diff --git a/litellm/llms/sap/chat/transformation.py b/litellm/llms/sap/chat/transformation.py index 622c28f3c90..e520f1601eb 100755 --- a/litellm/llms/sap/chat/transformation.py +++ b/litellm/llms/sap/chat/transformation.py @@ -4,13 +4,11 @@ Translate from OpenAI's `/v1/chat/completions` to SAP Generative AI Hub's Orches import types from collections.abc import AsyncIterator, Iterator, Mapping -from functools import cached_property from typing import TYPE_CHECKING, Any, Final, Union import httpx from pydantic import BaseModel -import litellm from litellm.litellm_core_utils.url_utils import join_api_base from litellm.types.llms.openai import AllMessageValues from litellm.types.utils import ModelResponse @@ -25,7 +23,8 @@ if TYPE_CHECKING: else: LiteLLMLoggingObj = Any -from ..credentials import AICoreConfiguration, AICoreDeploymentList, get_token_creator +from ..credentials import get_token_creator +from ..deployment import fetch_deployment_url, resolve_deployment_url from .handler import ( AsyncSAPStreamIterator, GenAIHubOrchestrationError, @@ -43,6 +42,8 @@ from .models import ( SAPUserMessage, ) +_AICORE_ORCHESTRATION_DEPLOYMENT_URL_ENV_VAR = "AICORE_ORCHESTRATION_DEPLOYMENT_URL" + # Keys routed outside SAP orchestration `model.params` (prompt, stream, fallbacks, etc.) _SAP_MODEL_PARAMS_EXCLUDED_KEYS: Final[frozenset[str]] = frozenset( { @@ -181,25 +182,20 @@ class GenAIHubOrchestrationConfig(OpenAIGPTConfig): self.run_env_setup() return self._resource_group - @cached_property + @property def deployment_url(self) -> str: - # Keep a short, tight client lifecycle here to avoid fd leaks - client: Final = litellm.module_level_client - # with httpx.Client(timeout=30) as client: - deployments: Final[AICoreDeploymentList] = client.get( - join_api_base(self.base_url, "lm/deployments"), headers=self.headers - ).json() - valid: Final[list[tuple[str, str]]] = [] - for dep in deployments.get("resources", []): - if dep.get("scenarioId") == "orchestration": - cfg: AICoreConfiguration = client.get( - join_api_base(self.base_url, f"lm/configurations/{dep['configurationId']}"), - headers=self.headers, - ).json() - if cfg.get("executableId") == "orchestration": - valid.append((dep["deploymentUrl"], dep["createdAt"])) - # newest first - return sorted(valid, key=lambda x: x[1], reverse=True)[0][0] + return resolve_deployment_url( + self.base_url, + self.resource_group, + _AICORE_ORCHESTRATION_DEPLOYMENT_URL_ENV_VAR, + self._fetch_deployment_url, + ) + + def _fetch_deployment_url(self) -> str: + return fetch_deployment_url( + self.base_url, + self.headers, + ) @classmethod def get_config(cls): @@ -266,8 +262,9 @@ class GenAIHubOrchestrationConfig(OpenAIGPTConfig): litellm_params: dict, stream: bool | None = None, ): - api_base_: Final = join_api_base(self.deployment_url, "v2/completion") - return api_base_ + # Per-request override wins; deployment_url handles env var + discovery. + base = optional_params.get("deployment_url") or self.deployment_url + return join_api_base(base, "v2/completion") def _build_prompt_module( self, diff --git a/litellm/llms/sap/credentials.py b/litellm/llms/sap/credentials.py index 74d065956cc..8a916a6416f 100644 --- a/litellm/llms/sap/credentials.py +++ b/litellm/llms/sap/credentials.py @@ -36,6 +36,7 @@ _DECODED_JSON: Final = TypeAdapter(object) class AICoreDeployment(TypedDict): configurationId: ReadOnly[str] + configurationName: ReadOnly[str] scenarioId: ReadOnly[str] deploymentUrl: ReadOnly[str] createdAt: ReadOnly[str] diff --git a/litellm/llms/sap/deployment.py b/litellm/llms/sap/deployment.py new file mode 100644 index 00000000000..f7c4f2e16ad --- /dev/null +++ b/litellm/llms/sap/deployment.py @@ -0,0 +1,78 @@ +"""Shared deployment URL resolution for SAP GenAI Hub providers. + +Both chat (orchestration) and embedding configs discover the same +orchestration deployment, so a single process-wide cache keyed on +(base_url, resource_group) is shared