in

# Introduction to Weight Quantization | In direction of Knowledge Science

The Python implementation is sort of easy:

`def zeropoint_quantize(X):# Calculate worth vary (denominator)x_range = torch.max(X) - torch.min(X)x_range = 1 if x_range == 0 else x_range# Calculate scalescale = 255 / x_range# Shift by zero-pointzeropoint = (-scale * torch.min(X) - 128).spherical()# Scale and around the inputsX_quant = torch.clip((X * scale + zeropoint).spherical(), -128, 127)# DequantizeX_dequant = (X_quant - zeropoint) / scalereturn X_quant.to(torch.int8), X_dequant`

As an alternative of counting on full toy examples, we will use these two features on an actual mannequin because of the `transformers`library.

We begin by loading the mannequin and tokenizer for GPT-2. It is a very small mannequin we in all probability don’t need to quantize, however it will likely be ok for this tutorial. First, we need to observe the mannequin’s dimension so we will evaluate it later and consider the reminiscence financial savings because of 8-bit quantization.

`!pip set up -q bitsandbytes>=0.39.0!pip set up -q git+https://github.com/huggingface/speed up.git!pip set up -q git+https://github.com/huggingface/transformers.git`
`from transformers import AutoModelForCausalLM, AutoTokenizerimport torchtorch.manual_seed(0)# Set gadget to CPU for nowgadget = 'cpu'# Load mannequin and tokenizermodel_id = 'gpt2'mannequin = AutoModelForCausalLM.from_pretrained(model_id).to(gadget)tokenizer = AutoTokenizer.from_pretrained(model_id)# Print mannequin dimensionprint(f"Mannequin dimension: {mannequin.get_memory_footprint():,} bytes")`
`Mannequin dimension: 510,342,192 bytes`

The dimensions of the GPT-2 mannequin is roughly 487MB in FP32. The following step consists of quantizing the weights utilizing zero-point and absmax quantization. Within the following instance, we apply these strategies to the primary consideration layer of GPT-2 to see the outcomes.

`# Extract weights of the primary layerweights = mannequin.transformer.h[0].attn.c_attn.weight.informationprint("Unique weights:")print(weights)# Quantize layer utilizing absmax quantizationweights_abs_quant, _ = absmax_quantize(weights)print("nAbsmax quantized weights:")print(weights_abs_quant)# Quantize layer utilizing absmax quantizationweights_zp_quant, _ = zeropoint_quantize(weights)print("nZero-point quantized weights:")print(weights_zp_quant)`
`Unique weights:tensor([[-0.4738, -0.2614, -0.0978,  ...,  0.0513, -0.0584,  0.0250],[ 0.0874,  0.1473,  0.2387,  ..., -0.0525, -0.0113, -0.0156],[ 0.0039,  0.0695,  0.3668,  ...,  0.1143,  0.0363, -0.0318],...,[-0.2592, -0.0164,  0.1991,  ...,  0.0095, -0.0516,  0.0319],[ 0.1517,  0.2170,  0.1043,  ...,  0.0293, -0.0429, -0.0475],[-0.4100, -0.1924, -0.2400,  ..., -0.0046,  0.0070,  0.0198]])Absmax quantized weights:tensor([[-21, -12,  -4,  ...,   2,  -3,   1],[  4,   7,  11,  ...,  -2,  -1,  -1],[  0,   3,  16,  ...,   5,   2,  -1],...,[-12,  -1,   9,  ...,   0,  -2,   1],[  7,  10,   5,  ...,   1,  -2,  -2],[-18,  -9, -11,  ...,   0,   0,   1]], dtype=torch.int8)Zero-point quantized weights:tensor([[-20, -11,  -3,  ...,   3,  -2,   2],[  5,   8,  12,  ...,  -1,   0,   0],[  1,   4,  18,  ...,   6,   3,   0],...,[-11,   0,  10,  ...,   1,  -1,   2],[  8,  11,   6,  ...,   2,  -1,  -1],[-18,  -8, -10,  ...,   1,   1,   2]], dtype=torch.int8)`

The distinction between the unique (FP32) and quantized values (INT8) is obvious, however the distinction between absmax and zero-point weights is extra delicate. On this case, the inputs look shifted by a price of -1. This implies that the load distribution on this layer is sort of symmetric.

We are able to evaluate these strategies by quantizing each layer in GPT-2 (linear layers, consideration layers, and so on.) and create two new fashions: `model_abs` and `model_zp`. To be exact, we are going to truly change the unique weights with de-quantized ones. This has two advantages: it permits us to 1/ evaluate the distribution of our weights (identical scale) and a couple of/ truly run the fashions.

Certainly, PyTorch doesn’t permit INT8 matrix multiplication by default. In an actual state of affairs, we might dequantize them to run the mannequin (in FP16 for instance) however retailer them as INT8. Within the subsequent part, we are going to use the `bitsandbytes` library to unravel this situation.

`import numpy as npfrom copy import deepcopy# Retailer authentic weightsweights = [param.data.clone() for param in model.parameters()]# Create mannequin to quantizemodel_abs = deepcopy(mannequin)# Quantize all mannequin weightsweights_abs = []for param in model_abs.parameters():_, dequantized = absmax_quantize(param.information)param.information = dequantizedweights_abs.append(dequantized)# Create mannequin to quantizemodel_zp = deepcopy(mannequin)# Quantize all mannequin weightsweights_zp = []for param in model_zp.parameters():_, dequantized = zeropoint_quantize(param.information)param.information = dequantizedweights_zp.append(dequantized)`

Now that our fashions have been quantized, we need to verify the impression of this course of. Intuitively, we need to guarantee that the quantized weights are near the unique ones. A visible method to verify it’s to plot the distribution of the dequantized and authentic weights. If the quantization is lossy, it could drastically change the load distribution.

