fix(router): pass custom_llm_provider to get_llm_provider for unprefixed model names

[Sameerlite]

Summary

Fixes 'LLM Provider NOT provided' errors when models are configured with custom_llm_provider but model names lack provider prefix (e.g., gpt-4.1-mini instead of azure/gpt-4.1-mini).

Root Cause

The router was calling get_llm_provider(model=data['model']) without passing custom_llm_provider parameter, causing provider inference to fail for unprefixed model names.

Changes

• Router now passes deployment's custom_llm_provider to get_llm_provider() in 6 code paths:
  • File creation (_acreate_file)
  • File content (_ageneric_api_call_with_fallbacks_helper)
  • Batch creation (acreate_batch)
  • Batch cancellation (acancel_batch)
  • Vector store creation (avector_store_create)
  • Generic API calls (sync version)
• File delete endpoint now explicitly passes custom_llm_provider
• Added 2 regression tests

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 8, 2026 10:56am

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_fix-llm-provider-inference-for-unprefixed-models (ee885f1) with main (62757ff)}

[greptile-apps]

Greptile Summary

This PR fixes a LLM Provider NOT provided error in the LiteLLM Router when deployments configure custom_llm_provider explicitly but use unprefixed model names (e.g. gpt-4.1-mini instead of azure/gpt-4.1-mini). Previously, six code paths in the router called get_llm_provider(model=data[\"model\"]) without forwarding the deployment's stored custom_llm_provider, causing provider inference to fail. The fix consistently reads custom_llm_provider from deployment[\"litellm_params\"] before calling get_llm_provider, then falls back to the inferred value. A parallel fix in the proxy delete_file endpoint ensures the resolved provider is explicitly threaded through to litellm.afile_delete.

• Applies the same two-step lookup (data.get(\"custom_llm_provider\") → get_llm_provider inference → or fallback) across six router methods: _acreate_file, _ageneric_api_call_with_fallbacks_helper, acreate_batch, acancel_batch, avector_store_create, and the sync generic_api_call.
• Adds explicit custom_llm_provider=credentials[\"custom_llm_provider\"] to the afile_delete proxy endpoint (safe because get_deployment_credentials_with_provider always guarantees this key).
• Two regression tests added for acreate_file and afile_content; the remaining four fixed paths (acreate_batch, acancel_batch, avector_store_create, sync generic) are not covered by new tests.

Confidence Score: 5/5

Safe to merge — fix is correct and consistent across all six paths; only remaining findings are P2 test coverage suggestions.

The two-step provider resolution pattern is applied uniformly and matches how other parts of the router already handle provider lookup. The files_endpoints.py change is safe because get_deployment_credentials_with_provider guarantees custom_llm_provider is always present. The two new tests are mock-only (matching project rules) and the first one strongly validates the core fix. The remaining gaps are P2 style improvements that do not affect merge safety.

tests/test_litellm/test_router.py — missing coverage for acreate_batch, acancel_batch, avector_store_create, and sync generic_api_call paths.

Vulnerabilities

No security concerns identified. The change only affects provider resolution logic within the router and does not touch authentication, authorization, credential storage, or user-controlled input handling. The credentials[\"custom_llm_provider\"] access in files_endpoints.py is safe because get_deployment_credentials_with_provider always populates this key before returning.

Important Files Changed

Filename	Overview
litellm/router.py	Six call-sites updated to read custom_llm_provider from deployment params before inferring via get_llm_provider; fix is correct and consistent, but four of the six paths lack regression tests.
litellm/proxy/openai_files_endpoints/files_endpoints.py	Explicitly passes credentials["custom_llm_provider"] to afile_delete; safe because get_deployment_credentials_with_provider always guarantees this key, and prepare_data_with_credentials already pops it from data to prevent duplication.
tests/test_litellm/test_router.py	Two new mock-only tests added; acreate_file test is thorough (asserts custom_llm_provider == 'azure'), but afile_content test only checks call count rather than the forwarded provider value, and four other fixed paths are untested.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Router method called\ne.g. acreate_file, acreate_batch] --> B[Select deployment\nfrom model_list]
    B --> C[data = deployment.litellm_params.copy]
    C --> D{data has\ncustom_llm_provider?}
    D -- Yes --> E[custom_llm_provider =\ndata.get custom_llm_provider]
    D -- No --> F[custom_llm_provider = None]
    E --> G[get_llm_provider\nmodel=data.model\ncustom_llm_provider=custom_llm_provider]
    F --> G
    G -- success --> H[inferred_custom_llm_provider]
    G -- exception --> I[custom_llm_provider = None]
    H --> J[custom_llm_provider =\ncustom_llm_provider OR inferred]
    J --> K[Pass custom_llm_provider\nexplicitly to litellm API call]
    I --> L{async path?}
    L -- Yes --> M[Skip setting provider\nin response_kwargs]
    L -- No --> N[Set provider = None\nin kwargs]
    K --> O[litellm routes to\ncorrect provider handler]

Loading

_{Reviews (1): Last reviewed commit: "fix(router): pass custom_llm_provider to..." | Re-trigger Greptile}