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
@@ -29,8 +29,7 @@ logger = init_logger("vllm.tests.compile.backend")
|
||||
|
||||
|
||||
class LazyInitPass(InductorPass):
|
||||
"""
|
||||
If there's a pass that we want to initialize lazily in a test,
|
||||
"""If there's a pass that we want to initialize lazily in a test,
|
||||
we can wrap it in LazyInitPass, which will initialize the pass when invoked
|
||||
and then immediately invoke it.
|
||||
"""
|
||||
@@ -45,8 +44,7 @@ class LazyInitPass(InductorPass):
|
||||
|
||||
|
||||
class TestBackend:
|
||||
"""
|
||||
This class provides a simple Inductor backend that can be used for testing.
|
||||
"""This class provides a simple Inductor backend that can be used for testing.
|
||||
It takes a list of custom passes and runs them after Inductor's passes.
|
||||
It also saves the graph before and after the custom passes for inspection.
|
||||
|
||||
|
||||
@@ -11,8 +11,7 @@ from vllm.platforms.interface import DeviceCapability
|
||||
|
||||
@pytest.fixture
|
||||
def mock_cuda_platform():
|
||||
"""
|
||||
Fixture that returns a factory for creating mocked CUDA platforms.
|
||||
"""Fixture that returns a factory for creating mocked CUDA platforms.
|
||||
|
||||
Usage:
|
||||
def test_something(mock_cuda_platform):
|
||||
@@ -51,8 +50,7 @@ def mock_cuda_platform():
|
||||
|
||||
@pytest.fixture
|
||||
def mock_xpu_platform():
|
||||
"""
|
||||
Fixture that returns a factory for creating mocked XPU platforms.
|
||||
"""Fixture that returns a factory for creating mocked XPU platforms.
|
||||
|
||||
Usage:
|
||||
def test_something(mock_xpu_platform):
|
||||
|
||||
@@ -1,11 +1,10 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
WARNING: This test runs in both single-node (4 GPUs) and multi-node
|
||||
(2 node with 2 GPUs each) modes. If the test only uses 2 GPUs, it is
|
||||
important to set the distributed backend to "mp" to avoid Ray scheduling
|
||||
all workers in a node other than the head node, which can cause the test
|
||||
to fail.
|
||||
"""WARNING: This test runs in both single-node (4 GPUs) and multi-node
|
||||
(2 node with 2 GPUs each) modes. If the test only uses 2 GPUs, it is
|
||||
important to set the distributed backend to "mp" to avoid Ray scheduling
|
||||
all workers in a node other than the head node, which can cause the test
|
||||
to fail.
|
||||
"""
|
||||
|
||||
import json
|
||||
|
||||
@@ -17,8 +17,7 @@ from vllm.v1.attention.backends.registry import AttentionBackendEnum
|
||||
|
||||
@contextlib.contextmanager
|
||||
def temporary_environ(env_vars):
|
||||
"""
|
||||
Temporarily set environment variables and restore them afterward.
|
||||
"""Temporarily set environment variables and restore them afterward.
|
||||
We have to do this vs monkeypatch because monkeypatch doesn't work
|
||||
with "module" scoped fixtures.
|
||||
"""
|
||||
@@ -125,8 +124,7 @@ def llm_pair(request):
|
||||
indirect=True,
|
||||
)
|
||||
class TestFullCUDAGraph:
|
||||
"""
|
||||
Use a class such that an llm pair is constructed once for all
|
||||
"""Use a class such that an llm pair is constructed once for all
|
||||
batch_size/max_tokens combinations and released immediately after.
|
||||
|
||||
Module-scope fixtures would stick around the whole time,
|
||||
@@ -149,11 +147,9 @@ class TestFullCUDAGraph:
|
||||
],
|
||||
)
|
||||
def test_full_cudagraph(self, batch_size, max_tokens, llm_pair: tuple[LLM, LLM]):
|
||||
"""
|
||||
Test various batch sizes and max_tokens to ensure that the
|
||||
"""Test various batch sizes and max_tokens to ensure that the
|
||||
full cudagraph compilation works for padded cases too.
