VLLM_MAX_SEQ_LEN_EXCEEDED warning

ValueError: prompt exceeds the model's maximum context length

The request's prompt + max_tokens exceeds max_model_len, often after that value was manually raised past what KV cache memory supports.

Symptom

ValueError: This model's maximum context length is X tokens. However, you requested Y tokens in the messages...

Cause

Either a genuinely long prompt exceeded the model’s real context window, or --max-model-len was manually set higher than the model actually supports (or higher than available KV cache memory can serve) when starting the vLLM server.

Fix

If the prompt is legitimately long, reduce it or truncate history before sending.
If --max-model-len was raised manually, confirm it doesn’t exceed the model’s actual trained context length (check the model card) — setting it higher doesn’t extend real context, it just changes what vLLM accepts before rejecting the request outright.
If context length is correct but VRAM is the real constraint, lowering --max-model-len reduces reserved KV cache memory, which may resolve this alongside CUDA OOM errors on the same server.

Sources

vLLM docs: Conserving Memory (limiting max_model_len and max_num_seqs reduces memory use)

Frequently asked questions

If I raise --max-model-len, will that let me use a longer context?

No — raising it past what the model was actually trained on doesn't extend real context, it only changes the threshold at which vLLM rejects a request. The model's usable context length is fixed by training, not by this flag.

Can this error be related to CUDA out-of-memory errors on the same server?

Yes. --max-model-len directly controls how much KV cache memory vLLM reserves — lowering it to match your actual context needs can resolve both this error and a separate KV-cache out-of-memory error on the same setup.