From 99cf403c9e84272b80cf36e7cca953d83be4b1d5 Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Wed, 21 Oct 2026 15:18:34 -0700 Subject: [PATCH] [Fix] dsv4: Write idle SWA stores to an idle request-window layout (#45586) --- .../layers/attention/deepseek_v4_backend.py | 9 +++++++-- .../srt/mem_cache/dsv41_request_window.py | 19 +++++++++++++++++++ 2 files changed, 26 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py index 6220301e9b..e76753aba0 100644 --- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py +++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py @@ -3948,10 +3948,15 @@ class DeepseekV4AttnBackend( ) def get_swa_out_cache_loc(self, forward_batch: ForwardBatch) -> torch.Tensor: - """Idle always re-translates at store time: its metadata may be stale, and - translating the zero-padded out_cache_loc writes to the dummy slot.""" + """Idle never trusts its metadata, which may be stale or unset: it writes + the request window's idle layout, or translates the zero-padded + out_cache_loc to the dummy slot.""" metadata = self.forward_metadata if self.token_to_kv_pool.request_window is not None: + if forward_batch.forward_mode.is_idle(): + return self.token_to_kv_pool.request_window.activate_idle( + forward_batch.out_cache_loc.shape[0] + ) layout = metadata.core_attn_metadata.request_window_layout self.token_to_kv_pool.request_window.activate(layout) return layout.write_loc diff --git a/python/sglang/srt/mem_cache/dsv41_request_window.py b/python/sglang/srt/mem_cache/dsv41_request_window.py index fd1c141702..9f8e7ed39a 100644 --- a/python/sglang/srt/mem_cache/dsv41_request_window.py +++ b/python/sglang/srt/mem_cache/dsv41_request_window.py @@ -222,6 +222,7 @@ class RequestWindow: self._ensure_workspace(workspace_rows) self.layout = None self.prepared = None + self._idle_layout = None def _ensure_workspace(self, rows: int) -> None: if self.workspace is not None and self.workspace.size >= rows: @@ -262,6 +263,24 @@ class RequestWindow: f"for a layout of {layout.size}" ) + def activate_idle(self, num_rows: int) -> torch.Tensor: + """Activate a layout for an idle forward, whose attention metadata is + stale or unset. Its rows write the workspace past an empty history, and + an idle forward never commits them.""" + layout = self._idle_layout + if layout is None or layout.pos.numel() != num_rows: + zeros = torch.zeros(num_rows, dtype=torch.int64, device=self.tags.device) + layout = window_layout( + zeros, + zeros, + zero_row=self.zero_row, + capacity=self.capacity, + num_groups=1, + ) + self._idle_layout = layout + self.activate(layout) + return layout.write_loc + def initialize_dummy_history(self): layout = self.layout self.tags.fill_(-1)