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
+5 -3
View File
@@ -34,6 +34,7 @@ def human_readable_int(value: str) -> int:
- '1k' -> 1,000
- '1K' -> 1,024
- '25.6k' -> 25,600
"""
value = value.strip()
@@ -82,6 +83,7 @@ def human_readable_int_or_auto(value: str) -> int:
- '1K' -> 1,024
- '25.6k' -> 25,600
- '-1' or 'auto' -> -1 (special value for auto-detection)
"""
value = value.strip()
@@ -95,8 +97,7 @@ class SortedHelpFormatter(ArgumentDefaultsHelpFormatter, RawDescriptionHelpForma
"""SortedHelpFormatter that sorts arguments by their option strings."""
def _split_lines(self, text, width):
"""
1. Sentences split across lines have their single newlines removed.
"""1. Sentences split across lines have their single newlines removed.
2. Paragraphs and lists are split into separate lines.
3. Each line is wrapped to the specified width (width of terminal).
"""
@@ -466,7 +467,7 @@ class FlexibleArgumentParser(ArgumentParser):
The arguments in config file will be inserted between
the argument list.
example:
Example:
```yaml
port: 12323
tensor-parallel-size: 4
@@ -492,6 +493,7 @@ class FlexibleArgumentParser(ArgumentParser):
Please note how the config args are inserted after the sub command.
this way the order of priorities is maintained when these are args
parsed by super().
"""
assert args.count("--config") <= 1, "More than one config file specified!"
+3 -7
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Contains helpers related to asynchronous code.
"""Contains helpers related to asynchronous code.
This is similar in concept to the `asyncio` module.
"""
@@ -29,8 +28,7 @@ def make_async(
func: Callable[P, T],
executor: Executor | None = None,
) -> Callable[P, Awaitable[T]]:
"""
Take a blocking function, and run it on in an executor thread.
"""Take a blocking function, and run it on in an executor thread.
This function prevents the blocking function from blocking the
asyncio event loop.
@@ -49,8 +47,7 @@ def make_async_with_semaphore(
func: Callable[P, T],
executor: ThreadPoolExecutor,
) -> Callable[P, Awaitable[T]]:
"""
Take a blocking function, and run it on in an executor thread.
"""Take a blocking function, and run it on in an executor thread.
This function prevents the blocking function from blocking the
asyncio event loop.
@@ -59,7 +56,6 @@ def make_async_with_semaphore(
The function is wrapped in a semaphore to limit the number of
concurrent executions making it easier to cancel tasks before they start.
"""
semaphore = asyncio.Semaphore(executor._max_workers)
async def _async_wrapper(*args: P.args, **kwargs: P.kwargs) -> T:
+4 -6
View File
@@ -102,8 +102,7 @@ class LRUCache(cachetools.LRUCache[_K, _V]):
return self.currsize / self.maxsize
def stat(self, *, delta: bool = False) -> CacheInfo:
"""
Gets the cumulative number of hits and queries against this cache.
"""Gets the cumulative number of hits and queries against this cache.
If `delta=True`, instead gets these statistics
since the last call that also passed `delta=True`.
@@ -165,6 +164,7 @@ class LRUCache(cachetools.LRUCache[_K, _V]):
Returns:
`True` if the item was cached, otherwise `False`.
"""
try:
self[key] = value
@@ -175,8 +175,7 @@ class LRUCache(cachetools.LRUCache[_K, _V]):
return True
def pin(self, key: _K) -> None:
"""
Pins a key in the cache preventing it from being
"""Pins a key in the cache preventing it from being
evicted in the LRU order.
"""
if key not in self:
@@ -184,8 +183,7 @@ class LRUCache(cachetools.LRUCache[_K, _V]):
self.pinned_items.add(key)
def _unpin(self, key: _K) -> None:
"""
Unpins a key in the cache allowing it to be
"""Unpins a key in the cache allowing it to be
evicted in the LRU order.
"""
self.pinned_items.remove(key)
+3 -6
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Contains helpers that are applied to collections.
"""Contains helpers that are applied to collections.
This is similar in concept to the `collections` module.
"""
@@ -20,8 +19,7 @@ _V = TypeVar("_V")
class LazyDict(Mapping[str, _V], Generic[_V]):
"""
Evaluates dictionary items only when they are accessed.
"""Evaluates dictionary items only when they are accessed.
Adapted from: https://stackoverflow.com/a/47212782/5082708
"""
@@ -116,8 +114,7 @@ def flatten_2d_lists(lists: Iterable[Iterable[T]]) -> list[T]:
def full_groupby(values: Iterable[_V], *, key: Callable[[_V], _K]):
"""
Unlike [`itertools.groupby`][], groups are not broken by
"""Unlike [`itertools.groupby`][], groups are not broken by
non-contiguous data.
