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
@@ -437,7 +437,6 @@ class FlashMLASparseMetadataBuilder(
be the full batch or only decodes when prefills use dense MHA. This avoids
the BF16 prefill kernel's head-padding overhead at high TP.
"""
scheduler_metadata, _ = get_mla_metadata()
return FlashMLASparseMetadata.FP8KernelMetadata(
scheduler_metadata=scheduler_metadata,
@@ -848,7 +847,8 @@ class FlashMLASparseImpl(SparseMLACommonImpl[FlashMLASparseMetadata]):
prefill_meta: "FlashMLASparseMetadata.FP8SeparatePrefillDecode.Prefill",
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""All-gather this rank's upconverted KV shard so the chunk's rows attend
the whole context, and map their top-k onto the rank-major result."""
the whole context, and map their top-k onto the rank-major result.
"""
shard_rows = int(chunk.chunk_tot_seqlen)
assert self.gathered_kv_workspace is not None
gathered_kv = self.gathered_kv_workspace[: self.dcp_world_size * shard_rows]