Skip to content

export-matrix.md

Latest commit

 

History

History
83 lines (71 loc) · 5.53 KB

export-matrix.md

File metadata and controls

83 lines (71 loc) · 5.53 KB
Raw

Export Format Support Matrix

Which combinations of architecture family, export backend, and precision tier does OpenMed actually support today? This page is the single answer, kept honest against the code by a coherence test in tests/unit/test_export_matrix.py.

Use it before you plan a conversion so you know up front whether your architecture reaches your target runtime.

Support levels

  • ✅ Supported — a first-class path with a documented converter, unit tests, and (where relevant) a certification gate.
  • ⚠️ Partial / conditional — path exists but has a documented restriction (for example a family allowlist, or "certified only when recall holds").
  • ❌ Not supported — no converter, or explicitly rejected by the backend.

The matrix rows are the Hugging Face model-type strings and OpenMed family names most contributors search for. The MLX aliases column lists the token-classification aliases understood by openmed.mlx.models._SUPPORTED_TOKEN_CLASSIFICATION_MODEL_TYPES (or the MLX custom-family set). The coherence test asserts every entry in that allowlist appears here as a backtick-wrapped alias, so the matrix cannot drift silently when a new family is added to the code.

Matrix

Family MLX aliases MLX fp MLX INT8 MLX INT4 CoreML fp16 CoreML INT8 ONNX (default) ONNX Android Transformers.js GGUF (embedding)
BERT bert ✅ ✅ ⚠️ INT4-if-recall-holds ✅ ✅ ✅ ✅ ✅ ⚠️ backbone only (see limitations)
DistilBERT distilbert ✅ (via bert) ✅ ⚠️ INT4-if-recall-holds ✅ ✅ ✅ ✅ ✅ ⚠️ backbone only
Electra electra ✅ (via bert) ✅ ⚠️ INT4-if-recall-holds ✅ ✅ ✅ ✅ ✅ ⚠️ backbone only
RoBERTa roberta ✅ (via bert) ✅ ⚠️ INT4-if-recall-holds ✅ ✅ ✅ ✅ ✅ ⚠️ backbone only
XLM-RoBERTa xlm-roberta, xlm_roberta ✅ (via bert) ✅ ⚠️ INT4-if-recall-holds ✅ ✅ ✅ ✅ ✅ ⚠️ backbone only
DeBERTa-v2 deberta, deberta-v2 ✅ ✅ ⚠️ INT4-if-recall-holds ✅ ✅ ✅ ✅ ✅ ⚠️ backbone only
ModernBERT modernbert, modern-bert ✅ ✅ ⚠️ INT4-if-recall-holds ❌ family not in CoreML allowlist ❌ ⚠️ upstream ONNX support required ⚠️ same as ONNX (default) ⚠️ same as ONNX (default) ❌
Longformer longformer ✅ ✅ ⚠️ INT4-if-recall-holds ❌ family not in CoreML allowlist ❌ ⚠️ upstream ONNX support required ⚠️ same as ONNX (default) ⚠️ same as ONNX (default) ❌
OpenAI Privacy Filter openai-privacy-filter, privacy-filter, privacy-filter-nemotron, nemotron-privacy-filter, privacy-filter-multilingual, multilingual-privacy-filter ✅ ✅ ⚠️ INT4-if-recall-holds ❌ family not in CoreML allowlist ❌ ❌ no ONNX converter path today ❌ ❌ ❌
GLiNER (span) gliner-uni-encoder-span ✅ (MLX custom family) ✅ ⚠️ INT4-if-recall-holds ❌ ❌ ❌ ❌ ❌ ❌
GLiClass gliclass-uni-encoder ✅ (MLX custom family) ✅ ⚠️ INT4-if-recall-holds ❌ ❌ ❌ ❌ ❌ ❌
GLiNER (token-relex) gliner-uni-encoder-token-relex ✅ (MLX custom family) ✅ ⚠️ INT4-if-recall-holds ❌ ❌ ❌ ❌ ❌ ❌

Known limitations

The cells above encode a handful of load-bearing constraints; the source of truth is the backend code, but the essentials are:

  • GGUF is embedding-only. openmed.gguf.convert deliberately rejects token-classification heads: llama.cpp can run the BERT-family encoder embedding path but does not expose the token classifier head required to produce OpenMed entity labels. Export the encoder backbone (for example a checkpoint whose config.json names BertModel rather than BertForTokenClassification) instead. See GGUF Embedding Export.
  • CoreML is token-classification-only and enforces a family allowlist. openmed.coreml.convert.SUPPORTED_TOKEN_CLASSIFICATION_MODEL_TYPES accepts bert, distilbert, electra, roberta, xlm-roberta, and deberta-v2. Any other architecture fails before tracing with an error that lists the supported families. See CoreML Packaging.
  • MLX enforces a family allowlist. The MLX dispatcher resolves architectures through _SUPPORTED_TOKEN_CLASSIFICATION_MODEL_TYPES (BERT-family, DeBERTa-v2, ModernBERT, and Longformer) plus the MLX-custom set (gliner-uni-encoder-span, gliclass-uni-encoder, gliner-uni-encoder-token-relex) and the privacy filter family. Architectures outside the allowlist raise a clear "unsupported family" error rather than falling through silently. See MLX Backend.
  • INT4 is certified only when recall holds. openmed.mlx.convert will emit an INT4 artifact for any allowlisted family, but the artifact is marked certified: false in recall_delta.json when the measured per-label recall delta exceeds the INT4 limit. Release gates block over-budget INT4 formats. See MLX Quantized Export Certification.

See also