chore(misc): make tests and starter faster (#3042)

A bunch of miscellaneous cleanup focusing on tests, but ended up
speeding up starter distro substantially.

- Pulled llama stack client init for tests into `pytest_sessionstart` so
it does not clobber output
- Profiling of that told me where we were doing lots of heavy imports
for starter, so lazied them
- starter now starts 20seconds+ faster on my Mac
- A few other smallish refactors for `compat_client`

llama_stack/providers/inline/datasetio/localfs/datasetio.py

@@ -5,8 +5,6 @@
 # the root directory of this source tree.
 from typing import Any
 
-import pandas
-
 from llama_stack.apis.common.responses import PaginatedResponse
 from llama_stack.apis.datasetio import DatasetIO
 from llama_stack.apis.datasets import Dataset
@@ -44,6 +42,8 @@ class PandasDataframeDataset:
         if self.dataset_def.source.type == "uri":
             self.df = await get_dataframe_from_uri(self.dataset_def.source.uri)
         elif self.dataset_def.source.type == "rows":
+            import pandas
+
             self.df = pandas.DataFrame(self.dataset_def.source.rows)
         else:
             raise ValueError(f"Unsupported dataset source type: {self.dataset_def.source.type}")
@@ -103,6 +103,8 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
         return paginate_records(records, start_index, limit)
 
     async def append_rows(self, dataset_id: str, rows: list[dict[str, Any]]) -> None:
+        import pandas
+
         dataset_def = self.dataset_infos[dataset_id]
         dataset_impl = PandasDataframeDataset(dataset_def)
         await dataset_impl.load()

llama_stack/providers/inline/post_training/huggingface/config.py

@@ -71,8 +71,13 @@ class HuggingFacePostTrainingConfig(BaseModel):
     dpo_beta: float = 0.1
     use_reference_model: bool = True
     dpo_loss_type: Literal["sigmoid", "hinge", "ipo", "kto_pair"] = "sigmoid"
-    dpo_output_dir: str = "./checkpoints/dpo"
+    dpo_output_dir: str
 
     @classmethod
     def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> dict[str, Any]:
-        return {"checkpoint_format": "huggingface", "distributed_backend": None, "device": "cpu"}
+        return {
+            "checkpoint_format": "huggingface",
+            "distributed_backend": None,
+            "device": "cpu",
+            "dpo_output_dir": __distro_dir__ + "/dpo_output",
+        }

llama_stack/providers/inline/post_training/huggingface/post_training.py

@@ -22,15 +22,8 @@ from llama_stack.apis.post_training import (
 from llama_stack.providers.inline.post_training.huggingface.config import (
     HuggingFacePostTrainingConfig,
 )
-from llama_stack.providers.inline.post_training.huggingface.recipes.finetune_single_device import (
-    HFFinetuningSingleDevice,
-)
-from llama_stack.providers.inline.post_training.huggingface.recipes.finetune_single_device_dpo import (
-    HFDPOAlignmentSingleDevice,
-)
 from llama_stack.providers.utils.scheduler import JobArtifact, Scheduler
 from llama_stack.providers.utils.scheduler import JobStatus as SchedulerJobStatus
-from llama_stack.schema_utils import webmethod
 
 
 class TrainingArtifactType(Enum):
@@ -85,6 +78,10 @@ class HuggingFacePostTrainingImpl:
         algorithm_config: AlgorithmConfig | None = None,
     ) -> PostTrainingJob:
         async def handler(on_log_message_cb, on_status_change_cb, on_artifact_collected_cb):
+            from llama_stack.providers.inline.post_training.huggingface.recipes.finetune_single_device import (
+                HFFinetuningSingleDevice,
+            )
+
             on_log_message_cb("Starting HF finetuning")
 
             recipe = HFFinetuningSingleDevice(
@@ -124,6 +121,10 @@ class HuggingFacePostTrainingImpl:
         logger_config: dict[str, Any],
     ) -> PostTrainingJob:
         async def handler(on_log_message_cb, on_status_change_cb, on_artifact_collected_cb):
+            from llama_stack.providers.inline.post_training.huggingface.recipes.finetune_single_device_dpo import (
+                HFDPOAlignmentSingleDevice,
+            )
+
             on_log_message_cb("Starting HF DPO alignment")
 
             recipe = HFDPOAlignmentSingleDevice(
@@ -168,7 +169,6 @@ class HuggingFacePostTrainingImpl:
         data = cls._get_artifacts_metadata_by_type(job, TrainingArtifactType.RESOURCES_STATS.value)
         return data[0] if data else None
 
