[Simulator][Compatibility] Adapt to latest KV cache pool interfaces (#40418)

tools/sglang-simulator/src/sglang_simulator/simulation/sglang/model_runner.py

@@ -176,6 +176,11 @@ class C_KVCacheConfiguratorHook(BaseHook):
                         )
                     except RuntimeError:
                         return 0
+                    except ValueError as error:
+                        # New runtimes reject deliberately undersized SWA probes.
+                        if "cannot hold even one request" not in str(error):
+                            raise
+                        return 0
                     return config.max_total_num_tokens
 
                 lower, upper = 0, 1
@@ -249,6 +254,20 @@ class C_KVCacheConfiguratorHook(BaseHook):
                 if hasattr(token_pool, name):
                     setattr(token_pool, name, value)
 
+            # Some pools derive their physical row widths while the dimensions
+            # above are compacted. Restore that derived geometry before host
+            # pools inspect the device-pool rows for HiCache sizing.
+            if all(
+                hasattr(token_pool, name)
+                for name in ("head_num", "head_dim", "row_dim")
+            ):
+                token_pool.row_dim = token_pool.head_num * token_pool.head_dim
+            if all(
+                hasattr(token_pool, name)
+                for name in ("head_num", "v_head_dim", "v_row_dim")
+            ):
+                token_pool.v_row_dim = token_pool.head_num * token_pool.v_head_dim
+
             if (
                 hasattr(token_pool, "kv_cache_dim")
                 and token_pool.kv_cache_dim == 2

tools/sglang-simulator/test/test_simulation_cache_hit_ratio.py

@@ -13,9 +13,9 @@ from test_simulation_sglang_serving import (
 def test_in_process_runner_reports_each_cache_tier(tmp_path):
     runner = make_sglang_runner(tmp_path)
     benchmark_config = BenchmarkConfig(request_rate=10, ignore_request_timestamp=False)
-    cached_ds = make_fixed_dataset(1000, 8)
+    cached_ds = make_fixed_dataset(1000, 3)
-    evict_l1_ds = make_fixed_dataset(2000, 10)
+    evict_l1_ds = make_fixed_dataset(2000, 5)
-    evict_l2_ds = make_fixed_dataset(3000, 20)
+    evict_l2_ds = make_fixed_dataset(3000, 10)
 
     try:
         metrics = runner.benchmark(benchmark_config, dataset=cached_ds)

tools/sglang-simulator/test/test_simulation_sglang_runner.py

@@ -78,7 +78,7 @@ def make_sglang_runner(tmp_path: Path):
             hicache_ratio=2,
             hicache_storage_backend="file",
             hicache_storage_prefetch_policy="wait_complete",
-            max_total_tokens=10 * 1024,
+            max_total_tokens=4 * 1024,
             page_size=256,
             skip_tokenizer_init=True,
         )