"""
groups = defaultdict[_K, list[_V]](list)
+4 -4
View File
@@ -19,10 +19,10 @@ class Counter:
class AtomicCounter:
"""An atomic, thread-safe counter"""
"""An atomic, thread-safe counter."""
def __init__(self, initial: int = 0) -> None:
"""Initialize a new atomic counter to given initial value"""
"""Initialize a new atomic counter to given initial value."""
super().__init__()
self._value = initial
@@ -33,13 +33,13 @@ class AtomicCounter:
return self._value
def inc(self, num: int = 1) -> int:
"""Atomically increment the counter by num and return the new value"""
"""Atomically increment the counter by num and return the new value."""
with self._lock:
self._value += num
return self._value
def dec(self, num: int = 1) -> int:
"""Atomically decrement the counter by num and return the new value"""
"""Atomically decrement the counter by num and return the new value."""
with self._lock:
self._value -= num
return self._value
+4 -2
View File
@@ -128,6 +128,7 @@ def check_cgroup_memory_available(
Low headroom logs a warning, but does not reject the allocation because
cgroup usage can include reclaimable memory. If the cgroup limit or usage
cannot be read, the check is skipped.
"""
cgroup_limit = get_cgroup_memory_limit()
cgroup_usage = get_cgroup_memory_usage()
@@ -171,7 +172,7 @@ def get_memory_affinity(pid: int = 0) -> list[int]:
def parse_id_list(raw_str: str) -> list[int]:
"""Parses strings like '0-2,4,7-8' into [0, 1, 2, 4, 7, 8]"""
"""Parses strings like '0-2,4,7-8' into [0, 1, 2, 4, 7, 8]."""
result: list[int] = []
if not raw_str:
return result
@@ -273,7 +274,8 @@ def get_visible_memory_node() -> list[int]:
def _synthesize_cpu_list() -> list[LogicalCPUInfo]:
"""Synthesize a flat CPU list: each logical CPU is its own core on
NUMA node 0. Used when lscpu output is unavailable or unparsable
(e.g. macOS, RISC-V)."""
(e.g. macOS, RISC-V).
"""
cpu_count = os.cpu_count()
assert cpu_count
return [LogicalCPUInfo(i, i, 0) for i in range(cpu_count)]
+1 -3
View File
@@ -1,9 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Contains replacement functions to fallback Triton usages in CPU backend
"""
"""Contains replacement functions to fallback Triton usages in CPU backend."""
import ctypes
from collections.abc import Callable
+14 -6
View File
@@ -64,7 +64,7 @@ class DeepGemmQuantScaleFMT(Enum):
@classmethod
def init_oracle_cache(cls) -> None:
"""Initialize the oracle decision and store it in the class cache"""
"""Initialize the oracle decision and store it in the class cache."""
cached = getattr(cls, "_oracle_cache", None)
if cached is not None:
return
@@ -426,7 +426,7 @@ def mk_alignment_scope(value: int):
def get_col_major_tma_aligned_tensor(x: torch.Tensor) -> torch.Tensor:
"""Wrapper for DeepGEMM's get_mn_major_tma_aligned_tensor"""
"""Wrapper for DeepGEMM's get_mn_major_tma_aligned_tensor."""
_lazy_init()
if _get_mn_major_tma_aligned_tensor_impl is None:
return _missing()
@@ -568,6 +568,7 @@ def fp8_fp4_mqa_logits(
Returns:
Logits tensor of shape [M, N], dtype `torch.float32`.
"""
_lazy_init()
if _fp8_fp4_mqa_logits_impl is None:
@@ -623,6 +624,7 @@ def get_paged_mqa_logits_metadata(
Returns:
Tensor of shape [slots + 1, 2] consumed by `fp8_fp4_paged_mqa_logits`
to schedule work across SMs.
"""
_lazy_init()
if _get_paged_mqa_logits_metadata_impl is None:
@@ -674,6 +676,7 @@ def fp8_fp4_paged_mqa_logits(
Returns:
Logits tensor of shape [B * next_n, max_model_len], dtype
`torch.float32`.
"""
_lazy_init()
if _fp8_fp4_paged_mqa_logits_impl is None:
@@ -701,7 +704,8 @@ def fp8_fp4_paged_mqa_logits(
def has_deep_gemm_sparse_mqa() -> bool:
"""Whether the installed DeepGEMM provides the sparse-indexer kernels
(``fp8_fp4_(paged_)sparse_mqa_logits``, added in DeepGEMM 2.8, SM100-only)."""
