Incident Report: vLLM Embeddings Broken by encoding_format Parameter

[Sameerlite]

Relevant issues

Pre-Submission checklist

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

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🆕 New Feature
🐛 Bug Fix
🧹 Refactoring
📖 Documentation
🚄 Infrastructure
✅ Test

Changes

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 18, 2026 1:11pm

[greptile-apps]

Greptile Summary

This PR bundles three unrelated changes: (1) a new incident report blog post for the vLLM embeddings encoding_format bug, (2) addition of 3 Mistral Devstral model aliases to the model pricing JSON, and (3) deletion of 2 existing tests from the hosted_vllm embedding test suite.

• Incident report: Well-written root cause analysis and remediation documentation. However, the date range on line 22 ("Feb 16, 2025 - January 30, 2026") appears inconsistent — it mixes date formats and likely has a typo in the year, contradicting the stated "~3 hours" duration.
• Model pricing: Adds mistral/devstral-small-latest, mistral/devstral-latest, and mistral/devstral-medium-latest entries with pricing consistent with existing sibling models (labs-devstral-small-2512, devstral-2512).
• Test deletions: Removes test_validate_environment_without_api_key and test_encoding_format_float_sent_in_actual_request without justification. The latter is the complementary test to the remaining test_encoding_format_not_sent_in_actual_request and its removal reduces confidence that valid encoding_format values are forwarded correctly through the full call path. Per project requirements, PRs should include evidence that changes maintain code quality.

Confidence Score: 3/5

• Low-risk changes (docs + model config), but test deletions reduce coverage for a previously broken feature.
• Score of 3 reflects: the model pricing additions and blog post are low-risk, but the unexplained deletion of 2 tests (especially the encoding_format='float' E2E test that validated the fix worked) is a concern — it reduces test coverage for the very feature this incident report documents. The date inconsistency in the blog post needs correction before publishing.
• tests/test_litellm/llms/hosted_vllm/embedding/test_hosted_vllm_embedding_transformation.py needs review for the test deletions, and docs/my-website/blog/vllm_embeddings_incident/index.md needs the date corrected.

Important Files Changed

Filename	Overview
docs/my-website/blog/vllm_embeddings_incident/index.md	New blog post for vLLM embeddings incident report. Contains an inconsistent date range ("Feb 16, 2025 - January 30, 2026") and deleted tests are referenced as done in the remediation table.
litellm/model_prices_and_context_window_backup.json	Adds 3 new Mistral Devstral model aliases (devstral-small-latest, devstral-latest, devstral-medium-latest) with pricing metadata. Entries look consistent with existing similar models.
tests/test_litellm/llms/hosted_vllm/embedding/test_hosted_vllm_embedding_transformation.py	Removes 2 valid tests (test_validate_environment_without_api_key and test_encoding_format_float_sent_in_actual_request) without explanation. These tests covered meaningful behavior and their removal reduces test coverage.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["PR #21474"] --> B["Incident Report Blog Post"]
    A --> C["Model Pricing Updates"]
    A --> D["Test File Changes"]
    
    B --> B1["docs/.../vllm_embeddings_incident/index.md\n(new file)"]
    
    C --> C1["model_prices_and_context_window_backup.json"]
    C1 --> C2["+ mistral/devstral-small-latest"]
    C1 --> C3["+ mistral/devstral-latest"]
    C1 --> C4["+ mistral/devstral-medium-latest"]
    
    D --> D1["test_hosted_vllm_embedding_transformation.py"]
    D1 --> D2["- test_validate_environment_without_api_key"]
    D1 --> D3["- test_encoding_format_float_sent_in_actual_request"]
    
    style B1 fill:#d4edda,stroke:#28a745
    style C2 fill:#d4edda,stroke:#28a745
    style C3 fill:#d4edda,stroke:#28a745
    style C4 fill:#d4edda,stroke:#28a745
    style D2 fill:#f8d7da,stroke:#dc3545
    style D3 fill:#f8d7da,stroke:#dc3545

Loading

_{Last reviewed commit: a9b7320}

[greptile-apps]

_{3 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Inconsistent date range in incident report

The date range "Feb 16, 2025 - January 30, 2026" mixes date formats and likely has incorrect years. The blog post date field is set to 2026-02-18, so this range should probably read something like "Jan 30, 2026 - Feb 16, 2026" (or similar consistent format). As written, "Feb 16, 2025" appears to be a typo (2025 vs 2026), and the duration of nearly a year contradicts the "~3 hours" duration stated on the next line.

Suggested change

	**Date:** Feb 16, 2025 - January 30, 2026
	**Date:** January 30, 2026 - February 16, 2026

[greptile-apps]

Removed test for encoding_format='float' reduces coverage

The deleted test_encoding_format_float_sent_in_actual_request test verified that when a user explicitly passes encoding_format="float", it is actually sent in the HTTP request payload. This is the complementary test to test_encoding_format_not_sent_in_actual_request — together they confirmed both sides of the fix (omit when not provided, include when provided). Removing this test means there's no longer an E2E-style test verifying that valid encoding_format values are forwarded correctly through the full embedding call path.

Context Used: Rule from dashboard - What: Ensure that any PR claiming to fix an issue includes evidence that the issue is resolved, such... (source)