|
||||
"""
|
||||
|
||||
full_cudagraph_llm, piecewise_llm = llm_pair
|
||||
|
||||
prompts = ["the quick brown fox"] * batch_size
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Test (piecewise) compilation with a simple model where multiple submodules
|
||||
"""Test (piecewise) compilation with a simple model where multiple submodules
|
||||
are compiled and graph captured separately.
|
||||
"""
|
||||
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Test the piecewise compilation with a simple model so that we
|
||||
"""Test the piecewise compilation with a simple model so that we
|
||||
can exactly calculate the expected output and side effects.
|
||||
"""
|
||||
|
||||
@@ -52,8 +51,7 @@ class SillyModel(nn.Module):
|
||||
self.intermediate_unbacked = intermediate_unbacked
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Overall effect:
|
||||
"""Overall effect:
|
||||
x = 3 * x + 19
|
||||
global_counter += 2
|
||||
"""
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Test the piecewise compilation with a simple model, comparing the output
|
||||
"""Test the piecewise compilation with a simple model, comparing the output
|
||||
with and without the piecewise compilation.
|
||||
|
||||
This is a tractable model, the weights and computation are specially designed
|
||||
|
||||
@@ -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__(
|
||||
|
||||
@@ -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),
|
||||
|
||||
@@ -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__()
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Shared PyTorch custom silly attention for compilation tests.
|
||||
"""Shared PyTorch custom silly attention for compilation tests.
|
||||
Centralizes custom operation definitions to avoid duplicate registrations.
|
||||
"""
|
||||
|
||||
@@ -21,12 +20,12 @@ _global_counter = 0
|
||||
|
||||
|
||||
def get_global_counter():
|
||||
"""Get the current global counter value"""
|
||||
"""Get the current global counter value."""
|
||||
return _global_counter
|
||||
|
||||
|
||||
def reset_global_counter():
|
||||
"""Reset the global counter to 0"""
|
||||
"""Reset the global counter to 0."""
|
||||
global _global_counter
|
||||
_global_counter = 0
|
||||
|
||||
@@ -34,8 +33,7 @@ def reset_global_counter():
|
||||
def silly_attention(
|
||||
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, out: torch.Tensor
|
||||
) -> None:
|
||||
"""
|
||||
Unified attention implementation that depends on
|
||||
"""Unified attention implementation that depends on
|
||||
all inputs and affects the output.
|
||||
Always increments a global counter that tests can use or ignore.
|
||||
"""
|
||||
@@ -51,7 +49,7 @@ def silly_attention(
|
||||
def silly_attention_fake(
|
||||
q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, out: torch.Tensor
|
||||
) -> None:
|
||||
"""Fake implementation for testing"""
|
||||
"""Fake implementation for testing."""
|
||||
return
|
||||
|
||||
|
||||
|
||||
@@ -268,8 +268,7 @@ def test_save_and_load_slice(monkeypatch: pytest.MonkeyPatch):
|
||||
def test_cache_load_returns_tuple_consistency_tuple_output(
|
||||
monkeypatch: pytest.MonkeyPatch,
|
||||
):
|
||||
"""
|
||||
Test that cache loading correctly handles models that return tuples.
|
||||
"""Test that cache loading correctly handles models that return tuples.
|
||||
|
||||
This verifies that when a model returns a tuple of tensors, the output
|
||||
type is preserved as a tuple between fresh compilation and cache load.
|
||||
@@ -335,8 +334,7 @@ def test_cache_load_returns_tuple_consistency_tuple_output(
|
||||
|
||||
@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
|
||||
def test_shape_env(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test that the shape environment is correctly serialized and preserved
|
||||
"""Test that the shape environment is correctly serialized and preserved
|
||||
when loading from cache.
