Skip to content

Latest commit

 

History

History
329 lines (247 loc) · 21.1 KB

PT_WeightOnlyQuant.md

File metadata and controls

329 lines (247 loc) · 21.1 KB

PyTorch Weight Only Quantization

Introduction

As large language models (LLMs) become more prevalent, there is a growing need for new and improved quantization methods that can meet the computational demands of these modern architectures while maintaining the accuracy. Compared to normal quantization like W8A8, weight only quantization is probably a better trade-off to balance the performance and the accuracy, since we will see below that the bottleneck of deploying LLMs is the memory bandwidth and normally weight only quantization could lead to better accuracy.

Model inference: Roughly speaking , two key steps are required to get the model's result. The first one is moving the model from the memory to the cache piece by piece, in which, memory bandwidth $B$ and parameter count $P$ are the key factors, theoretically the time cost is $P*4 /B$. The second one is computation, in which, the device's computation capacity $C$ measured in FLOPS and the forward FLOPs $F$ play the key roles, theoretically the cost is $F/C$.

Text generation: The most famous application of LLMs is text generation, which predicts the next token/word based on the inputs/context. To generate a sequence of texts, we need to predict them one by one. In this scenario, $F\approx P$ if some operations like bmm are ignored and past key values have been saved. However, the $C/B$ of the modern device could be to 100X, that makes the memory bandwidth as the bottleneck in this scenario.

Besides, as mentioned in many papers[1][2], activation quantization is the main reason to cause the accuracy drop. So for text generation task, weight only quantization is a preferred option in most cases.

Theoretically, round-to-nearest (RTN) is the most straightforward way to quantize weight using scale maps. However, when the number of bits is small (e.g. 3), the MSE loss is larger than expected. A group size is introduced to reduce elements using the same scale to improve accuracy.

Supported Matrix

Algorithms/Backend PyTorch eager mode
RTN
GPTQ
AutoRound
AWQ
TEQ
HQQ

RTN: A quantification method that we can think of very intuitively. It does not require additional datasets and is a very fast quantization method. Generally speaking, RTN will convert the weight into a uniformly distributed integer data type, but some algorithms, such as Qlora, propose a non-uniform NF4 data type and prove its theoretical optimality.

GPTQ: A new one-shot weight quantization method based on approximate second-order information, that is both highly-accurate and highly efficient[4]. The weights of each column are updated based on the fixed-scale pseudo-quantization error and the inverse of the Hessian matrix calculated from the activations. The updated columns sharing the same scale may generate a new max/min value, so the scale needs to be saved for restoration.

AutoRound: AutoRound is an advanced weight-only quantization algorithm for low-bits LLM inference. It's tailored for a wide range of models and consistently delivers noticeable improvements, often significantly outperforming SignRound[5] with the cost of more tuning time for quantization.

AWQ: Proved that protecting only 1% of salient weights can greatly reduce quantization error. the salient weight channels are selected by observing the distribution of activation and weight per channel. The salient weights are also quantized after multiplying a big scale factor before quantization for preserving.

TEQ: A trainable equivalent transformation that preserves the FP32 precision in weight-only quantization. It is inspired by AWQ while providing a new solution to search for the optimal per-channel scaling factor between activations and weights.

HQQ: The HQQ[6] method focuses specifically on minimizing errors in the weights rather than the layer activation. Additionally, by incorporating a sparsity-promoting loss, such as the $l_{p<1}$-norm, we effectively model outliers through a hyper-Laplacian distribution. This distribution more accurately captures the heavy-tailed nature of outlier errors compared to the squared error, resulting in a more nuanced representation of error distribution.

Usage

Get Started

WeightOnlyQuant quantization for PyTorch is using prepare and convert APIs.

Common arguments

Config Capability
dtype (str) ['int', 'nf4', 'fp4']
bits (int) [1, ..., 8]
group_size (int) [-1, 1, ..., $C_{in}$]
use_sym (bool) [True, False]
quant_lm_head (bool) [False, True]
use_double_quant (bool) [True, False]
double_quant_dtype (str) ['int']
double_quant_bits (int) [1, ..., bits]
double_quant_use_sym (bool) [True, False]
double_quant_group_size (int) [-1, 1, ..., $C_{in}$]

Notes:

  • group_size = -1 refers to per output channel quantization. Taking a linear layer (input channel = $C_{in}$, output channel = $C_{out}$) for instance, when group size = -1, quantization will calculate total $C_{out}$ quantization parameters. Otherwise, when group_size = gs quantization parameters are calculate with every $gs$ elements along with the input channel, leading to total $C_{out} \times (C_{in} / gs)$ quantization parameters.
  • 4-bit NormalFloat(NF4) is proposed in QLoRA[7]. 'fp4' includes fp4_e2m1 and fp4_e2m1_bnb. By default, fp4 refers to fp4_e2m1_bnb.
  • quant_lm_head defaults to False. This means that, except for transformer blocks, the last layer in transformer models will not be quantized by default. The last layer may be named "lm_head", "output_layer" or "embed_out".
  • Only RTN and GPTQ support double quant.

