Real-world validation

Verified end-to-end on an AMD Strix Halo (Ryzen AI) box against a live Lemonade backend (OpenAI-compatible), installed via pip install -e ".[litellm]" (litellm 1.88.1) on this branch (f2eb9f3):

• Live chat — create_client("litellm").chat(...) routed through litellm → Lemonade and echoed a planted, unguessable token verbatim: CHAT_RESP: 'VIOLET-OTTER-92' ✅
• embed() model override (the bug) — provider constructed with a chat model, then embed(text, model="…nomic-embed…") — the exact path that raised TypeError: embedding() got multiple values for keyword argument 'model' before this fix — now returns a real 768-dim vector: EMBED_OVERRIDE_OK: True dim: 768 ✅
• Real inference, not a stub — two distinct sentences yield near-orthogonal embeddings (cosine 0.15) ✅

Unit: pytest tests/unit/test_litellm_provider.py → 12 passed (incl. 2 new embed() tests; the override test reproduces the crash before the fix). Lint: util/lint.py --black --isort --flake8 → all pass.

Summary

Clean, focused fix that closes out the remaining #1593 review items. The headline change is a real bug: LiteLLMProvider.embed() passed model both as an explicit arg and inside the spread **call_kwargs, so any caller overriding the model hit TypeError: embedding() got multiple values for keyword argument 'model' — and that path had zero coverage. Popping model from kwargs before building call_kwargs (litellm.py:82) fixes it cleanly, and the new test_embed_uses_override_model reproduces the original crash. The drop_params cleanup (global module mutation → per-call setdefault) is a nice side improvement: it drops a global side-effect on import and lets callers override the default. Docs and the create_client docstring now list the provider. Nothing blocking here.

Issues Found

🟢 Minor — shell-globbing in install hints (docs/sdk/sdks/llm.mdx:10, litellm.py:24)
pip install gaia[litellm] fails under zsh (the default macOS shell), which globs the brackets: zsh: no matches found: gaia[litellm]. Quoting it works in both bash and zsh. Same string appears in the ImportError message — worth fixing both for consistency.

In the docs:

# Requires: pip install 'gaia[litellm]'

And in litellm.py:23-25:

            raise ImportError(
                "litellm is not installed. Install it with: pip install 'gaia[litellm]'"
            ) from e

Strengths

• Test reproduces the bug before fixing it. test_embed_uses_override_model would have raised the original TypeError; the inline comment naming the failure mode is exactly the kind of context that prevents a future regression.
• drop_params change is strictly better. Moving from litellm.drop_params = True (global, set on import) to a per-call call_kwargs.setdefault("drop_params", True) removes a hidden global side-effect and keeps caller override intact — matches the no-silent-global-state direction of the codebase.
• Scope-clean and well-documented. Four files, all directly tied to the stated fix; no drive-by churn. PR description leads with before/after impact and a concrete test plan.

Verdict

Approve with suggestions — the embed fix is correct and tested, and the rest is tidy cleanup. The only items are the optional zsh-quoting nits in the install hints; safe to merge once those are applied (or consciously skipped).

Applied the zsh-quoting suggestion in d6c0e199 — both the docs hint (docs/sdk/sdks/llm.mdx) and the ImportError message now read pip install 'gaia[litellm]', so it copy-pastes cleanly under zsh and bash. Lint + the 12 unit tests still green. Thanks for the catch!