Quark

Collect and normalize environment facts (OS, Python, GPU, CUDA/ROCm, container state) before Quark installation or PTQ planning. Use this skill whenever a downstream skill needs confirmed hardware and toolchain facts, when the user mentions their setup, when you need to decide between CUDA/ROCm/CPU install paths, or when any accelerator-related assumption is unconfirmed. Also trigger when the user says things like "check my environment", "what GPU do I have", "is my setup ready for Quark", or before any install/quantization step where hardware facts are missing.

Validate workspace shape, model paths, output directories, and repo structure before downstream Quark skills proceed. Use this skill whenever a skill needs confirmed file paths, when the user provides a model path or output directory, when you need to distinguish a local model from a HuggingFace ID, or when any path assumption is unverified. Also trigger for "where should I put outputs", "is my model path right", or before model-intake/export steps.

Diagnose failed Quark ONNX installation, calibration, quantization, custom-op compilation, or export attempts. Use when the user reports an error, stack trace, invalid artifact, missing dependency, ORT execution-provider mismatch, silent CPU fallback, OOM during calibration, custom-op load failure (BFPQuantizeDequantize / MXQuantizeDequantize / Extended*), or unexpected quantization results from the ONNX flow. Trigger for "Quark ONNX error", "onnxruntime error", "quantize_static failed", "calibration crashed", "CUDAExecutionProvider not available", "ROCMExecutionProvider not available", "custom op library load failed", "model.onnx larger than 2GB", "external data not found", "AdaRound diverged", "GPTQ ONNX failed", "QuaRot failed", "NPU power-of-2 scale", any Python traceback mentioning quark.onnx / onnxruntime / onnx, or when the user pastes an error message related to Quark ONNX workflows.

Install or verify the correct ONNX Runtime build (and the `onnx` package) for a user's accelerator backend before Quark's ONNX-to-ONNX flow. Use when the user needs ONNX Runtime set up, reports onnxruntime version conflicts, CPU vs GPU variant mix-ups (only one variant of `onnxruntime*` may be installed at a time), missing CUDA/ROCm execution providers, or when `import onnxruntime` / `import onnx` fails. Trigger for "install onnxruntime", "pip install onnxruntime", "set up onnxruntime for ROCm", "set up onnxruntime for CUDA", "onnxruntime-gpu vs onnxruntime", "onnx version mismatch", "CPU-only onnxruntime installed", "ORT providers list missing CUDAExecutionProvider/ROCMExecutionProvider", or any request to get the correct ONNX Runtime build running. Also trigger when `quark-install` reports that ONNX Runtime is missing or mismatched before proceeding with the ONNX flow.

Inspect a target ONNX model and prepare metadata for Quark ONNX PTQ planning. Use when the user needs `.onnx` path validation, opset / IR version detection, input/output shape and dtype discovery, op-type histogram, quantizable-op counting, deployment-target compatibility checks (CPU / CUDA / ROCm / AMD NPU CNN / AMD NPU Transformer), or risk assessment before planning PTQ. Trigger for "analyze my ONNX model", "check this onnx model", "what opset is this", "can Quark quantize this .onnx", "is my model NPU-compatible", "does my model already have QDQ", "is my model larger than 2 GB", or when any ONNX quantization step needs model facts that are missing.

Build a Quark ONNX PTQ quantization plan from `model_analysis.json` and user intent. Use when the user needs preset selection (XINT8 / A8W8 / A16W8 / BF16 / BFP16 / MX* / MXFP* …), calibration method choice (MinMax / Entropy / Percentile / Distribution / NonOverflow / MinMSE / LayerWisePercentile), algorithm selection (CLE / AdaRound / AdaQuant / BiasCorrection / AutoMixprecision), deployment-target gating (CPU / CUDA / ROCm / AMD NPU CNN / AMD NPU Transformer), op-type include/exclude lists, weights-only INT4 (MatMulNBits) decisions, dynamic vs static quantization, or a draft `quant_plan.json`. Trigger for "what preset should I use for my ONNX model", "choose XINT8 vs A8W8", "plan ONNX PTQ", "INT4 "BFP16 / MXFP4 for my model", "calibration method for Ryzen AI", "should I use AdaRound or AdaQuant", "SmoothQuant alpha", or when the user has a model analysis and needs to decide how to quantize an ONNX model.

Validate Quark ONNX quantization output using four lightweight checks: auxiliary file copy alignment, expected non-quantized initializer MD5 byte-identity (inline `raw_data` + external-data byte ranges), model metadata equality after stripping quantization-only opset entries / Quark domains, and fuzzy node-pattern + op-type + dtype summaries with QDQ / `com.amd.quark` custom-op presence. Intended for post-quantization inspection of `model.onnx` (with or without `model.onnx_data`). Trigger for "validate ONNX quantization result", "check quantized .onnx output", "verify ONNX initializers", "did QDQ insertion happen", "are the non-quantized weights byte-identical".

Route Quark ONNX user goals to the correct atomic skill. Use when a user describes an ONNX quantization task in plain language — such as "install onnxruntime", "analyze my .onnx model", "choose a preset for my YOLO model", "plan ONNX PTQ", "quantize this .onnx with XINT8/BFP16/MXFP4", "validate my quantized .onnx", "debug a failed ONNX quantization", or any request that involves Quark's ONNX-to-ONNX flow. This is the ONNX-side entry point — trigger it whenever the user's intent involves Quark ONNX and the correct downstream skill is not immediately obvious.

Install or verify the AMD Quark package and its dependencies. Use when the user needs Quark package installation, dependency setup, or post-install verification — after PyTorch is already set up. Trigger for "install Quark", "set up Quark", "pip install amd-quark", "install the Quark package", dependency errors, import failures for quark modules, or any request to get Quark running. Also trigger when the user reports ModuleNotFoundError for quark or missing C++ compiler errors. For PyTorch installation or torch version issues, use quark-torch-install instead.

Diagnose failed Quark installation, PTQ execution, script generation, or export attempts. Use when the user reports an error, stack trace, invalid artifact, missing dependency, CUDA OOM, version mismatch, or unexpected PTQ results. Trigger for "Quark error", "PTQ failed", "quantization crashed", "CUDA out of memory", "import error", "model loading failed", "wrong results", any Python traceback mentioning quark/torch/transformers, or when the user pastes an error message related to Quark workflows.

Prepare export and downstream evaluation handoff for a planned or completed Quark PTQ run. Use when the user wants to export a quantized model to HuggingFace SafeTensors, ONNX, or GGUF format, package for deployment, or set up post-quantization evaluation. Trigger for "export model", "save quantized model", "convert to GGUF", "export quantized model", "export to HuggingFace format", or when the user has a completed or planned PTQ run and needs deployment outputs.

Low-memory file2file quantization for very large safetensors LLMs that cannot be loaded whole. Use when the user wants to run file2file quantization, adapt a new safetensors checkpoint without loading the full model, register an external LLMTemplate, inspect sharded checkpoint naming, generate wrapper or conversion scripts, or validate low-memory sharded quantization outputs. Trigger for "run file2file quantization", "quantize without loading the model", "large safetensors low-memory quantization", "file2file for DeepSeek/Qwen/MoE", "safetensors naming incompatible", "register LLMTemplate externally".

Install or verify the correct PyTorch build for a user's accelerator backend before Quark installation. Use when the user needs PyTorch set up, reports torch version conflicts, CUDA/ROCm package mismatches, or when torch.cuda.is_available() returns False. Trigger for "install PyTorch", "pip install torch", "set up torch for ROCm", "set up torch for CUDA", "torch version mismatch", "CPU-only torch installed", or any request to get the correct PyTorch build running. Also trigger when quark-install reports that PyTorch is missing or mismatched before proceeding.

End-to-end LLM accuracy evaluation on AMD ROCm (ROCm-only) — container setup, vLLM/SGLang/ATOM serving, lm-eval / lighteval / evalscope benchmarks. Use when the user wants to evaluate, benchmark, or compare an LLM's accuracy. Trigger for "evaluate this model", "run gsm8k/mmlu/mmlu_pro/aime/gpqa/hellaswag/arc", "test accuracy", "measure perplexity", "compare quantized model accuracy", "does this mxfp4 model lose accuracy". For evaluating Quark Agent Skills themselves, use quark-torch-eval-runner instead.

Inspect a target model and prepare metadata for Quark PTQ planning. Use when the user needs model path validation, architecture detection, quantization target discovery, layer counting, risk assessment, or transformer compatibility checks before planning PTQ. Trigger for "analyze my model", "check this model", "what architecture is this", "can Quark quantize X", "is this model supported", or when any quantization step needs model facts that are missing.

Build a Quark Torch LLM PTQ quantization plan from model analysis and user intent. Use when the user needs quantization scheme recommendations, exclusion lists, algorithm selection, KV cache decisions, per-layer overrides, or a draft quant_plan. Trigger for "quantize with FP8", "what scheme should I use", "plan PTQ", "INT4 quantization", "choose quantization config", "quantization plan", or when the user has a model analysis and needs to decide how to quantize it.

Validate Quark quantization output using four lightweight checks: auxiliary file copy alignment, excluded tensor MD5 byte-identity, config.json deep comparison after stripping quantization keys, and safetensors header pattern/dtype summaries. Intended for post-export or file2file validation. Trigger for "validate quantization result", "check quantized model output", "verify exported weights".

Route Quark user goals to the correct atomic skill or workflow. Use when a user describes a Quark task in plain language — such as "install Quark", "quantize a model", "analyze my model", "build a PTQ plan", "export the quantized model", "debug a failed run", "run the full PTQ pipeline", "run model quantization", "install Quark", or any request that involves AMD Quark model quantization. This is the entry point skill — trigger it whenever the user's intent involves Quark and the correct downstream skill is not immediately obvious.

End-to-end ONNX PTQ workflow for AMD Quark — from a `.onnx` file (and calibration data) to a quantized `.onnx` output. Use when the user wants a complete ONNX-to-ONNX PTQ pipeline: model intake, quantization planning, calibration-script generation, manifest, and confirmed execution. Trigger for "quantize my .onnx", "run ONNX PTQ end to end", "full ONNX quantization pipeline", "quantize yolov8/resnet50/yolo_nas with XINT8/A8W8/BFP16/MXFP*", "weights-only INT4 for my .onnx LLM", or any request that spans more than one ONNX PTQ step. When in doubt between routing to an atomic skill vs. the workflow, prefer this workflow if the user's request implies they want to go from a `.onnx` file to a quantized output.

Torch LLM PTQ workflow for AMD Quark — from model selection to quantized output. Use when the user wants a complete PTQ pipeline: model inspection, quantization planning, script generation, and optional execution. Stops at the quantized output. Trigger for "quantize my model", "run PTQ", "run model quantization", "full quantization pipeline", "quantize Llama/Qwen/Mistral with FP8/INT4", or any request that spans more than one PTQ step. When in doubt between routing to an atomic skill vs. the workflow, prefer this workflow if the user's request implies they want to go from model to quantized output.

L3 recipe that runs `quark.onnx.AutoSearchPro` end-to-end on a user `.onnx` model: intake → preset selection (or custom search space) → calibration / eval data reader → standalone autosearch script generation → confirmed execution → best_params reporting. Wraps the Optuna-driven hyperparameter search exposed in `quark/onnx/quantization/auto_search/auto_search_pro.py`. Use when the user wants to "auto search", "tune quantization", "find best quant config", "sweep AdaRound/AdaQuant", "two-stage search", or pick `ADVANCED_SEARCH` / `XINT8_SEARCH` / `A8W8_SEARCH` / `A16W8_SEARCH` for an ONNX model. Not for single-shot PTQ (use `quark-onnx-ptq`). Not for safetensors models (use `quark-torch-ptq`).

L3 recipe that runs a Torch LLM PTQ end-to-end for AMD Quark — for PyTorch / HuggingFace transformers models (safetensors input): quantize → validate → evaluate. Phase 1 delegates the full PTQ path (model intake → quantization planning → manifest generation → confirmed execution) to the `quark-torch-ptq` workflow; Phase 2 runs mandatory structural validation via `quark-torch-result-validator`; Phase 3 runs opt-in accuracy evaluation via `quark-torch-llm-eval`. Use when the user wants to "quantize and validate", "quantize and evaluate", "run PTQ end to end with accuracy check", "full PTQ pipeline including validation and eval", or "quantize Llama/Qwen/Mistral with FP8/INT4 and measure accuracy". For PTQ only (stop at the quantized output, no validation/eval), use `quark-torch-ptq`. Not for .onnx input models — use `quark-onnx-ptq` / `quark-onnx-autosearch-pro`.

Compare ONNX skill contracts and guidance against current Quark ONNX documentation and source entry points. Use when maintainers need to verify that ONNX install docs, custom-op registry, QConfig fields, preset and calibration lists, AutoSearchPro presets, deployment-target gates, or example-script invocations still match upstream Quark ONNX reality. Trigger for "check ONNX doc drift", "are ONNX skills still accurate", "verify against Quark ONNX docs", "fact-check quark-onnx-* skills", or after a Quark release when upstream ONNX documentation may have changed.

Manually verify that the Quark ONNX skill family behaves correctly across the four contract categories (routing, planning, artifact, recovery). Use when maintainers need to confirm that ONNX routing, planning, artifact generation, or error recovery skills still work as expected. Trigger for "verify the ONNX skills", "smoke-test ONNX routing", "check ONNX skill behavior", or before tagging a release that touches `quark-onnx-*` skills. This is a governance tool for skill maintainers, not for end users running ONNX model accuracy evaluation — for the latter use the upstream Quark ONNX evaluation tooling.

Detect upstream Quark ONNX changes that affect the ONNX skill family and classify required updates. Use when Quark ONNX docs, custom-op registry, quantization config presets, calibration methods, AutoSearchPro presets, ONNX Runtime install matrix, or source behavior under `quark/onnx/` may have drifted from the skill contracts. Trigger for "check if ONNX skills are up to date", "sync ONNX skills with Quark", "has Quark ONNX changed", "update ONNX skills after Quark upgrade", or when ONNX debugging reveals a mismatch between skill instructions and actual `quark.onnx` behavior.

Author or restructure a Quark Agent Skill so it conforms to this project's template, contracts, and layer rules. Use when a maintainer says "create a new skill", "add a skill for X", "scaffold a skill", "draft a SKILL.md", or when an existing skill needs a structural rewrite to match the format contract. Walks the author through capture, draft, test, iterate, and governance handoff. This is a maintainer-facing meta skill, not an end-user PTQ tool.

Compare skill contracts and guidance against current Quark documentation and source entry points. Use when maintainers need to verify that installation docs, artifact schemas, PTQ planning assumptions, CLI flags, or supported model lists still match upstream Quark reality. Trigger for "check doc drift", "are skills still accurate", "verify against Quark docs", or after a Quark release when upstream documentation may have changed.

Manually verify that the Quark Agent Skills system behaves correctly across the four contract categories (routing, planning, artifact, recovery). Use when maintainers need to confirm that routing, planning, artifact generation, or error recovery skills still work as expected. Trigger for "verify the skills", "smoke-test routing", "check skill behavior", or before tagging a release. This is a governance tool for skill maintainers, not for end users running model evaluation.

Detect upstream Quark changes that affect the skill system and classify required updates. Use when Quark docs, CLI flags, quantization templates, model support, or source behavior may have drifted from the skill contracts. Trigger for "check if skills are up to date", "sync skills with Quark", "has Quark changed", "update skills after Quark upgrade", or when debugging reveals a mismatch between skill instructions and actual Quark behavior.

Collect and normalize environment facts (OS, Python, GPU, CUDA/ROCm, container state) before Quark installation or PTQ planning. Trigger for "check my environment", "what GPU do I have", "is my setup ready for Quark", or when any accelerator-related assumption is unconfirmed. Also trigger before any install/quantization step where hardware facts are missing.

Install or verify the AMD Quark package and its dependencies. Trigger for "install Quark", "set up Quark", "pip install amd-quark", dependency errors, import failures for quark modules, ModuleNotFoundError for quark, or missing C++ compiler errors. For PyTorch installation or torch version issues, use quark-torch-install instead.

End-to-end Quark ONNX AutoSearchPro recipe — drives `quark.onnx.AutoSearchPro` (Optuna-based hyperparameter search) on a `.onnx` model to find the best quantization config (activation/weight spec, calibration method, CLE, AdaRound / AdaQuant, FastFinetune params). Use when the user wants to "auto search", "tune quantization", "find the best quant config", "sweep AdaRound/AdaQuant", "run AutoSearchPro / AutoSearch", "two-stage search", or pick one of the built-in presets (`ADVANCED_SEARCH`, `XINT8_SEARCH`, `A8W8_SEARCH`, `A16W8_SEARCH`) for their `.onnx` model. Not for HuggingFace / safetensors / PyTorch input models — use quark-torch-ptq instead. Not for single-shot ONNX PTQ without search — use quark-onnx-ptq.

Diagnose failed Quark ONNX installation, calibration, quantization, custom-op compilation, or export attempts. Use when the user reports an error, stack trace, invalid artifact, missing dependency, ORT execution-provider mismatch, silent CPU fallback, OOM during calibration, custom-op load failure (BFPQuantizeDequantize / MXQuantizeDequantize / Extended*), or unexpected quantization results from the ONNX flow. Trigger for "Quark ONNX error", "onnxruntime error", "quantize_static failed", "calibration crashed", "CUDAExecutionProvider not available", "ROCMExecutionProvider not available", "custom op library load failed", "model.onnx larger than 2GB", "external data not found", "AdaRound diverged", "GPTQ ONNX failed", "QuaRot failed", "NPU power-of-2 scale", any Python traceback mentioning quark.onnx / onnxruntime / onnx.

Install or verify the correct ONNX Runtime build (and the matching `onnx` package) for a user's accelerator backend before Quark ONNX-flow usage. Trigger for "install onnxruntime", "pip install onnxruntime", "set up onnxruntime for ROCm", "set up onnxruntime for CUDA", "onnxruntime-gpu vs onnxruntime", "onnx version mismatch", "CPU-only onnxruntime installed", "onnxruntime providers list missing CUDAExecutionProvider/ROCMExecutionProvider", failures importing `onnxruntime`, or any request to get the correct ONNX Runtime build running. Also trigger when quark-install reports that ONNX Runtime is missing or mismatched before proceeding with the ONNX-to-ONNX flow.

Inspect a target ONNX model and prepare metadata for Quark ONNX PTQ planning. Use for `.onnx` path validation, opset / IR version detection, input-output shape and dtype discovery, op-type histogram, quantizable-op counting, deployment-target compatibility checks (CPU / CUDA / ROCm / AMD NPU CNN / AMD NPU Transformer), and risk assessment. Trigger for "analyze my ONNX model", "check this onnx model", "what opset is this", "can Quark quantize this .onnx", "is my model NPU-compatible", "does my model already have QDQ", "is my model larger than 2 GB", or before any ONNX quantization step that needs model facts that are missing.

End-to-end ONNX PTQ workflow for AMD Quark — for `.onnx` input models (with optional sibling `.onnx_data` external-weights file). Use when the user wants a complete ONNX-to-ONNX pipeline: model intake, quantization planning, calibration-script generation, manifest, and confirmed execution. Trigger for "quantize my .onnx", "run ONNX PTQ end to end", "full ONNX quantization pipeline", "quantize yolov8/resnet50/yolo_nas with XINT8/A8W8/BFP16/MXFP*", "weights-only INT4 for my .onnx LLM", or any request that spans more than one ONNX PTQ step. Not for HuggingFace / safetensors / PyTorch input models — use quark-torch-ptq instead.

Validate Quark ONNX quantization output using four lightweight checks: auxiliary file copy alignment, expected non-quantized initializer MD5 byte-identity (inline `raw_data` + external-data byte ranges), model metadata equality after stripping quantization-only opset entries / Quark domains, and fuzzy node-pattern + op-type + dtype summaries with QDQ / `com.amd.quark` custom-op presence. Intended for post-quantization inspection of `model.onnx` (with or without `model.onnx_data`). Trigger for "validate ONNX quantization result", "check quantized .onnx output", "verify ONNX initializers", "did QDQ insertion happen", "are the non-quantized weights byte-identical".

Diagnose failed Quark Torch PTQ installation, execution, script generation, or export attempts. Use when the user reports a torch-side error, stack trace, invalid artifact, missing dependency, CUDA OOM, version mismatch, or unexpected PTQ results. Trigger for "Quark error", "PTQ failed", "quantization crashed", "CUDA out of memory", "import error", "model loading failed", "wrong results", any Python traceback mentioning quark.torch / torch._dynamo / transformers / accelerate. Not for onnxruntime tracebacks — use quark-onnx-debug.

Prepare export and downstream evaluation handoff for a planned or completed Quark Torch PTQ run. Input is a PyTorch / HuggingFace transformers model; output formats include HF safetensors, GGUF, and ONNX. Trigger for "export model", "save quantized model", "convert to GGUF", "export to HuggingFace format", "export to ONNX", or when the user has a completed or planned Torch PTQ run and needs deployment outputs. Not for exporting from an .onnx input model — use quark-onnx-export.

Low-memory file2file quantization for very large safetensors LLMs that cannot be loaded whole. Use when the user wants to run file2file quantization, adapt a new safetensors checkpoint without loading the full model, register an external LLMTemplate, inspect sharded checkpoint naming, generate wrapper or conversion scripts, or validate low-memory sharded quantization outputs. Trigger for "run file2file quantization", "quantize without loading the model", "large safetensors low-memory quantization", "file2file for DeepSeek/Qwen/MoE", "safetensors naming incompatible", "register LLMTemplate externally".