quantization/precision_map.json

{
  "vision_model": "bfloat16",
  "projector": "bfloat16",
  "language_model.model.embed_tokens": "bfloat16",
  "language_model.model.norm": "bfloat16",
  "language_model.lm_head": "bfloat16",
  "language_model.model.layers.*.input_layernorm": "bfloat16",
  "language_model.model.layers.*.post_attention_layernorm": "bfloat16",
  "language_model.model.layers.*.self_attn.q_proj": "mxfp8",
  "language_model.model.layers.*.self_attn.k_proj": "mxfp8",
  "language_model.model.layers.*.self_attn.v_proj": "mxfp8",
  "language_model.model.layers.*.self_attn.o_proj": "mxfp8",
  "language_model.model.layers.*.mlp.gate_proj": "mxfp8",
  "language_model.model.layers.*.mlp.up_proj": "mxfp8",
  "language_model.model.layers.*.mlp.down_proj": "mxfp8",
  "language_model.model.layers.*.mlp.shared_experts.gate_proj": "mxfp8",
  "language_model.model.layers.*.mlp.shared_experts.up_proj": "mxfp8",
  "language_model.model.layers.*.mlp.shared_experts.down_proj": "mxfp8",
  "language_model.model.layers.*.mlp.switch_mlp.gate_proj": "mxfp8",
  "language_model.model.layers.*.mlp.switch_mlp.up_proj": "mxfp8",
  "language_model.model.layers.*.mlp.switch_mlp.down_proj": "mxfp8",
  "language_model.model.layers.*.mlp.gate": "bfloat16",
  "_comment": "MLX module paths for DeepseekOCR. Refine via layer_sensitivity.py before final release.",
  "_sensitivity_override": {
    "_description": "Layers identified as OCR-sensitive by sensitivity analysis should be promoted to bfloat16. Run layer_sensitivity.py to populate this section.",
    "examples": [
      "language_model.model.layers.11.self_attn.q_proj -> bfloat16 (if digit CER degrades >2%)",
      "language_model.model.layers.1.mlp.switch_mlp -> bfloat16 (if table score degrades >1%)"
    ]
  }
}