revert: perf: lazy-load SDK symbols so import litellm stays under 60 MB RSS (#39121)

[yuneng-berri]

TLDR

Problem this solves:

• perf: lazy-load SDK symbols so import litellm stays under 60 MB RSS #39121 made import litellm lazy and turned five CircleCI jobs red on staging
• litellm.cost_calculator, the Bedrock token counter and proxy auth checks now hit circular imports
• The Interactions API bridge fails on every call because litellm.responses and litellm.aresponses are no longer in the module namespace
• Staging has been red on these since 2026-09-05 19:58 UTC

How it solves it:

• Reverts perf: lazy-load SDK symbols so import litellm stays under 60 MB RSS #39121 so litellm/__init__.py eagerly wires the SDK symbols again
• Restores the pre-PR import order and the module-level responses / aresponses attributes

User Flow

Before: a developer importing a litellm submodule on its own gets a circular ImportError, and the Interactions API bridge crashes on its first request

1. They run python -c "from litellm.cost_calculator import batch_cost_calculator"
2. It exits with ImportError: cannot import name 'RealtimeAPITokenUsageProcessor' from partially initialized module 'litellm.cost_calculator' (most likely due to a circular import)
3. They call litellm.interactions.create(model="gpt-5.5", input="What is 2 + 2?")
4. It raises APIConnectionError: OpenAIException - 'module' object is not callable instead of returning an interaction

After: the same import and the same call work as they did before #39121

1. They run python -c "from litellm.cost_calculator import batch_cost_calculator"
2. It exits 0
3. They call litellm.interactions.create(model="gpt-5.5", input="What is 2 + 2?")
4. They get an InteractionsAPIResponse with status="completed"

Relevant issues

Reverts #39121

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Both sides ran from a fresh worktree with OPENAI_API_KEY set. The Interactions API call is a real request to OpenAI. The pytest runs are the two CircleCI test files that fail at collection time on staging (the batches_testing and litellm_utils_testing jobs)

Before (a0c9095)

Importing a submodule first

1. python -c "from litellm.cost_calculator import batch_cost_calculator"
2. Output: ImportError: cannot import name 'RealtimeAPITokenUsageProcessor' from partially initialized module 'litellm.cost_calculator' (most likely due to a circular import)

Interactions API bridged to /v1/responses

1. python -c 'import litellm; r = litellm.interactions.create(model="gpt-5.5", input="What is 2 + 2? Answer with one number."); print(type(r).__name__, "status", r.status)'
2. Output: litellm.exceptions.APIConnectionError: litellm.APIConnectionError: APIConnectionError: OpenAIException - 'module' object is not callable

The two CircleCI test files

1. pytest tests/litellm_utils_tests/test_bedrock_token_counter.py tests/batches_tests/test_batch_custom_pricing.py -q
2. Output: 4 warnings, 2 errors in 0.69s

After (eee162b)

Importing a submodule first

1. python -c "from litellm.cost_calculator import batch_cost_calculator"
2. Output: exit 0

Interactions API bridged to /v1/responses

1. python -c 'import litellm; r = litellm.interactions.create(model="gpt-5.5", input="What is 2 + 2? Answer with one number."); print(type(r).__name__, "status", r.status)'
2. Output: InteractionsAPIResponse status completed

The two CircleCI test files

1. pytest tests/litellm_utils_tests/test_bedrock_token_counter.py tests/batches_tests/test_batch_custom_pricing.py -q
2. Output: 12 passed, 1 skipped in 0.10s

Type

🐛 Bug Fix

Caveats (if any)

Medium

• The import-time RSS win from perf: lazy-load SDK symbols so import litellm stays under 60 MB RSS #39121 is gone again until it is re-landed with the three import cycles broken and the bridge reading litellm.responses / litellm.aresponses through getattr, not vars()

Low

• This also removes test_star_import_exports_public_api, so test(lazy_imports): check star-import exports against globals(), not dir() #39958 (which only fixed that test's dir() scoping bug) should be closed if this merges
• tests/proxy_admin_ui_tests/test_access_group_team_sync.py collects again after the revert; its tests then need a Postgres DATABASE_URL locally, which CircleCI provides

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[linear-code]

Issue reopened: LIT-6607 import litellm costs ~200 MB RSS

[greptile-apps]

Greptile Summary

This PR reverts SDK symbol lazy loading and restores eager package initialization to repair import cycles and interactions response bindings

• Restores eager SDK, configuration, integration, and endpoint imports in litellm/__init__.py
• Removes the SDK symbol registry, module alias resolver, and package module subclass used by the reverted implementation
• Restores the earlier proxy package initialization and removes lazy-loading tests that no longer match the implementation

Confidence Score: 4/5

The functional revert appears sound, but repository requirements for focused regression coverage and source comments must be satisfied before merging

The eager imports restore the intended namespace and no behavioral failure was established, but the exact interactions failure lacks deterministic regression coverage and restored comments violate repository guidance

Files Needing Attention: litellm/init.py, tests/test_litellm/test_lazy_imports.py

Important Files Changed

Filename	Overview
litellm/init.py	Restores eager SDK imports and module-level response symbols, with repository-prohibited explanatory comments
litellm/_lazy_imports.py	Removes SDK symbol, module alias, submodule fallback, and shadowable-function lazy-loading machinery
litellm/_lazy_imports_registry.py	Removes the large SDK import maps and public star-import registry introduced by the reverted change
litellm/proxy/init.py	Replaces dynamic proxy submodule resolution with the previous package initialization
tests/test_litellm/test_lazy_imports.py	Removes tests for reverted lazy behavior without adding deterministic coverage for the interactions bridge regression

_{Reviews (1): Last reviewed commit: "Revert "perf: lazy-load SDK symbols so i..." | Re-trigger Greptile}

[greptile-apps]

Interactions Regression Untested

This bug fix lacks a deterministic interactions bridge regression test, violating the repository requirement that exact failures remain covered. Add one before merging

Context Used: CLAUDE.md (source)

[greptile-apps]

Prohibited Source Comments

These routine import-order comments, also at lines 1282-1286 and 2189-2198, violate the repository's source-comment restrictions and must be removed before merging

Context Used: CLAUDE.md (source)

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing revert-39121-litellm_lazy_sdk_import (eee162b) with litellm_internal_staging (587311d)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!