[CB] bug fix: account for prefill token

[yannicks1]

[CB] bug fix: account for prefill token when asserting context length

Prefill already provides one new token (without requiring any KV cache for it).

Example: for max model length 2048 it is possible to do a prefill on a prompt of size 2048 to (32 blocks in Spyre) when only requesting 1 token. A 33rd block is only needed if a 2nd output token was requested and that would violate the max model length.

changes:

• correct assertion in platform.py: allow any requests that satisfy: prompt_padding_len + max_tokens - 1 <= max_model_len
• correct long_context.py: allowing prompt_len <= max_model_len while setting
  new_tokens = max_model_len + 1 - prompt_padding_len

[github-actions]

👋 Hi! Thank you for contributing to vLLM support on Spyre.
Just a reminder: Make sure that your code passes all the linting checks, otherwise your PR won't be able to be merged. To do so, first install the linting requirements, then run format.sh and commit the changes. This can be done with uv directly:

uv sync --frozen --group lint --active --inexact

Or this can be done with pip:

uv pip compile --group lint > requirements-lint.txt
pip install -r requirements-lint.txt
bash format.sh

Now you are good to go 🚀

[sducouedic]

I cannot help but read plen as if it was one word, can we rename that?

Suggested change

	args.max_model_len + 1 - round_up(plen) for plen in prompt_lens
	args.max_model_len + 1 - round_up(p_len) for p_len in prompt_lens

or

Suggested change

	args.max_model_len + 1 - round_up(plen) for plen in prompt_lens
	args.max_model_len + 1 - round_up(prompt_len) for prompt_lenin prompt_lens

[joerunde]

+1 for prompt_len, we don't need to save the bytes

[joerunde]

prompt_lenin