Add sp_min_token_num=0 to E2E correctness tests for SP and AsyncTP

Co-authored-by: ProExpertProg <11367180+ProExpertProg@users.noreply.github.com>
Agent-Logs-Url: https://github.com/vllm-project/vllm/sessions/f8e4f0ad-7a0f-4253-8a28-7a0100eacae5

tests/compile/correctness_e2e/test_async_tp.py

@@ -60,7 +60,12 @@ def test_async_tp_pass_correctness(
         "mode": CompilationMode.VLLM_COMPILE,
         "compile_sizes": [2, 4, 8],
         "splitting_ops": [],
-        "pass_config": {"fuse_gemm_comms": async_tp_enabled},
+        "pass_config": {
+            "fuse_gemm_comms": async_tp_enabled,
+            # Override threshold to ensure SP (required by async TP) is
+            # applied even for small models.
+            "sp_min_token_num": 0,
+        },
     }
 
     async_tp_args = [

tests/compile/correctness_e2e/test_sequence_parallel.py

@@ -258,6 +258,9 @@ def _compare_sp(
             "fuse_norm_quant": fuse_norm_quant,
             "fuse_act_quant": fuse_act_quant,
             "eliminate_noops": True,
+            # Override threshold to ensure SP is applied even for small
+            # models (e.g. tiny-random-LlamaForCausalLM).
+            "sp_min_token_num": 0,
         },
         "use_inductor_graph_partition": use_inductor_graph_partition,
     }