[DSV4] Remove per-step seqlen D2H from speculative to make overlap scheduler work (#30365)

Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>

python/sglang/test/kits/attention_unittest/attention_methods/dsv4_attention.py

@@ -1490,6 +1490,7 @@ def run_dsv4_target_verify_attention_case(
     fixture = build_dsv4_attention_fixture(testcase, case, dtype=dtype, device=device)
     runner = fixture.runner
     max_context_len = runner.req_to_token_pool.req_to_token.shape[1]
+    testcase.assertEqual(fixture.backend.max_context_len, max_context_len)
 
     _populate_swa_kv_cache(fixture, max_context_len=max_context_len, device=device)
     if case.compress_ratio in (4, 128):
@@ -1530,6 +1531,7 @@ def run_dsv4_draft_extend_attention_case(
     *,
     dtype: torch.dtype = torch.bfloat16,
     device: str = "cuda",
+    force_gpu_only_seq_lens: bool = False,
 ) -> None:
     """Math-faithful EAGLE `DRAFT_EXTEND` test for DSV4.
 
@@ -1567,6 +1569,10 @@ def run_dsv4_draft_extend_attention_case(
         fixture.forward_batch,
         device=device,
     )
+    if force_gpu_only_seq_lens:
+        fixture.forward_batch.seq_lens_cpu = None
+        fixture.forward_batch.seq_lens_sum = None
+        fixture.forward_batch.spec_info.seq_lens_cpu = None
 
     q_input, _ = fixture.actual_module.project(fixture.input_hidden)
     with torch.no_grad(), forward_context(ForwardContext(attn_backend=fixture.backend)):

python/sglang/test/kits/attention_unittest/runner_modes/speculative_draft_runner.py

@@ -1484,6 +1484,7 @@ def run_dsv4_eagle_draft_cuda_graph_runner_case(
     vocab_size: int = 64,
     dtype: torch.dtype = torch.bfloat16,
     device: str = "cuda",
+    force_gpu_only_seq_lens: bool = False,
 ):
     settings = EagleDraftRunnerSettings(
         topk=topk,
@@ -1498,12 +1499,20 @@ def run_dsv4_eagle_draft_cuda_graph_runner_case(
         atol=DSV4_ATOL,
         rtol=DSV4_RTOL,
     )
+
+    def _make_forward_batch(case, draft_inputs, settings):
+        batch = _make_dsv4_eagle_draft_forward_batch(case, draft_inputs, settings)
+        if force_gpu_only_seq_lens:
+            batch.seq_lens_cpu = None
+            batch.seq_lens_sum = None
+        return batch
+
     adapter = EagleDraftCudaGraphRunnerAdapter(
         build_fixture=build_dsv4_attention_fixture,
         make_model_forward=_make_dsv4_model_forward,
         make_draft_inputs=_make_dsv4_draft_inputs,
         prepare_replay_state=_prepare_dsv4_draft_replay_state,
-        make_forward_batch=_make_dsv4_eagle_draft_forward_batch,
+        make_forward_batch=_make_forward_batch,
         check_case=_check_dsv4_draft_cache_layout,
         init_eager_metadata=_init_dsv4_eager_metadata,
     )

python/sglang/test/kits/attention_unittest/runner_modes/speculative_target_verify_runner.py

@@ -890,6 +890,7 @@ def run_dsv4_eagle_verify_cuda_graph_case(
     dtype: torch.dtype = torch.bfloat16,
     device: str = "cuda",
     cuda_graph_capture_batch_size: int = 2,
+    force_gpu_only_seq_lens: bool = False,
 ):
     """DSV4 EAGLE target_verify CUDA-graph capture/replay. Chain only —
     `DeepseekV4AttnBackend.__init__` asserts `self.topk in [0, 1]` at
@@ -936,6 +937,11 @@ def run_dsv4_eagle_verify_cuda_graph_case(
         batch.spec_info = _make_eagle_verify_input(
             spec_case, batch, topk=topk, device=device
         )
+        if force_gpu_only_seq_lens:
+            batch.seq_lens_cpu = None
+            batch.seq_lens_sum = None
+            batch.spec_info.seq_lens_cpu = None
+            batch.spec_info.seq_lens_sum = None
 
     def _make_capture_case(base, name, capture_prefix_len: int, bs: int):
         # Capture uses uniform prefixes per request; each request still