local-ai: Clarify llama-server parallelism and context sizing for concurrent sub-agent requests

Question

OpenClaw's current managed llama-server recipe sets parallelRequests: 1, which is emitted as parallel = 1 in the generated server preset:

• LocalModelCatalog.cs
• LlamaServerRouterConfiguration.cs

I would like to confirm how this behaves in an OpenClaw workflow with multiple concurrent sub-agents.

Suppose one sub-agent is generating a long response and is in steady-state autoregressive decode, normally producing one token per forward pass or a few committed tokens when speculative MTP succeeds. Other sub-agents then submit inference requests concurrently.

1. With parallel = 1 (--parallel 1 / -np 1), does llama-server permit only one active inference sequence, leaving every other request queued until that request completes or releases the slot?

2. If so, are the queued requests unable to participate in continuous batching, regardless of how large ctx-size is?

3. If parallel is increased, how should ctx-size be interpreted and budgeted? Is it a total server context budget shared or partitioned across the slots, or can each active request use the configured context size?

4. Would increasing parallelism therefore require reducing the maximum context per sub-agent to keep KV-cache and overall GPU-memory usage feasible?

5. For OpenClaw's multi-agent workload, should the qualified recipe select parallelism and per-request context together for each model/GPU tier, or is intentionally serializing requests the preferred behavior?

Clarifying this would help determine whether the current value is an intentional memory/latency tradeoff or a limitation for concurrent local-agent workflows.

CC: @jacobtomlinson, Ryan Hill

Start with LocalModelCatalog.cs lines 225-237 and LlamaServerRouterConfiguration.cs lines 112-126 to trace how parallelRequests becomes the generated parallel setting. Verify the llama-server behavior for concurrent requests, context sizing, and memory tradeoffs, then document whether OpenClaw should pair parallelism with per-request context or retain serialization.