transformers/tests/quantization/autoawq/test_awq.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import gc
import tempfile
import unittest

from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, AwqConfig, OPTForCausalLM
from transformers.testing_utils import (
    backend_empty_cache,
    require_accelerate,
    require_gptqmodel,
    require_torch_accelerator,
    require_torch_gpu,
    require_torch_multi_accelerator,
    slow,
    torch_device,
)
from transformers.utils import is_torch_available
from transformers.utils.quantization_config import AwqBackend


if is_torch_available():
    import torch


@require_torch_accelerator
class AwqConfigTest(unittest.TestCase):
    def test_wrong_backend(self):
        """
        Simple test that checks if a user passes a wrong backend an error is raised
        """
        # This should work fine
        _ = AwqConfig(bits=4)

        with self.assertRaises(ValueError):
            AwqConfig(bits=4, backend="")

        # These should work fine
        _ = AwqConfig(bits=4, version="GEMM")
        _ = AwqConfig(bits=4, version="gemm")

        with self.assertRaises(ValueError):
            AwqConfig(bits=4, backend="unexisting-backend")

    def test_to_dict(self):
        """
        Simple test that checks if one uses a config and converts it to a dict, the dict is the same as the config object
        """
        quantization_config = AwqConfig(bits=4)
        config_to_dict = quantization_config.to_dict()

        for key in config_to_dict:
            if key == "version":
                # "version" is legacy filed.
                # It will be written in to_dict() for compatibility, but AwqConfig will not have this field.
                self.assertFalse(hasattr(quantization_config, key))
            else:
                self.assertEqual(getattr(quantization_config, key), config_to_dict[key])

    def test_from_dict(self):
        """
        Simple test that checks if one uses a dict and converts it to a config object, the config object is the same as the dict
        """
        dict = {"bits": 2, "zero_point": False, "backend": "auto"}
        quantization_config = AwqConfig.from_dict(dict)

        self.assertEqual(dict["bits"], quantization_config.bits)
        self.assertEqual(dict["zero_point"], quantization_config.zero_point)
        self.assertEqual(dict["backend"], quantization_config.backend)


@slow
@require_torch_accelerator
@require_gptqmodel
@require_accelerate
class AwqTest(unittest.TestCase):
    model_name = "TheBloke/Mistral-7B-v0.1-AWQ"
    dummy_transformers_model_name = "bigscience/bloom-560m"
    model_with_no_k_proj_quantized = "hf-internal-testing/opt-125m-awq-no-k-proj"

    input_text = "Hello my name is"

    EXPECTED_OUTPUT = set()
    EXPECTED_OUTPUT.add(
        "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am a junior and I am majoring in Journalism and minoring in Spanish"
    )
    EXPECTED_OUTPUT.add(
        "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am a junior and I am majoring in Journalism and minoring in Spanish. I am"
    )
    EXPECTED_OUTPUT.add(
        "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am a junior and I am majoring in Exercise and Sport Science with a"
    )
    EXPECTED_OUTPUT.add(
        "Hello my name is Katie and I am a 20 year old student from the UK. I am currently studying for a degree in English Literature and History at the University of York. I am a very out"
    )
    EXPECTED_OUTPUT.add(
        "Hello my name is Katie and I am a 20 year old student from the UK. I am currently studying for a degree in English Literature and History at the University of York. I am a very creative"
    )

    EXPECTED_OUTPUT_BF16 = [
        "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am a junior and I am majoring in Journalism and minoring in Spanish",
        "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am a junior and I am majoring in Exercise and Sport Science with a",
    ]

    EXPECTED_OUTPUT_EXLLAMA = [
        "Hello my name is Katie and I am a 20 year old student from the UK. I am currently studying for a degree in English Literature and History at the University of York. I am a very out",
        "Hello my name is Katie and I am a 20 year old student from the UK. I am currently studying for a degree in English Literature and History at the University of York. I am a very creative",
        "Hello my name is Katie and I am a 20 year old student at the University of North Carolina at Chapel Hill. I am a junior and I am majoring in Journalism and minoring in Spanish",
    ]
    device_map = torch_device

    # called only once for all test in this class
    @classmethod
    def setUpClass(cls):
        """
        Setup quantized model
        """
        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_name)
        cls.quantized_model = AutoModelForCausalLM.from_pretrained(cls.model_name, device_map=cls.device_map)

    def tearDown(self):
        gc.collect()
        backend_empty_cache(torch_device)
        gc.collect()

    def test_quantized_model_conversion(self):
        """
        Simple test that checks if the quantized model has been converted properly
        """
        from gptqmodel.nn_modules.qlinear import BaseQuantLinear

        from transformers.integrations.awq import replace_with_awq_linear

        model_id = "facebook/opt-350m"
        config = AutoConfig.from_pretrained(model_id, revision="cb32f77e905cccbca1d970436fb0f5e6b58ee3c5")
        quantization_config = AwqConfig(bits=4)

        with torch.device("meta"):
            model = OPTForCausalLM(config)

        nb_linears = 0
        for module in model.modules():
            if isinstance(module, torch.nn.Linear):
                nb_linears += 1

        model = replace_with_awq_linear(model, quantization_config=quantization_config)
        nb_awq_linear = 0
        for module in model.modules():
            if isinstance(module, BaseQuantLinear):
                nb_awq_linear += 1

        self.assertEqual(nb_linears, nb_awq_linear)

