[2.0] Default resources.gpu_type and gpus_per_node from srtslurm.yaml

docs/config-reference.md

@@ -113,7 +113,8 @@ The `srtslurm.yaml` file can contain the following fields:
 | `default_account`               | string | Default SLURM account                                 |
 | `default_partition`             | string | Default SLURM partition                               |
 | `default_time_limit`            | string | Default job time limit                                |
-| `gpus_per_node`                 | int    | Default GPUs per node                                 |
+| `gpus_per_node`                 | int    | Default GPUs per node (applied to recipes that omit `resources.gpus_per_node`) |
+| `default_gpu_type`              | string | Default `resources.gpu_type` for recipes that omit it |
 | `network_interface`             | string | Network interface for NCCL                            |
 | `srtctl_root`                   | string | Root directory for srtctl                             |
 | `output_dir`                    | string | Custom output directory (overrides srtctl_root/outputs) |
@@ -226,8 +227,8 @@ resources:
 
 | Field             | Type   | Default            | Description                           |
 | ----------------- | ------ | ------------------ | ------------------------------------- |
-| `gpu_type`        | string | -                  | GPU type: "gb200", "gb300", or "h100" |
-| `gpus_per_node`   | int    | 4                  | GPUs per node                         |
+| `gpu_type`        | string | `default_gpu_type` | GPU type, e.g. "gb200", "gb300", "h100". Optional; inherits `default_gpu_type` from `srtslurm.yaml` when omitted |
+| `gpus_per_node`   | int    | cluster / 4        | GPUs per node; inherits the cluster `gpus_per_node` when omitted, else 4 |
 | `prefill_nodes`   | int    | null               | Nodes dedicated to prefill            |
 | `decode_nodes`    | int    | null               | Nodes dedicated to decode             |
 | `prefill_workers` | int    | null               | Number of prefill workers             |

docs/schema-reference.md

@@ -55,7 +55,7 @@ Resource allocation configuration.
 
 | Key | Type | Default | Description |
 |---|---|---|---|
-| `gpu_type` | str | required |  |
+| `gpu_type` | str \| None | `None` | GPU type (h100, gb200, ...). Cluster fact, not a topology choice. Optional: a recipe that omits it inherits `default_gpu_type` from srtslurm.yaml, and `gpus_per_node` inherits the cluster `gpus_per_node`. Both are still worth setting in a recipe so it is self-describing for result rollups. |
 | `gpus_per_node` | int | `4` |  |
 | `prefill_nodes` | int \| None | `None` | Disaggregated mode |
 | `decode_nodes` | int \| None | `None` |  |
@@ -543,6 +543,7 @@ Top-level keys of `srtslurm.yaml`. Recipes inherit these defaults and resolve al
 | `default_partition` | str \| None | `None` |  |
 | `default_time_limit` | str \| None | `None` |  |
 | `gpus_per_node` | int \| None | `None` |  |
+| `default_gpu_type` | str \| None | `None` | Default for ``ResourceConfig.gpu_type`` when the recipe omits it. Lets one recipe move between clusters of different GPU types without an edit. |
 | `network_interface` | str \| None | `None` |  |
 | `use_gpus_per_node_directive` | bool | `True` |  |
 | `use_segment_sbatch_directive` | bool | `True` |  |

src/srtctl/core/config.py

@@ -180,6 +180,18 @@ def resolve_config_with_defaults(user_config: dict[str, Any], cluster_config: di
         slurm["time_limit"] = cluster_config["default_time_limit"]
         logger.debug(f"Applied default time_limit: {slurm['time_limit']}")
 
