[fine_tuning] Add Cluster scale up for Ilab testing

docs/toolbox.generated/Fine_Tuning.run_fine_tuning_job.rst

@@ -76,8 +76,6 @@ Parameters
 
 * The delimiter marking the beginning of the response in the dataset samples
 
-* default value: ``\n### Label:``
-
 
 ``container_image``  
 
@@ -143,3 +141,8 @@ Parameters
 
 * If true, sleeps forever instead of running the fine-tuning command.
 
+
+``ephemeral_output_pvc_size``  
+
+* If a size (with units) is passed, use an ephemeral volume claim for storing the fine-tuning output. Otherwise, use an emptyDir.
+

projects/fine_tuning/testing/command_args.yml.j2

@@ -20,15 +20,15 @@ sutest/cluster set_scale:
 gpu_operator enable_time_sharing:
   replicas: {{ gpu.time_sharing.replicas }}
 
-sutest/cluster preload_image/kserve-runtime:
+sutest/cluster preload_image:
   namespace: "{{ tests.fine_tuning.namespace }}"
   name: fine-tuning-image
 {% if tests.fine_tuning.fms.enabled %}
   image: {{ tests.fine_tuning.fms.image }}
 {% elif tests.fine_tuning.ray.enabled %}
-  image: {{ tests.fine_tuning.ray.enabled }}
+  image: {{ tests.fine_tuning.ray.image }}
 {% elif tests.fine_tuning.ilab.enabled %}
-  image: {{ tests.fine_tuning.ilab.enabled }}
+  image: {{ tests.fine_tuning.ilab.image }}
 {% else %}
   image: invalid_configuration
 {% endif %}

projects/fine_tuning/testing/config.yaml

@@ -17,6 +17,7 @@ ci_presets:
 
   light:
     extends: [light_cluster]
+    clusters.sutest.compute.dedicated: false
     tests.fine_tuning.matbenchmarking.enabled: false
     tests.fine_tuning.test_settings.gpu: null
     tests.fine_tuning.test_settings.dataset_replication: 1
@@ -26,6 +27,7 @@ ci_presets:
   gpu:
     gpu.prepare_cluster: true
     clusters.sutest.compute.machineset.type: g4dn.2xlarge
+    clusters.sutest.compute.machineset.count: 1
     tests.fine_tuning.test_settings.gpu: 1
 
   # ---
@@ -266,13 +268,10 @@ ci_presets:
 
   metal:
     clusters.sutest.is_metal: true
-    clusters.driver.is_metal: true
     clusters.sutest.compute.dedicated: false
-    clusters.driver.compute.dedicated: false
 
   not_metal:
     clusters.sutest.is_metal: false
-    clusters.driver.is_metal: false
 
   use_intlab_os:
     matbench.lts.opensearch.index_prefix: "psap-rhoai."
@@ -307,12 +306,17 @@ ci_presets:
   #
 
   fms:
-    tests.fine_tuning.fms.enabled: false
+    tests.fine_tuning.fms.enabled: true
     matbench.workload: projects.fine_tuning.visualizations.fms_hf_tuning
     matbench.prom_workload: projects.fine_tuning.visualizations.fms_prom
+    fine_tuning.default_response_template: "\n### Label:"
 
-  ray:
+  not_fms:
     tests.fine_tuning.fms.enabled: false
+    fine_tuning.default_response_template: null
+
+  ray:
+    extends: [not_fms]
     tests.fine_tuning.ray.enabled: true
     tests.capture_prom: false # not needed for the time being
     tests.fine_tuning.test_settings.hyper_parameters: {}
@@ -341,9 +345,14 @@ ci_presets:
   # ---
 
   ilab:
-    ci_presets.light["tests.fine_tuning.test_settings.dataset_name"]: sdg_data.jsonl
+    extends: [not_fms, gpu]
+    ci_presets.light["tests.fine_tuning.test_settings.dataset_name"]: ilab_skills_data.jsonl
     ci_presets.light["tests.fine_tuning.test_settings.model_name"]: ibm-granite/granite-3b-code-instruct@hf
-    tests.fine_tuning.fms.enabled: false
+    ci_presets.light["tests.fine_tuning.test_settings.gpu"]: 1
+    ci_presets.gpu["clusters.sutest.compute.machineset.type"]: g5.xlarge
+    clusters.sutest.compute.machineset.type: g5.xlarge
+
+
     tests.fine_tuning.ilab.enabled: true
     tests.fine_tuning.test_settings.name: ilab
     tests.fine_tuning.test_settings.dataset_name: ilab_skills_data.jsonl
@@ -360,20 +369,27 @@ ci_presets:
     extends: [ilab]
 
