diff --git a/python/sglang/srt/arg_groups/deepseek_v4_hook.py b/python/sglang/srt/arg_groups/deepseek_v4_hook.py index e9e04ec9b2..74a6837ed9 100644 --- a/python/sglang/srt/arg_groups/deepseek_v4_hook.py +++ b/python/sglang/srt/arg_groups/deepseek_v4_hook.py @@ -238,15 +238,19 @@ def validate_deepseek_v41_encoder_replay_budget(server_args: ServerArgs) -> None cfg = resolving_view(server_args) if not cfg.enable_encoder_swa_bounded_replay: return + # A hit's first chunk also runs up to 128 replay rows and floors its extend to + # whole pages, so a smaller chunk leaves the hit no page and it never admits. + min_chunk = 128 + (cfg.page_size or 1) if ( cfg.max_running_requests is None or cfg.max_running_requests <= 0 or not cfg.chunked_prefill_size - or cfg.chunked_prefill_size < 128 + or cfg.chunked_prefill_size < min_chunk ): raise ValueError( "encoder SWA replay requires --max-running-requests > 0 and at least " - "128 chunked prefill tokens per attention DP rank" + f"{min_chunk} chunked prefill tokens per attention DP rank (128 replay " + "tokens plus one page)" ) diff --git a/python/sglang/srt/arg_groups/model_hook.py b/python/sglang/srt/arg_groups/model_hook.py index 15babcaf42..eeb90fceed 100644 --- a/python/sglang/srt/arg_groups/model_hook.py +++ b/python/sglang/srt/arg_groups/model_hook.py @@ -1026,6 +1026,8 @@ def handle_model_capability_adjustments(server_args: Any): ) validate_deepseek_v41_encoder_replay_budget(server_args) + elif cfg.enable_encoder_swa_bounded_replay: + raise ValueError("--enable-encoder-swa-bounded-replay requires DeepSeek-V4.1") def handle_mamba_radix_cache(server_args: Any, hf_config: Any): diff --git a/test/registered/unit/managers/test_prefill_adder.py b/test/registered/unit/managers/test_prefill_adder.py index 4a0f9e1afd..ae828ee284 100644 --- a/test/registered/unit/managers/test_prefill_adder.py +++ b/test/registered/unit/managers/test_prefill_adder.py @@ -389,6 +389,7 @@ class TestPrefillAdder(CustomTestCase): (0, 1024, 1024), (64, 1024, 960), (4096, 1024, 896), + (4096, 192, 64), # the startup floor: replay rows plus one page (4096, 128, None), ): with self.subTest(prefix_len=prefix_len, rem_chunk=rem_chunk): diff --git a/test/registered/unit/model_executor/test_encoder_swa_replay.py b/test/registered/unit/model_executor/test_encoder_swa_replay.py index ce084f28c0..f61698570f 100644 --- a/test/registered/unit/model_executor/test_encoder_swa_replay.py +++ b/test/registered/unit/model_executor/test_encoder_swa_replay.py @@ -614,6 +614,23 @@ class _View(SimpleNamespace): raise _PastEncoderReplayChecks(name) +class TestEncoderReplayChunkFloor(CustomTestCase): + def _validate(self, chunked_prefill_size): + view = _View( + enable_encoder_swa_bounded_replay=True, + max_running_requests=8, + chunked_prefill_size=chunked_prefill_size, + page_size=256, + ) + with patch.object(deepseek_v4_hook, "resolving_view", return_value=view): + deepseek_v4_hook.validate_deepseek_v41_encoder_replay_budget(object()) + + def test_chunk_holds_the_replay_rows_and_a_page(self): + self._validate(384) + with self.assertRaisesRegex(ValueError, "at least 384 chunked prefill"): + self._validate(383) + + class TestEncoderReplayDpStartupCheck(CustomTestCase): """ROCm replays hits in a forward of their own, so it must refuse attention DP."""