[Bugfix][MRV2] Initialize DCP metadata after batch partitioning (#55212)

Signed-off-by: Raphael Rialland <raphael.rialland@mistral.ai>
Signed-off-by: Ziyang Zhang <hafuhafu@qq.com>
Signed-off-by: Sy03 <1370724210@qq.com>
Co-authored-by: Raphael Rialland <raphael.rialland@mistral.ai>
Co-authored-by: Ziyang Zhang <hafuhafu@qq.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Sy03
2026-09-10 06:50:09 +00:00
committed by GitHub
co-authored by Raphael Rialland Ziyang Zhang Codex mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 912d2e79fe
commit 83990f5bcc
12 changed files with 460 additions and 77 deletions
+7 -4
View File
@@ -5,17 +5,19 @@ import torch
from vllm.triton_utils import tl, triton
def prepare_dcp_local_seq_lens(
def maybe_prepare_dcp_local_seq_lens(
dcp_local_seq_lens: torch.Tensor,
seq_lens: torch.Tensor,
num_reqs: int,
dcp_size: int,
dcp_rank: int,
cp_interleave: int,
) -> None:
"""Populate the persistent DCP local seq_lens buffer (CUDA graph safe)."""
*,
num_reqs_padded: int | None = None,
) -> torch.Tensor | None:
"""Populate caller-owned storage and return its padded view, or None without DCP."""
if dcp_size == 1:
return
return None
max_num_reqs = dcp_local_seq_lens.shape[0]
BLOCK_SIZE = 128
@@ -30,6 +32,7 @@ def prepare_dcp_local_seq_lens(
max_num_reqs,
BLOCK_SIZE,
)
return dcp_local_seq_lens[:num_reqs_padded]
@triton.jit