[Bug] Incorrect default argument and KeyError for llama

[ZisIsNotZis]

Currently, if kv_quant_granularity not given:

from omniserve import LLMEngine, EngineArgs
M = 'mit-han-lab/Llama-3-8B-Instruct-QServe-g128'
M = LLMEngine.from_engine_args(EngineArgs(M, ifb_mode=1, precision='w4a8kv4', quant_path=M))

results in

NotImplementedError: Unsupported kv_quant_granularity None

After adding one:

from omniserve import LLMEngine, EngineArgs
M = 'mit-han-lab/Llama-3-8B-Instruct-QServe-g128'
M = LLMEngine.from_engine_args(EngineArgs(M, ifb_mode=1, precision='w4a8kv4', quant_path=M, kv_quant_granularity='per_tensor'))

still results in

KeyError: 'model.layers.0.self_attn.qkv_proj.s2_scales'

I tried llama and mistral, non of them work. Maybe the model uploaded and readme are already too outdated?