Automates iterative esp-ppq quantization tuning to recover accuracy on ESP-DL models.
npx skills add https://github.com/espressif/esp-dl --skill espdl-quantize
Please help me install this Agent Skill. Skill: espdl-quantize Source: https://github.com/espressif/esp-dl/tree/main/tools/agents/skills/espdl-quantize Command: npx skills add https://github.com/espressif/esp-dl --skill espdl-quantize
Manually tuning esp-ppq quantization settings (calibration algorithms, TQT, equalization, bias correction, mixed precision) to recover post-quantization accuracy on ESP-DL targets is slow and error-prone. This Skill replaces the guess-and-rerun loop with a structured, distribution-aware search that iterates toward a target metric automatically.
Provide a user_quant.py module with your calibration dataloader and evaluate function, then ask the agent to run the espdl-quantize skill to tune my esp-ppq quantization settings until the target accuracy is reached.
High-intent search queries and answers about installing and using this skill.
Run the espdl-quantize skill with a user_quant.py contract module that provides a calibration dataloader and evaluate function. The skill iterates over calibration algorithms, TQT schedules, equalization, bias correction, and mixed precision until your target metric is reached or a plateau is detected.
The skill runs a mandatory Phase-2 cartesian product pairing each of kl, mse, and percentile calibration with the default TQT schedule. Calibration is not separable from the training pass, so the comparison of all three legs reveals which pairing performs best for your model's distribution.
No. All esp-dl targets use a POWER_OF_2 quantization policy, and esp-ppq's LSQDelegator silently disables scale training under POWER_OF_2, so LSQ degenerates to weight-only tuning. The harness auto-disables LSQ on these targets and recommends TQT, which trains log2_scale natively.
Yes, but with caution. esp-ppq officially marks equalization as not recommended for esp32p4's per-channel weight quantization, yet some MobileNet-family networks still benefit. The harness runs it with a warning and suggests trying it only after the calibration × TQT sweep has settled.
A single Python module exporting QUANT_CONFIG (model path, input shape, target chip, primary metric), create_calib_dataloader(), and evaluate(quant_graph). Torch flows also need get_torch_model(). The skill never edits this module and stores all iteration state under outputs/.
Fusion alignment (lever 3c) only helps when the top elementwise op's max_snr falls in the 0.20-0.30 band or its input float std ratio exceeds 5. Outside that band, alignment can overwrite an already-tight TQT scale and regress; the state machine skips it when the R8 trigger does not fire.
NVIDIATrain, evaluate, export, and run inference for TAO NVPanoptix3D panoptic 3D reconstruction models.
synqingBuild no-std Rust firmware for ESP32 microcontrollers using ESP-RS and Embassy.
AevonixEnable CPU-based LLM inference on Apple Silicon and non-NVIDIA hardware with GGUF quantization via llama.cpp.
espressifImplement, test and optimize neural network operators for the ESP-DL framework.
