[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
@@ -76,8 +76,7 @@ class EncoderCudaGraphManager:
|
||||
model: SupportsEncoderCudaGraph,
|
||||
):
|
||||
"""Initialize CUDA graph manager with provided token budgets
|
||||
and max batch size.
|
||||
"""
|
||||
and max batch size."""
|
||||
self.vllm_config = vllm_config
|
||||
self.device = device
|
||||
self.dtype = dtype
|
||||
|
||||
@@ -62,8 +62,7 @@ class AttentionState(NamedTuple):
|
||||
@dataclass(frozen=True)
|
||||
class BatchExecutionDescriptor:
|
||||
"""Describes the shape of the batch and CG mode to run; this is used to make shape
|
||||
matches between the capture and runtime.
|
||||
"""
|
||||
matches between the capture and runtime."""
|
||||
|
||||
cg_mode: CUDAGraphMode
|
||||
num_tokens: int
|
||||
@@ -91,8 +90,7 @@ def make_cudagraph_stats(
|
||||
class CreateForwardFn(Protocol):
|
||||
"""Factory that prepares inputs (OUTSIDE the graph) and returns a
|
||||
forward_fn. Called with warmup=True for the warmup pass and warmup=False
|
||||
for the captured pass.
|
||||
"""
|
||||
for the captured pass."""
|
||||
|
||||
def __call__(
|
||||
self,
|
||||
@@ -955,8 +953,7 @@ def profile_cudagraph_memory(runner: "GPUModelRunner") -> int:
|
||||
def _extrapolate_full_graph_memory(mem_samples: list[int], total_graphs: int) -> int:
|
||||
"""Extrapolate the total FULL capture cost from samples of the largest
|
||||
graphs. The first capture allocates the pool baseline; later graphs mostly
|
||||
reuse it, so the second sample is taken as the per-graph cost.
|
||||
"""
|
||||
reuse it, so the second sample is taken as the per-graph cost."""
|
||||
if not mem_samples:
|
||||
return 0
|
||||
first_capture = mem_samples[0]
|
||||
@@ -993,8 +990,7 @@ def _init_minimal_kv_cache_for_profiling(runner: "GPUModelRunner") -> None:
|
||||
|
||||
def _teardown_profiling_state(runner: "GPUModelRunner") -> None:
|
||||
"""Release the profiling KV cache and captured graphs while keeping model
|
||||
weights, so the real ``initialize_kv_cache`` starts from a clean slate.
|
||||
"""
|
||||
weights, so the real ``initialize_kv_cache`` starts from a clean slate."""
|
||||
torch.accelerator.synchronize()
|
||||
if hasattr(runner.model_state, "_mamba_ctx"):
|
||||
runner.model_state._mamba_ctx = None
|
||||
|
||||
@@ -2228,8 +2228,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
|
||||
def shutdown(self) -> None:
|
||||
"""Release GPU tensors (model weights, KV caches, workspace) so that
|
||||
memory is reclaimable when running in the same process.
|
||||
"""
|
||||
memory is reclaimable when running in the same process."""
|
||||
torch.accelerator.synchronize()
|
||||
self.cudagraph_manager = None
|
||||
self.fast_prefill = None
|
||||
|
||||
@@ -144,8 +144,7 @@ class ModelState(ABC):
|
||||
) -> None:
|
||||
"""Hook run on real batches before the forward pass (after block tables
|
||||
are gathered). Used by mamba "align" prefix caching to pre-copy state
|
||||
across block boundaries. No-op by default.
|
||||
"""
|
||||
across block boundaries. No-op by default."""
|
||||
return None
|
||||
|
||||
def postprocess_state(
|
||||
|
||||
@@ -38,8 +38,7 @@ def compute_need_sampled_mask(input_batch: InputBatch) -> np.ndarray | None:
|
||||
"""Return a bool array of shape `[input_batch.num_reqs]` marking requests
|
||||
that produce a sampled token this step, and therefore must have that token
|
||||
(and the draft block proposed from it) propagated to the earlier PP stages.
|
||||
Returns None if no request in the batch produces a sample.
|
||||
"""
|
||||
Returns None if no request in the batch produces a sample."""
|
||||
old_computed = input_batch.num_computed_tokens_np
|
||||
prefill_len = input_batch.prefill_len_np
|
||||
# Exclude non-final prefill chunks (they don't produce a sample).
|
||||
@@ -179,8 +178,7 @@ class PPHandler:
|
||||
|
||||
def receive(self, input_batch: InputBatch) -> bool:
|
||||
"""Returns True iff sampled tokens need to be gathered from *all*
|
||||
requests in the batch.
|
||||
"""
|
||||
requests in the batch."""
