fix(vertex_ai): keep the old batch create error text and Google object names

A failed Vertex batch create keeps httpx's error line first and adds Vertex's body after it, and logs the status and body again. Only Claude object names keep the @ version, so Google models are stored under the same names as before
This commit is contained in:
mateo-berri 2026-10-29 13:34:13 -07:00
parent 757021ab19
commit 79cc204c54
4 changed files with 37 additions and 8 deletions

View file

@ -86,7 +86,7 @@ def _vertex_batch_list_payload(response: _VertexBatchListJsonSource) -> dict[str
def _vertex_batch_create_error(error: httpx.HTTPStatusError) -> MaskedHTTPStatusError:
body: Final = mask_sensitive_info(error.response.text)
return MaskedHTTPStatusError(
httpx.HTTPStatusError(f"Vertex AI batch create failed: {body}", request=error.request, response=error.response),
httpx.HTTPStatusError(f"{error}\n{body}", request=error.request, response=error.response),
message=body,
text=body,
)
@ -184,6 +184,11 @@ class VertexAIBatchPrediction(VertexLLM):
data=json.dumps(vertex_batch_request),
)
except httpx.HTTPStatusError as e:
litellm.verbose_logger.error(
"Vertex AI batch create failed: status=%s, body=%s",
e.response.status_code,
e.response.text[:1000],
)
raise _vertex_batch_create_error(e) from e
payload_view: Final[_VertexBatchPayloadView] = {"payload": response.json()}
@ -280,6 +285,11 @@ class VertexAIBatchPrediction(VertexLLM):
data=json.dumps(vertex_batch_request),
)
except httpx.HTTPStatusError as e:
litellm.verbose_logger.error(
"Vertex AI batch create failed: status=%s, body=%s",
e.response.status_code,
e.response.text[:1000],
)
raise _vertex_batch_create_error(e) from e
payload_view: Final[_VertexBatchPayloadView] = {"payload": response.json()}

View file

@ -1115,11 +1115,16 @@ def _publisher_model_path(raw_model: str) -> str:
def _managed_batch_object_name(raw_model: str, *, passthrough: bool) -> str:
endpoint_id: Final = get_vertex_ai_fine_tuned_endpoint_id(raw_model)
model_path: Final = f"endpoints/{endpoint_id}" if endpoint_id is not None else _publisher_model_path(raw_model)
safe_model_path: Final = "@".join(
prefix: Final = _PASSTHROUGH_MANAGED_GCS_PREFIX if passthrough else VERTEX_AI_MANAGED_GCS_PREFIX
return f"{prefix}{_safe_managed_model_path(model_path)}/{uuid.uuid4()}"
def _safe_managed_model_path(model_path: str) -> str:
if not is_vertex_anthropic_batch_model(model_path):
return sanitize_cloud_object_path(model_path, fallback="model")
return "@".join(
sanitize_cloud_object_path(version_part, fallback="model") for version_part in model_path.split("@")
)
prefix: Final = _PASSTHROUGH_MANAGED_GCS_PREFIX if passthrough else VERTEX_AI_MANAGED_GCS_PREFIX
return f"{prefix}{safe_model_path}/{uuid.uuid4()}"
class VertexAIFilesConfig(VertexBase, BaseFilesConfig):

View file

@ -46,6 +46,7 @@ from litellm.proxy.utils import handle_exception_on_proxy # noqa: E402
from litellm.types.utils import LiteLLMBatch # noqa: E402
HMOD = "litellm.llms.vertex_ai.batches.handler"
_CLIENT_ERROR_429: Final = "Client error '429 Too Many Requests' for url 'https://x/batchPredictionJobs'"
TOKEN = "ya29.fake-access-token"
PROJECT = "my-project"
LOCATION = "us-central1"
@ -380,11 +381,11 @@ def test_create_batch_sync_httpstatuserror_carries_vertex_status_body_and_header
status_code=429, request=request, text="boom", headers={"retry-after": "30"}
)
client.post.side_effect = httpx.HTTPStatusError(
"boom", request=request, response=err_response
_CLIENT_ERROR_429, request=request, response=err_response
)
with patch(f"{HMOD}.get_httpx_client", return_value=client):
with pytest.raises(httpx.HTTPStatusError, match="Vertex AI batch create failed: boom") as exc_info:
with pytest.raises(httpx.HTTPStatusError) as exc_info:
h.create_batch(
_is_async=False,
create_batch_data=CREATE_DATA,
@ -399,6 +400,7 @@ def test_create_batch_sync_httpstatuserror_carries_vertex_status_body_and_header
assert exc_info.value.response.headers["retry-after"] == "30"
assert handle_exception_on_proxy(exc_info.value).code == "429"
assert (getattr(exc_info.value, "message", None), getattr(exc_info.value, "text", None)) == ("boom", "boom")
assert str(exc_info.value).splitlines() == [_CLIENT_ERROR_429, "boom"]
def test_create_batch_input_file_id_without_model_raises_400_before_post():
@ -815,7 +817,7 @@ def test_create_batch_async_httpstatuserror_carries_vertex_status_body_and_heade
)
async_client.post = AsyncMock(
side_effect=httpx.HTTPStatusError(
"boom", request=request, response=err_response
_CLIENT_ERROR_429, request=request, response=err_response
)
)
@ -833,12 +835,13 @@ def test_create_batch_async_httpstatuserror_carries_vertex_status_body_and_heade
timeout=600.0,
max_retries=None,
)
with pytest.raises(httpx.HTTPStatusError, match="Vertex AI batch create failed: boom") as exc_info:
with pytest.raises(httpx.HTTPStatusError) as exc_info:
_run(coro)
assert exc_info.value.response.status_code == 429
assert exc_info.value.response.headers["retry-after"] == "30"
assert handle_exception_on_proxy(exc_info.value).code == "429"
assert (getattr(exc_info.value, "message", None), getattr(exc_info.value, "text", None)) == ("boom", "boom")
assert str(exc_info.value).splitlines() == [_CLIENT_ERROR_429, "boom"]
def test_async_retrieve_batch_non_200_raises():

View file

@ -169,6 +169,17 @@ class TestBatchObjectNaming:
object_name = config._get_gcs_object_name_from_batch_jsonl([{"body": {"model": "gemini-2.5-flash"}}])
assert object_name.startswith("litellm-vertex-files/publishers/google/models/gemini-2.5-flash/")
@pytest.mark.parametrize(
("model", "stored_model"),
[
("multimodalembedding@001", "publishers/google/models/multimodalembedding_001/"),
("claude-haiku-4-5@20251001", "publishers/anthropic/models/claude-haiku-4-5@20251001/"),
],
)
def test_only_claude_models_keep_the_at_version_in_their_object_name(self, config, model, stored_model):
object_name: Final = config._get_gcs_object_name_from_batch_jsonl([{"body": {"model": model}}])
assert object_name.startswith(f"litellm-vertex-files/{stored_model}")
def test_should_store_fine_tuned_endpoint_under_endpoints_path(self, config):
"""A numeric endpoint id must not be filed under publishers/google/models/gemini/<id>,
which the batch transformation later mangles into a nonexistent publisher model (LIT-6899)."""