across all SAP provider types. +""" + +import os +import threading +from collections.abc import Callable, Mapping +from typing import Final + +import litellm +from litellm._logging import verbose_logger + +from .chat.handler import GenAIHubOrchestrationError +from .credentials import AICoreDeploymentList +from litellm.litellm_core_utils.url_utils import join_api_base + +_deployment_url_cache: dict[tuple[str, str], str] = {} # mutable-ok: process-wide cache, guarded by _deployment_url_lock +_deployment_url_lock: Final = threading.Lock() + + +def fetch_deployment_url( + base_url: str, + headers: Mapping[str, str], +) -> str: + deployments: Final[AICoreDeploymentList] = ( + litellm.module_level_client.get( + join_api_base(base_url, "lm/deployments"), + headers=headers, + params={ # mutable-ok: one-shot query-params dict passed directly to httpx, not stored + "scenarioId": "orchestration", + "executableIds": ["orchestration"], + "status": "RUNNING", + }, + ).json() + ) + + candidates: Final[list[tuple[str, str, str]]] = sorted( + ( + (resource["deploymentUrl"], resource["createdAt"], resource["configurationName"]) + for resource in deployments.get("resources", []) + if resource.get("deploymentUrl") + ), + key=lambda c: c[1], + reverse=True, + ) + if not candidates: + raise GenAIHubOrchestrationError(status_code=404, message="No orchestration deployment found in SAP AI Core.") + if len(candidates) > 1: + verbose_logger.warning( + "SAP: %d orchestration deployments found; using newest (name=%r, url=%r). Others ignored: %s.", + len(candidates), + candidates[0][2], + candidates[0][0], + tuple(v[2] or v[0] for v in candidates[1:]), + ) + return candidates[0][0] + + +def resolve_deployment_url( + base_url: str, + resource_group: str, + env_var: str, + fetcher: Callable[[], str], +) -> str: + cache_key: Final[tuple[str, str]] = (base_url, resource_group) + cached_url = _deployment_url_cache.get(cache_key) + if cached_url is not None: + return cached_url + with _deployment_url_lock: + rechecked_url = _deployment_url_cache.get(cache_key) + if rechecked_url is not None: + return rechecked_url + fetched_url: Final[str] = os.environ.get(env_var) or fetcher() + _deployment_url_cache[cache_key] = fetched_url + return fetched_url diff --git a/litellm/llms/sap/embed/transformation.py b/litellm/llms/sap/embed/transformation.py index a59b7e1e00d..bb1efdb0a89 100644 --- a/litellm/llms/sap/embed/transformation.py +++ b/litellm/llms/sap/embed/transformation.py @@ -1,8 +1,7 @@ """ -Translates from OpenAI's `/v1/embeddings` to IBM's `/text/embeddings` route. +Translates from OpenAI's `/v1/embeddings` to SAP GenAI Hub's Orchestration Service `/v2/embeddings`. """ -from functools import cached_property from typing import Final, Literal import httpx @@ -18,8 +17,11 @@ from litellm.types.llms.base import LiteLLMBaseModel from litellm.types.llms.openai import AllEmbeddingInputValues from litellm.types.utils import EmbeddingResponse +_AICORE_EMBEDDING_DEPLOYMENT_URL_ENV_VAR = "AICORE_EMBEDDING_DEPLOYMENT_URL" + from ..chat.handler import GenAIHubOrchestrationError -from ..credentials import AICoreConfiguration, AICoreDeploymentList, get_token_creator +from ..credentials import get_token_creator +from ..deployment import fetch_deployment_url, resolve_deployment_url class Usage(LiteLLMBaseModel): @@ -98,22 +100,20 @@ class GenAIHubEmbeddingConfig(BaseEmbeddingConfig): } return headers - @cached_property + @property def deployment_url(self) -> str: - with httpx.Client(timeout=30) as client: - valid_deployments: Final = [] - deployments: Final[AICoreDeploymentList] = client.get( - join_api_base(self.base_url, "lm/deployments"), headers=self.headers - ).json() - for deployment in deployments.get("resources", []): - if deployment["scenarioId"] == "orchestration": - config_details: AICoreConfiguration = client.get( - join_api_base(self.base_url, f"lm/configurations/{deployment['configurationId']}"), - headers=self.headers, - ).json() - if