AIMET repo skill

AIMET (AI Model Efficiency Toolkit) is a toolkit for quantizing, compressing, analyzing, and exporting trained PyTorch and ONNX models, with repository-side GenAILab scorecard tooling and Qualcomm target-deployment guidance. Use this skill when a task names AIMET, aimet-torch, aimet-onnx, aimet_torch, aimet_onnx, QuantizationSimModel, AIMET encodings, AIMET compression, GenAILab, QNN/QAIRT/AI Hub, or AIMET repository build/test workflows.

This skill is a router. Open the focused sub-skill and bundled references before changing code, adjusting build settings, or running expensive examples.

Start here

1. Identify the surface. Decide whether the task is about installation/build, PyTorch quantization, ONNX quantization, GenAILab LLM/VLM evaluation, model/download credentials, cluster/Pod execution, or Qualcomm target deployment.
2. Verify the environment before debugging AIMET. For installed-package tasks, run python scripts/quick_smoke.py --framework both from this skill directory or copy it into a clean working area. For GenAILab configs, run python scripts/genai_config_preflight.py <config.yaml> --framework <torch|onnx|both> --print-command before long runs.
3. Do not assume CUDA is required. Core PyTorch and ONNX QuantSim workflows are CPU-valid; CUDA is required only for GPU provider checks, CUDA-marked tests, source CUDA builds, GenAILab runs whose model size needs GPU memory, or user workloads that cannot run on CPU.
4. Treat examples as evidence, not mandatory runtime. AIMET examples often expect ImageNet-scale data, Hugging Face models, cluster credentials, or target SDKs. Use the distilled workflows and preflight scripts here unless the user has explicitly supplied those assets.

Route map

User task	Read first	Why
Install aimet-torch/aimet-onnx, repair dependency conflicts, choose CPU vs CUDA, build from source, or run focused repo checks	install-and-build	Captures package names, build flags, dependency variants, and safe smoke checks.
Quantize a PyTorch model, prepare a model for AIMET, run calibration, use QAT, or export a Torch QuantSim model	torch-quantization	Covers aimet_torch workflows, model preparation, BatchNorm folding, QuantSim, encodings, and Torch export.
Quantize an ONNX model, choose ONNX Runtime providers, compute/load encodings, export QDQ, or run ONNX PTQ utilities such as SeqMSE/AdaRound	onnx-quantization	Covers aimet_onnx graph/session workflows, precision controls, PTQ utilities, and provider decisions.
Diagnose accuracy loss, use QuantAnalyzer/visualization, mixed precision, compression, deployment artifacts, or on-target inference handoff	optimization-analysis-deployment	Covers cross-framework optimization, compression, debugging workflow, export artifacts, and target-runtime boundaries.
Configure or run AIMET GenAILab LLM/VLM scorecards, recipe chains, local Torch/ONNX GenAI tests, online scorecards, exports, caches, or summaries	genai-lab	Covers python -m GenAILab, YAML contracts, framework choice, cache/export/result directories, and safe preflight.
Use Hugging Face tokens, gated/private models, GitHub Actions dispatch/download, AWS/S3 checkpoint artifacts, SAML login, or metric-version-aware result comparisons	model-access-and-credentialed-evaluation	Covers model downloads, credential boundaries, artifact download, cache planning, and comparability rules.
Launch or reuse Argo/Kubernetes pods, sync AIMET to /scratch, run GenAILab/source builds on remote GPU pods, list workflows, or stop pods	cluster-pod-workflows	Covers resource requests, launch/sync/exec/list/stop helper commands, and remote-state safety.
Send AIMET ONNX/encoding exports to Qualcomm AI Hub, QAIRT, QNN, HTP, DLC conversion, profiling, inference, or SDK command generation	qualcomm-sdk-deployment	Covers export validation, AI Hub compile/profile/inference, local QAIRT/QNN command generation, and target-runtime boundaries.

Bundled references

• references/install-and-build.md summarizes install commands, source-build variants, dependency files, and safe repo-maintenance test selection.
• references/api-overview.md records verified public entry points and signatures for the main aimet_torch, aimet_onnx, and compression APIs.
• references/workflows.md distills the main Torch, ONNX, compression, analysis, and export workflows into short recipes.
• references/backend-compatibility.md records CPU/CUDA/provider expectations and which backend checks are evidence-bearing.
• references/troubleshooting.md maps common symptoms to likely causes and recovery steps.
• references/genai-lab.md records self-contained GenAILab local/online/config/cache/export workflows.
• references/model-access-and-credentialed-evaluation.md records Hugging Face, GitHub Actions, AWS/S3, SAML, cache, and metric-comparability workflows.
• references/cluster-pod-workflows.md records Argo/Kubernetes launch, sync, exec, list, and stop workflows.
• references/qualcomm-sdk-workflows.md records Qualcomm AI Hub and local QAIRT/QNN deployment sequences.
• references/repo-provenance.md records the source snapshot and included/excluded evidence paths used to create this skill.

Bundled scripts

• scripts/quick_smoke.py performs a tiny installed-package import plus PyTorch and ONNX QuantSim calibration smoke check.
• scripts/build_from_source.sh is a safe adapter for source builds in a user-created environment; it checks build prerequisites, applies CMAKE_ARGS, and can run the bundled smoke check after install.
• scripts/inspect_export.py validates exported ONNX and AIMET encodings artifacts before handing them to target-runtime tooling.
• scripts/genai_config_preflight.py validates GenAILab YAML documents without downloading models/datasets or importing heavy backends.
• scripts/genai_results_summary.py summarizes GenAILab profiling_data.json files and warns on mixed metric scoring versions.
• scripts/download_genai_checkpoint.sh safely validates/downloads S3 GenAI checkpoint/export zips without auto-installing AWS/SAML tooling.
• scripts/cluster_pod_helper.sh provides self-contained Argo/Kubernetes preflight, launch, sync-once, exec, list, and stop commands.
• scripts/qairt_command_builder.py generates local QAIRT/QNN conversion, quantization, context, and inference commands from AIMET exports.
• scripts/qai_hub_qnn_job.py provides a dry-run-safe AI Hub QNN compile/profile/inference entry point.

Scope and limits

This repaired skill covers the public AIMET package surfaces under aimet_torch and aimet_onnx, common quantization/compression/analysis workflows, maintainer install/build/test orientation, GenAILab LLM/VLM scorecards, model/checkpoint access, credentialed evaluation boundaries, cluster/Pod execution patterns, and Qualcomm AI Hub / QAIRT / QNN target-deployment handoff. Heavy or credentialed actions remain gated: do not run large Hugging Face, ImageNet, S3, GitHub Actions, cluster, or SDK workflows until the user supplies assets/credentials and approves the runtime or remote-state effect.

Operating rules

• Prefer installed-package introspection and small QuantSim smoke checks before blaming repo source code.
• Use source-build commands only inside a dedicated environment; never mutate Conda base or a user production environment without approval.
• Do not run ImageNet, Hugging Face, GenAILab, online scorecards, cluster/Pod, S3/AWS, or on-target SDK examples unless the user has supplied datasets/models/credentials/SDKs and approved the runtime or remote-state cost.
• If CUDA behavior matters, verify an actual CUDA tensor allocation or ONNX Runtime CUDA provider; a CPU import is not GPU evidence.
• Keep exported artifacts together: an .onnx model plus the matching AIMET .encodings JSON are both needed for downstream quantized deployment flows. Use QDQ exports only when the downstream runtime explicitly expects ONNX QuantizeLinear/DequantizeLinear nodes.
• Do not import this generated skill into the live router unless a future user explicitly requests import.