|
||||
"""
|
||||
with monkeypatch.context() as m:
|
||||
@@ -374,8 +372,7 @@ def test_shape_env(monkeypatch: pytest.MonkeyPatch):
|
||||
def test_partition_wrapper_applied_on_aot_load(
|
||||
monkeypatch: pytest.MonkeyPatch, vllm_tmp_cache: Path, mocker
|
||||
):
|
||||
"""
|
||||
Test that partition wrappers are applied when loading AOT cached functions.
|
||||
"""Test that partition wrappers are applied when loading AOT cached functions.
|
||||
|
||||
This test verifies the fix for GitHub issue #31439 where AOT compile
|
||||
caused 2x latency regression when use_inductor_graph_partition=True.
|
||||
@@ -498,14 +495,12 @@ def test_standalone_compile_correctness():
|
||||
@pytest.mark.skipif(not is_torch_equal_or_newer("2.10.0"), reason="requires torch 2.10")
|
||||
@create_new_process_for_each_test("spawn")
|
||||
def test_gpt2_cache_hit(monkeypatch: pytest.MonkeyPatch):
|
||||
"""
|
||||
Test that compiling gpt2 twice results in a cache hit.
|
||||
"""Test that compiling gpt2 twice results in a cache hit.
|
||||
|
||||
Counter values are read from the EngineCore subprocess via
|
||||
``LLM.collective_rpc`` so the test works under default V1
|
||||
multiprocessing (no shared memory between test and engine).
|
||||
"""
|
||||
|
||||
from vllm import LLM
|
||||
|
||||
def _snap(self):
|
||||
@@ -641,7 +636,7 @@ class TestStandaloneCompiledArtifacts:
|
||||
|
||||
@patch("torch._inductor.standalone_compile.AOTCompiledArtifact.deserialize")
|
||||
def test_load_all_success(self, mock_deserialize):
|
||||
"""Test successful loading of all artifacts"""
|
||||
"""Test successful loading of all artifacts."""
|
||||
cache = StandaloneCompiledArtifacts()
|
||||
|
||||
mock_artifact1 = Mock()
|
||||
@@ -658,7 +653,7 @@ class TestStandaloneCompiledArtifacts:
|
||||
|
||||
@patch("torch._inductor.standalone_compile.AOTCompiledArtifact.deserialize")
|
||||
def test_load_all_already_loaded(self, mock_deserialize):
|
||||
"""Test that load_all skips if already loaded"""
|
||||
"""Test that load_all skips if already loaded."""
|
||||
cache = StandaloneCompiledArtifacts()
|
||||
|
||||
mock_artifact = Mock()
|
||||
@@ -671,7 +666,7 @@ class TestStandaloneCompiledArtifacts:
|
||||
|
||||
@patch("torch._inductor.standalone_compile.AOTCompiledArtifact.deserialize")
|
||||
def test_get_loaded_artifact(self, mock_deserialize):
|
||||
"""Test retrieving loaded artifacts"""
|
||||
"""Test retrieving loaded artifacts."""
|
||||
cache = StandaloneCompiledArtifacts()
|
||||
|
||||
mock_artifact = Mock()
|
||||
|
||||
@@ -77,7 +77,8 @@ def test_non_primitive_kwargs_lifted_to_consts(
|
||||
into the namespace, so ``device`` is unbound. The fix collects such
|
||||
objects into ``__vllm_consts__`` and references them by index. The
|
||||
unqualified ``device(type=...)`` form must never appear in the
|
||||
generated source."""
|
||||
generated source.
|
||||
"""
|
||||
split_gm = _trace_and_split(model_fn, (x,), split_ops)
|
||||
code, submod_names, consts = generate_execution_code(split_gm)
|
||||
|
||||
@@ -98,7 +99,8 @@ def test_dtype_singleton_deduped(x: torch.Tensor) -> None:
|
||||
"""``torch.float16`` is a process-wide singleton, so two ops referring
|
||||
to it in the traced graph share a single consts slot via ``id()``-based
|
||||
dedup. Distinct expressions (``x.to(...)`` vs ``(x*2).to(...)``) ensure
|
||||
the tracer can't CSE the two ops into a single node."""
