@@ -1010,101 +1010,36 @@ class TestJumpStartAdditionalModelDataSourcesUserFlow(unittest.TestCase):
10101010 """End-to-end tests for additional model data sources, asserting the
10111011 CreateModel request shape.
10121012
1013- Each test uses the documented single-call customer API:
1014-
1015- builder = ModelBuilder.from_jumpstart_config(
1016- jumpstart_config=JumpStartConfig(
1017- model_id="...", accept_eula=True, inference_config_name="..."),
1018- compute=Compute(instance_type="..."), ...)
1019- builder.build()
1020-
1021- and asserts on what the SDK sends to the CreateModel API -- the
1022- ``container_defs`` captured from ``sagemaker_session.create_model`` -- not
1023- on builder internals. This is the request boto receives, so the assertions
1024- cover the full downstream impact: primary container image, model data
1025- source, and AdditionalModelDataSources.
1013+ Each test drives the documented single-call customer API
1014+ (ModelBuilder.from_jumpstart_config(...) then build()) and asserts on the
1015+ ``container_defs`` captured from ``sagemaker_session.create_model`` -- the
1016+ request boto receives -- never on builder internals. Expected values are
1017+ hand-written literals so the wire contract is stated, not computed.
10261018
10271019 The spec fixtures under tests/unit/servers/data/jumpstart_specs/ are real,
1028- unmodified specs captured from the production JumpStart content bucket
1029- (jumpstart-cache-prod-us-west-2):
1020+ unmodified specs captured from the production JumpStart content bucket:
10301021
10311022 - pytorch-ic-mobilenet-v2: public model, no additional data sources.
1032- - openai-reasoning-gpt-oss-20b: public model whose default (lmi-optimized)
1033- config carries an ungated EAGLE speculative-decoding source -- the model
1034- that surfaced the propagation bug this suite guards.
1035- - meta-textgeneration-llama-3-1-70b: gated model whose lmi-optimized
1036- config carries a GATED draft_model source
1037- (hosting_eula_key=fmhMetadata/eula/llama3_2Eula.txt).
1038-
1039- Patching is centralized in setUp on the same seams the legacy master-v2
1040- JumpStartModel tests patched:
1041-
1042- 1. The spec-fetch boundary: JumpStartModelsAccessor.get_model_specs routes
1043- by model id into the captured spec files (master-v2's
1044- PROTOTYPICAL_MODEL_SPECS_DICT pattern); _get_manifest serves the
1045- captured manifest headers.
1046- 2. The AWS boundary: a mock session (whose create_model call is the
1047- assertion target), IAM role validation, artifact staging
1048- (_prepare_for_mode), and the post-create Model.get describe call.
1049-
1050- Everything in between runs for real: JumpStart model-id detection,
1051- JumpStartModelSpecs parsing, inference-config resolution, the
1052- get_init_kwargs factory (image URI resolution, content-bucket injection,
1053- PascalCase shaping), the propagation in _build_for_jumpstart, the
1054- accept_eula application in container_def (shared with the primary model),
1055- and _prepare_container_def_base assembling the CreateModel request.
1023+ - openai-reasoning-gpt-oss-20b: default (lmi-optimized) config carries an
1024+ ungated EAGLE speculative-decoding source -- the model that surfaced the
1025+ propagation bug this suite guards.
1026+ - meta-textgeneration-llama-3-1-70b: lmi-optimized config carries a GATED
1027+ draft_model source (hosting_eula_key=fmhMetadata/eula/llama3_2Eula.txt).
1028+
1029+ setUp patches only the spec-fetch boundary (get_model_specs/_get_manifest
1030+ serve the captured files, master-v2's PROTOTYPICAL_MODEL_SPECS_DICT
1031+ pattern) and the AWS boundary (mock session whose create_model call is the
1032+ assertion target, role validation, artifact staging, post-create
1033+ describe). Everything in between runs for real: model-id detection, spec
1034+ parsing, inference-config resolution, the get_init_kwargs factory
1035+ (image URI resolution, content-bucket injection, PascalCase shaping), and
1036+ the HostingEulaKey strip + accept_eula fold in _build_for_jumpstart.
10561037 """
10571038
10581039 SPEC_DIR = Path (__file__ ).parent / "data" / "jumpstart_specs"
10591040
10601041 ROLE_ARN = "arn:aws:iam::123456789012:role/SageMakerRole"
10611042
1062- # The content buckets the sources must resolve into (the raw specs publish
1063- # bucket-less key prefixes): public sources land in the public content
1064- # bucket, gated sources in the private one.
1065- PUBLIC_CONTENT_BUCKET = "jumpstart-cache-prod-us-west-2"
1066- PRIVATE_CONTENT_BUCKET = "jumpstart-private-cache-prod-us-west-2"
1067-
1068- @classmethod
1069- def _spec_additional_source (cls , model_id , config_name = "lmi-optimized" ):
1070- """The raw additional data source exactly as published in the captured
1071- spec fixture: snake_case keys, bucket-less s3_uri key prefix, and
1072- hosting_eula_key present when the source is gated."""
