[VickyXAI]

What

llama-index-llms-blockrun — a LlamaIndex LLM for any BlockRun chat model, living at integrations/llama-index-llms-blockrun and published as its own PyPI package.

from llama_index.llms.blockrun import BlockRun

llm = BlockRun(model="openai/gpt-5.5")                  # USDC on Base
llm = BlockRun(model="openai/gpt-5.5", chain="solana")  # USDC on Solana

Why a package and not an OpenAILike snippet

Context: run-llama/llama_index#23236. OpenAILike(api_base=..., api_key="fake") reaches the gateway and stops at the first 402 — nothing in the OpenAI client can sign a payment. LlamaIndex's CONTRIBUTING now asks for new provider integrations to live in their own repo and publish independently, so it lives here, next to the SDK it wraps, and CI tests it against this checkout.

Design

• chain="base" → LLMClient / AsyncLLMClient; chain="solana" → SolanaLLMClient / AsyncSolanaLLMClient ([solana] extra). Payment, max_cost_per_call and retry rules stay in the SDK.
• FunctionCallingLLM: chat, stream, async, tool calling, predict_and_call, FunctionAgent. Serialization reuses llama-index-llms-openai's to_openai_message_dicts.
• Private key is a PrivateAttr — never serialized by to_dict() / callbacks.
• temperature unset by default (reasoning models reject non-default values).
• context_window read once from /v1/models; cached, falls back with a warning.
• Solana's client has no **extra: unknown params are refused up front by name.
• One async client per event loop (httpx.AsyncClient across asyncio.run() → "Event loop is closed").
• ThinkingBlocks stripped before serializing (llama-index-llms-openai 0.6.5 raises on them).

CI / release

• New llama-index CI job: floor pins (core 0.14.5, openai 0.6.5, py3.10) and current (py3.12).
• publish-llama-index.yml publishes on releases tagged llama-index-llms-blockrun-v*; publish.yml now skips those tags. SDK sdist excludes integrations/.
• Needs once on PyPI: a trusted publisher for llama-index-llms-blockrun → repo BlockRunAI/blockrun-llm, workflow publish-llama-index.yml, environment pypi.

Verification

• 51 unit tests, passing on current deps (py3.13) and at the floor pins (py3.10); black, ruff, mypy clean; SDK unit suite 977 passed.
• Live, both chains, deepseek/deepseek-chat: chat, stream, acomplete, astream_chat, predict_and_call, and a FunctionAgent run (1234 × 4567 = 5,635,678).

🤖 Generated with Claude Code