Set materialize_only_last_token_logits=False when log_probs = True

[athitten]

MCore's DynamicInferenceEngine introduces materialize_only_last_token_logits in InferenceConfig with a default value of True to avoid computing logprobs for all tokens (prompt + completion) to save memory and optimize performance. By default in deployment, when the dynamic engine is built from the static engine it builds a DynamicInferenceContext with an InferenceConfig(...) that has the default value of True for materialize_only_last_token_logits.

This raises an error while evaluating the benchmarks that requires logprobs:

AssertionError: Prompt log probs cannot be calculated if only last token logits are materialized.
Set materialize_only_last_token_logits to False in DynamicInferenceContext
or skip_prompt_log_probs to True in SamplingParams.

As a fix this PR sets materialize_only_last_token_logits to False when logprobs = True which is the case for logrob benchmarks.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[athitten]

/ok to test a887964