feat/sarvam-llm-openai-compatible-integration

[dhruvladia-sarvam]

Summary

• Add Sarvam LLM integration under the Sarvam plugin as a dedicated llm module, with model support for sarvam-30b and sarvam-105b.
• Add OpenAI-compatible factory support via openai.LLM.with_sarvam(...) so Sarvam can be used like other OpenAI-compatible providers.

Why

• Enables first-class Sarvam LLM usage in both:
  • livekit.plugins.sarvam.LLM(...)
  • livekit.plugins.openai.LLM.with_sarvam(...)
• Keeps integration aligned with existing LiveKit OpenAI-compatible provider patterns.

Files changed

• livekit-plugins/livekit-plugins-sarvam/livekit/plugins/sarvam/llm/__init__.py
• livekit-plugins/livekit-plugins-sarvam/livekit/plugins/sarvam/llm/client.py
• livekit-plugins/livekit-plugins-sarvam/livekit/plugins/sarvam/llm/models.py
• livekit-plugins/livekit-plugins-sarvam/livekit/plugins/sarvam/__init__.py
• livekit-plugins/livekit-plugins-sarvam/pyproject.toml
• livekit-plugins/livekit-plugins-sarvam/README.md
• livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/llm.py
• livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/models.py

Validation / testing performed

• Import and integration smoke checks:
  • from livekit.plugins import sarvam, openai
  • openai.LLM.with_sarvam(...) availability and usage
• Functional checks across both models (sarvam-30b, sarvam-105b) for:
  • basic generation
  • tool calling (tool_choice: none, auto, required)
  • parameter pass-through (temperature, top_p, reasoning_effort, seed, stop, n, frequency_penalty, presence_penalty, max_tokens, wiki_grounding)
• Confirmed LLM metrics emission includes required fields (duration, completion_tokens, prompt_tokens, prompt_cached_tokens, total_tokens, tokens_per_second, ttft, speech_id).

Notes

• speech_id is expected to be null in direct standalone llm.chat(...).collect() tests; it is populated in full AgentSession voice flow context.