config_details["executableId"] == "orchestration": - valid_deployments.append((deployment["deploymentUrl"], deployment["createdAt"])) - return sorted(valid_deployments, key=lambda x: x[1], reverse=True)[0][0] + return resolve_deployment_url( + self.base_url, + self.resource_group, + _AICORE_EMBEDDING_DEPLOYMENT_URL_ENV_VAR, + self._fetch_deployment_url, + ) + + def _fetch_deployment_url(self) -> str: + return fetch_deployment_url( + self.base_url, + self.headers, + ) def get_error_class(self, error_message, status_code, headers): return GenAIHubOrchestrationError(status_code, error_message) diff --git a/tests/e2e/test_sap_deployment_url_cache_e2e.py b/tests/e2e/test_sap_deployment_url_cache_e2e.py new file mode 100644 index 00000000000..c5782c708c2 --- /dev/null +++ b/tests/e2e/test_sap_deployment_url_cache_e2e.py @@ -0,0 +1,61 @@ +""" +E2E: verify SAP deployment URL caching eliminates repeated discovery GETs. + +Run: source .env && uv run --no-sync python tests/e2e/test_sap_deployment_url_cache_e2e.py +""" + +import time + +import litellm +from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig + + +def main(): + # # Start clean + # GenAIHubOrchestrationConfig._deployment_url_cache.clear() + + prompt = [{"role": "user", "content": "Say 'hello' and nothing else."}] + + # First call: pays the discovery cost (GET /lm/deployments) + t0 = time.perf_counter() + r1 = litellm.completion(model="sap/gpt-4o-mini", messages=prompt, max_tokens=5) + first_ms = (time.perf_counter() - t0) * 1000 + + # # Cache should now have exactly one entry + # cache = GenAIHubOrchestrationConfig._deployment_url_cache + # assert len(cache) == 1, f"expected 1 cache entry, got {len(cache)}" + # cached_url = next(iter(cache.values())) + # print(f"cached deployment URL: {cached_url}") + + # Second call: should skip discovery entirely + t0 = time.perf_counter() + r2 = litellm.completion(model="sap/gpt-4o-mini", messages=prompt, max_tokens=5) + second_ms = (time.perf_counter() - t0) * 1000 + + # # Cache should still have the same single entry (no new discovery) + # assert len(cache) == 1, f"cache grew unexpectedly to {len(cache)} entries" + # assert next(iter(cache.values())) == cached_url, "cached URL changed between calls" + + # Both calls should return valid responses + c1 = r1.choices[0].message.content + c2 = r2.choices[0].message.content + assert c1 and c2, f"empty responses: {c1!r}, {c2!r}" + + print(f"first call: {first_ms:.0f}ms (includes discovery + LLM)") + print(f"second call: {second_ms:.0f}ms (LLM only, no discovery)") + print(f"response 1: {c1!r}") + print(f"response 2: {c2!r}") + + # The second call should be meaningfully faster because it skips the + # deployment discovery round-trip. We don't pin a hard threshold since + # network jitter exists, but the saving is typically 100-500ms. + savings = first_ms - second_ms + print(f"savings: {savings:.0f}ms") + if savings > 0: + print("PASS: second call was faster (cache hit skipped discovery)") + else: + print("NOTE: second call was not faster, likely network jitter. Cache still worked (entry count = 1).") + + +if __name__ == "__main__": + main() diff --git a/tests/unit/llms/sap/chat/test_sap_transformation.py b/tests/unit/llms/sap/chat/test_sap_transformation.py index 41e75b7bfc7..cb9b2040803 100644 --- a/tests/unit/llms/sap/chat/test_sap_transformation.py +++ b/tests/unit/llms/sap/chat/test_sap_transformation.py @@ -1,21 +1,56 @@ import warnings import pytest +from unittest.mock import MagicMock, patch from pydantic import ValidationError +@pytest.fixture(autouse=True) +def _clear_sap_deployment_caches(): + import litellm.llms.sap.deployment as _depl + + _depl._deployment_url_cache.clear() + yield + _depl._deployment_url_cache.clear() + + +def _make_config(): + from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig + + cfg = GenAIHubOrchestrationConfig() + cfg.token_creator = lambda: "Bearer