RTN

rtn_args comments default value
group_dim (int) Dimension for grouping 1
use_full_range (bool) Enables full range for activations False
use_mse_search (bool) Enables mean squared error (MSE) search False
use_layer_wise (bool) Enables quantize model per layer False
model_path (str) Model path that is used to load state_dict per layer

Notes: model_path is only used when use_layer_wise=True. layer-wise is stay-tuned.

# Quantization code
from neural_compressor.torch.quantization import prepare, convert, RTNConfig

quant_config = RTNConfig()
model = prepare(model, quant_config)
model = convert(model)

GPTQ

gptq_args comments default value
use_mse_search (bool) Enables mean squared error (MSE) search False
use_layer_wise (bool) Enables quantize model per layer False
model_path (str) Model path that is used to load state_dict per layer
use_double_quant (bool) Enables double quantization False
act_order (bool) Whether to sort Hessian's diagonal values to rearrange channel-wise quantization order False
percdamp (float) Percentage of Hessian's diagonal values' average, which will be added to Hessian's diagonal to increase numerical stability 0.01
block_size (int) Execute GPTQ quantization per block, block shape = [C_out, block_size] 128
static_groups (bool) Whether to calculate group wise quantization parameters in advance. This option mitigate actorder's extra computational requirements. False
true_sequential (bool) Whether to quantize layers within a transformer block in their original order. This can lead to higher accuracy but slower overall quantization process. False

Note: model_path is only used when use_layer_wise=True. layer-wise is stay-tuned.

# Quantization code
from neural_compressor.torch.quantization import prepare, convert, GPTQConfig

quant_config = GPTQConfig()
model = prepare(model, quant_config)
run_fn(model)  # calibration
model = convert(model)

AutoRound

autoround_args comments default value
enable_full_range (bool) Whether to enable full range quantization False
batch_size (int) Batch size for training 8
lr_scheduler The learning rate scheduler to be used None
enable_quanted_input (bool) Whether to use quantized input data True
enable_minmax_tuning (bool) Whether to enable min-max tuning True
lr (float) The learning rate 0
minmax_lr (float) The learning rate for min-max tuning None
low_gpu_mem_usage (bool) Whether to use low GPU memory True
iters (int) Number of iterations 200
seqlen (int) Length of the sequence 2048
n_samples (int) Number of samples 512
sampler (str) The sampling method "rand"
seed (int) The random seed 42
n_blocks (int) Number of blocks 1
gradient_accumulate_steps (int) Number of gradient accumulation steps 1
not_use_best_mse (bool) Whether to use mean squared error False
dynamic_max_gap (int) The dynamic maximum gap -1
scale_dtype (str) The data type of quantization scale to be used, different kernels have different choices "float16"
# Quantization code
from neural_compressor.torch.quantization import prepare, convert, AutoRoundConfig

quant_config = AutoRoundConfig()
model = prepare(model, quant_config)
run_fn(model)  # calibration
model = convert(model)

AWQ

awq_args comments default value
group_dim (int) Dimension for grouping 1
use_full_range (bool) Enables full range for activations False
use_mse_search (bool) Enables mean squared error (MSE) search False
use_layer_wise (bool) Enables quantize model per layer False
use_auto_scale (bool) Enables best scales search based on activation distribution True
use_auto_clip (bool) Enables clip range search True
folding(bool) Allow insert mul before linear when the scale cannot be absorbed by last layer False.

Notes: layer-wise is stay-tuned.

# Quantization code
from neural_compressor.torch.quantization import prepare, convert, AWQConfig

quant_config = AWQConfig()
model = prepare(model, quant_config, example_inputs=example_inputs)
run_fn(model)  # calibration
model = convert(model)

TEQ

teq_args comments default value
group_dim (int) Dimension for grouping 1
use_full_range (bool) Enables full range for activations False
use_mse_search (bool) Enables mean squared error (MSE) search False
use_layer_wise (bool) Enables quantize model per layer False
use_double_quant (bool) Enables double quantization False
folding(bool) Allow insert mul before linear when the scale cannot be absorbed by last layer False

Notes: layer-wise is stay-tuned.

# Quantization code
from neural_compressor.torch.quantization import prepare, convert, TEQConfig

quant_config = TEQConfig()
model = prepare(model, quant_config, example_inputs=example_inputs)
train_fn(model)  # calibration
model = convert(model)

HQQ

hqq_args comments default value
quant_zero (bool) Whether to quantize zero point True
quant_scale: (bool) Whether to quantize scale: point False
scale_quant_group_size (int) The group size for quantizing scale 128
# Quantization code
from neural_compressor.torch.quantization import prepare, convert, HQQConfig

quant_config = HQQConfig()
model = prepare(model, quant_config)
run_fn(model)  # calibration
model = convert(model)

Specify Quantization Rules

Intel(R) Neural Compressor support specify quantization rules by operator name or operator type. Users can set local in dict or use set_local method of config class to achieve the above purpose.

  1. Example of setting local from a dict
quant_config = {
    "rtn": {
        "global": {
            "dtype": "int",
            "bits": 4,
            "group_size": -1,
            "use_sym": True,
        },
        "local": {
            "lm_head": {
                "dtype": "fp32",
            },
        },
    }
}
  1. Example of using set_local
quant_config = RTNConfig()
lm_head_config = RTNConfig(dtype="fp32")
quant_config.set_local("lm_head", lm_head_config)

Saving and Loading

The saved_results folder contains two files: quantized_model.pt and qconfig.json, and the generated model is a quantized model. The quantitative model will include WeightOnlyLinear. To support low memory inference, Intel(R) Neural Compressor implemented WeightOnlyLinear, a torch.nn.Module, to compress the fake quantized fp32 model. Since torch does not provide flexible data type storage, WeightOnlyLinear combines low bits data into a long date type, such as torch.int8 and torch.int32. Low bits data includes weights and zero points. When using WeightOnlyLinear for inference, it will restore the compressed data to float32 and run torch linear function.

# Quantization code
from neural_compressor.torch.quantization import prepare, convert, RTNConfig

quant_config = RTNConfig()
model = prepare(model, quant_config)
model = convert(model)

# save
model.save("saved_results")

# load
from neural_compressor.torch.quantization import load

orig_model = YOURMODEL()
loaded_model = load(
    "saved_results", original_model=orig_model
)  # Please note that the original_model parameter passes the original model.

Layer Wise Quantization

As the size of LLMs continues to grow, loading the entire model into a single GPU card or the RAM of a client machine becomes impractical. To address this challenge, we introduce Layer-wise Quantization (LWQ), a method that quantizes LLMs layer by layer or block by block. This approach significantly reduces memory consumption. The diagram below illustrates the LWQ process.

Figure 1: The process of layer-wise quantization for PyTorch model. The color grey means empty parameters and the color blue represents parameters need to be quantized. Every rectangle inside model represents one layer.

Currently, we support LWQ for RTN, AutoRound, and GPTQ.

Here, we take the RTN algorithm as example to demonstrate the usage of LWQ.

from neural_compressor.torch.quantization import RTNConfig, convert, prepare
from neural_compressor.torch import load_empty_model

model_state_dict_path = "/path/to/model/state/dict"
float_model = load_empty_model(model_state_dict_path)
quant_config = RTNConfig(use_layer_wise=True)
prepared_model = prepare(float_model, quant_config)
quantized_model = convert(prepared_model)

Efficient Usage on Client-Side

For client machines with limited RAM and cores, we offer optimizations to reduce computational overhead and minimize memory usage. For detailed information, please refer to Quantization on Client.

Examples

Users can also refer to examples on how to quantize a model with WeightOnlyQuant.

Reference

[1]. Xiao, Guangxuan, et al. "Smoothquant: Accurate and efficient post-training quantization for large language models." arXiv preprint arXiv:2211.10438 (2022).

[2]. Wei, Xiuying, et al. "Outlier suppression: Pushing the limit of low-bit transformer language models." arXiv preprint arXiv:2209.13325 (2022).

[3]. Lin, Ji, et al. "AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration." arXiv preprint arXiv:2306.00978 (2023).

[4]. Frantar, Elias, et al. "Gptq: Accurate post-training quantization for generative pre-trained transformers." arXiv preprint arXiv:2210.17323 (2022).

[5]. Cheng, Wenhua, et al. "Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs" arXiv preprint arXiv:2309.05516 (2023).

[6]. Badri, Hicham and Shaji, Appu. "Half-Quadratic Quantization of Large Machine Learning Models." [Online] Available: https://mobiusml.github.io/hqq_blog/ (2023).

[7]. Dettmers, Tim, et al. "Qlora: Efficient finetuning of quantized llms." arXiv preprint arXiv:2305.14314 (2023).