generate() silently misaligned results[i] with prompts[i] when
sampling_params / multi_modal_data list lengths did not match the number
of prompts: extra entries were silently dropped, missing entries raised
an uninformative IndexError, and a bare multi_modal_data dict paired with
several prompts raised an obscure TypeError/KeyError deep inside the
request loop.
Now the mismatches raise a clear ValueError up front (matching vLLM
upstream behavior), and a bare multi_modal_data dict is only accepted
for a single prompt. Also fixes the hidden nesting bug where a str
prompt + list multi_modal_data became [[...]].
DeepEP now records an HTTPS DeepJIT URL, so build_vendor no longer needs the git@github.com rewrite.
Co-authored-by: Shaoyuan Chen <chensy@deepseek.com>