tedzhouhk · Sep 14, 2026 · Sep 9, 2026 · Sep 9, 2026 · Sep 14, 2026 · Sep 14, 2026
@@ -55,7 +55,7 @@ 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" }

@@ -40,7 +40,7 @@ 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==0.10.0",
     "matplotlib",
     "networkx",

@@ -40,7 +40,7 @@ def _config() -> PlannerConfig:
             hf_id="Qwen/Qwen3-32B",
             system="h200_sxm",
             backend="vllm",
-            backend_version="0.19.0",
+            backend_version="0.24.0",
             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="0.24.0",
         isl=128,
         osl=8,
         ttft=100_000.0,

@@ -310,21 +310,21 @@ 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": "0.24.0",
+    "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 (
@@ -57,15 +60,9 @@
     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
@@ -57,7 +57,7 @@ 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"])

@@ -99,7 +99,7 @@ def fake_estimate_num_gpu_blocks(**kwargs):
             "gpu_memory_utilization": 0.8,
             "mem_fraction_static": None,
             "free_gpu_memory_fraction": None,
-            "backend_version": None,
+            "backend_version": "current",
             "pp_size": 1,
             "moe_tp_size": None,
             "moe_ep_size": None,

@@ -1,12 +1,14 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Offline, network-free coverage of Dynamo replay against AIC 0.11."""
+"""Offline, network-free coverage of Dynamo replay against the pinned AISimulate release."""
 
 from __future__ import annotations
 
 import pytest
 
+from dynamo.replay.config import load_engine_args
+
 pytestmark = [
     pytest.mark.aiconfigurator,
     pytest.mark.gpu_0,
@@ -17,7 +19,7 @@
 
 AIC_MODEL = "Qwen/Qwen3-32B"
 AIC_SYSTEM = "h200_sxm"
-AIC_BACKEND_VERSION = "0.19.0"
+AIC_BACKEND_VERSION = "0.24.0"
 
 
 @pytest.fixture(autouse=True)
@@ -61,6 +63,23 @@ def test_real_aic_memory_estimates_gpu_blocks() -> None:
     assert blocks > 0
 
 
+def test_default_aic_capacity_uses_queryable_version() -> None:
+    args = load_engine_args(
+        {
+            "aic_backend": "vllm",
+            "aic_system": AIC_SYSTEM,
+            "aic_model_path": AIC_MODEL,
+            "aic_tp_size": 1,
+            "block_size": 64,
+            "max_num_batched_tokens": 4096,
+            "max_num_seqs": 128,
+        }
+    )
+    assert args is not None
+    assert args.num_gpu_blocks > 0
+    assert args.aic_backend_version == "current"
+
+
 def test_aggregated_replay_uses_native_aic_engine() -> None:
     from aiconfigurator_core.sdk.engine import compile_engine
 

@@ -264,6 +264,11 @@ def broken_memory_module(name, *args, **kwargs):
     assert exc_info.value is missing_dependency
 
 
+@pytest.mark.parametrize("backend", ["vllm", "sglang", "trtllm"])
+def test_default_version_uses_queryable_slot(backend):
+    assert resolve_backend_version(backend, None) == "current"
+
+
 def test_trtllm_version_resolution():
     assert resolve_backend_version("trtllm", "0.20.0") == "0.20.0"
 

@@ -461,7 +461,9 @@ struct MockEngineArgsSerde {
     enable_local_indexer: OptionalConfigValue<bool>,
     bootstrap_port: OptionalConfigValue<u16>,
     handoff_session_timeout_ms: OptionalConfigValue<u64>,
+    #[serde(alias = "kv_transfer_bytes_per_token")]
     kv_bytes_per_token: OptionalConfigValue<usize>,
+    kv_cache_bytes_per_token: OptionalConfigValue<usize>,
     kv_transfer_bandwidth: OptionalConfigValue<f64>,
     kv_transfer_timing_mode: OptionalConfigValue<String>,
     reasoning: OptionalConfigValue<ReasoningConfig>,
@@ -703,11 +705,16 @@ pub struct MockEngineArgs {
     #[validate(range(min = 1))]
     pub handoff_session_timeout_ms: u64,
 
-    /// KV cache bytes per token, auto-computed from model config by Python CLI.
+    /// Bytes transferred per token, auto-computed from model config by Python CLI.
     /// Formula: num_layers * 2 * num_kv_heads * head_dim * dtype_bytes
     #[builder(default = "None")]
     pub kv_bytes_per_token: Option<usize>,
 
+    /// Physical KV-cache bytes occupied by one token, independent of transfer geometry.
+    #[builder(default = "None")]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub kv_cache_bytes_per_token: Option<usize>,
+
     /// KV cache transfer bandwidth in GB/s for disaggregated serving latency simulation.
     /// Default: 64.0 (inter-node InfiniBand). Set to 0 to disable KV transfer delay.
     /// For intra-node NVLink, typical value is ~450.
@@ -1047,6 +1054,9 @@ impl TryFrom<MockEngineArgsSerde> for MockEngineArgs {
         if let Some(kv_bytes_per_token) = compat.kv_bytes_per_token.into_nullable() {
             builder = builder.kv_bytes_per_token(kv_bytes_per_token);
         }
+        if let Some(kv_cache_bytes_per_token) = compat.kv_cache_bytes_per_token.into_nullable() {
+            builder = builder.kv_cache_bytes_per_token(kv_cache_bytes_per_token);
+        }
         if let Some(kv_transfer_bandwidth) = compat.kv_transfer_bandwidth.into_nullable() {
             builder = builder.kv_transfer_bandwidth(kv_transfer_bandwidth);
         }