alexmarques commited on
Commit
fb5da44
1 Parent(s): 59e165b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +0 -2
README.md CHANGED
@@ -46,7 +46,6 @@ Weight quantization also reduces disk size requirements by approximately 50%.
46
  Only weights and activations of the linear operators within transformers blocks are quantized.
47
  Weights are quantized with a symmetric static per-channel scheme, where a fixed linear scaling factor is applied between INT8 and floating point representations for each output channel dimension.
48
  Activations are quantized with a symmetric dynamic per-token scheme, computing a linear scaling factor at runtime for each token between INT8 and floating point representations.
49
- Linear scaling factors are computed via by minimizing the mean squarred error (MSE).
50
  The [SmoothQuant](https://arxiv.org/abs/2211.10438) algorithm is used to alleviate outliers in the activations, whereas rhe [GPTQ](https://arxiv.org/abs/2210.17323) algorithm is applied for quantization.
51
  Both algorithms are implemented in the [llm-compressor](https://github.com/vllm-project/llm-compressor) library.
52
  GPTQ used a 1% damping factor and 512 sequences sequences taken from Neural Magic's [LLM compression calibration dataset](https://huggingface.co/datasets/neuralmagic/LLM_compression_calibration).
@@ -124,7 +123,6 @@ recipe = [
124
  scheme="W8A8",
125
  ignore=["lm_head"],
126
  dampening_frac=0.01,
127
- observer="mse",
128
  )
129
  ]
130
 
 
46
  Only weights and activations of the linear operators within transformers blocks are quantized.
47
  Weights are quantized with a symmetric static per-channel scheme, where a fixed linear scaling factor is applied between INT8 and floating point representations for each output channel dimension.
48
  Activations are quantized with a symmetric dynamic per-token scheme, computing a linear scaling factor at runtime for each token between INT8 and floating point representations.
 
49
  The [SmoothQuant](https://arxiv.org/abs/2211.10438) algorithm is used to alleviate outliers in the activations, whereas rhe [GPTQ](https://arxiv.org/abs/2210.17323) algorithm is applied for quantization.
50
  Both algorithms are implemented in the [llm-compressor](https://github.com/vllm-project/llm-compressor) library.
51
  GPTQ used a 1% damping factor and 512 sequences sequences taken from Neural Magic's [LLM compression calibration dataset](https://huggingface.co/datasets/neuralmagic/LLM_compression_calibration).
 
123
  scheme="W8A8",
124
  ignore=["lm_head"],
125
  dampening_frac=0.01,
 
126
  )
127
  ]
128