     tests.fine_tuning.test_settings.model_name: ibm-granite/granite-7b-base@hf
-    tests.fine_tuning.test_settings.dataset_name: [ilab_skills_data.jsonl, ilab_knowledge_data.jsonl]
+    tests.fine_tuning.test_settings.dataset_name: [ilab_large_10000samples_skills_data.jsonl, ilab_large_knowledge_data.jsonl]
 
-    tests.fine_tuning.test_settings.pod_count: 2
-    tests.fine_tuning.test_settings.gpu: 1
-
-    tests.fine_tuning.test_settings.hyper_parameters.num_epochs: 8
+    tests.fine_tuning.test_settings.pod_count: [1, 2, 4]
+    tests.fine_tuning.test_settings.gpu: 2
+    clusters.sutest.compute.machineset.count: 4
+    tests.fine_tuning.test_settings.hyper_parameters.num_epochs: 1
     tests.fine_tuning.matbenchmarking.enabled: true
     tests.fine_tuning.matbenchmarking.stop_on_error: true
 
   # ---
 
   cluster_instructlab:
-    clusters.sutest.compute.machineset.type: gx3-48x240x2l40s
     fine_tuning.pvc.storage_class_name: nfs-csi
+    fine_tuning.pvc.access_mode: ReadWriteMany
+    fine_tuning.pvc.size: 500Gi
+    clusters.sutest.is_metal: false
+    clusters.sutest.compute.machineset.type: gx3-48x240x2l40s
+    clusters.sutest.compute.machineset.name: instructlab-standalon-6rjg8-worker-1
+    clusters.sutest.compute.machineset.taint: null
+    clusters.sutest.compute.machineset.rest_count: 1
+    tests.fine_tuning.test_settings.ephemeral_output_pvc_size: 500Gi
 
   cluster_ibm_dgx:
     clusters.sutest.compute.machineset.type: "IBM-DGX A100-80GB"
@@ -445,22 +461,11 @@ clusters:
         name: workload-pods
         type: m6i.2xlarge
         count: null
+        rest_count: 0
         taint:
           key: only-workload-pods
           value: "yes"
           effect: NoSchedule
-  driver:
-    is_metal: false
-    compute:
-      dedicated: true
-      machineset:
-        name: test-pods
-        count: null
-        type: m6i.2xlarge
-        taint:
-          key: only-test-pods
-          value: "yes"
-          effect: NoSchedule
   cleanup_on_exit: false
 
 rhods:
@@ -488,6 +493,8 @@ fine_tuning:
     size: 80Gi
     storage_class_name: null
   model_registry: null # if set to a fine_tuning_sources.* model registry, all the lookups will be done in this registry
+  default_response_template: "\n### Label:"
+
   sources:
     dmf:
       type: model-registry
@@ -515,6 +522,8 @@ fine_tuning:
       source_dir: /manually/populated
       response_template: "\n### Response:"
 
+    # ---
+
     ilab_skills_data.jsonl:
       type: dataset
       source_dir: 's3://instructlab-standalone/data'
@@ -530,11 +539,18 @@ fine_tuning:
       source_dir: 's3://instructlab-standalone/data'
       secret_key: secrets.aws_credentials
 
+    ilab_large_10000samples_skills_data.jsonl:
+      type: dataset
+      source_dir: 's3://instructlab-standalone/data'
+      secret_key: secrets.aws_credentials
+
     ilab_large_knowledge_data.jsonl:
       type: dataset
       source_dir: 's3://instructlab-standalone/data'
       secret_key: secrets.aws_credentials
 
