vLLM fixes context errors caused by oversized output limits

vLLM 0.32.0 fixes a context-window bug that could reject valid prompts when clients requested large output limits. Generation is now capped at the space actually left after the prompt.

The merged PR illustrates the problem with a one-token discrepancy. With a 128,000-token context window, a 62,466-token prompt and max_tokens: 65535, the old calculation allowed only 62,465 input tokens. The request failed. After the change, it is accepted with an output ceiling of 65,534 tokens.

The PR cites clients such as Zed, Factory and opencode to explain why large output limits are common. A high ceiling does not mean the client expects every requested token to be available.

There was a less visible consequence: the calculation also reached the tokenizer. According to the PR, some models configured to truncate from the left could silently lose prompt text and disrupt prefix-cache reuse.

The regression tests cover accepting a fitting prompt despite a large output request, rejecting genuinely oversized input, and preserving the generation budget when tokenization settings are copied. Automatic truncation also retains one output token rather than filling the entire window with input.

The 0.32.0 release notes, published October 18, list the fix among the breaking changes. It does not enlarge the context window or guarantee the requested response length. It makes the existing window usable without treating the maximum answer length as space that must be reserved in advance.