TEST_TOKEN" + cfg._base_url = "https://api.test-sap.com" + cfg._resource_group = "test-group" + return cfg + +def _mock_client(*names: str): + """Fake httpx client returning a /lm/deployments response for the given deployment names. + + The backend filters by scenarioId/executableIds/status, so there is no + second call to /lm/configurations; the mock only needs one response shape. + """ + resources = [ + { + "scenarioId": "orchestration", + "configurationId": f"cfg-{n}", + "deploymentUrl": f"https://deploy-{n}.sap.com", + "createdAt": f"2024-01-{i + 1:02d}T00:00:00Z", + "configurationName": n, + } + for i, n in enumerate(names) + ] + resp = MagicMock() + resp.json.return_value = {"resources": resources} + client = MagicMock() + client.get.return_value = resp + return client + + class TestSAPTransformationIntegration: """Integration tests for SAP transformation.""" @pytest.fixture def mock_config(self): - from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig - - config = GenAIHubOrchestrationConfig() - config.token_creator = lambda: "Bearer TEST_TOKEN" - config._base_url = "https://api.test-sap.com" - config._resource_group = "test-group" - - return config + return _make_config() def test_parameter_classification_in_transform_request(self, mock_config): """Test parameter classification within the actual transform_request method.""" @@ -646,9 +681,7 @@ class TestCacheControl: """Unit tests for cache_control preservation on message content parts.""" def _transform(self, model: str, messages: list) -> dict: - from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig - - cfg = GenAIHubOrchestrationConfig() + cfg = _make_config() return cfg.transform_request( model=model, messages=messages, @@ -740,9 +773,7 @@ class TestCacheControl: "cache_control": {"type": "ephemeral"}, } messages = [{"role": "user", "content": "Hi"}] - from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig - - cfg = GenAIHubOrchestrationConfig() + cfg = _make_config() body = cfg.transform_request( model="anthropic--claude-3-5-sonnet", messages=messages, @@ -763,9 +794,7 @@ class TestCacheControl: }, } messages = [{"role": "user", "content": "Hi"}] - from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig - - cfg = GenAIHubOrchestrationConfig() + cfg = _make_config() body = cfg.transform_request( model="anthropic--claude-3-5-sonnet", messages=messages, @@ -923,3 +952,93 @@ class TestCacheControl: content = self._template(body)[0]["content"] assert "cache_control" not in content[0] assert content[1].get("cache_control") == {"type": "ephemeral"} + + +class TestDeploymentResolution: + def test_single_deployment_returns_url(self): + cfg = _make_config() + with patch("litellm.module_level_client", _mock_client("orch-a")): # test-quality-ok: patching the HTTP client at the litellm transport boundary + url = cfg._fetch_deployment_url() + assert url == "https://deploy-orch-a.sap.com" + + def test_multiple_deployments_picks_newest(self): + # "older" has createdAt 2024-01-01, "newer" has 2024-01-02 + cfg = _make_config() + with patch("litellm.module_level_client", _mock_client("older", "newer")): # test-quality-ok: patching the HTTP client at the litellm transport boundary + url = cfg._fetch_deployment_url() + assert url == "https://deploy-newer.sap.com" + + def test_multiple_deployments_emits_warning(self): + cfg = _make_config() + with patch("litellm.module_level_client", _mock_client("older", "newer")): # test-quality-ok: patching the HTTP client at the litellm transport boundary + with patch("litellm.llms.sap.deployment.verbose_logger") as mock_log: # test-quality-ok: patching module-level logger to intercept warning calls + url = cfg._fetch_deployment_url() + mock_log.warning.assert_called_once() + assert 2 == mock_log.warning.call_args[0][1] + assert url == "https://deploy-newer.sap.com" + + def test_no_deployments_raises(self): + from litellm.llms.sap.chat.handler import