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
@@ -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]
|
||||
|
||||
Reference in New Issue
Block a user