The next determine reveals this comparability, the place the blue histogram represents the unique (FP32) weights, and the crimson one represents the dequantized (from INT8) weights. Be aware that we solely show this plot between -2 and a couple of due to outliers with very excessive absolute values (extra on that later).

Each plots are fairly related, with a shocking spike round 0. This spike reveals that our quantization is sort of lossy since reversing the method doesn’t output the unique values. That is notably true for the absmax mannequin, which shows each a decrease valley and a better spike round 0.

Let’s evaluate the efficiency of the unique and quantized fashions. For this goal, we outline a `generate_text()` perform to generate 50 tokens with top-k sampling.

`def generate_text(mannequin, input_text, max_length=50):input_ids = tokenizer.encode(input_text, return_tensors='pt').to(gadget)output = mannequin.generate(inputs=input_ids,max_length=max_length,do_sample=True,top_k=30,pad_token_id=tokenizer.eos_token_id,attention_mask=input_ids.new_ones(input_ids.form))return tokenizer.decode(output[0], skip_special_tokens=True)# Generate textual content with authentic and quantized fashionsoriginal_text = generate_text(mannequin, "I've a dream")absmax_text   = generate_text(model_abs, "I've a dream")zp_text       = generate_text(model_zp, "I've a dream")print(f"Unique mannequin:n{original_text}")print("-" * 50)print(f"Absmax mannequin:n{absmax_text}")print("-" * 50)print(f"Zeropoint mannequin:n{zp_text}")`
`Unique mannequin:I've a dream, and it's a dream I consider I'd get to reside in my future. I really like my mom, and there was that one time I had been informed that my household wasn't even that sturdy. After which I bought the--------------------------------------------------Absmax mannequin:I've a dream to search out out the origin of her hair. She loves it. However there isn't any means you may be sincere about how her hair is made. She should be loopy.We discovered a photograph of the coiffure posted on--------------------------------------------------Zeropoint mannequin:I've a dream of making two full-time jobs in America—one for individuals with psychological well being points, and one for individuals who don't undergo from psychological sickness—or a minimum of have an employment and household historical past of substance abuse, to work half`

As an alternative of attempting to see if one output makes extra sense than the others, we will quantify it by calculating the perplexity of every output. It is a widespread metric used to judge language fashions, which measures the uncertainty of a mannequin in predicting the subsequent token in a sequence. On this comparability, we make the widespread assumption that the decrease the rating, the higher the mannequin is. In observe, a sentence with a excessive perplexity is also appropriate.

We implement it utilizing a minimal perform because it doesn’t want to contemplate particulars just like the size of the context window since our sentences are brief.

`def calculate_perplexity(mannequin, textual content):# Encode the textual contentencodings = tokenizer(textual content, return_tensors='pt').to(gadget)# Outline input_ids and target_idsinput_ids = encodings.input_idstarget_ids = input_ids.clone()with torch.no_grad():outputs = mannequin(input_ids, labels=target_ids)# Loss calculationneg_log_likelihood = outputs.loss# Perplexity calculationppl = torch.exp(neg_log_likelihood)return pplppl     = calculate_perplexity(mannequin, original_text)ppl_abs = calculate_perplexity(model_abs, absmax_text)ppl_zp  = calculate_perplexity(model_zp, absmax_text)print(f"Unique perplexity:  {ppl.merchandise():.2f}")print(f"Absmax perplexity:    {ppl_abs.merchandise():.2f}")print(f"Zeropoint perplexity: {ppl_zp.merchandise():.2f}")`
`Unique perplexity:  15.53Absmax perplexity:    17.92Zeropoint perplexity: 17.97`

We see that the perplexity of the unique mannequin is barely decrease than the 2 others. A single experiment will not be very dependable, however we might repeat this course of a number of instances to see the distinction between every mannequin. In idea, zero-point quantization must be barely higher than absmax, however can also be extra pricey to compute.

On this instance, we utilized quantization strategies to whole layers (per-tensor foundation). Nevertheless, we might apply it at totally different granularity ranges: from the complete mannequin to particular person values. Quantizing the complete mannequin in a single move would severely degrade the efficiency, whereas quantizing particular person values would create an enormous overhead. In observe, we regularly want the vector-wise quantization, which considers the variability of values in rows and columns within the identical tensor.

Nevertheless, even vector-wise quantization doesn’t remedy the issue of outlier options. Outlier options are excessive values (damaging or constructive) that seem in all transformer layers when the mannequin attain a sure scale (>6.7B parameters). This is a matter since a single outlier can scale back the precision for all different values. However discarding these outlier options will not be an possibility since it could tremendously degrade the mannequin’s efficiency.

Launched by Dettmers et al. (2022), LLM.int8() is an answer to the outlier downside. It depends on a vector-wise (absmax) quantization scheme and introduces mixed-precision quantization. Which means outlier options are processed in a FP16 format to retain their precision, whereas the opposite values are processed in an INT8 format. As outliers characterize about 0.1% of values, this successfully reduces the reminiscence footprint of the LLM by virtually 2x.

LLM.int8() works by conducting matrix multiplication computation in three key steps:

1. Extract columns from the enter hidden states X containing outlier options utilizing a customized threshold.
2. Carry out the matrix multiplication of the outliers utilizing FP16 and the non-outliers utilizing INT8 with vector-wise quantization (row-wise for the hidden state X and column-wise for the load matrix W).
3. Dequantize the non-outlier outcomes (INT8 to FP16) and add them to the outlier outcomes to get the complete end in FP16.