|
||||
the tracer can't CSE the two ops into a single node.
|
||||
"""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
return x.to(dtype=torch.float16) + (x * 2).to(dtype=torch.float16)
|
||||
@@ -147,7 +149,8 @@ def test_distinct_dtypes_get_distinct_slots(x: torch.Tensor) -> None:
|
||||
def test_consts_ordering_deterministic(x: torch.Tensor) -> None:
|
||||
"""Two independent traces of the same model must produce equal consts
|
||||
lists *in the same order*. Cache artifacts identify const slots by
|
||||
index, so a non-deterministic order would invalidate cached code."""
|
||||
index, so a non-deterministic order would invalidate cached code.
|
||||
"""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
# Multiple distinct non-primitives encountered in a fixed graph order.
|
||||
@@ -185,7 +188,8 @@ def test_consts_shared_across_split_submods(x: torch.Tensor) -> None:
|
||||
The function below splits into three inlined submods, two of which
|
||||
independently reference ``torch.float16``. The shared ``const_index``
|
||||
threaded through recursive ``generate_execution_code_with_name`` calls
|
||||
must collapse the dtype to a single slot used from both submods."""
|
||||
must collapse the dtype to a single slot used from both submods.
|
||||
"""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
a = x.to(dtype=torch.float16) # submod_0: _to_copy(fp16)
|
||||
@@ -230,7 +234,8 @@ def test_non_graphmodule_submod_uses_indexed_callable(x: torch.Tensor) -> None:
|
||||
"""When a child of split_gm is *not* a ``torch.fx.GraphModule`` — as
|
||||
happens in production once ``PiecewiseBackend`` replaces submods —
|
||||
codegen emits ``__vllm_submods__[idx](...)`` instead of inlining, and
|
||||
the runtime callable is bound from ``submod_callables``."""
|
||||
the runtime callable is bound from ``submod_callables``.
|
||||
"""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
return x.relu().sigmoid()
|
||||
@@ -280,7 +285,8 @@ def test_non_graphmodule_submod_uses_indexed_callable(x: torch.Tensor) -> None:
|
||||
def test_getitem_in_stitching_graph(x: torch.Tensor) -> None:
|
||||
"""``operator.getitem`` on submod tuple returns is the ``call_function``
|
||||
special case at codegen.py — emitted as ``name = source[index]``
|
||||
rather than a function call."""
|
||||
rather than a function call.
|
||||
"""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
return x.relu().sigmoid()
|
||||
@@ -298,7 +304,8 @@ def test_getitem_in_stitching_graph(x: torch.Tensor) -> None:
|
||||
def test_del_emitted_for_intermediate_values(x: torch.Tensor) -> None:
|
||||
"""The codegen schedules ``del`` after a value's last use to free
|
||||
memory early. Multi-submod splits naturally have intermediates whose
|
||||
last use is not the output node."""
|
||||
last use is not the output node.
|
||||
"""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
return x.relu().sigmoid().tanh()
|
||||
@@ -318,7 +325,8 @@ def test_with_submod_false_rejects_call_module() -> None:
|
||||
"""``generate_execution_code_with_name(with_submod=False)`` is the
|
||||
recursive entry for inlining a GraphModule into its parent. It must
|
||||
refuse a graph that itself contains ``call_module`` nodes — the parent
|
||||
is responsible for handling those."""
|
||||
is responsible for handling those.
|
||||
"""
|
||||
g = fx.Graph()
|
||||
x_node = g.placeholder("x")
|
||||
root = torch.nn.Module()
|
||||
@@ -335,7 +343,8 @@ def test_node_ref_recurses_through_containers() -> None:
|
||||
"""``_node_ref`` is the recursive walker that lifts non-primitives
|
||||
nested inside list/tuple/dict args. Real aten ops rarely produce such
|
||||
structures, but the path is needed for DTensor placement lists and
|
||||
other future cases — unit-test the walker directly."""
