BitNet model outputs repetitive gibberish ('Breis') regardless of input during text generation

[EladWarshawsky]

Problem: The model generates repetitive, nonsensical outputs like "Breis" regardless of the input provided. This happens even with different generation settings (e.g., temperature, top_k, top_p).

from transformers import AutoModelForCausalLM, AutoTokenizer

Load the model and tokenizer from Hugging Face

model = AutoModelForCausalLM.from_pretrained("1bitLLM/bitnet_b1_58-3B") Loading checkpoint shards: 0%| | 0/3 [00:00<?, ?it/s]

from transformers import AutoConfig

Load model configuration

config = AutoConfig.from_pretrained("1bitLLM/bitnet_b1_58-3B") print(config) LlamaConfig { "_name_or_path": "1bitLLM/bitnet_b1_58-3B", "architectures": [ "BitnetForCausalLM" ], "attention_bias": false, "attention_dropout": 0.0, "bos_token_id": 1, "eos_token_id": 2, "head_dim": 100, "hidden_act": "silu", "hidden_size": 3200, "initializer_range": 0.02, "input_bits": 8, "intermediate_size": 8640, "max_position_embeddings": 2048, "mlp_bias": false, "model_type": "llama", "num_attention_heads": 32, "num_hidden_layers": 26, "num_key_value_heads": 32, "pad_token_id": 32000, "pretraining_tp": 1, "rms_norm_eps": 1e-05, "rope_scaling": null, "rope_theta": 10000.0, "tie_word_embeddings": true, "torch_dtype": "float16", "transformers_version": "4.45.1", "use_cache": true, "vocab_size": 32002, "weight_bits": 1 }

from tokenization_bitnet import BitnetTokenizer

Load the custom tokenizer

tokenizer = BitnetTokenizer( vocab_file="bitnet_tokenizer/tokenizer.model", special_tokens_map_file="bitnet_tokenizer/special_tokens_map.json", tokenizer_config_file="bitnet_tokenizer/tokenizer_config.json", unk_token='', bos_token='', eos_token='', )

Example input for inference

input_text = "?" inputs = tokenizer(input_text, return_tensors="pt")

Perform inference

outputs = model.generate(**inputs, max_length=20, temperature=0.7, top_k=50, top_p=0.9) print(tokenizer.decode(outputs[0]))

The output: ? suppis Breis Breis Breis Breis Breis Breis Breis Breis

[chinguetti]

I have the same issue.

[DatCaptainHorse]

The listed models are "proof-of-concept" at most, they are not meant for real use, someone would need to train a new model from beginning with BitNet in mind for coherent output.

[EladWarshawsky]

The listed models are "proof-of-concept" at most, they are not meant for real use, someone would need to train a new model from beginning with BitNet in mind for coherent output.

Yeah I guess 100B tokens isn't enough to even have autocomplete. That sucks. I will see about training it potentially.