[lmi] log warnings for unused generation parameters across all rollin…

[siddvenk]

…g batch backends instead of throwing errors

Description

This PR standardizes the behavior across all rolling batch backends to log warnings for unused generation parameters. Before, some backends would throw errors, and some would succeed.

I have also logged the unused parameters, and the supported set of parameters. The supported set of parameters does include some stuff that the user should not set, so it's not perfect. I'm ok to remove that portion for now and just log the unused params to the warning.

Testing Samples

Using the following request:

curl -X POST http://localhost:8080/invocations -H 'Content-Type: application/json' -d '{"inputs": "Hello there", "parameters": {"badkwarg": 4}, "stream": false}'

vllm (would fail before)

response: {"generated_text": ", I’m a 20 year old guy from the UK. I’ve been playing guitar for about 10 years now and"}
logs: The following parameters are not supported by vllm: {'badkwarg'}. The supported parameters are {'include_stop_str_in_output', 'length_penalty', 'min_p', 'presence_penalty', 'temperature', 'prompt_logprobs', 'best_of', 'skip_special_tokens', 'early_stopping', 'stop_token_ids', 'repetition_penalty', 'use_beam_search', 'logprobs', 'logits_processors', 'top_k', 'stop', 'top_p', 'ignore_eos', 'n', 'frequency_penalty', 'seed', 'max_tokens', 'spaces_between_special_tokens'}

lmi-dist (would fail before)

response: {"generated_text": ", I’m a 20 year old guy from the UK. I’ve been playing guitar for about 10 years now and"}
logs: The following parameters are not supported by lmi-dist: {'badkwarg'}. The supported parameters are {'typical_p', 'max_tokens', 'truncate', 'ignore_eos', 'length_penalty', 'top_p', 'repetition_penalty', 'best_of', 'include_stop_str_in_output', 'stop_token_ids', 'top_k', 'prompt_logprobs', 'use_beam_search', 'spaces_between_special_tokens', 'logprobs', 'stop', 'skip_special_tokens', 'seed', 'frequency_penalty', 'temperature', 'min_p', 'logits_processors', 'sampling_params', 'presence_penalty', 'n', 'early_stopping'}

hf-accelerate w/scheduler

response: {"generated_text": ", I’m a 20 year old guy from the UK. I’ve been playing guitar for about 10 years now and"}
logs: The following parameters are not supported by huggingface accelerate rolling batch: {'seed', 'badkwarg'}. The supported parameters are {'topp', 'eos_token_id', 'use_lru_kv_cache', 'sampling', 'topk', 'alpha', '_max_seqlen', 'beam', 'max_new_seqlen', 'temperature', 'pad_token_id'}

deepspeed

response: {"generated_text": ", I’m a newbie here. I’m a 20 year old guy from the UK. I’ve been a fan"}
logs: The following parameters are not supported by deepspeed: {'badkwarg'}. The supported parameters are {'watermark', 'repetition_penalty', 'temperature', 'max_new_tokens', 'top_p', 'do_sample', 'seed', 'typical_p', 'stop_sequences', 'top_k', 'ignore_eos_token'}

tnx

response: {"generated_text": ", I’m a newbie here. I’m a 20 year old guy from the UK. I’ve been a fan"}
logs: The following parameters are not supported by neuron: {'badkwarg'}. The supported parameters are {'epsilon_cutoff', 'bos_token_id', 'begin_suppress_tokens', 'num_assistant_tokens_schedule', 'forced_eos_token_id', 'do_sample', '_commit_hash', 'max_new_tokens', 'output_attentions', 'num_assistant_tokens', 'encoder_no_repeat_ngram_size', 'use_cache', 'length_penalty', 'low_memory', 'min_length', 'encoder_repetition_penalty', 'output_scores', 'exponential_decay_length_penalty', 'return_dict_in_generate', '_from_model_config', 'num_beam_groups', 'decoder_start_token_id', 'pad_token_id', 'top_k', 'repetition_penalty', 'top_p', 'num_beams', 'force_words_ids', 'forced_decoder_ids', 'num_return_sequences', 'diversity_penalty', 'sequence_bias', 'min_new_tokens', 'renormalize_logits', 'guidance_scale', 'temperature', 'remove_invalid_values', 'eos_token_id', 'early_stopping', 'no_repeat_ngram_size', 'generation_kwargs', 'constraints', 'bad_words_ids', 'output_hidden_states', 'eta_cutoff', 'transformers_version', 'penalty_alpha', 'max_time', 'max_length', 'forced_bos_token_id', 'suppress_tokens', 'typical_p'}

optimum:

response: {"generated_text": ", I’m a newbie here. I’m a 20 year old guy from the UK. I’ve been a fan"}
logs: The following parameters are not supported by neuron: {'badkwarg'}. The supported parameters are {'forced_eos_token_id', 'penalty_alpha', 'num_beams', 'forced_decoder_ids', 'sequence_bias', 'output_hidden_states', 'typical_p', 'pad_token_id', 'temperature', '_commit_hash', 'bad_words_ids', 'top_p', 'suppress_tokens', 'generation_kwargs', 'constraints', 'exponential_decay_length_penalty', 'begin_suppress_tokens', 'length_penalty', 'num_return_sequences', 'min_length', 'encoder_no_repeat_ngram_size', 'force_words_ids', 'max_time', 'num_beam_groups', 'low_memory', 'no_repeat_ngram_size', 'top_k', 'guidance_scale', 'max_new_tokens', 'diversity_penalty', 'encoder_repetition_penalty', 'do_sample', 'max_length', 'decoder_start_token_id', 'output_attentions', 'min_new_tokens', 'early_stopping', 'output_scores', 'num_assistant_tokens_schedule', 'repetition_penalty', 'forced_bos_token_id', 'bos_token_id', 'eos_token_id', 'use_cache', 'renormalize_logits', 'remove_invalid_values', 'return_dict_in_generate', 'num_assistant_tokens', '_from_model_config', 'eta_cutoff', 'transformers_version', 'epsilon_cutoff'}