HDCharles
HDCharles
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #83 * #104 Summary: this makes gptq work Test Plan: Reviewers: Subscribers: Tasks: Tags:
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #2266 Summary: Test Plan: python run_benchmark.py group_bench -c /home/cdhernandez/local/benchmark/userbenchmark/group_bench/configs/torch_ao.yaml Reviewers: Subscribers: Tasks: Tags:
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #2265 * __->__ #2264 Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
Air Gradient One home assistant values don't match what i see in the air gradient dashboard when they are very small i.e.  it looks like its showing 0 decimal...
Background: The [spin quant paper](https://arxiv.org/pdf/2405.16406) introduces a method of improving quantization by adding additional rotation matrices to the model weights that improve quantization performance. While spin-quant is a fairly sophisticated...
fixed the device swapping, the previous code was mega slow without it changed the point where we get tensor ids and check the name so we can safely to device...
A lot of code for tensor subclasses can likely be conslidated together into a base class that other classes can utilize _get_to_kwargs: https://github.com/pytorch/ao/blob/main/torchao/dtypes/affine_quantized_tensor.py#L64 https://github.com/pytorch/ao/blob/main/torchao/dtypes/affine_quantized_tensor.py#L276 (also needed for https://github.com/pytorch/ao/blob/main/torchao/quantization/autoquant.py#L40) to: https://github.com/pytorch/ao/blob/main/torchao/dtypes/affine_quantized_tensor.py#L594...
Summary: 1) integrated gemlite kernels from https://github.com/mobiusml/gemlite 2) updated kernel wrappers to work with compile (needed to change acc_dtype to torch.dtype equivalent and added basic op/fake_op registration) 3) added bs>1...
Reverts pytorch/ao#1061 seems this caused an issue relevant to: https://github.com/pytorch/ao/issues/1063 https://github.com/pytorch/ao/issues/1062 hopefully https://github.com/pytorch/ao/pull/1061 does not resurface and it was an env issue