sima-model-quantize-compile

name	sima-model-quantize-compile
description	Use when quantizing and compiling a standard ONNX model for the SiMa platform, including dependency bootstrap via sima-cli, optional real-data calibration, verification, and compilation for MLSoC or Modalix.

Quantize and Compile Standard ONNX Models for SiMa

Purpose

Use skills/quantize_compile/scripts/quantize_compile.py to quantize and compile standard ONNX models for SiMa (mlsoc or modalix).

Target Devices

• --device mlsoc selects gen1_target for MLSoC.
• --device modalix selects gen2_target for Modalix.

Use When

• You have an ONNX model and need SiMa quantized/compiled artifacts.
• You want a default, repeatable flow for Codex or Claude agents.

Prerequisites

• you must activate the Model Compiler virtual environment first:

activate-model-compiler

• Ensure sima-frontend installed (provides afe modules).
• Python deps: onnx, onnxsim, torch, numpy, Pillow must be available in the environment.

If required deps are missing:

sima-cli login
# amd64 host
sima-cli install tools/model-compiler/amd64
# arm64 host
sima-cli install tools/model-compiler/arm64

Default Workflow

1. Run operator audit first (required gate before quantize/compile; use model_surgery policy).
2. Validate model path and input/output interface.
3. If model has symbolic dimensions, staticify/simplify it with model_surgery helper.
4. Run quantization (with ONNX simplification enabled by default).
5. Compile for target device.
6. Optionally run verification (--verify).

Pre-Compile Audit (Required)

Run graph compatibility audit before quantization/compilation, following skills/model_surgery/SKILL.md target/dtype policy.

Typical command:

python3 skills/model_surgery/scripts/model_surgery_guard.py audit-model \
  --model /abs/path/model.onnx \
  --dtype int8

Default Command

python3 skills/quantize_compile/scripts/quantize_compile.py \
  --model_path /abs/path/model.onnx \
  --model_format onnx \
  --device modalix \
  --build_dir ./build

Recommended Reproducible Command

python3 skills/quantize_compile/scripts/quantize_compile.py \
  --model_path /abs/path/model.onnx \
  --model_format onnx \
  --input_names input \
  --input_shapes 1,3,224,224 \
  --output_names output \
  --device mlsoc \
  --build_dir ./build \
  --real_data \
  --dataset_images /abs/path/calib_images \
  --num_calib_samples 50 \
  --calib_method mse \
  --requant_mode sima \
  --verify

Key Flags

• --device {modalix|mlsoc}
• --input_names --input_shapes --output_names
• --real_data --dataset_images --num_calib_samples
• --bf16-weights --bf16-activations
• --calib_method --requant_mode
• --verify, --analyse-error
• --no-compile for quantize-only runs

Output

Artifacts are written to:

• <build_dir>/<model_basename>/

Notes

• Auto-shape detection may fail on dynamic ONNX inputs; pass explicit --input_shapes.
• For symbolic/dynamic ONNX dimensions, run:

python3 skills/model_surgery/scripts/onnx_static_simplify.py \
  --input /abs/path/model.onnx \
  --output /abs/path/model.static.sim.onnx \
  --replace batch=1

• --mla-tesselation exists for advanced MLA direct mode (argument spelling is tesselation in the script).