tedzhouhk · Sep 14, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 14, 2026 · Sep 14, 2026
@@ -56,7 +56,7 @@ keywords = ["llm", "genai", "inference", "nvidia", "distributed"]
 
 [workspace.dependencies]
 # Published separately by https://github.com/ai-dynamo/aisimulate.
-aisimulate-core = { version = "=0.1.0-dev.2" }
+aisimulate-core = { version = "=0.12.0-dev.1" }
 
 # Local crates
 dynamo-runtime = { path = "lib/runtime", version = "1.5.0" }

@@ -39,7 +39,7 @@
 MODEL = "Qwen/Qwen3-32B"
 SYSTEM = "h200_sxm"
 BACKEND = "vllm"
-BACKEND_VERSION = "0.19.0"
+BACKEND_VERSION = "current"
 
 BS, ISL = 16, 2048
 THREAD_COUNTS = (1, 2, 4, 8, 12, 16, 24)

@@ -65,7 +65,7 @@
 MODEL = "Qwen/Qwen3-32B"
 SYSTEM = "h200_sxm"
 BACKEND = "vllm"
-BACKEND_VERSION = "0.19.0"
+BACKEND_VERSION = "current"
 
 # Workloads chosen to stress the predict hot path differently. ``num_workers`` > 1
 # is where the Python GIL serialises predictions, so the Rust win is largest there.

@@ -39,7 +39,7 @@ classifiers = [
 ]
 
 dependencies = [
-    "aisimulate==0.1.0.dev2; python_version >= '3.11' and python_version < '3.14'",
+    "aisimulate==0.12.0.dev1; python_version >= '3.11' and python_version < '3.14'",
     "aiperf==0.12.0",
     "matplotlib",
     "networkx",

@@ -344,8 +344,9 @@ def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
         "--aic-backend-version",
         type=str,
         default=None,
-        help="AIC backend engine version (e.g., '0.19.0' for vLLM, '0.5.10' for SGLang, "
-        "'1.3.0rc10' for TRT-LLM). If not set, uses the default version for the backend.",
+        help="AIC performance-database version: 'current', 'previous', or 'next' "
+        "when available, or a version assigned to one of those slots. "
+        "Defaults to the release database's 'current' slot.",
     )
     parser.add_argument(
         "--aic-tp-size",

@@ -40,7 +40,7 @@ def _config() -> PlannerConfig:
             hf_id="Qwen/Qwen3-32B",
             system="h200_sxm",
             backend="vllm",
-            backend_version="0.19.0",
+            backend_version="current",
             prefill_pick=pick,
             decode_pick=pick,
         ),

@@ -31,7 +31,7 @@ def test_rapid_task_v2_build_and_execute_contract() -> None:
         total_gpus=2,
         system="h200_sxm",
         backend="vllm",
-        backend_version="0.19.0",
+        backend_version="current",
         isl=128,
         osl=8,
         ttft=100_000.0,

@@ -313,21 +313,21 @@ def test_injects_all_required_flags(self):
         assert out[out.index("--aic-attention-dp-size") + 1] == "8"
         # trtllm is not a mocker engine_type; leave --engine-type alone.
         assert "--engine-type" not in out
-        assert out[out.index("--aic-backend-version") + 1] == "1.3.0rc10"
+        assert out[out.index("--aic-backend-version") + 1] == "current"
 
     def test_matches_engine_type_for_vllm(self):
         spec = self._spec("vllm")
         out = _inject_mocker_aic_args([], spec, spec.prefill_pick)
         assert out[out.index("--engine-type") + 1] == "vllm"
         assert out[out.index("--aic-backend") + 1] == "vllm"
-        assert out[out.index("--aic-backend-version") + 1] == "0.14.0"
+        assert out[out.index("--aic-backend-version") + 1] == "current"
 
     def test_matches_engine_type_for_sglang(self):
         spec = self._spec("sglang")
         out = _inject_mocker_aic_args([], spec, spec.decode_pick)
         assert out[out.index("--engine-type") + 1] == "sglang"
         assert out[out.index("--aic-backend") + 1] == "sglang"
-        assert out[out.index("--aic-backend-version") + 1] == "0.5.6.post2"
+        assert out[out.index("--aic-backend-version") + 1] == "current"
 
 
 class TestBuildPlannerConfigEmbedsAicSpec:

@@ -19,8 +19,8 @@
 AIC_PARITY_MODEL = "Qwen/Qwen3-32B"
 AIC_PARITY_SYSTEM = "h200_sxm"
 AIC_PARITY_VERSIONS = {
-    "vllm": "0.19.0",
-    "sglang": "0.5.6.post2",
+    "vllm": "current",
+    "sglang": "current",
 }
 AIC_PARITY_BACKENDS = [
     pytest.param("vllm", id="vllm"),

@@ -23,6 +23,9 @@
 import numpy as np
 import yaml
 
+from dynamo._internal.aic import (
+    DEFAULT_BACKEND_VERSIONS as _MOCKER_AIC_BACKEND_VERSIONS,
+)
 from dynamo.common.utils.paths import get_workspace_dir
 from dynamo.planner.config.aic_interpolation_spec import AICInterpolationSpec
 from dynamo.planner.config.backend_components import MockerComponentName
@@ -54,15 +57,9 @@
     needs_mocker_aic_perf_model,
     needs_profile_data,
 )
-from dynamo.profiler.utils.replay_optimize.constants import AIC_BACKEND_VERSIONS
 
 logger = logging.getLogger(__name__)
 
-_MOCKER_AIC_BACKEND_VERSIONS = {
-    **AIC_BACKEND_VERSIONS,
-    "trtllm": "1.3.0rc10",
-}
-
 
 def _load_latest_database_version() -> Optional[Callable[..., Optional[str]]]:
     try:

@@ -5,9 +5,10 @@
 
 import os
 
+from dynamo._internal.aic import DEFAULT_BACKEND_VERSIONS
+
 AIC_BACKEND_VERSIONS = {
-    "vllm": "0.14.0",
-    "sglang": "0.5.6.post2",
+    backend: DEFAULT_BACKEND_VERSIONS[backend] for backend in ("vllm", "sglang")
 }
 
 DEFAULT_OVERLAP_SCORE_CREDITS = (1.0,)

@@ -13,6 +13,7 @@
 
 from aisimulate.aic import materialize_aic_num_gpu_blocks
 
+from dynamo._internal.aic import resolve_backend_version
 from dynamo.mocker import MockEngineArgs
 from dynamo.mocker.args import (
     resolve_planner_profile_data as _resolve_mocker_planner_profile_data,
@@ -26,6 +27,13 @@ class PlannerProfileDataResult(Protocol):
 def resolve_aic_num_gpu_blocks(raw: dict[str, Any]) -> None:
     """Materialize AIC KV capacity in-place for SDK compatibility."""
 
+    if (
+        raw.get("aic_backend") is not None
+        or raw.get("aic_attention_dp_size") is not None
+    ):
+        raw["aic_backend_version"] = resolve_backend_version(
+            raw.get("aic_backend") or "vllm", raw.get("aic_backend_version")
+        )
     lowered = materialize_aic_num_gpu_blocks(raw)
     raw.clear()
     raw.update(lowered)

@@ -17,7 +17,6 @@
 from numbers import Real
 from typing import Any
 
-from aisimulate.aic import materialize_aic_num_gpu_blocks
 from aisimulate.sweeper.provider import JSONValue, RuntimeHookSpec
 from aisimulate.sweeper.replay import (
     HookCapability,
@@ -30,6 +29,7 @@
 from dynamo.llm import AicPerfConfig, KvRouterConfig
 from dynamo.mocker import MockEngineArgs
 from dynamo.replay.api import run_synthetic_trace_replay, run_trace_replay
+from dynamo.replay.config import resolve_aic_num_gpu_blocks
 
 _PLANNER_HOOK = HookCapability(
     provider="dynamo.planner",
@@ -260,7 +260,7 @@ def _engine_args(payload: dict[str, JSONValue] | None) -> MockEngineArgs:
                 "aic_model_path",
             ):
                 lowered.pop(name, None)
-        lowered = materialize_aic_num_gpu_blocks(lowered)
+        resolve_aic_num_gpu_blocks(lowered)
         # Pipeline parallelism is already represented in the public parallel
         # mapping and used for AIC capacity. MockEngineArgs has no PP field.
         lowered.pop("aic_pp_size", None)

@@ -369,7 +369,7 @@ def test_fixed_timing_keeps_aic_identity_out_of_runtime_args(monkeypatch) -> Non
     monkeypatch.setattr(simulation, "MockEngineArgs", _FakeEngineArgs)
     monkeypatch.setattr(
         simulation,
-        "materialize_aic_num_gpu_blocks",
+        "resolve_aic_num_gpu_blocks",
         lambda payload: payload,
     )
     engine_args = simulation.DynamoReplayRunner._engine_args(

@@ -235,7 +235,7 @@ def test_sweeper_runs_real_dynamo_replay_in_spawned_workers() -> None:
         include_router=True,
     )
 
-    candidates = Sweeper(
+    result = Sweeper(
         runner_factory=DynamoReplayRunnerFactory(),
         providers={
             "dynamo.planner": create_planner_provider(),
@@ -245,7 +245,9 @@ def test_sweeper_runs_real_dynamo_replay_in_spawned_workers() -> None:
         show_progress=False,
     ).run(config)
 
+    candidates = result.candidates
     assert len(candidates) == 2
+    assert all(candidate.status == "feasible" for candidate in candidates)
     assert all(
         candidate.metrics["output_throughput_tok_s"] > 0.0 for candidate in candidates
     )

@@ -2,4 +2,4 @@
 # SPDX-License-Identifier: Apache-2.0
 
 # Published AISimulate wheel staged into planner and framework runtime wheelhouses.
-aisimulate==0.1.0.dev2
+aisimulate==0.12.0.dev1
@@ -639,10 +639,22 @@ COPY container/deps/requirements.aisimulate.txt /opt/dynamo/container/deps/requi
 # wheel consumed by ai-dynamo instead of rebuilding it from vendored source.
 # Download only this distribution; runtime images own dependency installation
 # through their requirements files and local wheels.
+# AISimulate targets glibc 2.34+, which all consuming runtime images support.
+# Select that target explicitly: the manylinux_2_28 builder only stages the
+# wheel and must not filter it using the builder's older glibc version.
 RUN --mount=type=cache,id=uv-root-{{ context.dynamo.uv_version }},target=/root/.cache/uv,sharing=shared \
     export UV_CACHE_DIR=/root/.cache/uv && \
     source ${VIRTUAL_ENV}/bin/activate && \
+    case "${TARGETARCH}" in \
+        amd64) AISIMULATE_WHEEL_ARCH=x86_64 ;; \
+        arm64) AISIMULATE_WHEEL_ARCH=aarch64 ;; \
+        *) echo "Unsupported AISimulate target architecture: ${TARGETARCH}" >&2; exit 1 ;; \
+    esac && \
     python -m pip download \
+        --platform "manylinux_2_34_${AISIMULATE_WHEEL_ARCH}" \
+        --python-version "${PYTHON_VERSION}" \
+        --implementation cp \
+        --abi abi3 \
         --only-binary=:all: \
         --no-deps \
         --dest /opt/dynamo/dist \

@@ -167,8 +167,9 @@ For the timing-model design, see [Mocker Engine Architecture](../../developer-gu
 </ParamField>
 
 <ParamField path="--aic-backend-version" type="string" default="auto">
-  AIC backend engine version, for example `0.12.0` for vLLM. When unset, uses the default version for
-  the backend.
+  AIC performance-database version. Use `current`, `previous`, or `next` when the slot is available
+  for the selected system and backend, or specify a version assigned to one of those slots.
+  When unset, uses the release database's `current` slot.
 </ParamField>
 
 <ParamField path="--aic-tp-size" type="integer" default="1">

@@ -60,7 +60,7 @@ request-trace-s3 = ["dynamo-llm/request-trace-s3"]
 # former AIC crate surface. The `python` feature is required for
 # AicEngineBuilder; this extension supplies the interpreter, so the dependency
 # must not enable `extension-module`.
-aisimulate-core = { version = "=0.1.0-dev.2", optional = true, features = ["python"] }
+aisimulate-core = { version = "=0.12.0-dev.1", optional = true, features = ["python"] }
 dynamo-runtime = { path = "../../runtime" }
 dynamo-mocker = { path = "../../mocker" }
 dynamo-parsers = { version = "8.1.0" }

@@ -15,11 +15,11 @@
 # Dynamo's historical default when conditional acceptance rates are omitted.
 _DEFAULT_NEXTN_ACCEPT_RATES = [0.85, 0.3, 0.0, 0.0, 0.0]
 
-# Default backend versions match the AIC-core v0.11.0 perf DB.
+# Resolve defaults through the queryable slots in the pinned AISimulate perf DB.
 DEFAULT_BACKEND_VERSIONS = {
-    "vllm": "0.19.0",
-    "sglang": "0.5.10",
-    "trtllm": "1.3.0rc10",
+    "vllm": "current",
+    "sglang": "current",
+    "trtllm": "current",
 }
 _KV_CAPACITY_BACKENDS = frozenset(DEFAULT_BACKEND_VERSIONS)
 DEFAULT_STATIC_STRIDE = 32
@@ -39,7 +39,7 @@ def _validate_kv_capacity_backend(backend_name: str) -> None:
 
 
 def resolve_backend_version(backend_name: str, backend_version: str | None) -> str:
-    """Return the pinned backend version used for AIC perf lookups."""
+    """Preserve explicit versions; otherwise use the release database current slot."""
     if backend_version is not None:
         return backend_version
     return DEFAULT_BACKEND_VERSIONS.get(backend_name, DEFAULT_BACKEND_VERSIONS["vllm"])