1073- with open (cls .SPEC_DIR / f"{ model_id } .json" ) as f :
1074- spec = json .load (f )
1075- sources = spec ["inference_config_components" ][config_name ][
1076- "hosting_additional_data_sources"
1077- ]["speculative_decoding" ]
1078- assert len (sources ) == 1 , f"expected one source in { model_id } /{ config_name } "
1079- return sources [0 ]
1080-
1081- @classmethod
1082- def _expected_create_model_source (cls , model_id , bucket , accept_eula = None ):
1083- """The shape the CreateModel request must carry for the fixture's
1084- source. Data values (channel name, compression, data type, key prefix)
1085- come from the spec fixture so it stays the single source of truth; the
1086- transformation under test is spelled out declaratively here instead of
1087- reusing production code (which would make the assertion a tautology):
1088-
1089- - snake_case spec keys become the PascalCase API keys
1090- - the bucket-less key prefix resolves into the given content bucket
1091- - HostingEulaKey never appears (the API rejects it)
1092- - accept_eula, when set, is folded into S3DataSource.ModelAccessConfig
1093- exactly like the primary model's ModelDataSource
1094- """
1095- raw = cls ._spec_additional_source (model_id )
1096- expected = {
1097- "ChannelName" : raw ["channel_name" ],
1098- "S3DataSource" : {
1099- "CompressionType" : raw ["s3_data_source" ]["compression_type" ],
1100- "S3DataType" : raw ["s3_data_source" ]["s3_data_type" ],
1101- "S3Uri" : f"s3://{ bucket } /{ raw ['s3_data_source' ]['s3_uri' ]} " ,
1102- },
1103- }
1104- if accept_eula is not None :
1105- expected ["S3DataSource" ]["ModelAccessConfig" ] = {"AcceptEula" : accept_eula }
1106- return expected
1107-
11081043 def setUp (self ):
11091044 import sagemaker .serve .model_builder as model_builder_module
11101045 from sagemaker .core .jumpstart .accessors import JumpStartModelsAccessor
@@ -1182,120 +1117,94 @@ def test_model_without_additional_sources_sends_none_to_create_model(self):
11821117 self .assertNotIn ("AdditionalModelDataSources" , container_def )
11831118
11841119 def test_ungated_additional_source_reaches_create_model_without_eula (self ):
1185- """The bug this suite guards: an ungated speculative-decoding source
1186- in the model's default config must reach CreateModel even though the
1187- customer never touches accept_eula."""
1120+ """The bug this suite guards: an ungated speculative-decoding source in
1121+ the model's default config must reach CreateModel even though the
1122+ customer never touches accept_eula. This is the one exact-shape pin:
1123+ PascalCase keys, key prefix resolved into the public content bucket,
1124+ no HostingEulaKey, no ModelAccessConfig."""
11881125 self ._build ("openai-reasoning-gpt-oss-20b" , instance_type = "ml.g7e.2xlarge" )
11891126
11901127 container_def = self ._create_model_container_def ()
11911128 self .assertEqual (
11921129 container_def ["AdditionalModelDataSources" ],
11931130 [
1194- self ._expected_create_model_source (
1195- "openai-reasoning-gpt-oss-20b" , bucket = self .PUBLIC_CONTENT_BUCKET
1196- )
1131+ {
1132+ "ChannelName" : "eagle" ,
1133+ "S3DataSource" : {
1134+ "CompressionType" : "None" ,
1135+ "S3DataType" : "S3Prefix" ,
1136+ "S3Uri" : "s3://jumpstart-cache-prod-us-west-2/"
1137+ "lmi-eagle-heads/models/gpt-oss-20b-p-eagle/" ,
1138+ },
1139+ }
11971140 ],
11981141 )
11991142
12001143 def test_accept_eula_applies_to_every_model_data_source_uniformly (self ):
12011144 """The single accept_eula knob folds the same ModelAccessConfig into
1202- the primary model AND each additional source, exactly like the primary
1203- model's ModelDataSource handling. Gatedness is not the SDK's call: the
1204- service determines it from the artifact's bucket and ignores the config
1205- on ungated sources."""
1145+ the primary model AND the (ungated) additional source. Gatedness is
1146+ not the SDK's call: the service resolves it from the bucket and
1147+ ignores the config on ungated sources."""
12061148 self ._build (
12071149 "openai-reasoning-gpt-oss-20b" , accept_eula = True , instance_type = "ml.g7e.2xlarge"
12081150 )
12091151
12101152 container_def = self ._create_model_container_def ()
1211- self .assertEqual (
1212- container_def ["AdditionalModelDataSources" ],
1213- [
1214- self ._expected_create_model_source (
1215- "openai-reasoning-gpt-oss-20b" ,
1216- bucket = self .PUBLIC_CONTENT_BUCKET ,
1217- accept_eula = True ,
1218- )
1219- ],
1220- )
1153+ (source ,) = container_def ["AdditionalModelDataSources" ]
1154+ self .assertEqual (source ["S3DataSource" ]["ModelAccessConfig" ], {"AcceptEula" : True })
12211155 self .assertEqual (
12221156 container_def ["ModelDataSource" ]["S3DataSource" ]["ModelAccessConfig" ],
12231157 {"AcceptEula" : True },
12241158 )
12251159
12261160 def test_gated_additional_source_with_accepted_eula_sends_model_access_config (self ):
1227- """Accepting the EULA folds ModelAccessConfig into the gated
1228- additional source AND the primary model in the CreateModel request."""
