sttts · Oct 29, 2026 · Nov 2, 2026
@@ -18,6 +18,10 @@ checks emitted manifests and `index.json` against the selected Candidates in
 runtime image, namespace, runtime override, and total GPU use. The Pareto checks
 reject rendering the same Candidate repeatedly under different names.
 
+The aggregated case omits the runtime block size, and the router case explicitly
+sets it to `null`; both must omit `--block-size` from the generated worker args.
+The other cases retain explicit block sizes.
+
 The fixed aggregated and disaggregated cases also compare the complete DGD to
 `goldens/`. Run the table from the repository root:
 

@@ -17,7 +17,7 @@ engine:
         preset:
           - {replicas: 1, tensor: 1, pipeline: 1, attention_data: 1, moe_tensor: 1, moe_expert: 1}
       scheduler: {max_batched_tokens: 8192, max_sequences: 4}
-      kv_cache: {block_size: 64, prefix_caching: true, capacity: {type: fixed, blocks: 256}}
+      kv_cache: {prefix_caching: true, capacity: {type: fixed, blocks: 256}}
       timing: {type: fixed, prefill_ms: 1, decode_ms: 1}
 optimization:
   target: throughput

@@ -16,7 +16,7 @@ engine:
       parallelism:
         preset:
           - {replicas: 1, tensor: 1, pipeline: 1, attention_data: 1, moe_tensor: 1, moe_expert: 1}
-      scheduler: {max_batched_tokens: 32, max_sequences: {choices: [1, 4]}}
+      scheduler: {max_batched_tokens: 64, max_sequences: {choices: [1, 4]}}
       kv_cache: {block_size: 64, prefix_caching: true, capacity: {type: fixed, blocks: 256}}
       timing: {type: fixed, prefill_ms: 1, decode_ms: 1}
 optimization:

@@ -17,7 +17,7 @@ engine:
         preset:
           - {replicas: 2, tensor: 1, pipeline: 1, attention_data: 1, moe_tensor: 1, moe_expert: 1}
       scheduler: {max_batched_tokens: 8192, max_sequences: 256}
-      kv_cache: {block_size: 64, prefix_caching: true, capacity: {type: fixed, blocks: 256}}
+      kv_cache: {block_size: null, prefix_caching: true, capacity: {type: fixed, blocks: 256}}
       timing: {type: fixed, prefill_ms: 1, decode_ms: 1}
 router:
   policy: {choices: [round_robin, kv_router]}

@@ -44,8 +44,6 @@ spec:
           - '1'
           - --data-parallel-size
           - '1'
-          - --block-size
-          - '64'
           - --num-gpu-blocks-override
           - '256'
           - --max-num-seqs

@@ -75,8 +75,9 @@ def _assert_dgd_golden(golden: Path, manifests: list[dict[str, Any]]) -> None:
 
 
 def _assert_dgd_outputs(
-    config_path: Path, options: dict[str, Any], output: Path
+    config_path: Path, input_config: dict[str, Any], output: Path
 ) -> None:
+    options = input_config["dgd"]
     result = SweepResult.model_validate_json(
         (output / "recommendation.json").read_text()
     )
@@ -183,6 +184,13 @@ def _assert_dgd_outputs(
             role = roles[component["type"]]
             prefix = "" if role == "agg" else f"{role}_"
             assert component["replicas"] == config[f"{prefix}replicas"]
+
+            # Preserve the requested block size through Candidate creation and rendering.
+            worker_config = input_config["engine"]["workers"][
+                "aggregated" if role == "agg" else role
+            ]
+            block_size = worker_config["kv_cache"].get("block_size")
+            assert config.get(f"{role}_block_size") == block_size
             args = container["args"]
             for flag, expected in {
                 "--model": config["model_name"],
@@ -192,11 +200,14 @@ def _assert_dgd_outputs(
                 "--max-model-len": config["context_length"],
                 "--max-num-seqs": config[f"{role}_max_num_seqs"],
                 "--max-num-batched-tokens": config[f"{role}_max_num_batched_tokens"],
-                "--block-size": config[f"{role}_block_size"],
+                "--block-size": block_size,
                 "--num-gpu-blocks-override": config[f"{role}_num_gpu_blocks"],
             }.items():
-                assert args.count(flag) == 1
-                assert args[args.index(flag) + 1] == str(expected)
+                if expected is None:
+                    assert flag not in args
+                else:
+                    assert args.count(flag) == 1
+                    assert args[args.index(flag) + 1] == str(expected)
             assert ("--no-enable-prefix-caching" in args) != config[
                 f"{role}_enable_prefix_caching"
             ]
@@ -249,4 +260,4 @@ def test_dgd_cli_cases(config_path: Path, tmp_path: Path) -> None:
     )
     assert json.loads(result.stdout)
     assert "router AIC payload requires" not in result.stderr
-    _assert_dgd_outputs(config_path, config["dgd"], output)
+    _assert_dgd_outputs(config_path, config, output)
@@ -13,8 +13,6 @@
 
 import pytest
 import yaml
-from aisimulate.generator.request.schema import ModelFacts
-from aisimulate.sweeper.config import Workload
 
 from dynamo.aisimulate.output.dgd.renderers import (
     CandidateMaterializationError,
@@ -24,7 +22,6 @@
 from dynamo.aisimulate.output.dgd.renderers import render_dgd
 from dynamo.aisimulate.output.dgd.renderers.aic import renderer as aic_renderer
 from dynamo.frontend.frontend_args import FrontendArgGroup, FrontendConfig
-from dynamo.profiler.utils.config_modifiers import vllm
 
 pytestmark = [
     pytest.mark.unit,
@@ -161,6 +158,12 @@ def _without_transfer_buffer_size(rendered: str) -> dict:
 def test_installed_aic_renders_trace_workloads_like_synthetic_ones(
     monkeypatch, tmp_path, backend, backend_version, image
 ) -> None:
+    pytest.importorskip("aisimulate.generator")
+    from aisimulate.generator.request.schema import ModelFacts
+    from aisimulate.sweeper.config import Workload
+
+    from dynamo.profiler.utils.config_modifiers import vllm
+
     # Keep the renders offline: model facts come from the request below.
     monkeypatch.setattr(vllm, "get_model_context_length", lambda _model: None)
     monkeypatch.setattr(vllm, "get_mamba_cache_align_block_size", lambda _model: None)

@@ -26,10 +26,6 @@ spec:
         mode: aggregated
         context_length: 32768
         workers:
-          # TODO(#17581): Remove this explicit value after AISimulate preserves an unset runtime block size.
-          aggregated:
-            kv_cache:
-              block_size: 64
           prefill: null
           decode: null
       traffic: