lfm2.5-1.2b-intel-lora

LFM Intel LoRA Smoke Artifacts

This repository contains tiny synthetic artifacts for validating an Intel/OpenVINO LoRA optimization pipeline.

It does not contain full LiquidAI/lfm2.5-1.2b-instruct model weights. The current Windows environment blocks the IPEX-backed path, so the files here validate the mechanics of LoRA adapter training, ONNX export, OpenVINO IR conversion, INT8 quantization, and CPU inference latency on a small synthetic model.

Artifacts

• adapter_smoke.pt: synthetic LoRA adapter checkpoint from a 5-step smoke run
• smoke_result.json: smoke training metrics
• model.onnx: ONNX export of the tiny smoke model
• model.json: ONNX export metadata
• openvino-ir/model.xml: FP OpenVINO IR definition
• openvino-ir/model.bin: FP OpenVINO IR weights
• openvino-ir/model_ir.json: FP OpenVINO IR conversion metadata
• openvino-int8/model_int8.xml: INT8 OpenVINO IR definition
• openvino-int8/model_int8.bin: INT8 OpenVINO IR weights
• openvino-int8/model_int8_quantization.json: INT8 quantization metadata
• cpu_latency.json: CPU latency validation metrics

CPU Validation

Measured on CPU with 50 iterations after 5 warmups:

• FP mean latency: 0.053568 ms
• FP p95 latency: 0.065855 ms
• INT8 mean latency: 0.061286 ms
• INT8 p95 latency: 0.087240 ms
• Max FP-vs-INT8 absolute difference: 0.64642227
• INT8 predictions finite: true

Environment Notes

• openvino==2024.6.0
• openvino-dev==2024.6.0
• nncf==2.8.0
• onnx==1.21.0
• onnxscript==0.7.0

nncf==3.1.0 was incompatible with openvino==2024.6.0 in this environment because openvino.Node is not available. nncf==2.8.0 successfully produced the INT8 IR.

Intended Use

Use these files as a small reproducible smoke package for pipeline validation. Do not treat them as a real fine-tuned LFM checkpoint.