Fix create_completion capping all batch prompts to the shortest context's max_tokens

[Chessing234]

Bug

In create_completion, a pre-generation loop validates context length for each prompt:

for text in request.prompt:
    max_tokens, error_check_ret = await check_length(
        request, text, request.max_tokens, worker_addr
    )
    if error_check_ret is not None:
        return error_check_ret

    if isinstance(max_tokens, int) and max_tokens < request.max_tokens:
        request.max_tokens = max_tokens  # ← mutates shared state

check_length returns min(request.max_tokens, context_len - token_num) — the maximum tokens that can be generated given that prompt's context usage. The mutation request.max_tokens = max_tokens carries this reduced value into subsequent iterations.

Root cause

If prompt A is long (leaves 148 tokens available), request.max_tokens is reduced from 1000 to 148. Prompt B (short, leaves 1948 tokens available) is then checked against the already-reduced 148 and the generation loop uses 148 for both. Prompt B is silently limited to 148 tokens even though the user requested 1000 and the context easily fits them.

Why the fix is correct

Each prompt's available tokens are a function of its own length, not of other prompts in the batch. check_length already raises an error if a prompt overflows context, so the mutation's validation purpose is redundant. Without the mutation, generation uses request.max_tokens (the user's requested value) for all prompts; model workers naturally cap output at available context, producing correct results independently per prompt.