[Perf][GDN] Internal prefill checkpoints for the FlashInfer GDN prefill backend (#60050)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Jiangyun Zhu
2026-10-14 14:49:46 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent a2fb482732
commit badfcdbf91
12 changed files with 794 additions and 151 deletions
+9
View File
@@ -947,6 +947,9 @@ class Platform:
from math import lcm
from vllm.config.vllm import set_current_vllm_config
from vllm.model_executor.layers.mamba.abstract import (
get_mamba_prefill_checkpoint_alignment,
)
from vllm.model_executor.models import ModelRegistry
from vllm.utils.math_utils import cdiv
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
@@ -1080,6 +1083,12 @@ class Platform:
# For hybrid MLA/Mamba models, make the manager block size a
# multiple of 128 so split kernel blocks keep that invariant.
kernel_block_alignment_size = max(kernel_block_alignment_size, 128)
if cache_config.mamba_cache_mode == "align":
# Internal prefill checkpoints land on block boundaries.
kernel_block_alignment_size = lcm(
kernel_block_alignment_size,
get_mamba_prefill_checkpoint_alignment(vllm_config) or 1,
)
# Use minimum block size that satisfies both backend alignment and
# mamba page size compatibility