Dokumentationerreichbar
vLLM Recipes, GPT-OSS, beim Commit 2338e9e
Die Anleitung des v
geprüft 10.11.2026
Worauf sich diese Seite beruft, wörtlich, abgerufen am 20.10.2026:
To achieve higher throughput, tokens from multiple requests must be batched and processed together, but that increases the latencies.
bestätigt 10.11.2026Set this to 1 to achieve the best throughput per GPU, and set this to 2, 4, or 8 to achieve better per-user latencies.
bestätigt 10.11.2026gpt-oss-120b will fit on a single A100 (80GB), but scaling it to higher TP sizes can help with throughput
bestätigt 10.11.2026By default, it uses port 8000.
bestätigt 10.11.2026By default, this is set to a large number like 1024 on GPUs with large memory sizes.
bestätigt 10.11.2026endpoint can perform tool use (browsing, python, mcp) in between chain-of-thought and deliver a final response
bestätigt 10.11.2026v
bestätigt 10.11.2026LLM itself can act as MCP client to multiple services. On H100 using tensor parallel size 1, default gpu memory utilization, and batched token will cause CUDA Out-of-memory.
bestätigt 10.11.2026