load_in_4bit does not work

[PatchouliPatch]

the following is a modified version of the code from the Phi3 huggingface page.
For some reason, load_in_4bit does not work but load_in_8bit does.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, BitsAndBytesConfig
import time

torch.random.manual_seed(0)

nf4_config = BitsAndBytesConfig(
   #load_in_8bit=True,
   load_in_4bit=True,
   #bnb_4bit_quant_type="nf4",
   #bnb_4bit_use_double_quant=True,
   bnb_4bit_compute_dtype=torch.float16
)

model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct", 
    device_map="cuda", 
    torch_dtype="auto", 
    trust_remote_code=True,
    quantization_config=nf4_config
)
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")

messages = [
    {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"},
    {"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."},
    {"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"},
]

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)

generation_args = {
    "max_new_tokens": 500,
    "return_full_text": False,
    "temperature": 0.0,
    "do_sample": False,
}
t = time.time()
output = pipe(messages, **generation_args)
print('===========================================')
print(f"inference took {time.time() - t} seconds")
print('===========================================')
print(output[0]['generated_text'])

the inference does not seem to finish.
at full precision the output is:

===========================================
inference took 3.9708006381988525 seconds
===========================================
 To solve the equation 2x + 3 = 7, follow these steps:

1. Subtract 3 from both sides of the equation: 2x + 3 - 3 = 7 - 3
2. Simplify: 2x = 4
3. Divide both sides by 2: 2x/2 = 4/2
4. Simplify: x = 2

and at 8-bit quant:

===========================================
inference took 18.259739637374878 seconds
===========================================
 To solve the equation 2x + 3 = 7, follow these steps:

1. Subtract 3 from both sides of the equation: 2x + 3 - 3 = 7 - 3
2. Simplify the equation: 2x = 4
3. Divide both sides of the equation by 2: 2x / 2 = 4 / 2
4. Simplify the equation: x = 2

So, the solution to the equation 2x + 3 = 7 is x = 2.

im using ROCm 6.0.3 and the stable branch of torch after finding out that the latest nightly builds of 6.1.x are having problems with quantization

any possible debugging steps?