README for Modifiers (#165)

* WIP readme for modifiers

* fill out modifier summaries

* fix typo

* fix typo

---------

Co-authored-by: Kyle Sayers <[email protected]>

Author: Sara Adkins

src/llmcompressor/modifiers/README.md

@@ -0,0 +1,75 @@
+# Modifiers Overview
+
+A `Modifier` in `llm-compressor` is an algorithm that can be applied to a model to change 
+its state in some way. Some modifiers can be applied during one-shot, while others 
+are relevant only during training. Below is a summary of the key modifiers available.
+
+## Pruning Modifiers
+
+Modifiers that introduce sparsity into a model
+
+### [SparseGPT](./obcq/base.py)
+One-shot algorithm that uses calibration data to introduce unstructured or structured 
+sparsity into weights. Implementation based on [SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot](https://arxiv.org/abs/2301.00774). A small amount of calibration data is used 
+to calculate a Hessian for each layers input activations, this Hessian is then used to 
+solve a regression problem that minimizes the error introduced by a target sparsity. This algorithm 
+has a good amount of memory overhead introduced by storing the Hessians.
+
+### [WANDA](./pruning/wanda/base.py)
+One-shot algorithm that uses calibration data to introduce unstructured or structured sparsity. Implementation is
+based on [A Simple and Effective Pruning Approach for Large Language Models](https://arxiv.org/pdf/2306.11695).
+Calibration data is used to calculate the magnitude of input activations for each layer, and weights 
+are pruned based on this magnitude combined with their distance from 0. This requires less 
+memory overhead and computation than SparseGPT, but reduces accuracy in many cases.
+
+### [Magnitude Pruning](./pruning/magnitude/base.py)
+Naive one-shot pruning algorithm that does not require any calibration data. Weights are 
+pruned based solely on their distance from 0 up to the target sparsity.
+
+## Quantization Modifiers
+
+Modifiers that quantize weights or activations of a model
+
+### [Basic Quantization](./quantization/quantization/base.py)
+One-shot algorithm that quantizes weights, input activations and/or output activations by 
+calculating a range from weights or calibration data. All data is quantized to the closest 
+bin using a scale and (optional) zero point. This basic quantization algorithm is 
+suitable for FP8 quantization. A variety of quantization schemes are supported via the 
+[compressed-tensors](https://github.com/neuralmagic/compressed-tensors) library. 
+
+### [GPTQ](./quantization/gptq/base.py)
+One-shot algorithm that uses calibration data to select the ideal bin for weight quantization. 
+This algorithm is applied on top of the basic quantization algorithm, and affects weights only.
+The implementation is based on [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://arxiv.org/pdf/2210.17323). The algorithm is very similar to SparseGPT: A small amount of calibration data is used 
+to calculate a Hessian for each layers input activations, this Hessian is then used to 
+solve a regression problem that minimizes the error introduced by a given quantization configuration. This algorithm 
+has a good amount of memory overhead introduced by storing the Hessians.
+
+## "Helper" Modifiers
+
+These modifiers do not introduce sparsity or quantization themselves, but are used 
+in conjunction with one of the above modifiers to improve their accuracy.
+
+### [SmoothQuant](./smoothquant/base.py)
+The modifier is intended to be used prior to a `QuantizationModifier` or `GPTQModifier`. Its purpose is 
+to make input activations easier to quantize by smoothing away outliers in the inputs, and applying the inverse 
+smoothing operation to the following weights. This makes weights slightly harder to quantize, but the inputs much
+easier to quantize. The implementation is based on [SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models](https://arxiv.org/pdf/2211.10438) and requires calibration data. 
+
+### [Logarithmic Equalization](./logarithmic_equalization/base.py)
+Very similar to `SmoothQuantModifier`, but applies smoothing on an inverse log scale 
+rather than the linear smoothing done by SmoothQuant. The implementation is based on 
+[FPTQ: Fine-grained Post-Training Quantization for Large Language Models](https://arxiv.org/pdf/2308.15987)
+
+### [Constant Pruning](./pruning/constant/base.py)
+One-shot pruning algorithms often introduce accuracy degradation that can be recovered with finetuning. This 
+modifier ensures that the sparsity mask of the model is maintained during finetuning, allowing a sparse 
+model to recover accuracy while maintaining its sparsity structure. It is intended to be used after a pruning modifier
+such as `SparseGPT` or `WANDA` has already been applied.
+
+### [Distillation](./distillation/output/base.py)
+To better recover accuracy of sparse models during finetuning, we can also use a teacher model of the same architecture
+to influence the loss. This modifier is intended to be used in conjunction with `ConstantPruning` modifier on a 
+pruned model, with the dense version of the model being used as the teacher. Both output distillation loss and 
+layer-by-layer distillation loss are supported. The layer-by-layer implementation follows the Square Head distillation 
+algorithm presented in [Sparse Fine-tuning for Inference Acceleration of Large Language Models](https://arxiv.org/pdf/2310.06927).