Fix sparsity arg in Engine/ModelArgs

[mgoin]

The sparsity argument was being ignored because it was switched around with the new quantization_param_path arg.

>>> from vllm import LLM
>>> model = LLM("nm-testing/OpenHermes-2.5-Mistral-7B-pruned50", sparsity="sparse_w16a16", max_model_len=1024)
INFO 04-10 18:02:12 llm_engine.py:81] Initializing an LLM engine (v0.2.0) with config: model='nm-testing/OpenHermes-2.5-Mistral-7B-pruned50', speculative_config=None, tokenizer='nm-testing/OpenHermes-2.5-Mistral-7B-pruned50', tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.bfloat16, max_seq_len=1024, download_dir=None, load_format=auto, tensor_parallel_size=1, disable_custom_all_reduce=True, quantization=None, sparsity=None, enforce_eager=8192, kv_cache_dtype=auto, quantization_param_path=False, device_config=cuda, seed=0)

Notice the sparsity=None in the log, which doesn't match the arg we specified.

Because the model config construction in create_engine_config() doesn't use named arguments, these args got flipped around, see the mass of:

    def create_engine_config(self, ) -> EngineConfig:
        device_config = DeviceConfig(self.device)
        model_config = ModelConfig(
            self.model, self.tokenizer, self.tokenizer_mode,
            self.trust_remote_code, self.download_dir, self.load_format,
            self.dtype, self.seed, self.revision, self.code_revision,
            self.tokenizer_revision, self.max_model_len, self.quantization,
            self.quantization_param_path, self.sparsity, self.enforce_eager,
            self.max_context_len_to_capture, self.max_logprobs)
            ...

[bnellnm]

considering how this broke, should we be using names for some/most of these arguments?

[mgoin]

i do want to, but that would further diverge from the upstream so i didn't go for it

[robertgshaw2-neuralmagic]

lets do an upstream PR for this (i.e. for bills idea of making these named args)