| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests AutoGPTQ (GPTQ with Marlin kernels) output correctness. |
| 4 | |
| 5 | Note: Marlin internally uses locks to synchronize the threads. This can |
| 6 | result in very slight nondeterminism for Marlin. As a result, we re-run the test |
| 7 | up to 3 times to see if we pass. |
| 8 | """ |
| 9 | |
| 10 | import os |
| 11 | |
| 12 | import pytest |
| 13 | |
| 14 | from tests.quantization.utils import is_quant_method_supported |
| 15 | from vllm.model_executor.layers.rotary_embedding import _ROPE_DICT |
| 16 | from vllm.platforms import current_platform |
| 17 | |
| 18 | from ..utils import check_logprobs_close |
| 19 | |
| 20 | os.environ["TOKENIZERS_PARALLELISM"] = "true" |
| 21 | |
| 22 | MAX_MODEL_LEN = 1024 |
| 23 | |
| 24 | MODELS = [ |
| 25 | ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", "main"), |
| 26 | # desc_act=True is a no-op for channelwise quantization. |
| 27 | ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-8bit--1g-actorder_True"), |
| 28 | ("TechxGenus/gemma-2b-it-GPTQ", "main"), |
| 29 | ] |
| 30 | |
| 31 | |
| 32 | @pytest.mark.flaky(reruns=3) |
| 33 | @pytest.mark.skipif( |
| 34 | not is_quant_method_supported("auto_gptq") |
| 35 | or current_platform.is_rocm() |
| 36 | or not current_platform.is_cuda(), |
| 37 | reason="auto_gptq is not supported on this GPU type.", |
| 38 | ) |
| 39 | @pytest.mark.parametrize("model", MODELS) |
| 40 | @pytest.mark.parametrize("dtype", ["half", "bfloat16"]) |
| 41 | @pytest.mark.parametrize("max_tokens", [32]) |
| 42 | @pytest.mark.parametrize("num_logprobs", [5]) |
| 43 | def test_models( |
| 44 | vllm_runner, |
| 45 | example_prompts, |
| 46 | model, |
| 47 | dtype: str, |
| 48 | max_tokens: int, |
| 49 | num_logprobs: int, |
| 50 | ) -> None: |
| 51 | model_name, revision = model |
| 52 | |
| 53 | # Run marlin. |
| 54 | with vllm_runner( |
| 55 | model_name=model_name, |
| 56 | revision=revision, |
| 57 | dtype=dtype, |
| 58 | quantization="marlin", |
| 59 | max_model_len=MAX_MODEL_LEN, |
| 60 | tensor_parallel_size=1, |
| 61 | ) as gptq_marlin_model: |
| 62 | gptq_marlin_outputs = gptq_marlin_model.generate_greedy_logprobs( |
| 63 | example_prompts[:-1], max_tokens, num_logprobs |
| 64 | ) |
| 65 | _ROPE_DICT.clear() # clear rope cache to avoid rope dtype error |
| 66 | |
| 67 | # Run gptq. |
| 68 | # The naive gptq kernel doesn't support bf16 yet. |
| 69 | # Here we always compare fp16/bf16 gpt marlin kernel |
| 70 | # to fp16 gptq kernel. |
| 71 | with vllm_runner( |
| 72 | model_name=model_name, |
| 73 | revision=revision, |
| 74 | dtype="half", |
| 75 | quantization="gptq", |
| 76 | max_model_len=MAX_MODEL_LEN, |
| 77 | tensor_parallel_size=1, |
| 78 | ) as gptq_model: |
| 79 | gptq_outputs = gptq_model.generate_greedy_logprobs( |
| 80 | example_prompts[:-1], max_tokens, num_logprobs |
| 81 | ) |
| 82 | |
| 83 | check_logprobs_close( |
| 84 | outputs_0_lst=gptq_outputs, |
| 85 | outputs_1_lst=gptq_marlin_outputs, |
| 86 | name_0="gptq", |
| 87 | name_1="gptq_marlin", |
| 88 | ) |