -304,7 +303,7 @@ def test_attention_quant_pattern(
dist_init,
disable_vllm_compile_cache,
):
"""Test AttentionStaticQuantPattern fusion pass"""
"""Test AttentionStaticQuantPattern fusion pass."""
if backend == AttentionBackendEnum.FLASHINFER and (
not current_platform.is_device_capability((10, 0)) or not has_flashinfer()
):
@@ -141,7 +141,6 @@ class MLAAttentionQuantPatternModel(torch.nn.Module):
the full output[:num_actual_toks] buffer after both forward_mha and
forward_mqa have written their results.
"""
batch_spec = BatchSpec(seq_lens=[1] * batch_size, query_lens=[1] * batch_size)
common_attn_metadata = create_common_attn_metadata(
batch_spec, self.block_size, self.device, arange_block_indices=True
@@ -408,7 +407,7 @@ def test_mla_attention_quant_pattern(
dist_init,
disable_vllm_compile_cache,
):
"""Test MLA AttentionQuantPattern fusion pass"""
"""Test MLA AttentionQuantPattern fusion pass."""
if (
model_class is TestMLAAttentionNvfp4QuantPatternModel
and not is_nvfp4_supported()
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Tests for manual fusion via maybe_fused_act_quant.
"""Tests for manual fusion via maybe_fused_act_quant.
Tests all fusion paths in _FUSED_ACT_QUANT:
- kFp8StaticTensorSym: all platforms
@@ -15,7 +15,8 @@ DEVICE_TYPE = current_platform.device_type
class SplitCoalescingModel(torch.nn.Module):
"""Model with 3 separate split_with_sizes calls on the same input,
simulating the B200+FP8 graph where CSE fails to merge them."""
simulating the B200+FP8 graph where CSE fails to merge them.
"""
def __init__(self, q_size: int, kv_size: int) -> None:
super().__init__()