1161+ """Accepting the EULA folds ModelAccessConfig into the gated draft
1162+ model source, whose URI resolves into the private content bucket. The
1163+ spec-internal HostingEulaKey never reaches the request."""
12291164 self ._build (
12301165 "meta-textgeneration-llama-3-1-70b" ,
12311166 accept_eula = True ,
12321167 inference_config_name = "lmi-optimized" ,
12331168 instance_type = "ml.p4d.24xlarge" ,
12341169 )
12351170
1236- container_def = self ._create_model_container_def ()
1237- self .assertEqual (
1238- container_def ["AdditionalModelDataSources" ],
1239- [
1240- self ._expected_create_model_source (
1241- "meta-textgeneration-llama-3-1-70b" ,
1242- bucket = self .PRIVATE_CONTENT_BUCKET ,
1243- accept_eula = True ,
1244- )
1245- ],
1246- )
1171+ (source ,) = self ._create_model_container_def ()["AdditionalModelDataSources" ]
1172+ self .assertEqual (source ["ChannelName" ], "draft_model" )
1173+ self .assertEqual (source ["S3DataSource" ]["ModelAccessConfig" ], {"AcceptEula" : True })
12471174 self .assertEqual (
1248- container_def ["ModelDataSource" ]["S3DataSource" ]["ModelAccessConfig" ],
1249- {"AcceptEula" : True },
1175+ source ["S3DataSource" ]["S3Uri" ],
1176+ "s3://jumpstart-private-cache-prod-us-west-2/meta-textgeneration/"
1177+ "meta-textgeneration-llama-3-2-1b/artifacts/inference-prepack/v1.0.0/" ,
12501178 )
1179+ self .assertNotIn ("HostingEulaKey" , source )
12511180
12521181 def test_gated_additional_source_without_eula_sends_no_model_access_config (self ):
1253- """No EULA decision: the request carries the gated source without any
1254- ModelAccessConfig. Enforcement is the service's: the SageMaker control
1255- plane resolves the private- bucket URI as gated and rejects CreateModel
1256- with an EULA validation error, the same as for a gated primary model ."""
1182+ """No EULA decision: the gated source goes out with no
1183+ ModelAccessConfig (and no HostingEulaKey) . Enforcement is the
1184+ service's: the control plane resolves the private bucket as gated and
1185+ rejects CreateModel with an EULA validation error."""
12571186 self ._build (
12581187 "meta-textgeneration-llama-3-1-70b" ,
12591188 inference_config_name = "lmi-optimized" ,
12601189 instance_type = "ml.p4d.24xlarge" ,
12611190 )
12621191
1263- container_def = self ._create_model_container_def ()
1264- self .assertEqual (
1265- container_def ["AdditionalModelDataSources" ],
1266- [
1267- self ._expected_create_model_source (
1268- "meta-textgeneration-llama-3-1-70b" , bucket = self .PRIVATE_CONTENT_BUCKET
1269- )
1270- ],
1271- )
1192+ (source ,) = self ._create_model_container_def ()["AdditionalModelDataSources" ]
1193+ self .assertNotIn ("ModelAccessConfig" , source ["S3DataSource" ])
1194+ self .assertNotIn ("HostingEulaKey" , source )
12721195
12731196 def test_gated_additional_source_with_rejected_eula_sends_acceptance_false (self ):
12741197 """Explicit accept_eula=False is transmitted faithfully as
1275- ModelAccessConfig={"AcceptEula": False} on every source, exactly like
1276- the primary model. The service rejects the gated sources."""
1198+ ModelAccessConfig={"AcceptEula": False}; the service rejects it."""
12771199 self ._build (
12781200 "meta-textgeneration-llama-3-1-70b" ,
12791201 accept_eula = False ,
12801202 inference_config_name = "lmi-optimized" ,
12811203 instance_type = "ml.p4d.24xlarge" ,
12821204 )
12831205
1284- container_def = self ._create_model_container_def ()
1285- self .assertEqual (
1286- container_def ["AdditionalModelDataSources" ],
1287- [
1288- self ._expected_create_model_source (
1289- "meta-textgeneration-llama-3-1-70b" ,
1290- bucket = self .PRIVATE_CONTENT_BUCKET ,
1291- accept_eula = False ,
1292- )
1293- ],
1294- )
1295- self .assertEqual (
1296- container_def ["ModelDataSource" ]["S3DataSource" ]["ModelAccessConfig" ],
1297- {"AcceptEula" : False },
1298- )
1206+ (source ,) = self ._create_model_container_def ()["AdditionalModelDataSources" ]
1207+ self .assertEqual (source ["S3DataSource" ]["ModelAccessConfig" ], {"AcceptEula" : False })
12991208
13001209 def test_unselected_config_sources_do_not_leak_into_create_model (self ):
13011210 """Config resolution gates which sources apply: the same gated model on
0 commit comments