(``fp8_fp4_(paged_)sparse_mqa_logits``, added in DeepGEMM 2.8, SM100-only).
"""
_lazy_init()
return (
_fp8_fp4_sparse_mqa_logits_impl is not None
@@ -740,6 +744,7 @@ def get_sparse_mqa_logits_metadata(
qk_dtype: dtype of the packed Q values (``torch.int8`` for MXFP4,
``torch.float8_e4m3fn`` for FP8).
sparse_block_kv: Tokens per sparse block, 8 or 16.
"""
_lazy_init()
if _get_sparse_mqa_logits_metadata_impl is None:
@@ -777,6 +782,7 @@ def get_paged_sparse_mqa_logits_metadata(
block ids, [num_q_tokens, num_max_sparse_blocks], int32.
qk_dtype: dtype of the packed Q values.
sparse_block_kv: Tokens per sparse block, 8 or 16.
"""
_lazy_init()
if _get_paged_sparse_mqa_logits_metadata_impl is None:
@@ -817,6 +823,7 @@ def fp8_fp4_sparse_mqa_logits(
bf16 logits of shape [M, num_max_sparse_blocks * sparse_block_kv];
column ``j * sparse_block_kv + o`` scores the token at
``sparse_kv_block_indices[row, j] * sparse_block_kv + ks % sbk + o``.
"""
_lazy_init()
if _fp8_fp4_sparse_mqa_logits_impl is None:
@@ -848,10 +855,13 @@ def fp8_fp4_paged_sparse_mqa_logits(
page stride 512B-aligned.
weights: [num_q_tokens, H] ``torch.bfloat16``.
metadata: From `get_paged_sparse_mqa_logits_metadata`.
num_max_sparse_blocks: Candidate blocks per row.
sparse_block_kv: Tokens per sparse block, 8 or 16.
Returns:
bf16 logits of shape
[num_q_tokens, num_max_sparse_blocks * sparse_block_kv].
"""
_lazy_init()
if _fp8_fp4_paged_sparse_mqa_logits_impl is None:
@@ -873,8 +883,7 @@ def tf32_hc_prenorm_gemm(
sqrsum: torch.Tensor,
num_split: int,
) -> torch.Tensor:
"""
Perform the following computation:
"""Perform the following computation:
out = x.float() @ fn.T
sqrsum = x.float().square().sum(-1)
@@ -950,7 +959,6 @@ def calc_diff(x: torch.Tensor, y: torch.Tensor):
and report `1 - sim`. Once kernel accuracy improves this helper can be
removed.
"""
x, y = x.double(), y.double()
denominator = (x * x + y * y).sum()
sim = 2 * (x * y).sum() / denominator
+2 -3
View File
@@ -589,8 +589,7 @@ def supports_trtllm_attention(is_prefill: bool = False) -> bool:
def force_use_trtllm_attention() -> bool | None:
"""
This function should only be called during initialization stage when vllm config
"""This function should only be called during initialization stage when vllm config
is set.
Return `None` if --attention-config.use_trtllm_attention is not set,
return `True` if TRTLLM attention is forced to be used,
@@ -628,7 +627,6 @@ def use_trtllm_attention(
has_spec: bool = False,
) -> bool:
"""Return `True` if TRTLLM attention is used."""
# CLI argument is set to 0 - respect it
if force_use_trtllm is not None and not force_use_trtllm:
return False
@@ -736,6 +734,7 @@ if has_flashinfer():
k_nope: The nope part of k, shape [num_tokens, num_heads, nope_dim].
k_pe: The rope part of k (shared), shape [num_tokens, 1, rope_dim].
This is broadcast to all heads.
"""
from flashinfer.concat_ops import concat_mla_k
+4 -4
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Contains helpers that are applied to functions.
"""Contains helpers that are applied to functions.
This is similar in concept to the `functools` module.
"""
@@ -127,8 +126,7 @@ def get_allowed_kwarg_only_overrides(
requires_kw_only: bool = True,
allow_var_kwargs: bool = False,
) -> dict[str, Any]:
"""
Given a callable which has one or more keyword only params and a dict
"""Given a callable which has one or more keyword only params and a dict
mapping param names to values, drop values that can be not be kwarg
expanded to overwrite one or more keyword-only args. This is used in a
few places to handle custom processor overrides for multimodal models,
@@ -139,12 +137,14 @@ def get_allowed_kwarg_only_overrides(
callable: Callable which takes 0 or more keyword only arguments.
If None is provided, all overrides names are allowed.
overrides: Potential overrides to be used when invoking the callable.
requires_kw_only: If True, only keyword-only params may be overridden.
allow_var_kwargs: Allows overrides that are expandable for var kwargs.
