fix: resolve 'auto' backend to concrete value in every situation (#7158)

Signed-off-by: Hannah Zhang <hannahz@nvidia.com>

components/src/dynamo/profiler/rapid.py

@@ -99,7 +99,7 @@ def _generate_dgd_from_pick(
     return None
 
 
-# in naive mode, use vllm as the default backend
+# Fallback backend when AIC simulation is unavailable and no concrete backend is specified.
 _DEFAULT_NAIVE_BACKEND = "vllm"
 
 
@@ -110,19 +110,10 @@ def _run_naive_fallback(
     system: str,
     backend: str,
 ) -> dict:
-    """Handle the AIC-unsupported path via naive config generation.
-
-    Builds naive generator params (CLI args, parallelism) and then
-    assembles the DGD via ``build_dgd_config`` — the same route used
-    by the normal simulation path — so the output always uses the
-    clean base DGD YAMLs with actual ``command``/``args`` arrays.
-    """
+    """Handle the AIC-unsupported path via naive config generation."""
     if backend == "auto":
         backend = _DEFAULT_NAIVE_BACKEND
-        logger.info(
-            "Auto backend resolved to '%s' for naive fallback.",
-            backend,
-        )
+        logger.info("Auto backend resolved to '%s' for naive fallback.", backend)
     logger.info(
         "AIC does not support this combo — falling back to naive config generation."
     )
@@ -169,6 +160,13 @@ def _run_autoscale_sim(
     request_latency: float | None,
 ) -> dict:
     """Build a TaskConfig, run autoscale simulation, collect latencies, generate DGD."""
+    # TODO(AIC): the autoscale path constructs TaskConfig directly; BackendName("auto")
+    # is not a valid enum value, so resolve "auto" to a concrete backend here.
+    # AIC should add native auto-backend support in the autoscale path.
+    if backend == "auto":
+        backend = _DEFAULT_NAIVE_BACKEND
+        logger.info("Auto backend resolved to '%s' for autoscale simulation.", backend)
+
     planner_cfg = dgdr.features.planner if dgdr.features else None
     if planner_cfg and planner_cfg.enable_throughput_scaling:
         logger.warning(
@@ -205,6 +203,7 @@ def _run_autoscale_sim(
         "dgd_config": dgd_config,
         "chosen_exp": "disagg",
         "task_configs": task_configs,
+        "resolved_backend": backend,
     }
 
 
@@ -276,12 +275,24 @@ def _run_default_sim(
 
     dgd_config = _generate_dgd_from_pick(dgdr, best_config_df, chosen, task_configs)
 
+    # When backend="auto" AIC expands to per-backend task configs; the winning
+    # row carries the concrete backend name so downstream consumers (e.g.
+    # run_interpolation) can use it without re-encountering "auto".
+    resolved_backend = backend
+    if (
+        backend == "auto"
+        and not best_config_df.empty
+        and "backend" in best_config_df.columns
+    ):
+        resolved_backend = best_config_df.iloc[0]["backend"]
+
     return {
         "best_config_df": best_config_df,
         "best_latencies": best_latencies,
         "dgd_config": dgd_config,
         "chosen_exp": chosen,
         "task_configs": task_configs,
+        "resolved_backend": resolved_backend,
     }
 
 

components/src/dynamo/profiler/tests/test_profile_sla_auto_backend.py

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Unit tests for auto-backend resolution in the profiler pipeline."""
+
+import pytest
+
+from dynamo.profiler.rapid import _DEFAULT_NAIVE_BACKEND
+
+pytestmark = [
+    pytest.mark.unit,
+    pytest.mark.gpu_0,
+    pytest.mark.pre_merge,
+    pytest.mark.parallel,
+]
+
+
+def test_autoscale_sim_resolves_auto_to_default() -> None:
+    """_run_autoscale_sim must resolve 'auto' to _DEFAULT_NAIVE_BACKEND before
+    constructing TaskConfig, since BackendName('auto') is not a valid enum value.
+    """
+    import inspect
+
+    from dynamo.profiler.rapid import _run_autoscale_sim
+
+    src = inspect.getsource(_run_autoscale_sim)
+    # The function must guard against "auto" before TaskConfig is constructed.
+    assert (
+        'backend == "auto"' in src
+    ), "_run_autoscale_sim must resolve backend='auto' before constructing TaskConfig"
+    assert (
+        "_DEFAULT_NAIVE_BACKEND" in src
+    ), "_run_autoscale_sim must fall back to _DEFAULT_NAIVE_BACKEND when backend='auto'"
+
+
+def test_autoscale_sim_returns_resolved_backend() -> None:
+    """_run_autoscale_sim must include 'resolved_backend' in its result dict so
+    profile_sla.py can pass the concrete backend to run_interpolation.
+    """
+    import inspect
+
+    from dynamo.profiler.rapid import _run_autoscale_sim
+
+    src = inspect.getsource(_run_autoscale_sim)
+    assert (
+        '"resolved_backend"' in src
+    ), "_run_autoscale_sim must return 'resolved_backend' in its result dict"
+
+
+def test_naive_fallback_resolves_auto_to_default() -> None:
+    """_run_naive_fallback must resolve 'auto' to _DEFAULT_NAIVE_BACKEND.
+
+    The naive path is taken when AIC doesn't support the model/system combo;
+    it uses build_naive_generator_params and generate_backend_artifacts which
+    require a concrete backend string.
+    """
+    import inspect
+
+    from dynamo.profiler.rapid import _run_naive_fallback
+
+    src = inspect.getsource(_run_naive_fallback)
+    assert (
+        'backend == "auto"' in src
+    ), "_run_naive_fallback must resolve backend='auto' before calling AIC helpers"
+    assert "_DEFAULT_NAIVE_BACKEND" in src
+
+
+def test_default_sim_returns_resolved_backend() -> None:
+    """_run_default_sim must include 'resolved_backend' in its result dict.
+
+    When backend='auto', AIC expands to per-backend task configs and the
+    winning row carries the concrete backend name; _run_default_sim must
+    extract and surface it so run_interpolation never receives 'auto'.
+    """
+    import inspect
+
+    from dynamo.profiler.rapid import _run_default_sim
+
+    src = inspect.getsource(_run_default_sim)
+    assert (
+        '"resolved_backend"' in src
+    ), "_run_default_sim must return 'resolved_backend' in its result dict"
+
+
+def test_default_naive_backend_is_concrete() -> None:
+    """_DEFAULT_NAIVE_BACKEND must be a concrete backend string, not 'auto'."""
+    assert _DEFAULT_NAIVE_BACKEND != "auto"
+    assert _DEFAULT_NAIVE_BACKEND in ("vllm", "sglang", "trtllm")