Does setting torch_dtype=torch.float16 override 8-bit quantization in BitsAndBytes?

I'm trying to run the Qwen2.5-Coder-3B model locally with 8-bit quantization using BitsAndBytes.

While loading the model, I noticed that some examples also specify torch_dtype=torch.float16. From my understanding, torch_dtype mainly affects the activation and output dtypes, not the quantized weights themselves.

However, I’m not completely sure whether setting torch_dtype=torch.float16 actually overrides the quantization or if both can safely coexist.

With torch_dtype=torch.float16

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(load_in_8bit=True)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-3B")

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-3B",
    quantization_config=bnb_config,
    torch_dtype=torch.float16,   # <-- does this override quantization?
    device_map="auto"
)

Without specifying torch_dtype

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(load_in_8bit=True)

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-3B",
    quantization_config=bnb_config,
    device_map="auto"
)

What is the difference between these two setups in terms of:

1. How model weights are stored and loaded (INT8 vs FP16)
2. The dtype used for activations and outputs during inference
3. Whether setting torch_dtype=torch.float16 can override or interfere with 8-bit quantization applied by BitsAndBytes