Returns:
Dictionary containing the kwargs to be leveraged which may be used
to overwrite one or more keyword only arguments when invoking the
callable.
"""
if not overrides:
return {}
+7 -17
View File
@@ -14,8 +14,7 @@ logger = init_logger(__name__)
class GCDebugConfig:
"""
Config for GC Debugger.
"""Config for GC Debugger.
- 0: disable GC debugger
- 1: enable GC debugger with gc.collect elapsed times
- '{"top_objects":5}': enable GC debugger with top 5 collected objects
@@ -44,8 +43,7 @@ class GCDebugConfig:
class GCDebugger:
"""
Debugger for GC which logs helpful information for GC understanding.
"""Debugger for GC which logs helpful information for GC understanding.
To enable, you should call maybe_attach_gc_debug_callback in the process.
"""
@@ -59,9 +57,7 @@ class GCDebugger:
self.gc_top_collected_objects: str = ""
def handle(self, phase: str, info: dict[str, int]) -> None:
"""
Handles a GC event (e.g. GC start or GC finish)
"""
"""Handles a GC event (e.g. GC start or GC finish)."""
generation = info.get("generation")
if generation is None:
return
@@ -122,8 +118,7 @@ def freeze_gc_for_cudagraph_capture():
def freeze_gc_heap() -> None:
"""
Freeze all objects tracked by the garbage collector. It should be invoked
"""Freeze all objects tracked by the garbage collector. It should be invoked
after server init / warmup, to reduce GC overhead from static objects
during serving time.
"""
@@ -137,9 +132,7 @@ def freeze_gc_heap() -> None:
def maybe_attach_gc_debug_callback() -> None:
"""
Attached a callback for GC debug when VLLM_GC_DEBUG is enabled.
"""
"""Attached a callback for GC debug when VLLM_GC_DEBUG is enabled."""
config = GCDebugConfig(envs.VLLM_GC_DEBUG)
if config.enabled:
debugger: GCDebugger = GCDebugger(config)
@@ -151,8 +144,7 @@ def maybe_attach_gc_debug_callback() -> None:
def _compute_detailed_type(o: Any) -> str:
"""
Detailed object type.
"""Detailed object type.
TODO(Jialin): Further enhance the detailed type with element types for
easier debugging. We tried but occasionally it would run into signals
@@ -167,9 +159,7 @@ def _compute_detailed_type(o: Any) -> str:
def _compute_top_gc_collected_objects(objects: list[Any], top: int) -> str:
"""
Group collected objects by types.
"""
"""Group collected objects by types."""
if top <= 0:
return ""
object_types = [_compute_detailed_type(o) for o in objects]
+4 -2
View File
@@ -72,7 +72,8 @@ _sync_filter_head: tuple | None = None
def _sync_warning_hook(message, category, filename, lineno, file=None, line=None):
"""Turn torch's sync warning into an error, but only on a thread that is
being checked and is outside any allow region."""
being checked and is outside any allow region.
"""
if _TORCH_SYNC_WARNING in str(message):
mode = _checking.get()
if mode is None or _allow_depth.get():
@@ -251,7 +252,8 @@ def _is_dense(t: torch.Tensor) -> bool:
"""Whether `t`'s elements fill a contiguous storage range, possibly
permuted (e.g. a transposed matrix). Such layouts can be copied with
pitched cudaMemcpy2D/3DAsync directly from/to pinned memory, whereas
gapped layouts (e.g. strided slices) stage through a pageable temp."""
gapped layouts (e.g. strided slices) stage through a pageable temp.
"""
if t.is_contiguous():
# The common case, answered by a memoized TensorImpl flag.
return True
+4
View File
@@ -35,6 +35,7 @@ def sha256(input: Any) -> bytes:
Returns:
Bytes representing the SHA-256 hash of the serialized input.
"""
input_bytes = pickle.dumps(input, protocol=pickle.HIGHEST_PROTOCOL)
return hashlib.sha256(input_bytes).digest()
@@ -53,6 +54,7 @@ def sha256_cbor(input: Any) -> bytes:
Returns:
Bytes representing the SHA-256 hash of the CBOR serialized input.
"""
input_bytes = cbor2.dumps(input, canonical=True)
return hashlib.sha256(input_bytes).digest()
@@ -87,6 +89,7 @@ def get_hash_fn_by_name(hash_fn_name: str) -> Callable[[Any], bytes]:
Returns:
A hash function.
"""
if hash_fn_name == "sha256":
return sha256
@@ -110,6 +113,7 @@ def safe_hash(data: bytes, usedforsecurity: bool = True) -> HASH:
Returns:
Hash object
"""
try:
return hashlib.md5(data, usedforsecurity=usedforsecurity)
+10 -22
View File
@@ -1,8 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Contains helpers related to importing modules.