|
||||
assert not self.is_last_rank
|
||||
need_sampled_mask = compute_need_sampled_mask(input_batch)
|
||||
if need_sampled_mask is None:
|
||||
|
||||
@@ -76,8 +76,7 @@ MAX_COMPACT_SUPPORT = 2048
|
||||
|
||||
class SamplingMaskTensors(NamedTuple):
|
||||
"""Device-side masks pending async D2H: compact ids, plus the bitmask as
|
||||
the exact fallback for rows wider than ``max_num_kept``.
|
||||
"""
|
||||
the exact fallback for rows wider than ``max_num_kept``."""
|
||||
|
||||
# [num_requests, max_num_kept]
|
||||
token_ids: torch.Tensor
|
||||
|
||||
@@ -107,8 +107,7 @@ class Sampler:
|
||||
self, idx_mapping_np: np.ndarray, include_token_ids: bool = True
|
||||
) -> tuple[int, int] | None:
|
||||
"""(num_logprobs, max_per_req_token_ids) for the given requests, or
|
||||
None when none of them want logprobs.
|
||||
"""
|
||||
None when none of them want logprobs."""
|
||||
max_num_logprobs = self.sampling_states.max_num_logprobs(idx_mapping_np)
|
||||
max_token_ids = (
|
||||
self.logprob_token_ids_state.max_num_token_ids(idx_mapping_np)
|
||||
|
||||
@@ -190,8 +190,7 @@ class AdaptiveVerificationManager:
|
||||
"""Dummy-run kwargs whose step timings seed the cost tables.
|
||||
|
||||
Run these inside StepTimingCollector.collect(), then hand the block's
|
||||
timings to set_initial_cost_curves.
|
||||
"""
|
||||
timings to set_initial_cost_curves."""
|
||||
max_num_tokens = self.req_states.max_num_batched_tokens
|
||||
size = self._cudagraph_limit = capture_sizes[-1] if capture_sizes else 0
|
||||
# Also profile beyond the capture limit: real steps run there
|
||||
@@ -261,8 +260,7 @@ class AdaptiveVerificationManager:
|
||||
input_batch: "InputBatch",
|
||||
) -> None:
|
||||
"""Publish this step's raw confidences for the ranking kernel and start
|
||||
copying them to the CPU, where a later step's budget reads them.
|
||||
"""
|
||||
copying them to the CPU, where a later step's budget reads them."""
|
||||
num_reqs = input_batch.num_reqs
|
||||
ready_idx = self._stale_idx ^ 1
|
||||
with gpu_sync_allowed():
|
||||
|
||||
@@ -72,8 +72,7 @@ def _prepare_dflash_inputs_to_capture(
|
||||
|
||||
class DFlashCudaGraphManager(CudaGraphManager):
|
||||
"""DFlash CudaGraphManager for the parallel-drafting query forward,
|
||||
building its own attention metadata from scratch.
|
||||
"""
|
||||
building its own attention metadata from scratch."""
|
||||
|
||||
def capture(
|
||||
self,
|
||||
|
||||
@@ -28,8 +28,7 @@ def _should_share(eagle: nn.Module, flag: str, draft, target) -> bool:
|
||||
|
||||
def get_target_lm_head(target_model: nn.Module, target_language_model: nn.Module):
|
||||
"""The target's lm_head — from get_language_model() for
|
||||
*ForConditionalGeneration targets, else the top-level module.
|
||||
"""
|
||||
*ForConditionalGeneration targets, else the top-level module."""
|
||||
return getattr(target_language_model, "lm_head", None) or getattr(
|
||||
target_model, "lm_head", None
|
||||
)
|
||||
|
||||
@@ -268,8 +268,7 @@ class DraftModelSpeculator(BaseSpeculator):
|
||||
@property
|
||||
def attn_vllm_config(self) -> VllmConfig:
|
||||
"""Config for the draft's attention metadata builders. Overridden by
|
||||
speculators whose attention mode differs from the target's.
|
||||
"""
|
||||
speculators whose attention mode differs from the target's."""
|
||||
return self.vllm_config
|
||||
|
||||
def set_attn(
|
||||
|
||||
@@ -481,8 +481,7 @@ class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
|
||||
|
||||
class ExecuteModelState(NamedTuple):
|
||||
"""Ephemeral cached state transferred between execute_model() and
|
||||
sample_tokens(), after execute_model() returns None.
|
||||
"""
|
||||
sample_tokens(), after execute_model() returns None."""
|
||||
|
||||
scheduler_output: "SchedulerOutput"
|
||||
logits: torch.Tensor
|
||||
@@ -1072,8 +1071,7 @@ class GPUModelRunner(
|
||||
first scheduled token (column j is position start - 1 - j); -1 where
|
||||
the position is before the prompt or already past it. Generated
|
||||
positions are left to the model: under async scheduling the CPU token
|
||||
table holds placeholders for them.
|
||||
"""
|
||||
table holds placeholders for them."""