+    # ---
+
     alpaca_data.json:
       type: dataset
       source_dir: 'https://raw.githubusercontent.com/gururise/AlpacaDataCleaned/main'
@@ -569,6 +585,7 @@ tests:
       dataset_replication: 1
       pod_count: 1
       container_image: null
+      ephemeral_output_pvc_size: null
       # ---
       # https://huggingface.co/transformers/v3.0.2/main_classes/trainer.html
       hyper_parameters:

projects/fine_tuning/testing/prepare_finetuning.py

@@ -13,7 +13,7 @@
 def prepare():
     with run.Parallel("prepare1") as parallel:
         parallel.delayed(prepare_rhoai)
-        parallel.delayed(scale_up_sutest)
+        parallel.delayed(cluster_scale_up)
 
 
     test_settings = config.project.get_config("tests.fine_tuning.test_settings")
@@ -220,9 +220,9 @@ def prepare_namespace(test_settings):
     with env.NextArtifactDir("prepare_namespace"):
         set_namespace_annotations()
 
-        with run.Parallel("download") as parallel:
-            parallel.delayed(download_data_sources, test_settings)
-            parallel.delayed(preload_image)
+    with run.Parallel("prepare_data") as parallel:
+        parallel.delayed(download_data_sources, test_settings)
+        parallel.delayed(preload_image)
 
     if not dry_mode:
         run.run(f"oc delete pytorchjobs -n {namespace} --all")
@@ -237,18 +237,21 @@ def prepare_namespace(test_settings):
     prepare_kueue_queue(False, namespace, local_kueue_name)
 
 
-def scale_up_sutest():
+def cluster_scale_up(wait_gpu=False):
     if config.project.get_config("clusters.sutest.is_metal"):
         return
 
     node_count = config.project.get_config("clusters.sutest.compute.machineset.count")
 
     if node_count is None:
-        node_count = 1
+        logging.info("clusters.sutest.compute.machineset.count isn't set. Not touching the cluster scale.")
+        return
 
     extra = dict(scale=node_count)
     run.run_toolbox_from_config("cluster", "set_scale", prefix="sutest", extra=extra, artifact_dir_suffix="_sutest")
 
+    if wait_gpu:
+        prepare_gpu_operator.wait_ready(enable_time_sharing=False, wait_stack_deployed=True, wait_metrics=False)
 
 def cleanup_rhoai(mute=True):
     prepare_rhoai_mod.uninstall(mute)
@@ -262,27 +265,32 @@ def cleanup_cluster():
         cleanup_rhoai()
 
 
-def cleanup_sutest_ns():
-    cleanup_namespace_test()
-
-
 def cleanup_sutest_ns():
     namespace = config.project.get_config("tests.fine_tuning.namespace")
     # do not delete it ... (to save the PVC)
     # empty the namespace
 
 
-def cluster_scale_down(to_zero):
+def cluster_scale_down(to_zero=None):
     if config.project.get_config("clusters.sutest.is_metal"):
         return
 
+    if config.project.get_config("clusters.sutest.compute.machineset.count") is None:
+        logging.info("clusters.sutest.compute.machineset.count isn't set. Not touching the cluster scale.")
+        return
+
     machineset_name = config.project.get_config("clusters.sutest.compute.machineset.name")
     has_machineset = run.run(f"oc get machineset {machineset_name} -n openshift-machine-api -oname --ignore-not-found", capture_stdout=True).stdout
     if not has_machineset:
         logging.info(f"No {machineset_name} machineset. Nothing to scale down.")
         return
 
-    replicas = 0 if to_zero else 1
+    if to_zero:
+        replicas = 0
+    else:
+        replicas = config.project.get_config("clusters.sutest.compute.machineset.rest_count")
+        if replicas is None: replicas = 1
+
     run.run(f"oc scale --replicas={replicas} machineset/{machineset_name} -n openshift-machine-api")
 
 
@@ -297,6 +305,6 @@ def preload_image():
 
             break
         except Exception:
-            logging.warning(f"Preloading of '{image}' try #{i+1}/{RETRIES} failed :/")
+            logging.warning(f"Image preloading try #{i+1}/{RETRIES} failed :/")
             if i+1 == RETRIES:
                 raise

projects/fine_tuning/testing/test_finetuning.py

@@ -96,6 +96,8 @@ def _run_test(test_artifact_dir_p, test_override_values, job_index=None):
 
         if (response_template := dataset_source.get("response_template")) is not None:
             test_settings["dataset_response_template"] = response_template
+        elif (default_response_template := config.project.get_config("fine_tuning.default_response_template")) is not None:
+            test_settings["dataset_response_template"] = default_response_template
 
     remove_none_values(test_settings)
 