"""Contains helpers related to importing modules.
This is similar in concept to the `importlib` module.
"""
@@ -24,8 +23,7 @@ logger = init_logger(__name__)
def import_pynvml():
"""
Historical comments:
"""Historical comments:
libnvml.so is the library behind nvidia-smi, and
pynvml is a Python wrapper around it. We use it to get GPU
@@ -57,8 +55,7 @@ def import_pynvml():
@cache
def import_triton_kernels():
"""
For convenience, prioritize triton_kernels that is available in
"""For convenience, prioritize triton_kernels that is available in
`site-packages`. Use `vllm.third_party.triton_kernels` as a fall-back.
"""
if _has_module("triton_kernels"):
@@ -83,8 +80,7 @@ def import_triton_kernels():
def import_plugin(plugin_path: str) -> ModuleType | None:
"""
Import a user-defined plugin.
"""Import a user-defined plugin.
Plugin can be either:
* a module in site-packages
@@ -104,8 +100,7 @@ def import_plugin(plugin_path: str) -> ModuleType | None:
def import_from_path(module_name: str, file_path: str | os.PathLike) -> ModuleType:
"""
Import a Python file according to its file path.
"""Import a Python file according to its file path.
Based on the official recipe:
https://docs.python.org/3/library/importlib.html#importing-a-source-file-directly
@@ -123,9 +118,7 @@ def import_from_path(module_name: str, file_path: str | os.PathLike) -> ModuleTy
def resolve_obj_by_qualname(qualname: str) -> Any:
"""
Resolve an object by its fully-qualified class name.
"""
"""Resolve an object by its fully-qualified class name."""
module_name, obj_name = qualname.rsplit(".", 1)
module = importlib.import_module(module_name)
return getattr(module, obj_name)
@@ -148,8 +141,7 @@ def get_vllm_optional_dependencies():
class _PlaceholderBase:
"""
Disallows downstream usage of placeholder modules.
"""Disallows downstream usage of placeholder modules.
We need to explicitly override each dunder method because
[`__getattr__`][vllm.utils.import_utils._PlaceholderBase.__getattr__]
@@ -160,8 +152,7 @@ class _PlaceholderBase:
"""
def __getattr__(self, key: str) -> Never:
"""
The main class should implement this to throw an error
"""The main class should implement this to throw an error
for attribute accesses representing downstream usage.
"""
raise NotImplementedError
@@ -304,8 +295,7 @@ class _PlaceholderBase:
class PlaceholderModule(_PlaceholderBase):
"""
A placeholder object to use when a module does not exist.
"""A placeholder object to use when a module does not exist.
This enables more informative errors when trying to access attributes
of a module that does not exist.
@@ -360,8 +350,7 @@ class _PlaceholderModuleAttr(_PlaceholderBase):
class LazyLoader(ModuleType):
"""
`LazyLoader` module borrowed from [Tensorflow]
"""`LazyLoader` module borrowed from [Tensorflow]
(https://github.com/tensorflow/tensorflow/blob/main/tensorflow/python/util/lazy_loader.py)
with an addition of "module caching".
@@ -562,7 +551,6 @@ def has_tilelang() -> bool:
def has_arctic_inference() -> bool:
"""Whether the optional `arctic_inference` package is available."""
return _has_module("arctic_inference")
+1 -2
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Monitor unexpected kernel JIT compilation during inference.
"""Monitor unexpected kernel JIT compilation during inference.
After server warmup completes, any kernel JIT compilation or autotuning event
indicates a cache miss or unexpected input shape that causes a latency spike.
+1 -2
View File
@@ -142,8 +142,7 @@ def json_reduce_leaves(
initial: _U = ..., # type: ignore[assignment]
/,
) -> _T | _U:
"""
Apply a function of two arguments cumulatively to each leaf in a
"""Apply a function of two arguments cumulatively to each leaf in a
nested JSON structure, from left to right, so as to reduce the
sequence to a single value.
"""
+1 -1
View File
@@ -13,7 +13,7 @@ def cdiv(a: int, b: int) -> int:
def next_power_of_2(n: int) -> int:
"""The next power of 2 (inclusive)"""
"""The next power of 2 (inclusive)."""
return 1 if n < 1 else 1 << (n - 1).bit_length()
+2 -2
View File
@@ -54,6 +54,7 @@ def release_device_memory_under_pressure(device: torch.device) -> bool:
Returns:
True if memory was released.
