[CI] Ignore ruff D209, rejoin the docstrings it split, and silence incompatible-rule warnings (#57212)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
1085b64425
commit
651a88c09a
@@ -732,8 +732,7 @@ 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),
|
||||
|
||||
@@ -15,8 +15,7 @@ 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__()
|
||||
|
||||
@@ -77,8 +77,7 @@ 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)
|
||||
|
||||
@@ -99,8 +98,7 @@ 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)
|
||||
@@ -149,8 +147,7 @@ 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.
|
||||
@@ -188,8 +185,7 @@ 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)
|
||||
@@ -234,8 +230,7 @@ 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()
|
||||
@@ -285,8 +280,7 @@ 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()
|
||||
@@ -304,8 +298,7 @@ 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()
|
||||
@@ -325,8 +318,7 @@ 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()
|
||||
@@ -343,8 +335,7 @@ 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")
|
||||
@@ -368,8 +359,7 @@ 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"
|
||||
|
||||
@@ -129,8 +129,7 @@ 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
|
||||
@@ -171,8 +170,7 @@ 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(
|
||||
|
||||
@@ -1229,8 +1229,7 @@ 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
|
||||
@@ -1248,8 +1247,7 @@ 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
|
||||
@@ -1285,8 +1283,7 @@ 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
|
||||
|
||||
@@ -236,8 +236,7 @@ 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
|
||||
|
||||
Reference in New Issue
Block a user