v0.1.7

• Add support for nn.Embedding quantization.
• Add modules_quant_config to be able to use a different quant config on some layers.
• Make dynamic_loss_threshold auto selected by default.
• Add use_gram_ns option to Muon.
• Add Tensor Descriptor kernels to Triton MM for Intel Arc GPUs.
• Add more optimized Triton MM configs.
• Add Gemma4 keys.

Full Changelog: v0.1.6...v0.1.7

v0.1.6

• Add 9, 10, 11, 12, 13, 14, 15 and 16 bit integer and floating point format support.
  • SDNQ now supports every format all the way from 1 bit to 16 bit, totaling to 176 supported formats.
• Allow all supported types to be allowed in dynamic quantization.
• Add support for basic CPU offloading on optimizer states.
• Update the default optimizer betas to match PyTorch defaults.
• Set dequantize_fp32 to True by default and cast svd to torch_dtype.
• Handle FP64 inputs and don't downcast them to FP32 when quantizing.
• Fix FP16 MM is selected instead of FP8 MM on FP formats other than FP8.

v0.1.5

• Rename norm_mode none to clip on optimizers.
• Fix torch.compile on optimizers.
• Disable contiguous_mm for fp8 mm.
• Cast scales, zero_point and svd to torch_dtype before quantization.
• Add clamping to std normalization on dynamic quantization.
• Add CosmosTransformer3DModel keys.

Full Changelog: v0.1.4...v0.1.5

v0.1.4

• Add Kahan Summation with use_kahan to optimizers.
• Add basic CPU offload with offload_buffers to optimizers.
• Add torch.compile support with use_torch_compile to optimizers.
• Add minimum ndim and numel checks to optimizers.
• Add more guards against NaNs on optimizers.
• Add more broadcast, view, split and select ops to SDNQTensor for DeepSpeed compatibility.
• Add SDNQLayer, SDNQLinear and SDNQConv layer wrappers.
• Add GLM Image and LTX 2 keys to the model specific keys.
• Disable stochastic rounding on backward pass with quantized matmul.
• Set the default max_shard_size to 5GB.
• Set the minimum supported Python version to 3.10.
• Set skip compile on dynamic quantization.
• Register sdnq_training as a backend to diffusers and transformers.
• Refactor optimizers again to make them more easily expandable.
• Fix compatibility with outdated PyTorch versions.
• Fix quantized matmul with packed weights.
• Fix stochastic rounding with float type weights.

Full Changelog: v0.1.3...v0.1.4

v0.1.3

• Add new stack of custom floating point types totaling to 69 different floating point formats.
  New notable ones are: float7_e3m3fn, float6_e3m2fn, float5_e2m2fn, float4_e2m1fn, float3_e1m1fn, float2_e1m0fn
• Add use_dynamic_quantization option.
  Enabling this option will dynamically select a per layer quantization type based on the dynamic_loss_threshold.
  weights_dtype will be used as the minimum allowed quantization type when this option is enabled.
• Rename sdnq.training.sdnq_post_load_quant to sdnq.training.sdnq_training_post_load_quant.
• Check for incompatible weight shapes for matmul in apply_sdnq_options_to_module.
• Remove forced uint4 minimum from conv layers.
• Add integer stochastic rounding to copy_stochastic_.
• Refactor optimizer codes in more generalized way.
• Reduce optimizer memory usage.
• Fix wrong fp8 type is set as the matmul type.
• Fix CAME optimizer.

Full Changelog: v0.1.2...v0.1.3

v0.1.2

• Add Transformers V5 support.
• Add basic sanity check for Triton.
• Add nan check and grad clip to optimizers.
• Fix FP8 CKPT MM on training.
• Fix FP16 MM with FP8 weights.
• Fix end to end torch.compile on training.
• Fix use_stochastic_rounding getting ignored.
• Fix quantization_device getting ignored on post load quant.
• Improve integer stochastic rounding performance on backward pass.

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Add pypi workflow and minor fixes.

Full Changelog: v0.1.0...v0.1.1

v0.1.0

v0.1.0 release