"""
if device.type != "cuda" or not current_platform.is_integrated_gpu(device.index):
return False
@@ -230,8 +231,7 @@ def memory_profiling(
baseline_snapshot: MemorySnapshot,
weights_memory: int = 0,
) -> Generator[MemoryProfilingResult, None, None]:
"""
Memory profiling context manager.
"""Memory profiling context manager.
baseline_snapshot: the memory snapshot before the current vLLM instance.
weights_memory: memory used by PyTorch when loading the model weights.
+2
View File
@@ -41,6 +41,7 @@ def maybe_execute_in_parallel(
Returns:
Tuple of (fn0_result, fn1_result).
"""
if aux_stream is not None:
from vllm.compilation.breakable_cudagraph import BreakableCUDAGraphCapture
@@ -100,6 +101,7 @@ def execute_in_parallel(
Returns:
Tuple of (default_result, aux_results) where aux_results[i] is the
result of aux_fns[i] (or None when skipped).
"""
aux_results: list[Any]
if aux_streams is None or not enable:
+3 -5
View File
@@ -172,8 +172,7 @@ def _get_reserved_port_range() -> range:
def get_open_port() -> int:
"""
Get an open port for the vLLM process to listen on.
"""Get an open port for the vLLM process to listen on.
An edge case to handle, is when we run data parallel,
we need to avoid ports that are potentially used by
the data parallel master process.
@@ -298,6 +297,7 @@ def make_zmq_path(scheme: str, host: str, port: int | None = None) -> str:
Returns:
A properly formatted ZMQ path string.
"""
if port is None:
return f"{scheme}://{host}"
@@ -317,7 +317,6 @@ def make_zmq_socket(
router_handover: bool = False,
) -> zmq.Socket | zmq.asyncio.Socket: # type: ignore[name-defined]
"""Make a ZMQ socket with the proper bind/connect semantics."""
mem = psutil.virtual_memory()
socket = ctx.socket(socket_type)
@@ -377,8 +376,7 @@ def zmq_socket_ctx(
identity: bytes | None = None,
router_handover: bool = False,
) -> Iterator[zmq.Socket]:
"""Context manager for a ZMQ socket"""
"""Context manager for a ZMQ socket."""
ctx = zmq.Context() # type: ignore[attr-defined]
try:
yield make_zmq_socket(
+3 -2
View File
@@ -25,7 +25,7 @@ def print_tensor(tensor_obj, prefix, tensor_list=None):
def process_layer_params(module_obj):
"""Extract the static parameters from LLM and VLM relevant layer types"""
"""Extract the static parameters from LLM and VLM relevant layer types."""
param_info = {}
# Extract parameters for layers commonly used in LLMs and VLMs
if isinstance(module_obj, (torch.nn.Conv1d, torch.nn.Conv2d, torch.nn.Conv3d)):
@@ -186,6 +186,7 @@ def layerwise_nvtx_marker_context(module_name, module_obj, in_tensor=None, kwarg
kwargs=kwargs) as ctx:
ctx.result = module(*args, **kwargs)
return ctx.result
"""
holder = ResultHolder()
@@ -221,9 +222,9 @@ class PytHooks:
a PytHook object. Then call the register_hooks method.
Example:
my_hook = PytHook()
my_hook.register_hooks(my_network_model)
"""
def __init__(self):
+7 -6
View File
@@ -187,18 +187,19 @@ class OMPProcessManager:
def _get_autobind_cpu_ids(
self, cpu_selector: Callable[[list[LogicalCPUInfo]], list[LogicalCPUInfo]]
) -> tuple[list[list[LogicalCPUInfo]], list[LogicalCPUInfo]]:
"""
Return CPU ids to bind based on NUMA nodes, and CPU ids reserved for
"""Return CPU ids to bind based on NUMA nodes, and CPU ids reserved for
other processes.
Currently for rank N, only CPU ids on the N-th node in available NUMA
node list will be selected.
Args:
cpu_selector: a callable object to select CPUs from a CPU list
of a physical core. The input is a LogicalCPUInfo list contains
logical CPUs of a physical CPU, sorted by the LogicalCPUInfo.id.
A selected LogicalCPUInfo list should be returned.
"""
of a physical core. The input is a LogicalCPUInfo list contains
logical CPUs of a physical CPU, sorted by the
LogicalCPUInfo.id. A selected LogicalCPUInfo list should be
returned.
"""
# this memory node list has been sliced for DP offset
allowed_numa_nodes = cr_utils.get_visible_memory_node()
logical_cpu_list = cr_utils.get_allowed_cpu_list()
+2 -1
View File
@@ -29,7 +29,8 @@ def cuda_get_device_properties(
device, names: Sequence[str], init_cuda=False
) -> tuple[Any, ...]:
"""Get specified CUDA device property values without initializing CUDA in
the current process."""
the current process.