-    @webmethod(route="/post-training/job/status")
     async def get_training_job_status(self, job_uuid: str) -> PostTrainingJobStatusResponse | None:
         job = self._scheduler.get_job(job_uuid)
 
@@ -195,16 +195,13 @@ class HuggingFacePostTrainingImpl:
             resources_allocated=self._get_resources_allocated(job),
         )
 
-    @webmethod(route="/post-training/job/cancel")
     async def cancel_training_job(self, job_uuid: str) -> None:
         self._scheduler.cancel(job_uuid)
 
-    @webmethod(route="/post-training/job/artifacts")
     async def get_training_job_artifacts(self, job_uuid: str) -> PostTrainingJobArtifactsResponse | None:
         job = self._scheduler.get_job(job_uuid)
         return PostTrainingJobArtifactsResponse(job_uuid=job_uuid, checkpoints=self._get_checkpoints(job))
 
-    @webmethod(route="/post-training/jobs", method="GET")
     async def get_training_jobs(self) -> ListPostTrainingJobsResponse:
         return ListPostTrainingJobsResponse(
             data=[PostTrainingJob(job_uuid=job.id) for job in self._scheduler.get_jobs()]

llama_stack/providers/inline/post_training/torchtune/post_training.py

@@ -23,12 +23,8 @@ from llama_stack.apis.post_training import (
 from llama_stack.providers.inline.post_training.torchtune.config import (
     TorchtunePostTrainingConfig,
 )
-from llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device import (
-    LoraFinetuningSingleDevice,
-)
 from llama_stack.providers.utils.scheduler import JobArtifact, Scheduler
 from llama_stack.providers.utils.scheduler import JobStatus as SchedulerJobStatus
-from llama_stack.schema_utils import webmethod
 
 
 class TrainingArtifactType(Enum):
@@ -84,6 +80,10 @@ class TorchtunePostTrainingImpl:
         if isinstance(algorithm_config, LoraFinetuningConfig):
 
             async def handler(on_log_message_cb, on_status_change_cb, on_artifact_collected_cb):
+                from llama_stack.providers.inline.post_training.torchtune.recipes.lora_finetuning_single_device import (
+                    LoraFinetuningSingleDevice,
+                )
+
                 on_log_message_cb("Starting Lora finetuning")
 
                 recipe = LoraFinetuningSingleDevice(
@@ -144,7 +144,6 @@ class TorchtunePostTrainingImpl:
         data = cls._get_artifacts_metadata_by_type(job, TrainingArtifactType.RESOURCES_STATS.value)
         return data[0] if data else None
 
-    @webmethod(route="/post-training/job/status")
     async def get_training_job_status(self, job_uuid: str) -> PostTrainingJobStatusResponse | None:
         job = self._scheduler.get_job(job_uuid)
 
@@ -171,11 +170,9 @@ class TorchtunePostTrainingImpl:
             resources_allocated=self._get_resources_allocated(job),
         )
 
-    @webmethod(route="/post-training/job/cancel")
     async def cancel_training_job(self, job_uuid: str) -> None:
         self._scheduler.cancel(job_uuid)
 
-    @webmethod(route="/post-training/job/artifacts")
     async def get_training_job_artifacts(self, job_uuid: str) -> PostTrainingJobArtifactsResponse | None:
         job = self._scheduler.get_job(job_uuid)
         return PostTrainingJobArtifactsResponse(job_uuid=job_uuid, checkpoints=self._get_checkpoints(job))