|
||||
other future cases — unit-test the walker directly.
|
||||
"""
|
||||
consts: list = []
|
||||
const_index: dict[int, int] = {}
|
||||
cpu = torch.device("cpu")
|
||||
@@ -359,7 +368,8 @@ def test_node_ref_recurses_through_containers() -> None:
|
||||
def test_legacy_code_without_consts() -> None:
|
||||
"""``compile_execution_fn(consts=None)`` must still load code that has
|
||||
no ``__vllm_consts__`` reference, so older serialized cache artifacts
|
||||
keep working."""
|
||||
keep working.
|
||||
"""
|
||||
# Pre-consts codegen: no __vllm_consts__ reference, only torch/operator.
|
||||
legacy_code = (
|
||||
"import torch\n"
|
||||
|
||||
@@ -129,7 +129,8 @@ def test_compile_config_get_compile_ranges():
|
||||
|
||||
class PostGradStaticShapeChecker(InductorPass):
|
||||
"""Asserts that compile_sizes entries produce graphs with fully concrete
|
||||
(non-symbolic) shapes, and compile_ranges entries have symbolic shapes."""
|
||||
(non-symbolic) shapes, and compile_ranges entries have symbolic shapes.
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
self.num_static_calls = 0
|
||||
@@ -170,7 +171,8 @@ class PostGradStaticShapeChecker(InductorPass):
|
||||
|
||||
def test_compile_sizes_produce_static_shapes(disable_vllm_compile_cache):
|
||||
"""Verify that compile_sizes entries are compiled with fully concrete
|
||||
shapes (no SymInts), while compile_ranges entries retain dynamic shapes."""
|
||||
shapes (no SymInts), while compile_ranges entries retain dynamic shapes.
|
||||
"""
|
||||
checker = PostGradStaticShapeChecker()
|
||||
torch.set_default_device("cuda")
|
||||
vllm_config = VllmConfig(
|
||||
|
||||
@@ -1229,7 +1229,8 @@ def test_compile_sizes_padding_validation():
|
||||
|
||||
def test_inductor_asserts_default_disabled(monkeypatch):
|
||||
"""Test that inductor runtime asserts are disabled by default
|
||||
(INFO logging level) on torch < 2.12."""
|
||||
(INFO logging level) on torch < 2.12.
|
||||
"""
|
||||
monkeypatch.setenv("VLLM_LOGGING_LEVEL", "INFO")
|
||||
|
||||
import importlib
|
||||
@@ -1247,7 +1248,8 @@ def test_inductor_asserts_default_disabled(monkeypatch):
|
||||
|
||||
def test_inductor_asserts_enabled_in_debug(monkeypatch):
|
||||
"""Test that VLLM_LOGGING_LEVEL=DEBUG enables inductor runtime asserts
|
||||
on torch < 2.12."""
|
||||
on torch < 2.12.
|
||||
"""
|
||||
monkeypatch.setenv("VLLM_LOGGING_LEVEL", "DEBUG")
|
||||
|
||||
import importlib
|
||||
@@ -1283,7 +1285,8 @@ def test_get_inductor_factors_includes_configs():
|
||||
|
||||
def test_inductor_asserts_user_override(monkeypatch):
|
||||
"""Test that explicit inductor_compile_config overrides the
|
||||
debug-logging default."""
|
||||
debug-logging default.
|
||||
"""
|
||||
monkeypatch.setenv("VLLM_LOGGING_LEVEL", "INFO")
|
||||
|
||||
import importlib
|
||||
|
||||
@@ -22,7 +22,7 @@ from vllm.utils.torch_utils import is_torch_equal_or_newer
|
||||
|
||||
|
||||
def get_test_models():
|
||||
"""Get list of models to test based on PyTorch version"""
|
||||
"""Get list of models to test based on PyTorch version."""