        # Try with `modules_not_to_convert`
        with torch.device("meta"):
            model = OPTForCausalLM(config)

        model = replace_with_awq_linear(
            model, quantization_config=quantization_config, modules_to_not_convert=["lm_head"]
        )
        nb_awq_linear = 0
        for module in model.modules():
            if isinstance(module, BaseQuantLinear):
                nb_awq_linear += 1

        self.assertEqual(nb_linears - 1, nb_awq_linear)

    def test_quantized_model(self):
        """
        Simple test that checks if the quantized model is working properly
        """
        input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)

        output = self.quantized_model.generate(**input_ids, max_new_tokens=40)
        self.assertIn(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)

    def test_raise_if_non_quantized(self):
        model_id = "facebook/opt-125m"
        quantization_config = AwqConfig(bits=4)

        with self.assertRaises(ValueError):
            _ = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config)

    def test_quantized_model_bf16(self):
        """
        Simple test that checks if the quantized model is working properly with bf16
        """
        input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)

        quantized_model = AutoModelForCausalLM.from_pretrained(self.model_name, dtype=torch.bfloat16).to(torch_device)

        output = quantized_model.generate(**input_ids, max_new_tokens=40)
        self.assertIn(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT_BF16)

    @require_torch_gpu
    def test_quantized_model_exllama(self):
        """
        Simple test that checks if the quantized model is working properly with exllama backend
        """
        input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)

        quantization_config = AwqConfig(backend=AwqBackend.EXLLAMA_V2)
        quantized_model = AutoModelForCausalLM.from_pretrained(
            self.model_name, quantization_config=quantization_config, device_map=torch_device
        )

        output = quantized_model.generate(**input_ids, max_new_tokens=40)
        self.assertIn(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT_EXLLAMA)

    def test_quantized_model_no_device_map(self):
        """
        Simple test that checks if the quantized model is working properly
        """
        input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)

        quantized_model = AutoModelForCausalLM.from_pretrained(self.model_name).to(torch_device)
        output = quantized_model.generate(**input_ids, max_new_tokens=40)

        self.assertIn(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)

    def test_save_pretrained(self):
        """
        Simple test that checks if the quantized model is working properly after being saved and loaded
        """
        # Load a fresh model for saving — the shared self.quantized_model may have
        # already been in-place transformed by a prior generate() call, and saving
        # those transformed buffers then re-transforming on reload would corrupt data.
        fresh_model = AutoModelForCausalLM.from_pretrained(self.model_name)
        with tempfile.TemporaryDirectory() as tmpdirname:
            fresh_model.save_pretrained(tmpdirname)
            model = AutoModelForCausalLM.from_pretrained(tmpdirname, device_map=self.device_map)

            input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)

            output = model.generate(**input_ids, max_new_tokens=40)
            self.assertIn(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)

    @require_torch_multi_accelerator
    def test_quantized_model_multi_accelerator(self):
        """
        Simple test that checks if the quantized model is working properly with multiple GPUs
        """
        input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(torch_device)

        quantized_model = AutoModelForCausalLM.from_pretrained(self.model_name, device_map="auto")

        self.assertTrue(len(set(quantized_model.hf_device_map.values())) >= 2)

        output = quantized_model.generate(**input_ids, max_new_tokens=40)

        self.assertIn(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)

    def test_quantized_model_no_k_proj_quantized(self):
        """
        Simple test that checks if the quantized model is working properly with multiple GPUs
        """
        dummy_input = torch.LongTensor([[0, 1, 0]]).to(torch_device)

        quantized_model = AutoModelForCausalLM.from_pretrained(self.model_with_no_k_proj_quantized).to(torch_device)

        self.assertTrue(isinstance(quantized_model.model.decoder.layers[0].self_attn.k_proj, torch.nn.Linear))
        self.assertFalse(isinstance(quantized_model.model.decoder.layers[0].self_attn.v_proj, torch.nn.Linear))

        EXPECTED_OUTPUT = torch.LongTensor([[0, 1, 0, 50118, 50118, 133, 248, 12, 134, 16, 10, 372, 2031]]).to(
            torch_device
        )

        output = quantized_model.generate(dummy_input, max_new_tokens=10)
        self.assertTrue((output == EXPECTED_OUTPUT).all())


@slow
@require_torch_accelerator
@require_gptqmodel
@require_accelerate
class AwqScaleTest(unittest.TestCase):
    model_name = "TechxGenus/starcoder2-3b-AWQ"

    def test_load_quantized_model(self):
        from gptqmodel.quantization.awq.modules.act import ScaledActivation

        """
        Simple test that checks if the scales have been replaced in the quantized model
        """
        quantized_model = AutoModelForCausalLM.from_pretrained(
            "TechxGenus/starcoder2-3b-AWQ", dtype=torch.float16, device_map=torch_device
        )
        self.assertTrue(isinstance(quantized_model.model.layers[0].mlp.act, ScaledActivation))


@slow
@require_gptqmodel
@require_accelerate
class AwqTorchFusedTest(unittest.TestCase):
    def test_quantized_model_torch_fused(self):
        """
        Simple test that checks if the quantized model is working properly with torch_fused backend
        """
        quantization_config = AwqConfig(backend=AwqBackend.TORCH_FUSED_AWQ)

        model = AutoModelForCausalLM.from_pretrained(
            "TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ",
            quantization_config=quantization_config,
            device_map="cpu",
        )
        tokenizer = AutoTokenizer.from_pretrained("TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ")
        input_ids = tokenizer.encode("How to make a cake", return_tensors="pt")
        pad_token_id = tokenizer.eos_token_id
        output = model.generate(input_ids, do_sample=False, max_length=20, pad_token_id=pad_token_id)
        print(tokenizer.decode(output[0], skip_special_tokens=True))

        expected_output = (
            "How to make a cake with a round tin?\nHow to make a cake with a round tin?\n1. Preheat the oven to 180°"
        )
        self.assertIn(tokenizer.decode(output[0], skip_special_tokens=True), expected_output)