dependency version check & fix ignore_eos logic

[grimoire]

add version check

• torch [2.0.0, 2.1.2]
• triton [2.1.0, 2.2.0)
• transformers [4.33.0, 4.37.1]
• peft [0.7.1]

fix ignore_eos, model eos and stop words would be moved to bad words when ignore_eos is True.
Tested on chat (with manually set ignore_eos=True), benchmark_pytorch_throughput

[AllentDan]

No need to clear stop_words? I mean if a word is both a bad_word and a stop_word, what is the expected behavior?

[grimoire]

lmdeploy/lmdeploy/pytorch/engine/logits_process.py

Lines 63 to 68 in aaf3942

	if bad_words:
	bad_words = list(set(bad_words))
	bad_words_bias = new_scores.new_zeros(new_scores.size(1))
	bad_words_bias[bad_words] = filter_value
	new_scores += bad_words_bias[None]

bad words would be unique before sampling.

[lvhan028]

The following tests passed:

1. lmdeploy chat torch /workspace/models-142/models/internlm2-chat-7b/
2. lmdeploy serve api_server /workspace/models-142/models/internlm2-chat-7b/ --server-port 33337
3. profile_throughput.py