[PCP] Support prefill context parallelism with data parallelism (#57075)

Signed-off-by: QiuChunshuo <qiuchunshuo@huawei.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: QiuChunshuo <qiuchunshuo@huawei.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Lucas Wilkinson
2026-09-23 22:32:36 -04:00
committed by GitHub
co-authored by QiuChunshuo Claude Fable 5.1 Codex Robert Shaw Nick Hill mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 6697a7dd30
commit e6dc16cebd
15 changed files with 268 additions and 26 deletions
+14 -3
View File
@@ -58,8 +58,17 @@ class BatchDescriptor:
def _compute_sp_num_tokens(
num_tokens_across_dp_cpu: torch.Tensor, sequence_parallel_size: int
num_tokens_across_dp_cpu: torch.Tensor,
sequence_parallel_size: int,
pcp_size: int = 1,
use_ep: bool = True,
) -> list[int]:
if pcp_size > 1:
num_tokens_across_dp_cpu = (
num_tokens_across_dp_cpu.repeat_interleave(pcp_size)
if use_ep
else num_tokens_across_dp_cpu * pcp_size
)
sp_tokens = (
num_tokens_across_dp_cpu + sequence_parallel_size - 1
) // sequence_parallel_size
@@ -98,12 +107,14 @@ class DPMetadata:
return DPMetadata(num_tokens_across_dp_cpu)
@contextmanager
def sp_local_sizes(self, sequence_parallel_size: int):
def sp_local_sizes(
self, sequence_parallel_size: int, pcp_size: int = 1, use_ep: bool = False
):
"""Context manager for setting self.local_sizes. Same as self.chunked_sizes
but without any chunking.
"""
self.local_sizes = _compute_sp_num_tokens(
self.num_tokens_across_dp_cpu, sequence_parallel_size
self.num_tokens_across_dp_cpu, sequence_parallel_size, pcp_size, use_ep
)
try:
yield self.local_sizes