"""
if init_cuda or cuda_is_initialized():
props = torch.cuda.get_device_properties(device)
return tuple(getattr(props, name) for name in names)
+4 -11
View File
@@ -6,8 +6,7 @@ _T = TypeVar("_T", bound=type)
class ExtensionManager:
"""
A registry for managing pluggable extension classes.
"""A registry for managing pluggable extension classes.
This class provides a simple mechanism to register and instantiate
extension classes by name. It is commonly used to implement plugin
@@ -26,15 +25,11 @@ class ExtensionManager:
"""
def __init__(self) -> None:
"""
Initialize an empty extension registry.
"""
"""Initialize an empty extension registry."""
self.name2class: dict[str, type] = {}
def register(self, name: str):
"""
Decorator to register a class with the given name.
"""
"""Decorator to register a class with the given name."""
def wrap(cls_to_register: _T) -> _T:
self.name2class[name] = cls_to_register
@@ -43,9 +38,7 @@ class ExtensionManager:
return wrap
def load(self, cls_name: str, *args, **kwargs) -> Any:
"""
Instantiate and return a registered extension class by name.
"""
"""Instantiate and return a registered extension class by name."""
cls = self.name2class.get(cls_name)
assert cls is not None, f"Extension class {cls_name} not found"
return cls(*args, **kwargs)
+1 -2
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Safe deserialization helpers for client-supplied embedding tensors.
"""Safe deserialization helpers for client-supplied embedding tensors.
`check_sparse_tensor_invariants_threadsafe()` is a thread-safe wrapper for
sparse tensor invariant validation. `safe_to_dense()` bounds the memory a
+4 -5
View File
@@ -60,8 +60,7 @@ def set_env_var(key: str, value: str) -> Iterator[None]:
@contextlib.contextmanager
def suppress_stdout():
"""
Suppress stdout from C libraries at the file descriptor level.
"""Suppress stdout from C libraries at the file descriptor level.
Only suppresses stdout, not stderr, to preserve error messages.
Suppression is disabled when VLLM_LOGGING_LEVEL is set to DEBUG.
@@ -70,6 +69,7 @@ def suppress_stdout():
with suppress_stdout():
# C library calls that would normally print to stdout
torch.distributed.new_group(ranks, backend="gloo")
"""
# Don't suppress if logging level is DEBUG
if envs.VLLM_LOGGING_LEVEL == "DEBUG":
@@ -114,7 +114,6 @@ def unique_filepath(fn: Callable[[int], Path]) -> Path:
def _sync_visible_devices_env_vars():
"""Sync HIP/CUDA visibility env vars before spawning (ROCm only)."""
if not current_platform.is_rocm():
return
@@ -254,11 +253,11 @@ def decorate_logs(
def kill_process_tree(pid: int):
"""
Kills all descendant processes of the given pid by sending SIGKILL.
"""Kills all descendant processes of the given pid by sending SIGKILL.
Args:
pid (int): Process ID of the parent process
"""
try:
parent = psutil.Process(pid)
-1
View File
@@ -161,7 +161,6 @@ class TensorSchema:
dynamic_dims: set[str],
) -> None:
"""Validate that the actual tensor shape matches the expected shape."""
if len(actual_shape) != len(expected_shape):
raise ValueError(
f"{field_name} has rank {len(actual_shape)} "
+17 -27
View File
@@ -96,8 +96,7 @@ def is_meta_module(module: torch.nn.Module) -> bool:
def is_strictly_contiguous(t: torch.Tensor) -> bool:
"""
Check if tensor is contiguous AND has no degenerate strides.
"""Check if tensor is contiguous AND has no degenerate strides.
A degenerate stride occurs when a dimension has size 1 but the stride
doesn't match the canonical contiguous layout. This can cause issues
@@ -231,7 +230,8 @@ def _cgroup_cpu_limit() -> float | None:
def available_cpu_count() -> int:
"""CPUs actually usable by this process: scheduling affinity capped by
the cgroup CPU quota (unlike `os.cpu_count()`, which is quota-blind)."""
the cgroup CPU quota (unlike `os.cpu_count()`, which is quota-blind).
"""
if sys.platform != "linux":
return os.cpu_count() or 1
count = len(os.sched_getaffinity(0))
@@ -294,8 +294,7 @@ def set_torch_threads_for_runtime() -> None:
@contextlib.contextmanager
def set_default_torch_num_threads(num_threads: int | None = None):
"""
Sets the default number of threads for PyTorch to the given value.
"""Sets the default number of threads for PyTorch to the given value.
`None` means using the value of the environment variable `OMP_NUM_THREADS`
(or `1` if that is not available).