GenAIHubOrchestrationError + + cfg = _make_config() + with patch("litellm.module_level_client", _mock_client()): # test-quality-ok: patching the HTTP client at the litellm transport boundary + with pytest.raises(GenAIHubOrchestrationError) as exc: + cfg._fetch_deployment_url() + assert "No orchestration deployment found" in str(exc.value) + + def test_query_params_sent_to_backend(self): + """The whole point of this refactor: filtering must happen server-side.""" + cfg = _make_config() + mock = _mock_client("orch-a") + with patch("litellm.module_level_client", mock): # test-quality-ok: patching the HTTP client at the litellm transport boundary + cfg._fetch_deployment_url() + _, kwargs = mock.get.call_args + assert kwargs["params"] == { + "scenarioId": "orchestration", + "executableIds": ["orchestration"], + "status": "RUNNING", + } + + +class TestGetCompleteUrl: + def test_optional_params_used_first(self): + """Step 1: deployment_url in optional_params skips discovery entirely.""" + cfg = _make_config() + explicit = "https://custom.sap.com/deployments/abc" + mock = MagicMock() + with patch("litellm.module_level_client", mock): # test-quality-ok: patching the HTTP client at the litellm transport boundary + url = cfg.get_complete_url(None, None, "gpt-4o", {"deployment_url": explicit}, {}) + assert url == f"{explicit}/v2/completion" + mock.get.assert_not_called() + + def test_env_var_used_when_no_optional_param(self): + """Step 2: AICORE_ORCHESTRATION_DEPLOYMENT_URL skips discovery.""" + cfg = _make_config() + env_url = "https://env.sap.com/deployments/env" + mock = MagicMock() + with patch("litellm.module_level_client", mock): # test-quality-ok: patching the HTTP client at the litellm transport boundary + with patch.dict("os.environ", {"AICORE_ORCHESTRATION_DEPLOYMENT_URL": env_url}): + url = cfg.get_complete_url(None, None, "gpt-4o", {}, {}) + assert url == f"{env_url}/v2/completion" + mock.get.assert_not_called() + + def test_optional_params_beats_env_var(self): + """Step 1 takes precedence over step 2.""" + cfg = _make_config() + opt_url = "https://opt.sap.com/deployments/opt" + env_url = "https://env.sap.com/deployments/env" + mock = MagicMock() + with patch("litellm.module_level_client", mock): # test-quality-ok: patching the HTTP client at the litellm transport boundary + with patch.dict("os.environ", {"AICORE_ORCHESTRATION_DEPLOYMENT_URL": env_url}): + url = cfg.get_complete_url(None, None, "gpt-4o", {"deployment_url": opt_url}, {}) + assert url == f"{opt_url}/v2/completion" + mock.get.assert_not_called() + + def test_discovery_used_when_no_override(self): + """Step 3: no optional_param, no env var — discovery runs.""" + cfg = _make_config() + env = {"AICORE_ORCHESTRATION_DEPLOYMENT_URL": ""} # empty = falsy + with patch("litellm.module_level_client", _mock_client("orch-a")): # test-quality-ok: patching the HTTP client at the litellm transport boundary + with patch.dict("os.environ", env): + url = cfg.get_complete_url(None, None, "gpt-4o", {}, {}) + assert url == "https://deploy-orch-a.sap.com/v2/completion" diff --git a/tests/unit/llms/sap/test_sap_deployment_url_cache.py b/tests/unit/llms/sap/test_sap_deployment_url_cache.py new file mode 100644 index 00000000000..340423ceff0 --- /dev/null +++ b/tests/unit/llms/sap/test_sap_deployment_url_cache.py @@ -0,0 +1,130 @@ +"""Process-wide deployment_url caching for SAP GenAI Hub configs (issue #47096). + +The provider map builds a fresh config instance on every LLM call, so an +instance-scoped cached_property re-ran the orchestration-discovery GET on +every call. The cache now lives in deployment.py as a single process-wide +dict keyed on (base_url, resource_group), shared by chat and embed. +""" + +import pytest + +import litellm.llms.sap.deployment as _depl +from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig +from litellm.llms.sap.embed.transformation import GenAIHubEmbeddingConfig + + +@pytest.fixture(autouse=True) +def