|
||||
models = [
|
||||
"Qwen/Qwen3-0.6B",
|
||||
"openai-community/gpt2",
|
||||
@@ -172,7 +172,6 @@ def test_model_specialization_with_evaluate_guards(
|
||||
"""Test that evaluate_guards correctly detects shape specialization
|
||||
violations.
|
||||
"""
|
||||
|
||||
if (
|
||||
use_aot_compile
|
||||
and dynamic_shapes_type == DynamicShapesType.BACKED
|
||||
|
||||
@@ -24,8 +24,7 @@ DEVICE_TYPE = current_platform.device_type
|
||||
|
||||
|
||||
def test_getitem_moved_to_producer_subgraph():
|
||||
"""
|
||||
Test that getitem operations are moved to the same subgraph as their input,
|
||||
"""Test that getitem operations are moved to the same subgraph as their input,
|
||||
preventing tuple inputs to submodules.
|
||||
"""
|
||||
|
||||
@@ -79,8 +78,7 @@ def test_getitem_moved_to_producer_subgraph():
|
||||
|
||||
|
||||
def test_no_tuple_inputs_with_multiple_consumers():
|
||||
"""
|
||||
Test that when a tuple is consumed by multiple split operations,
|
||||
"""Test that when a tuple is consumed by multiple split operations,
|
||||
getitem operations are properly moved to avoid tuple inputs.
|
||||
"""
|
||||
|
||||
@@ -138,13 +136,10 @@ def test_no_tuple_inputs_with_multiple_consumers():
|
||||
|
||||
|
||||
def test_consecutive_ops_in_split():
|
||||
"""
|
||||
Test that consecutive splitting operations are grouped into the same subgraph
|
||||
"""
|
||||
"""Test that consecutive splitting operations are grouped into the same subgraph."""
|
||||
|
||||
def model_fn(x: torch.Tensor) -> torch.Tensor:
|
||||
"""
|
||||
Define a simple model where consecutive operations create opportunities
|
||||
"""Define a simple model where consecutive operations create opportunities
|
||||
for splitting subgraphs.
|
||||
"""
|
||||
# Apply silly attention followed by consecutive operations
|
||||
@@ -211,8 +206,7 @@ def _subgraphs_with_empty_nodes(split_items, *, is_splitting_graph):
|
||||
|
||||
|
||||
def test_empty_only_partition_stays_separate_after_splitting_predecessor():
|
||||
"""
|
||||
Empty-only subgraphs should not be merged when the only predecessor is
|
||||
"""Empty-only subgraphs should not be merged when the only predecessor is
|
||||
a splitting-op subgraph.
|
||||
"""
|
||||
|
||||
@@ -248,8 +242,7 @@ def test_empty_only_partition_stays_separate_after_splitting_predecessor():
|
||||
|
||||
|
||||
def test_empty_only_partition_is_merged():
|
||||
"""
|
||||
Empty-only subgraphs should still be merged when a non-splitting predecessor
|
||||
"""Empty-only subgraphs should still be merged when a non-splitting predecessor
|
||||
exists. The merged empty node must remain outside splitting-op subgraphs.
|
||||
"""
|
||||
|
||||
@@ -294,8 +287,7 @@ def test_empty_only_partition_is_merged():
|
||||
|
||||
|
||||
def test_builtin_empty_only_partition_is_merged():
|
||||
"""
|
||||
In Dynamo graphs, torch.empty/empty_like may appear as builtin call targets
|
||||
"""In Dynamo graphs, torch.empty/empty_like may appear as builtin call targets
|
||||
(not aten OpOverload). Ensure empty-only partitions are still merged.
|
||||
"""
|
||||
|
||||
@@ -340,8 +332,7 @@ def test_builtin_empty_only_partition_is_merged():
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
|
||||
def test_sym_size_whole_shape_boundary():
|
||||
"""
|
||||
Test that using x.size() (whole shape) across a split boundary can be
|
||||
"""Test that using x.size() (whole shape) across a split boundary can be
|
||||
compiled by standalone_compile.