@@ -245,6 +247,7 @@ def _run_test_quality_evaluation(test_settings):
 
 def _run_test_and_visualize(test_override_values=None):
     failed = True
+    dry_mode = config.project.get_config("tests.dry_mode")
     do_matbenchmarking = test_override_values is None and config.project.get_config("tests.fine_tuning.matbenchmarking.enabled")
     do_multi_model = config.project.get_config("tests.fine_tuning.multi_model.enabled")
 
@@ -285,6 +288,10 @@ def _run_test_and_visualize(test_override_values=None):
         logging.error(msg)
         raise RuntimeError(msg)
 
+    if not dry_mode:
+        with env.NextArtifactDir("prepare_nodes"):
+            prepare_finetuning.cluster_scale_up(wait_gpu=True)
+
     test_artifact_dir_p = [None]
     try:
         if do_multi_model:
@@ -302,7 +309,9 @@ def _run_test_and_visualize(test_override_values=None):
             failed = _run_test(test_artifact_dir_p, test_override_values)
 
     finally:
-        dry_mode = config.project.get_config("tests.dry_mode")
+        if not dry_mode:
+            prepare_finetuning.cluster_scale_down()
+
         if not config.project.get_config("tests.visualize"):
             logging.info(f"Visualization disabled.")
 

projects/fine_tuning/toolbox/fine_tuning.py

@@ -27,7 +27,7 @@ def run_fine_tuning_job(
             dataset_transform=None,
             dataset_prefer_cache=True,
             dataset_prepare_cache_only=False,
-            dataset_response_template="\n### Label:",
+            dataset_response_template=None,
             container_image="quay.io/modh/fms-hf-tuning:release-7a8ff0f4114ba43398d34fd976f6b17bb1f665f3",
 
             gpu=0,
@@ -44,6 +44,8 @@ def run_fine_tuning_job(
 
             capture_artifacts=True,
             sleep_forever=False,
+
+            ephemeral_output_pvc_size=None,
     ):
         """
         Run a simple fine-tuning Job.
@@ -77,6 +79,8 @@ def run_fine_tuning_job(
 
           capture_artifacts: if enabled, captures the artifacts that will help post-mortem analyses
           sleep_forever: if true, sleeps forever instead of running the fine-tuning command.
+
+          ephemeral_output_pvc_size: if a size (with units) is passed, use an ephemeral volume claim for storing the fine-tuning output. Otherwise, use an emptyDir.
         """
 
         return RunAnsibleRole(locals())

projects/fine_tuning/toolbox/fine_tuning_run_fine_tuning_job/defaults/main/config.yml

@@ -40,9 +40,7 @@ fine_tuning_run_fine_tuning_job_dataset_prefer_cache: true
 fine_tuning_run_fine_tuning_job_dataset_prepare_cache_only: false
 
 # the delimiter marking the beginning of the response in the dataset samples
-fine_tuning_run_fine_tuning_job_dataset_response_template: '
-
-  ### Label:'
+fine_tuning_run_fine_tuning_job_dataset_response_template: null
 
 # the image to use for the fine-tuning container
 fine_tuning_run_fine_tuning_job_container_image: quay.io/modh/fms-hf-tuning:release-7a8ff0f4114ba43398d34fd976f6b17bb1f665f3
@@ -76,3 +74,6 @@ fine_tuning_run_fine_tuning_job_capture_artifacts: true
 
 # if true, sleeps forever instead of running the fine-tuning command.
 fine_tuning_run_fine_tuning_job_sleep_forever: false
+
+# if a size (with units) is passed, use an ephemeral volume claim for storing the fine-tuning output. Otherwise, use an emptyDir.
+fine_tuning_run_fine_tuning_job_ephemeral_output_pvc_size: null

projects/fine_tuning/toolbox/fine_tuning_run_fine_tuning_job/tasks/main.yml

@@ -161,7 +161,7 @@
          -n {{ fine_tuning_run_fine_tuning_job_namespace }}
          --no-headers | awk '{print $3}'
     register: wait_pod_start
-    retries: 720
+    retries: 9999
     delay: 30
     until: wait_pod_start.stdout != "Running"