|
||||
buf = self.lookback_token_ids
|
||||
assert buf is not None
|
||||
buf.np.fill(-1)
|
||||
@@ -3206,8 +3204,7 @@ class GPUModelRunner(
|
||||
|
||||
def _get_encoder_output_from_cache(self, mm_hash: str) -> torch.Tensor | None:
|
||||
"""Return a cached encoder output for multimodal
|
||||
embedding gather.
|
||||
"""
|
||||
embedding gather."""
|
||||
return self.encoder_cache.get(mm_hash, None)
|
||||
|
||||
def _gather_mm_embeddings(
|
||||
@@ -4179,8 +4176,7 @@ class GPUModelRunner(
|
||||
|
||||
This is true when `discard_request_mask` is set for every scheduled
|
||||
request (e.g., for chunked prefill requests that are not the last
|
||||
prefill chunk).
|
||||
"""
|
||||
prefill chunk)."""
|
||||
num_reqs = self.input_batch.num_reqs
|
||||
return bool(self.discard_request_mask.np[:num_reqs].all())
|
||||
|
||||
@@ -6573,8 +6569,7 @@ class GPUModelRunner(
|
||||
|
||||
def shutdown(self) -> None:
|
||||
"""Release GPU tensors (model weights, KV caches, workspace) so that
|
||||
memory is reclaimable when running in the same process.
|
||||
"""
|
||||
memory is reclaimable when running in the same process."""
|
||||
from vllm.model_executor.layers.rotary_embedding import _ROPE_DICT
|
||||
from vllm.v1.worker.workspace import reset_workspace_manager
|
||||
|
||||
|
||||
@@ -336,8 +336,7 @@ class Worker(WorkerBase):
|
||||
def _scoped_allocator_max_split(self, max_split_size_mb: int):
|
||||
"""Temporarily set max_split_size_mb to reduce allocator fragmentation at the
|
||||
cost of more cudaMalloc calls (negligible in practice). Restores the original
|
||||
value on exit.
|
||||
"""
|
||||
value on exit."""
|
||||
if not current_platform.is_cuda():
|
||||
yield
|
||||
return
|
||||
|
||||
@@ -77,8 +77,7 @@ def compute_plan_fingerprint(
|
||||
def _plan_path(fingerprint: str) -> str:
|
||||
"""Plans are regenerable derived state, so they live under the standard
|
||||
vLLM cache root (like the torch.compile cache) and relocate with
|
||||
``VLLM_CACHE_ROOT`` instead of needing a location knob of their own.
|
||||
"""
|
||||
``VLLM_CACHE_ROOT`` instead of needing a location knob of their own."""
|
||||
# VLLM_CACHE_ROOT is already user-expanded by envs.py.
|
||||
return os.path.join(
|
||||
envs.VLLM_CACHE_ROOT, "startup_plan", f"startup_plan_{fingerprint}.json"
|
||||
@@ -135,8 +134,7 @@ def _applicable_kv_cache_memory_bytes(
|
||||
def maybe_apply_startup_plan(worker: "Worker") -> None:
|
||||
"""If enabled and ``--kv-cache-memory`` was not set explicitly, apply a
|
||||
persisted plan by setting ``worker.cache_config.kv_cache_memory_bytes``.
|
||||
No-op unless ``VLLM_ENABLE_STARTUP_PLAN=1``.
|
||||
"""
|
||||
No-op unless ``VLLM_ENABLE_STARTUP_PLAN=1``."""
|
||||
if (
|
||||
not envs.VLLM_ENABLE_STARTUP_PLAN
|
||||
or worker.cache_config.kv_cache_memory_bytes is not None
|
||||
@@ -169,8 +167,7 @@ def maybe_apply_startup_plan(worker: "Worker") -> None:
|
||||
def maybe_save_startup_plan(worker: "Worker", kv_cache_memory_bytes: int) -> None:
|
||||
"""Atomically persist this boot's profiling result for future boots.
|
||||
No-op unless ``VLLM_ENABLE_STARTUP_PLAN=1``; failures are logged,
|
||||
never raised.
|
||||
"""
|
||||
never raised."""
|
||||
if not envs.VLLM_ENABLE_STARTUP_PLAN:
|
||||
return
|
||||
fingerprint = compute_plan_fingerprint(
|
||||
|
||||
@@ -128,8 +128,7 @@ class WorkerBase:
|
||||
|
||||
def synchronize_device(self) -> None:
|
||||
"""Block until in-flight device work completes; backends outside
|
||||
``torch.accelerator`` must override with their own wait.
|
||||
"""
|
||||
``torch.accelerator`` must override with their own wait."""
|
||||
if torch.accelerator.is_available():
|
||||
torch.accelerator.synchronize()
|
||||
|
||||
|
||||
Reference in New Issue
Block a user