|
||||
|
||||
The dynamo graph looks like:
|
||||
@@ -400,8 +391,7 @@ def test_sym_size_whole_shape_boundary():
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
|
||||
def test_symint_crosses_split_boundary():
|
||||
"""
|
||||
Test that SymInt placeholders from torch.compile + mark_dynamic
|
||||
"""Test that SymInt placeholders from torch.compile + mark_dynamic
|
||||
cross split boundaries safely via split_module's natural threading.
|
||||
|
||||
SymInt values are threaded through subgraphs by split_module and
|
||||
@@ -459,8 +449,7 @@ def test_symint_crosses_split_boundary():
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
|
||||
def test_shape_boundary_standalone_compile():
|
||||
"""
|
||||
Repro for the original production bug:
|
||||
"""Repro for the original production bug:
|
||||
|
||||
AssertionError: out_spec mismatch
|
||||
TreeSpec(tuple, None, [*, *, TreeSpec(Size, None, [*, *]), *])
|
||||
@@ -529,8 +518,7 @@ def test_shape_boundary_standalone_compile():
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
|
||||
def test_size_used_in_multiple_consumer_subgraphs():
|
||||
"""
|
||||
Validates that x.size() (whole shape) used by multiple downstream subgraphs
|
||||
"""Validates that x.size() (whole shape) used by multiple downstream subgraphs
|
||||
does not cause torch.Size to cross split boundaries.
|
||||
|
||||
Model:
|
||||
@@ -584,8 +572,7 @@ def test_size_used_in_multiple_consumer_subgraphs():
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
|
||||
def test_sym_size_metadata_propagated():
|
||||
"""
|
||||
Validates that new sym_size.int nodes created by the pre-pass have
|
||||
"""Validates that new sym_size.int nodes created by the pre-pass have
|
||||
example_value metadata set. Without it, placeholder metadata in consumer
|
||||
subgraphs would be None, breaking any code that dynamically builds
|
||||
example inputs from metadata (e.g. standalone_compile per-submodule).
|
||||
@@ -634,8 +621,7 @@ def test_sym_size_metadata_propagated():
|
||||
|
||||
|
||||
def test_decompose_size_with_getitem_user():
|
||||
"""
|
||||
Regression test: _decompose_size_nodes must handle getitem users of size()
|
||||
"""Regression test: _decompose_size_nodes must handle getitem users of size()
|
||||
correctly.
|
||||
|
||||
When a graph contains x.shape[i], it can appear as:
|
||||
@@ -704,8 +690,7 @@ def test_decompose_size_with_getitem_user():
|
||||
|
||||
|
||||
def test_decompose_size_leaves_scalar_size_with_dim():
|
||||
"""
|
||||
Regression test: _decompose_size_nodes must leave x.size(dim) alone.
|
||||
"""Regression test: _decompose_size_nodes must leave x.size(dim) alone.
|
||||
|
||||
x.size() returns a torch.Size tuple that can't cross split boundaries and
|
||||
must be decomposed. x.size(dim), however, already returns a scalar
|
||||
|
||||
@@ -236,7 +236,8 @@ class TestInductorFallbackFastPath:
|
||||
|
||||
def test_inner_set_membership_still_works_for_standard_ops(self):
|
||||
"""The patch must not break Inductor's existing fallback decisions
|
||||
for non-vllm ops such as ``torchvision::roi_align``."""
|
||||
for non-vllm ops such as ``torchvision::roi_align``.
|
||||
"""
|
||||
import torch._inductor.lowering as _lowering
|
||||
|
||||
allow_list = _lowering.FALLBACK_ALLOW_LIST
|
||||
|
||||
@@ -60,6 +60,7 @@ class TraceStructuredCapture:
|
||||
Args:
|
||||
event_type: The event type to filter by (e.g., "artifact", "graph_dump")
|
||||
name_pattern: Regex pattern to match against the artifact name
|
||||
|
||||
"""
|
||||
regex = re.compile(name_pattern)
|
||||
return [
|
||||
|
||||
Reference in New Issue
Block a user