| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | """E2E tests for online FP8 per-channel quantization. |
| 5 | |
| 6 | Loads a BF16 model with ``--quantization fp8_per_channel`` (online |
| 7 | quantization) and compares log-probabilities against the same model served in |
| 8 | BF16 without quantization. This exercises the full pipeline: config parsing, |
| 9 | ``Fp8PtpcOnlineLinearMethod``, ``Fp8PtpcOnlineMoEMethod``, weight |
| 10 | loading, online quantization / shuffling, and inference. |
| 11 | |
| 12 | ``example_prompts`` is a pytest fixture (from conftest.py) that loads 8 |
| 13 | diverse prompts from ``tests/prompts/example.txt``. |
| 14 | """ |
| 15 | |
| 16 | import pytest |
| 17 | |
| 18 | from tests.quantization.utils import is_quant_method_supported |
| 19 | |
| 20 | from ..utils import check_logprobs_close |
| 21 | |
| 22 | # Small MoE model that fits on a single GPU and exercises both linear + MoE. |
| 23 | MOE_MODEL = "allenai/OLMoE-1B-7B-0125-Instruct" |
| 24 | # Small dense model (no MoE) to validate the linear-only path. |
| 25 | DENSE_MODEL = "Qwen/Qwen3-0.6B" |
| 26 | |
| 27 | MAX_MODEL_LEN = 1024 |
| 28 | MAX_TOKENS = 4 |
| 29 | NUM_LOG_PROBS = 8 |
| 30 | |
| 31 | |
| 32 | @pytest.mark.skipif( |
| 33 | not is_quant_method_supported("fp8"), |
| 34 | reason="fp8 is not supported on this GPU type.", |
| 35 | ) |
| 36 | @pytest.mark.quant_model |
| 37 | @pytest.mark.parametrize("model", [DENSE_MODEL, MOE_MODEL], ids=["dense", "moe"]) |
| 38 | def test_fp8_per_channel_logprobs( |
| 39 | vllm_runner, |
| 40 | quant_baseline_logprobs, |
| 41 | example_prompts, |
| 42 | model: str, |
| 43 | monkeypatch: pytest.MonkeyPatch, |
| 44 | ) -> None: |
| 45 | """Compare BF16 baseline logprobs against online per-channel-quantized |
| 46 | model. |
| 47 | |
| 48 | Runs the same model twice -- once in BF16 (baseline) and once with online |
| 49 | FP8 per-channel quantization -- then checks that the top log-probabilities |
| 50 | are close. Only 4 tokens are generated to keep the test fast while still |
| 51 | catching numerical divergence beyond expected per-channel error. |
| 52 | """ |
| 53 | with monkeypatch.context() as m: |
| 54 | m.setenv("TOKENIZERS_PARALLELISM", "true") |
| 55 | |
| 56 | baseline_outputs = quant_baseline_logprobs( |
| 57 | model, |
| 58 | example_prompts, |
| 59 | max_model_len=MAX_MODEL_LEN, |
| 60 | max_tokens=MAX_TOKENS, |
| 61 | num_logprobs=NUM_LOG_PROBS, |
| 62 | ) |
| 63 | |
| 64 | with vllm_runner( |
| 65 | model, |
| 66 | max_model_len=MAX_MODEL_LEN, |
| 67 | enforce_eager=True, |
| 68 | quantization="fp8_per_channel", |
| 69 | ) as vllm_model: |
| 70 | test_outputs = vllm_model.generate_greedy_logprobs( |
| 71 | example_prompts, MAX_TOKENS, NUM_LOG_PROBS |
| 72 | ) |
| 73 | |
| 74 | check_logprobs_close( |
| 75 | outputs_0_lst=baseline_outputs, |
| 76 | outputs_1_lst=test_outputs, |
| 77 | name_0="bf16", |
| 78 | name_1="fp8_per_channel", |
| 79 | ) |