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
@@ -62,7 +62,8 @@ 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
|
||||
@@ -90,7 +91,8 @@ 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,
|
||||
@@ -408,6 +410,8 @@ class CudaGraphManager:
|
||||
it is invoked once with warmup=True and again with warmup=False
|
||||
because attention backends may mutate or lazily initialize
|
||||
metadata during warmup.
|
||||
progress_bar_desc: Description shown on the capture progress bar.
|
||||
|
||||
"""
|
||||
with graph_capture(device=self.device), ExitStack() as stack:
|
||||
if self.ubatch_runner is not None:
|
||||
@@ -503,7 +507,6 @@ class CudaGraphManager:
|
||||
num_ubatches: int = 1,
|
||||
) -> BatchExecutionDescriptor:
|
||||
"""Find matching cudagraph descriptor from priority-ordered candidates."""
|
||||
|
||||
effective_loras = self._resolve_effective_loras(num_active_loras)
|
||||
key = (num_tokens, effective_loras)
|
||||
if self._graphs_captured and num_tokens > 0 and key in self._candidates:
|
||||
@@ -952,7 +955,8 @@ 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]
|
||||
@@ -989,7 +993,8 @@ 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
|
||||
|
||||
@@ -55,8 +55,7 @@ def sync_cudagraph_and_dp_padding(
|
||||
allow_ubatching: bool = False,
|
||||
uniform_decode: bool = False,
|
||||
) -> tuple[BatchExecutionDescriptor, DPSyncState | None]:
|
||||
"""
|
||||
Coordinates the batch descriptor and DP padding across all ranks.
|
||||
"""Coordinates the batch descriptor and DP padding across all ranks.
|
||||
|
||||
`parallel_config` is only needed to decide whether to microbatch, so callers
|
||||
that never do (`allow_ubatching=False`) can leave it out.
|
||||
@@ -248,6 +247,12 @@ def dispatch_cg_and_sync_dp(
|
||||
need_eager: Force `CUDAGraphMode.NONE` instead of dispatching.
|
||||
num_active_loras: Active LoRA count for this rank. Does not need
|
||||
cross-rank agreement; it never changes a bucket's token count.
|
||||
parallel_config: Only needed to decide whether to microbatch, so
|
||||
callers that never do (`allow_ubatching=False`) can leave it out.
|
||||
allow_ubatching: Whether this rank's batch may be split into
|
||||
microbatches. Agreed across ranks before it takes effect.
|
||||
uniform_decode: Whether this rank's batch is a uniform decode, used
|
||||
to pick the microbatching split.
|
||||
dp_sync: Agreement from a prior dispatch over this same batch, to reuse.
|
||||
Must come from a batch with this same padded `num_tokens` and the
|
||||
same `uniform_token_count`; `num_reqs` may differ, as neither
|
||||
@@ -257,6 +262,7 @@ def dispatch_cg_and_sync_dp(
|
||||
Returns:
|
||||
(batch_desc, sync), where `sync` is this batch's agreement for a later
|
||||
dispatch to reuse. It is None when `dp_size` is 1 or no rank has work.
|
||||
|
||||
"""
|
||||
reuse_eager = dp_sync is not None and dp_sync.eager
|
||||
|
||||
|
||||
@@ -21,8 +21,7 @@ def get_lora_capture_cases(
|
||||
lora_config: "LoRAConfig | None",
|
||||
compilation_config: "CompilationConfig",
|
||||
) -> list[int]:
|
||||
"""
|
||||
Return num_active_loras values for cudagraph capture.
|
||||
"""Return num_active_loras values for cudagraph capture.
|
||||
|
||||
When cudagraph_specialize_lora=True: powers of 2 up to max_loras, plus
|
||||
max_loras+1. When False: [0, max_loras+1]. When LoRA disabled: [0].
|
||||
|
||||
@@ -24,8 +24,7 @@ class EncoderCache:
|
||||
self.mm_features.pop(req_id, None)
|
||||
|
||||
def reset_mm_cache(self) -> None:
|
||||
"""
|
||||
Clear the multi-modal cache that was used during profiling,
|
||||
"""Clear the multi-modal cache that was used during profiling,
|
||||
but no longer needed during inference.
|
||||
"""
|
||||
# NOTE: v2 encoder cache profiling skips the multi-modal cache
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
NOTE: Coding style guide for this file:
|
||||
"""NOTE: Coding style guide for this file:
|
||||
This model runner is shared by all models: text and multimodal, generative
|
||||
and embedding, public and private. As a result, this file must only contain
|
||||
code that is common to every model. Model-specific behavior belongs in the
|
||||
@@ -2229,7 +2228,8 @@ 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
|
||||
|
||||
@@ -144,7 +144,8 @@ 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 +38,8 @@ 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).
|
||||
@@ -179,7 +179,8 @@ 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:
|
||||
|
||||
@@ -76,7 +76,8 @@ 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
|
||||
|
||||
@@ -107,7 +107,8 @@ 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)
|
||||
|
||||
@@ -163,6 +163,7 @@ def repeated_context_mask(
|
||||
include_prompt: Search the prompt as well as the generated tokens.
|
||||
skip_partial_context: Mark contexts containing start padding so they use
|
||||
ordinary sampling.
|
||||
|
||||
"""
|
||||
if max_history is not None and max_history < 1:
|
||||
raise ValueError("max_history must be positive or None")
|
||||
|
||||
@@ -88,9 +88,14 @@ def build_cost_tables_from_curves(
|
||||
"""Build cost tables: graph-padded below the capture limit, smooth above.
|
||||
|
||||
Args:
|
||||
draft_curve: (size, cost) samples for the draft model.
|
||||
verify_curve: (size, cost) samples for the verify model.
|
||||
max_num_reqs: Largest request count to build a table for.
|
||||
max_batch_tokens: Largest token count to build a table for.
|
||||
cudagraph_limit: Largest cudagraph-captured size. At or below it,
|
||||
execution pads up to the next captured size, so cost is a step
|
||||
function. Above it there is no padding, so cost is continuous.
|
||||
|
||||
"""
|
||||
|
||||
def build_table(limit: int, curve: list[tuple[int, float]]) -> np.ndarray:
|
||||
@@ -185,7 +190,8 @@ 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
|
||||
@@ -255,7 +261,8 @@ 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():
|
||||
|
||||
@@ -73,8 +73,7 @@ class AutoRegressiveSpeculator(DraftModelSpeculator):
|
||||
|
||||
@property
|
||||
def advance_draft_positions(self) -> bool:
|
||||
"""
|
||||
Whether to increment positions and seq_lens between draft steps.
|
||||
"""Whether to increment positions and seq_lens between draft steps.
|
||||
|
||||
True for Eagle/standard MTP (each step produces new KV).
|
||||
False for Gemma4 MTP (Q-only, shares target KV, constant positions).
|
||||
|
||||
@@ -72,7 +72,8 @@ 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,
|
||||
|
||||
@@ -13,7 +13,6 @@ from vllm.v1.worker.gpu.spec_decode.utils import get_pp_safe_draft_load_config
|
||||
|
||||
def _should_share(eagle: nn.Module, flag: str, draft, target) -> bool:
|
||||
"""Share when the draft has no own copy, or its copy matches the target."""
|
||||
|
||||
if not getattr(eagle, flag, False) or draft is None:
|
||||
return True
|
||||
if target is None:
|
||||
@@ -29,7 +28,8 @@ 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
|
||||
)
|
||||
|
||||
@@ -268,7 +268,8 @@ 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(
|
||||
|
||||
Reference in New Issue
Block a user