Add embeddings endpoint support

[maryamtahhan]

Adds support for benchmarking the /v1/embeddings endpoint, enabling performance testing of text embedding models.

• Add embeddings request type to schemas

• Implement EmbeddingsResponseHandler for processing embedding responses

• Add EmbeddingsRequestFormatter for request preparation

• Implement mock server handler with synthetic embedding generation

• Add e2e and unit tests for embeddings benchmarking

• Add embeddings guide documentation

  Summary

This PR adds support for benchmarking the /v1/embeddings endpoint, enabling performance testing of text embedding models.

The implementation includes full integration with guidellm's existing infrastructure: request formatting, response handling, metrics collection, synthetic data generation, and the mock server for testing.

Details

Core Implementation:

• Add embeddings to GenerativeRequestType literal in schemas
• Implement EmbeddingsResponseHandler for processing embedding API responses
• Implement EmbeddingsRequestFormatter for preparing embedding requests
• Add embeddings route to OpenAI backend (/v1/embeddings)
• Export EmbeddingsResponseHandler from backends module

Mock Server Support:

• Implement EmbeddingsHandler for mock server
• Add EmbeddingsRequest, EmbeddingObject, and EmbeddingsResponse models
• Register /v1/embeddings endpoint in mock server
• Generate synthetic normalized embedding vectors (configurable dimensions)
• Apply realistic timing delays based on token count

Testing:

• Add unit tests for EmbeddingsResponseHandler
• Add e2e tests: test_embeddings_max_requests_benchmark, test_embeddings_max_seconds_benchmark, test_embeddings_rate_benchmark
• Update existing tests to account for new request type
• All tests pass (1717 unit tests, 2 integration tests)

Documentation:

• Create comprehensive embeddings guide (docs/guides/embeddings.md, 284 lines)
• Update backends guide with embeddings endpoint information
• Update datasets guide with embeddings-specific synthetic data examples
• Add docstrings to all new classes and methods

Code Quality:

• All linting checks pass (tox -e quality)
• All type checks pass (tox -e types)
• Pre-commit hooks pass
• Code properly formatted with ruff and mdformat

Test Plan

Automated Tests:

1. Run unit tests: tox -e test-unit
   • Verify TestEmbeddingsResponseHandler tests pass
2. Run integration tests: tox -e test-integration
3. Run e2e tests: tox -e test-e2e
   • Tests will pass in CI with vllm-sim binary

Manual Testing:

1. Start the mock server:

   guidellm mock-server --port 8000
   

2. Test embeddings endpoint directly:
   curl -X POST http://localhost:8000/v1/embeddings
   -H "Content-Type: application/json"
   -d '{"input":"Test sentence","model":"test"}'
3. Expected: Returns JSON with embedding vectors (1536 dimensions by default)
4. Run an embeddings benchmark:
   guidellm benchmark run
   --target http://localhost:8000
   --request-type embeddings
   --rate 5
   --max-requests 20
   --data '{"type":"synthetic","prompt_tokens":128,"output_tokens":1}'
   --processor gpt2
5. Expected: Benchmark completes successfully with metrics for 20 requests
6. Test with vLLM serving an embedding model:
   vllm serve "BAAI/bge-small-en-v1.5"
   guidellm benchmark run
   --target http://localhost:8000
   --request-type embeddings
   --rate 10
   --max-requests 100
   --data '{"type":"synthetic","prompt_tokens":256,"output_tokens":1}'
   --processor gpt2

Related Issues

• Addresses need for embedding model benchmarking support
• Complements existing text completions, chat completions, and audio benchmarking

---

• "I certify that all code in this PR is my own, except as noted below."

Use of AI

• Includes AI-assisted code completion
• Includes code generated by an AI application
• Includes AI-generated tests (NOTE: AI written tests should have a docstring that includes ## WRITTEN BY AI ##)