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
|
||||
|
||||
Reference in New Issue
Block a user