Skip to content

RUNTIME_VALIDATION_RESULTS.md

Latest commit

 

History

History
133 lines (81 loc) · 3.22 KB

RUNTIME_VALIDATION_RESULTS.md

File metadata and controls

133 lines (81 loc) · 3.22 KB
Raw

Runtime Validation Results

Validation date: March 10, 2026

This document records what was actually validated after repository cleanup.

Environment Used

ADS automation:

  • outer Python: a local Windows orchestration environment validated with Python 3.13.2
  • ADS Python: the ADS 2026 Update1 bundled Python runtime
  • ADS install root: a local ADS 2026 Update1 installation
  • PDK: DemoKit_Non_Linear
  • PDK tech library: DemoKit_Non_Linear_tech
  • substrate: demo

ML workflow:

  • Python: a local ML environment validated with Python 3.12.11

Validated Pipeline Stages

ADS / EM Flow

Passed:

  • ADS runtime detection
  • JSON sample parsing
  • workspace creation
  • library creation
  • layout and rfpro_view creation
  • single-sample simulation
  • 3-sample parallel batch simulation
  • Touchstone export

Dataset And ML Flow

Passed:

  • HDF5 inspection
  • HDF5 generation from real .s2p results
  • 1-epoch training smoke test
  • direct verification / inference script
  • 5-epoch retraining of a new local 2-channel checkpoint on the 600-sample dataset

Important Fixes Confirmed During Validation

1. Reference Library Inference

Problem:

  • create-design-only could create a design whose substrate reference resolved to the target design library instead of the PDK tech library
  • this caused RFPro EM loading to fail with substrate-read errors

Fix:

  • parallel_version/subprocess_cli_parallel.py now infers the best reference library from workspace lib.defs

Result:

  • default design creation and simulation now work without manually passing --ref-library-name

2. Batch Config BOM Compatibility

Problem:

  • Windows-generated JSON config files with UTF-8 BOM failed in batch_config.py

Fix:

  • parallel_version/batch_config.py now reads config and JSON assets with utf-8-sig

Result:

  • both BOM and non-BOM config files validate successfully

3. Tool Script Import Reliability

Problem:

  • Pytorch_Model/src/tools/verify_model.py failed when executed directly because dataset was not importable from that working mode

Fix:

  • both verification tools now support direct execution and module execution

Result:

  • python verify_model.py ... and python -m tools.verify_model ... both work

Legacy Model Boundary

The former legacy checkpoint was renamed locally to:

Pytorch_Model/models/best_model_legacy_1ch_2port_16x16_pre20260310.pth

It is not compatible with the current 2-channel datasets and should remain a local legacy artifact.

See:

Local Replacement Checkpoint

A new local replacement checkpoint was generated during validation:

  • Pytorch_Model/models/best_model_2ch_2port_16x16_20260310.pth
  • matching manifest: Pytorch_Model/models/best_model_2ch_2port_16x16_20260310.json

Training summary:

  • dataset: Pytorch_Model/data/dataset_2port_16x16.h5
  • samples: 600
  • input channels: 2
  • output dimension: 42
  • epochs: 5
  • batch size: 16
  • seed: 42
  • best validation loss: about 0.004049

Single-sample verification:

  • sample 0 MSE: about 0.003831

This checkpoint is local-only because Pytorch_Model/models/ is intentionally ignored by Git.

See: