pypi / amd-quark
amd-quark
What changed in amd-quark 0.11, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-19.
13 breaking7 deprecations22 new APIs and options5 performance7 behavior and defaults1 tooling10 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initamd-quark 0.1165 changes
Breaking changes 13
The argument is removed from the example script and CLI; group size is now embedded in the quantization scheme name.
The argument is removed from the example script and CLI; group size is now embedded in the quantization scheme name.
The argument now takes repeated pattern-and-scheme pairs, allowing per-layer quantization scheme configuration.
quark.torch.export.api.ModelExporter is removed.
quark.torch.export.api.ModelImporter is removed.
freeze now permanently quantizes weights by running QDQ on them, avoiding repeated QDQ on static weights in subsequent calls.
OnnxSlim replaces ONNX Simplifier in preprocessing before quantization.
ONNX QLayerConfig renames activation to input_tensors; layer-specific quantization configuration now specifies quantization per tensor.
The quark_format export format is removed.
The CLI is refactored around LLMTemplate, removes redundant features and arguments, exports only Hugging Face safetensors, and evaluates perplexity on wikitext-2.
The LLM evaluation module moves to quark.contrib.llm_eval, where perplexity evaluation is shared between the CLI and examples.
ONNX quantization configuration classes and algorithms are imported uniformly from quark.onnx, and QLayerConfig uses input_tensors instead of activation.
Module replacement utilities are now located in quark.torch.utils.module_replacement.
Deprecations 7
Config remains available as a backward-compatible alias but is deprecated in favor of QConfig.
QConfig replaces Config as the quantization configuration class.
QLayerConfig replaces QuantizationConfig as the layer quantization configuration class.
QTensorConfig replaces QuantizationSpec as the quantization tensor configuration class.
Legacy scheme names are deprecated in the example script and CLI in favor of simplified built-in names, with group size encoded in names such as int4_wo_32.
QuantizationConfig remains available as a backward-compatible alias but is deprecated in favor of QLayerConfig.
QuantizationSpec remains available as a backward-compatible alias but is deprecated in favor of QTensorConfig.
New APIs and options 22
DeepSeek-OCR is now supported out of the box.
DeepSeek-V3.2 is now supported out of the box.
Adds file-to-file quantization for ultra-large models, supporting weight-only quantization or dynamic activation quantization plus weight quantization; it exports hf_format only and accepts pre-quantized inputs for re-quantization.
GLM-4.7 is now supported out of the box.
Kimi-K2-Instruct is now supported out of the box.
Kimi-K2-Thinking is now supported out of the box.
Kimi-K2.5 is now supported out of the box.
Minimax-M2.1 is now supported out of the box.
An optional config parameter is now available for onnxslim optimization.
Qwen3 Coder is now supported out of the box.
Qwen3 Coder-Next is now supported out of the box.
Qwen3 MoE is now supported out of the box.
A YAML file can now be used as input for custom preprocessing of float models before quantization.
Adds hierarchical and parallel hyperparameter search, custom objectives, sampler choices, checkpoint resume, visualization, and output saving.
Specific model inputs or outputs can now be selected for conversion from NCHW to NHWC.
Setting online_r1_rotation=True in RotationConfig enables online rotations, applying the activation rotation during inference rather than fusing it in a preceding layer.
Adds a graph transformation tool for preprocessing, including constant folding, operator fusion, redundant-node removal, and input/output streamlining.
The example script supports custom algorithm configurations through this argument.
This environment variable sets the logging level, with debug, warning, error, and critical given as examples.
Common model preparation and data preparation utilities are available from quark.torch.utils for reuse.
Joint rotations and SmoothQuant scales can be fine-tuned as a non-destructive transformation; support is tested for llama, qwen3, qwen3_moe, and gpt_oss architectures.
TensorQuantOverrides replaces MixedPrecisionTensor and supports per-tensor or per-channel, symmetric or asymmetric quantization, additional parameters, and precedence for multiple settings.
Performance 5
Memory optimization now applies to all calibration methods, particularly reducing memory use during activation data collection.
Parallel computation significantly speeds up the calibration process.
CPU/GPU communication is removed from quantization observers, improving quantization and runtime speed.
The default mode of MinMSE uses less memory to prevent out-of-memory errors.
An efficient activation quantization kernel fuses scale computation and QDQ when quantizing models.
Behavior and default changes 7
Dry-run compatibility checking is enabled by default, with clearer error messages when model loading fails.
Calibration now includes token distribution coverage warnings.
LLM PTQ workflows now perform a pre-quantization compatibility check for transformers.
Fast Finetune now uses a fixed seed.
The expert calibration token counter is disabled by default and can be enabled by setting this environment variable to 1.
The rotation and quarot pre-quantization algorithms are combined into one rotation algorithm configurable with RotationConfig; the default applies only R1 rotation.
The QDQ operator is now compiled with torch.compile by default, enabling speedups that depend on the quantization scheme.
Tooling 1
Profiling reports individual quantization-stage latency and supports CPU memory traces and real-time ROCm GPU memory monitoring.
Fixes and other changes 10
Scale extraction now handles int32_data in ONNX initializers.
Fast finetuning now infers kernel size from weights when Conv nodes omit the kernel_shape attribute.
Model inference output now handles NaN and Inf values.
Weight calibration now covers weights outside the forward path to ensure complete calibration.
The out-of-bounds axis issue for scalar weights or biases is fixed.
Memory duplication and out-of-memory issues when loading gpt_oss models for quantization are fixed.
The percentile value selection issue is fixed.
Quark ONNX adds support for Python 3.12.
AMD Quark 0.11 is tested against PyTorch 2.9, giving developers a tested PyTorch compatibility target.
AMD Quark 0.11 is compatible with upstream transformers 4.57.