Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

3.1 KB
Raw Blame
import copy
import shutil
from test.helpers import get_model_path, save_tiny_model

import pytest
import torch
import transformers
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

from auto_round import AutoRound


class TestAutoRound:
    save_dir = "./saved"

    @pytest.fixture(autouse=True, scope="class")
    def setup_and_teardown_class(self):
        # ===== SETUP (setup_class) =====
        print("[Setup] Running before any test in class")

        # Yield to hand control to the test methods
        yield

        # ===== TEARDOWN (teardown_class) =====
        print("[Teardown] Running after all tests in class")
        shutil.rmtree("./saved", ignore_errors=True)
        shutil.rmtree("runs", ignore_errors=True)

    def test_transform_mxfp4_quant_infer(self):
        model_name = get_model_path("qwen/Qwen3-0.6B")
        scheme = "MXFP4"

        ar = AutoRound(
            model=model_name,
            iters=0,
            seqlen=2,
            scheme=scheme,
            alg_configs=["rtn", "hadamard"],
        )
        compressed_model, quantized_model_path = ar.quantize_and_save(output_dir=self.save_dir, format="auto_round")

        model = AutoModelForCausalLM.from_pretrained(quantized_model_path, torch_dtype="auto", device_map="cuda")
        tokenizer = AutoTokenizer.from_pretrained(quantized_model_path)
        from test.helpers import generate_prompt

        generate_prompt(model, tokenizer)

    @pytest.mark.skip_ci(
        reason="Matrix: Default MXFP4 transform covers CUDA export/reload; MXFP4 tuning remains in the full test tier"
    )
    def test_transform_mxfp4_tuning_quant_infer(self):
        model_name = get_model_path("qwen/Qwen3-0.6B")
        scheme = "MXFP4"

        ar = AutoRound(
            model=model_name,
            iters=2,
            seqlen=2,
            scheme=scheme,
            alg_configs=["auto_round", "hadamard"],
        )
        compressed_model, quantized_model_path = ar.quantize_and_save(output_dir=self.save_dir, format="auto_round")

        model = AutoModelForCausalLM.from_pretrained(quantized_model_path, torch_dtype="auto", device_map="cuda")
        tokenizer = AutoTokenizer.from_pretrained(quantized_model_path)
        from test.helpers import generate_prompt

        generate_prompt(model, tokenizer)

    @pytest.mark.skip_ci(
        reason="Matrix: Random Hadamard behavior has dedicated unit coverage; keep the CUDA MXFP4 E2E smoke deterministic"
    )
    def test_random_transform_mxfp4_quant_infer(self):
        model_name = get_model_path("qwen/Qwen3-0.6B")
        scheme = "MXFP4"

        ar = AutoRound(
            model=model_name,
            iters=0,
            seqlen=2,
            scheme=scheme,
            alg_configs=["rtn", "random_hadamard"],
        )
        compressed_model, quantized_model_path = ar.quantize_and_save(output_dir=self.save_dir, format="auto_round")

        model = AutoModelForCausalLM.from_pretrained(quantized_model_path, torch_dtype="auto", device_map="cuda")
        tokenizer = AutoTokenizer.from_pretrained(quantized_model_path)
        from test.helpers import generate_prompt

        generate_prompt(model, tokenizer)