fix: add MLX weight remapping for openai_privacy_filter / nemotron architecture

[kswanjitsu]

Problem

The openai_privacy_filter model family (e.g. OpenMed/privacy-filter-nemotron) cannot be loaded via create_mlx_pipeline because its HuggingFace weight namespace does not match the MLX model's namespace. All 140 parameters are rejected with ValueError: Received 140 parameters not in model.

Root cause

The existing remap_key() function handles BERT/RoBERTa/DeBERTa/DistilBERT/ELECTRA architectures but has no case for openai_privacy_filter. The OPF architecture differs from all others:

• Different top-level prefix (score.* → unembedding.*)
• Different layer container (model.layers.N.* → block.N.*)
• QKV fusion: HF has separate q/k/v_proj, MLX has single attn.qkv
• Different sub-key names (input_layernorm → attn.norm, router → gate)
• RMSNorm uses scale not weight
• Expert weights already in [E, in, out] format (no transpose needed)

Fix

Add _convert_opf_weights() and wire into convert_weights() for model types openai-privacy-filter, privacy-filter-nemotron, nemotron-privacy-filter. Also set classifier_bias: True in the MLX config when score.bias is present.

Testing

Verified on OpenMed/privacy-filter-nemotron — model now loads and runs inference correctly:

pipe = create_mlx_pipeline("OpenMed/privacy-filter-nemotron")
result = pipe("Patient John Smith DOB 01/01/1980")
# Returns: first_name=John, last_name=Smith, date=01/01/1980

🤖 Generated with Claude Code

[maziyarpanahi]

Thanks again for chasing this down. I pushed one follow-up commit onto this PR so it now carries the PR #57 fix plus the OPF/Nemotron remapping in one place.

What changed in the follow-up:

• BF16 tensors are now cast to explicit float32 before .numpy(). This is still needed because PyTorch CPU cannot expose torch.bfloat16 tensors through NumPy. I used t.float() rather than t.to(float), because Python float promotes to float64 in PyTorch and would unnecessarily increase memory/artifact size.
• The OPF/privacy-filter converter now validates the converted MLX key set and tensor shapes before returning weights. That means a partial mapping, for example missing one Q/K/V tensor, fails early instead of producing a broken artifact.
• Added focused regression tests for BF16 conversion, OPF HF→MLX remapping, QKV fusion order, and partial-QKV rejection.

The reason this failed for local conversion but not for the existing OpenMed demo is that the demo uses the already-exported OpenMed/privacy-filter-nemotron-mlx artifact. That artifact is already in the OpenMed MLX runtime layout. The broken path was only raw HF → MLX conversion from OpenMed/privacy-filter-nemotron, where the source checkpoint uses model.layers.*, score.*, separate q/k/v_proj, RMSNorm weight, and MoE expert tensors.

I also ran the full suite locally on the updated PR branch: 1153 passed, 1 skipped.