@@ -363,8 +362,7 @@ def _get_precision_level(dtype: torch.dtype) -> int:
def is_lossless_cast(src_dtype: torch.dtype, tgt_dtype: torch.dtype):
"""
Test whether it is lossless to cast a tensor from
"""Test whether it is lossless to cast a tensor from
`src_dtype` to `tgt_dtype`.
"""
if src_dtype == tgt_dtype:
@@ -395,8 +393,7 @@ def is_lossless_cast(src_dtype: torch.dtype, tgt_dtype: torch.dtype):
def common_broadcastable_dtype(dtypes: Collection[torch.dtype]):
"""
Get the common `dtype` where all of the other `dtypes` can be
"""Get the common `dtype` where all of the other `dtypes` can be
cast to it without losing any information.
"""
return max(
@@ -567,6 +564,7 @@ def nvfp4_split_data_scale(
Returns:
``(data, scale)`` where *data* is uint8 and *scale* is
float8_e4m3fn, both views of the same storage.
"""
num_pages = kv_side.shape[0]
dim_1, dim_2 = kv_side.shape[1], kv_side.shape[2]
@@ -744,8 +742,7 @@ def make_ndarray_with_pad(
*,
max_len: int | None = None,
) -> npt.NDArray:
"""
Make a padded array from 2D inputs.
"""Make a padded array from 2D inputs.
The padding is applied to the end of each inner list until it reaches
`max_len`.
@@ -771,8 +768,7 @@ def make_tensor_with_pad(
device: str | torch.device | None = None,
pin_memory: bool = False,
) -> torch.Tensor:
"""
Make a padded tensor from 2D inputs.
"""Make a padded tensor from 2D inputs.
The padding is applied to the end of each inner list until it reaches
`max_len`.
@@ -806,8 +802,7 @@ class _StreamPlaceholder:
def current_stream() -> torch.cuda.Stream:
"""
replace `torch.cuda.current_stream()` with `vllm.utils.current_stream()`.
"""Replace `torch.cuda.current_stream()` with `vllm.utils.current_stream()`.
it turns out that `torch.cuda.current_stream()` is quite expensive,
as it will construct a new stream object at each call.
here we patch `torch.cuda.set_stream` to keep track of the current stream
@@ -858,9 +853,7 @@ _aux_stream: torch.cuda.Stream | None = None
def aux_stream() -> torch.cuda.Stream | None:
"""
Ensures aux_stream is initialized only once
"""
"""Ensures aux_stream is initialized only once."""
global _aux_stream
from vllm.platforms import current_platform
@@ -872,8 +865,7 @@ def aux_stream() -> torch.cuda.Stream | None:
def weak_ref_tensor(tensor: Any) -> Any:
"""
Create a weak reference to a tensor.
"""Create a weak reference to a tensor.
The new tensor will share the same data as the original tensor,
but will not keep the original tensor alive.
This ignores 0-size tensors as those don't allocate any memory.
@@ -890,8 +882,7 @@ def weak_ref_tensors(
| tuple[torch.Tensor]
| IntermediateTensors,
) -> torch.Tensor | list[Any] | tuple[Any] | Any:
"""
Convenience function to create weak references to tensors,
"""Convenience function to create weak references to tensors,
for single tensor, list of tensors or tuple of tensors.
"""
if isinstance(tensors, torch.Tensor):
@@ -913,9 +904,7 @@ def weak_ref_tensors(
def get_accelerator_view_from_cpu_tensor(cpu_tensor: torch.Tensor) -> torch.Tensor:
"""
Get an accelerator view of a CPU tensor using Unified Virtual Addressing (UVA).
"""
"""Get an accelerator view of a CPU tensor via Unified Virtual Addressing."""
from vllm.platforms import current_platform
if current_platform.is_xpu():
@@ -951,6 +940,7 @@ def is_torch_equal_or_newer(target: str) -> bool:
Returns:
Whether the condition meets.
"""
try:
return _is_torch_equal_or_newer(str(torch.__version__), target)
@@ -979,6 +969,7 @@ def is_torch_equal(target: str) -> bool:
Returns:
Whether the condition meets.
"""
try:
return _is_torch_equal(target)
@@ -1064,8 +1055,7 @@ def direct_register_custom_op(
dispatch_key: str | None = None,
tags: tuple[torch.Tag, ...] = (),
):
"""
`torch.library.custom_op` can have significant overhead because it
"""`torch.library.custom_op` can have significant overhead because it
needs to consider complicated dispatching logic. This function
directly registers a custom op and dispatches it to the CUDA backend.
See https://gist.github.com/youkaichao/ecbea9ec9fc79a45d2adce1784d7a9a5