fix: correct BitNetForCausalLM model registration and tokenizer type

[goodcomm74]

Problem

Two bugs in utils/convert-hf-to-gguf-bitnet.py prevent the microsoft/BitNet-b1.58-2B-4T model from being converted on Apple Silicon (and likely other platforms):

Bug 1: Architecture name mismatch (capital 'N')

# Before (wrong)
@Model.register("BitnetForCausalLM")

# After (correct)
@Model.register("BitNetForCausalLM")

The config.json of microsoft/BitNet-b1.58-2B-4T declares:

"architectures": ["BitNetForCausalLM"]

This mismatch caused:

NotImplementedError: Architecture 'BitNetForCausalLM' not supported!

Bug 2: Wrong tokenizer type

# Before (wrong - looks for tokenizer.model which doesn't exist)
def set_vocab(self):
    self._set_vocab_sentencepiece()

# After (correct - uses tokenizer.json which exists)
def set_vocab(self):
    self._set_vocab_gpt2()

microsoft/BitNet-b1.58-2B-4T uses a GPT-2 style tokenizer (tokenizer.json), not SentencePiece (tokenizer.model). This caused:

FileNotFoundError: File not found: models/BitNet-b1.58-2B-4T/tokenizer.model

Testing

Tested on Apple Silicon M4 Max (macOS 15.3, arm64).

After this fix, the model converts successfully and runs via mlx-lm:

• Speed: ~200 tokens/sec
• Memory: ~1.2 GB

Note: tl1 quantization still requires preset kernels for BitNet-b1.58-2B-4T (not included in this PR). i2_s conversion works after this fix.

Related Issues

• Similar to Patch for TL1 kernel generation setup on Arm64 (Mac M1 Pro). #324 (Patch for TL1 kernel generation setup on Arm64)