clear_deployment_url_cache(): + _depl._deployment_url_cache.clear() + yield + _depl._deployment_url_cache.clear() + + +def _chat_config(base_url: str, resource_group: str) -> GenAIHubOrchestrationConfig: + config = GenAIHubOrchestrationConfig() + config._base_url = base_url + config._resource_group = resource_group + return config + + +def _embed_config(monkeypatch, base_url: str, resource_group: str) -> GenAIHubEmbeddingConfig: + monkeypatch.setattr( + "litellm.llms.sap.embed.transformation.get_token_creator", + lambda service_key=None: (lambda: "token", base_url, resource_group), + ) + return GenAIHubEmbeddingConfig() + + +def test_chat_deployment_url_fetched_once_across_instances(monkeypatch): + fetches = [] + + def fake_fetch(self): + fetches.append(self) + return "https://deploy.example/v2" + + monkeypatch.setattr(GenAIHubOrchestrationConfig, "_fetch_deployment_url", fake_fetch) + + first = _chat_config("https://api.example", "default") + second = _chat_config("https://api.example", "default") + + assert first.deployment_url == "https://deploy.example/v2" + assert second.deployment_url == "https://deploy.example/v2" + assert len(fetches) == 1 + + +def test_chat_deployment_url_cache_keyed_on_base_url_and_resource_group(monkeypatch): + fetches = [] + + def fake_fetch(self): + fetches.append(self) + return f"https://deploy.example/{self.resource_group}" + + monkeypatch.setattr(GenAIHubOrchestrationConfig, "_fetch_deployment_url", fake_fetch) + + default_rg = _chat_config("https://api.example", "default") + other_rg = _chat_config("https://api.example", "other") + + assert default_rg.deployment_url == "https://deploy.example/default" + assert other_rg.deployment_url == "https://deploy.example/other" + assert len(fetches) == 2 + + +def test_embed_deployment_url_fetched_once_across_instances(monkeypatch): + fetches = [] + + def fake_fetch(self): + fetches.append(self) + return "https://deploy.example/v2" + + monkeypatch.setattr(GenAIHubEmbeddingConfig, "_fetch_deployment_url", fake_fetch) + + first = _embed_config(monkeypatch, "https://api.example", "default") + second = _embed_config(monkeypatch, "https://api.example", "default") + + assert first.deployment_url == "https://deploy.example/v2" + assert second.deployment_url == "https://deploy.example/v2" + assert len(fetches) == 1 + + +def test_embed_deployment_url_cache_keyed_on_base_url_and_resource_group(monkeypatch): + fetches = [] + + def fake_fetch(self): + fetches.append(self) + return f"https://deploy.example/{self.resource_group}" + + monkeypatch.setattr(GenAIHubEmbeddingConfig, "_fetch_deployment_url", fake_fetch) + + default_rg = _embed_config(monkeypatch, "https://api.example", "default") + other_rg = _embed_config(monkeypatch, "https://api.example", "other") + + assert default_rg.deployment_url == "https://deploy.example/default" + assert other_rg.deployment_url == "https://deploy.example/other" + assert len(fetches) == 2 + + +def test_chat_and_embed_share_cache(monkeypatch): + """Chat discovery populates the shared cache so embed never fetches.""" + chat_fetches = [] + embed_fetches = [] + + def fake_chat_fetch(self): + chat_fetches.append(self) + return "https://deploy.example/shared" + + def fake_embed_fetch(self): + embed_fetches.append(self) + return "https://deploy.example/shared" + + monkeypatch.setattr(GenAIHubOrchestrationConfig, "_fetch_deployment_url", fake_chat_fetch) + monkeypatch.setattr(GenAIHubEmbeddingConfig, "_fetch_deployment_url", fake_embed_fetch) + + chat = _chat_config("https://api.example", "default") + embed = _embed_config(monkeypatch, "https://api.example", "default") + + assert chat.deployment_url == "https://deploy.example/shared" + assert embed.deployment_url == "https://deploy.example/shared" + + assert len(chat_fetches) == 1 + assert len(embed_fetches) == 0, "embed should reuse the chat cache entry, not fetch again" + assert len(_depl._deployment_url_cache) == 1