[LLM] Custom models

doc/workloads/sglang.rst

@@ -70,6 +70,28 @@ supported under ``[[Tests]]`` in a test scenario. Define them in a test definiti
 ``test_name`` when custom benchmark or semantic-evaluation arguments are needed.
 
 
+Local Models
+------------
+Set ``cmd_args.model`` to an absolute, container-visible path to load an existing model directory instead of
+downloading a model from Hugging Face. Use ``extra_container_mounts`` to mount a host or shared-filesystem directory.
+
+.. code-block:: toml
+   :caption: test.toml (local model)
+
+   name = "sglang_local_model"
+   description = "SGLang with a local model"
+   test_template_name = "sglang"
+   extra_container_mounts = [
+     "/lustre/models/custom:/models/custom:ro",
+   ]
+
+   [cmd_args]
+   docker_image_url = "lmsysorg/sglang:dev-cu13"
+   model = "/models/custom"
+
+The mounted directory must contain a model, configuration, and tokenizer supported by SGLang.
+
+
 Semantic Validation
 -------------------
 To run GSM8K semantic validation after the serving benchmark, add ``semantic_eval_cmd_args``. CloudAI reports

doc/workloads/vllm.rst

@@ -70,6 +70,28 @@ supported under ``[[Tests]]`` in a test scenario. Define them in a test definiti
 ``test_name`` when custom benchmark or semantic-evaluation arguments are needed.
 
 
+Local Models
+------------
+Set ``cmd_args.model`` to an absolute, container-visible path to load an existing model directory instead of
+downloading a model from Hugging Face. Use ``extra_container_mounts`` to mount a host or shared-filesystem directory.
+
+.. code-block:: toml
+   :caption: test.toml (local model)
+
+   name = "vllm_local_model"
+   description = "vLLM with a local model"
+   test_template_name = "vllm"
+   extra_container_mounts = [
+     "/lustre/models/custom:/models/custom:ro",
+   ]
+
+   [cmd_args]
+   docker_image_url = "vllm/vllm-openai:v0.14.0-cu130"
+   model = "/models/custom"
+
+The mounted directory must contain a model, configuration, and tokenizer supported by vLLM.
+
+
 Semantic Validation
 -------------------
 To run GSM8K semantic validation after the serving benchmark, add ``semantic_eval_cmd_args``. CloudAI reports

src/cloudai/workloads/common/llm_serving.py

@@ -190,7 +190,10 @@ def extra_installables(self) -> list[Installable]:
 
     @property
     def installables(self) -> list[Installable]:
-        return [*self.git_repos, self.docker_image, self.hf_model, *self.extra_installables]
+        installables: list[Installable] = [*self.git_repos, self.docker_image]
+        if not Path(self.cmd_args.model).is_absolute():
+            installables.append(self.hf_model)
+        return [*installables, *self.extra_installables]
 
     def model_post_init(self, __context: Any) -> None:
         super().model_post_init(__context)

tests/workloads/sglang/test_workload.py

@@ -14,8 +14,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from cloudai.workloads.sglang import SglangArgs
+from cloudai.workloads.sglang import SglangArgs, SglangCmdArgs, SglangTestDefinition
 
 
 def test_sglang_serve_args_exclude_internal_fields() -> None:
     assert SglangArgs(gpu_ids="0", disaggregation_transfer_backend="nccl").serve_args == []
+
+
+def test_local_model_is_not_installed_from_hugging_face() -> None:
+    tdef = SglangTestDefinition(
+        name="test",
+        description="test",
+        test_template_name="sglang",
+        cmd_args=SglangCmdArgs(
+            docker_image_url="test_url",
+            model="/models/custom-model",
+        ),
+    )
+
+    assert tdef.installables == [tdef.docker_image]

tests/workloads/vllm/test_workload.py

@@ -54,6 +54,22 @@ def test_installables_include_proxy_script_repo() -> None:
     assert tdef.installables == [tdef.docker_image, tdef.hf_model, proxy_script_repo]
 
 
+def test_local_model_is_not_installed_from_hugging_face() -> None:
+    proxy_script_repo = GitRepo(url="./proxy_script_repo", commit="commit")
+    tdef = VllmTestDefinition(
+        name="test",
+        description="test",
+        test_template_name="vllm",
+        cmd_args=VllmCmdArgs(
+            docker_image_url="test_url",
+            model="/models/custom-model",
+        ),
+        proxy_script_repo=proxy_script_repo,
+    )
+
+    assert tdef.installables == [tdef.docker_image, proxy_script_repo]
+
+
 def test_constraint_check_rejects_tp_pp_dp_above_available_gpus(tmp_path) -> None:
     tdef = VllmTestDefinition(
         name="test",