From 7f1edbba80a5f0b691bc6bf216facc01f22ba2e2 Mon Sep 17 00:00:00 2001 From: "minke.yu" Date: Sat, 10 Oct 2026 01:20:15 +0800 Subject: [PATCH] =?UTF-8?q?cherry-pick=20=E4=B8=8A=E6=B8=B8=20#40429?= =?UTF-8?q?=EF=BC=9AFlashMLA=20sched=20meta=20=E8=B6=85=2048KiB=20smem=20?= =?UTF-8?q?=E6=97=B6=E8=B7=B3=E8=BF=87=E9=A2=84=E8=AE=A1=E7=AE=97=EF=BC=88?= =?UTF-8?q?encoder/decoder=20replay=20=E7=9A=84=E5=A4=A7=20batch=20replay?= =?UTF-8?q?=20pass=20=E4=BC=9A=E8=A7=A6=E5=8F=91=20smem=20=E6=BA=A2?= =?UTF-8?q?=E5=87=BA=E5=B4=A9=E6=BA=83=EF=BC=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- python/sglang/srt/layers/attention/deepseek_v4_backend.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py index 5929ca396d..e7931dd90b 100644 --- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py +++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py @@ -206,6 +206,7 @@ def _create_flashmla_metadata(): # only for that shape and go stale silently if FlashMLA retunes it. _FLASHMLA_SCHED_BLOCK_SIZE_N = 64 _FLASHMLA_SCHED_FIXED_OVERHEAD = 5 +_FLASHMLA_SCHED_MAX_SHARED_BYTES = 48 << 10 @functools.lru_cache(maxsize=None) @@ -242,6 +243,10 @@ def _maybe_precompute_flashmla_sched_meta( b, s_q = q.shape[0], q.shape[1] num_sm_parts = max(_num_sms(q.device.index) // s_q, 1) + # Skip when shared memory exceeds the limit. + smem_bytes = 4 * (b * 5 + 1 + num_sm_parts * 8) + if smem_bytes > _FLASHMLA_SCHED_MAX_SHARED_BYTES: + return meta = torch.empty((num_sm_parts, META_INTS), dtype=torch.int32, device=q.device) num_splits = torch.empty((b + 1,), dtype=torch.int32, device=q.device) flashmla_sched_meta(