+    # GPU-topology facts inherited from the cluster when the recipe omits them.
+    # gpu_type and gpus_per_node describe the machine, not the deployment, so a
+    # recipe can move between clusters by leaving them to srtslurm.yaml.
+    resources_defaults = config.get("resources")
+    if isinstance(resources_defaults, dict):
+        if not resources_defaults.get("gpu_type") and cluster_config.get("default_gpu_type"):
+            resources_defaults["gpu_type"] = cluster_config["default_gpu_type"]
+            logger.debug("Applied default gpu_type: %s", resources_defaults["gpu_type"])
+        if "gpus_per_node" not in resources_defaults and cluster_config.get("gpus_per_node") is not None:
+            resources_defaults["gpus_per_node"] = cluster_config["gpus_per_node"]
+            logger.debug("Applied cluster gpus_per_node: %s", resources_defaults["gpus_per_node"])
+
     default_sbatch_directives = cluster_config.get("default_sbatch_directives")
     if isinstance(default_sbatch_directives, dict):
         sbatch_directives = config.setdefault("sbatch_directives", {})

src/srtctl/core/schema.py

@@ -247,6 +247,9 @@ class ClusterConfig:
     default_partition: str | None = None
     default_time_limit: str | None = None
     gpus_per_node: int | None = None
+    # Default for ``ResourceConfig.gpu_type`` when the recipe omits it. Lets one
+    # recipe move between clusters of different GPU types without an edit.
+    default_gpu_type: str | None = None
     network_interface: str | None = None
     use_gpus_per_node_directive: bool = True
     use_segment_sbatch_directive: bool = True
@@ -548,7 +551,11 @@ class HetComponent:
 class ResourceConfig:
     """Resource allocation configuration."""
 
-    gpu_type: str
+    # GPU type (h100, gb200, ...). Cluster fact, not a topology choice. Optional:
+    # a recipe that omits it inherits `default_gpu_type` from srtslurm.yaml, and
+    # `gpus_per_node` inherits the cluster `gpus_per_node`. Both are still worth
+    # setting in a recipe so it is self-describing for result rollups.
+    gpu_type: str | None = None
     gpus_per_node: int = 4
 
     # Disaggregated mode

tests/test_configs.py

@@ -5035,3 +5035,41 @@ def fake_start_endpoint(ep_procs):
 
             # Each node has only 1 worker — no wait should be triggered
             assert wait_called == []
+
+
+class TestClusterGpuDefaults:
+    """resources.gpu_type / gpus_per_node inherit from srtslurm.yaml when omitted."""
+
+    def _recipe(self, resources: dict) -> dict:
+        return {
+            "name": "gpu-defaults",
+            "model": {"path": "/m", "container": "/c.sqsh", "precision": "fp8"},
+            "resources": resources,
+        }
+
+    def test_recipe_without_gpu_type_inherits_default_gpu_type(self):
+        from srtctl.core.config import resolve_config_with_defaults
+
+        resolved = resolve_config_with_defaults(
+            self._recipe({"agg_nodes": 1, "agg_workers": 1}),
+            {"default_gpu_type": "gb200", "gpus_per_node": 4},
+        )
+        assert resolved["resources"]["gpu_type"] == "gb200"
+        assert resolved["resources"]["gpus_per_node"] == 4
+
+    def test_recipe_gpu_fields_win_over_cluster_defaults(self):
+        from srtctl.core.config import resolve_config_with_defaults
+
+        resolved = resolve_config_with_defaults(
+            self._recipe({"gpu_type": "h100", "gpus_per_node": 8, "agg_nodes": 1}),
+            {"default_gpu_type": "gb200", "gpus_per_node": 4},
+        )
+        assert resolved["resources"]["gpu_type"] == "h100"
+        assert resolved["resources"]["gpus_per_node"] == 8
+
+    def test_recipe_without_gpu_type_and_no_cluster_default_loads(self):
+        from srtctl.core.schema import SrtConfig
+
+        config = SrtConfig.Schema().load(self._recipe({"agg_nodes": 1, "agg_workers": 1}))
+        assert config.resources.gpu_type is None
+        assert config.resources.gpus_per_node == 4

tests/test_schema_docs.py

@@ -46,7 +46,12 @@ def test_marshmallow_data_key_wins_over_private_attribute_name() -> None:
     assert "gpus_per_decode" in rows
     assert "_explicit_gpus_per_prefill" not in rows
     assert rows["gpus_per_node"].default == "`4`"
-    assert rows["gpu_type"].default == "required"
+    assert rows["gpu_type"].default == "`None`"
+
+
+def test_required_field_renders_as_required() -> None:
+    rows = {row.key: row for row in field_docs(SrtConfig)}
+    assert rows["name"].default == "required"
 
 
 def test_docstring_attributes_become_descriptions() -> None: