diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py index 43e586266b..352e819bb6 100644 --- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py +++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py @@ -2420,7 +2420,7 @@ class DeepseekV4AttnBackend( req_to_token=self.req_to_token, full_to_swa=self.token_to_kv_pool.full_to_swa_index_mapping, swa_window_size=SWA_WINDOW, - swa_page_size=self.token_to_kv_pool.swa_kv_pool.page_size, + swa_page_size=self.token_to_kv_pool.get_swa_key_page_size(), num_qo_tokens=num_qo_tokens, max_seq_len=max(seq_lens_cpu_list), total_swa=total_swa, @@ -3340,7 +3340,7 @@ class DeepseekV4AttnBackend( compress_ratio ) - swa_kv_page_size = token_to_kv_pool.swa_kv_pool.page_size + swa_kv_page_size = token_to_kv_pool.get_swa_key_page_size() assert swa_k_cache.ndim == 2 # The kernel detects each cache's format from the last dim of this # view: 584 (V4), 528 (V4.1 fp8) or 288 (V4.1 fp4, extra cache only). diff --git a/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py b/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py index 75cb02c0a2..f535c75f63 100644 --- a/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py +++ b/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py @@ -1836,6 +1836,11 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool): # swa_kv_pool is None under the request window and unified_kv. return self.kv_layout + def get_swa_key_page_size(self) -> int: + if self.request_window is not None: + return self.request_window.page_size + return self.swa_kv_pool.page_size + def get_swa_key_bytes_per_token(self) -> int: """Last dim of the ``(pages, page_size, 1, bytes)`` view the attention kernel detects the SWA cache's format from.""" diff --git a/test/registered/unit/mem_cache/test_dsv4_compressed_pools.py b/test/registered/unit/mem_cache/test_dsv4_compressed_pools.py index a83bb4baac..413be2c431 100644 --- a/test/registered/unit/mem_cache/test_dsv4_compressed_pools.py +++ b/test/registered/unit/mem_cache/test_dsv4_compressed_pools.py @@ -24,6 +24,21 @@ register_cpu_ci(est_time=5, suite="base-a-test-cpu") class TestDSV4CompressedPools(CustomTestCase): + def test_swa_key_page_size_uses_physical_paged_size(self): + pool = DeepSeekV4TokenToKVPool.__new__(DeepSeekV4TokenToKVPool) + pool.request_window = None + pool.swa_page_size = 256 + for physical_page_size in (256, 64): + with self.subTest(physical_page_size=physical_page_size): + pool.swa_kv_pool = SimpleNamespace(page_size=physical_page_size) + self.assertEqual(pool.get_swa_key_page_size(), physical_page_size) + + def test_swa_key_page_size_without_paged_pool(self): + pool = DeepSeekV4TokenToKVPool.__new__(DeepSeekV4TokenToKVPool) + pool.swa_kv_pool = None + pool.request_window = SimpleNamespace(page_size=256) + self.assertEqual(pool.get_swa_key_page_size(), 256) + def test_physical_kv_pages_cover_reserved_logical_page(self): size = 8192 self.assertEqual(_num_dsv4_physical_kv_pages(size, 256, 256), 33)