sgl-project--sglang/sgl-kernel/tests/test_per_token_quant_fp8.py

import itertools
import sys
from typing import Tuple

import pytest
import torch
from sgl_kernel import sgl_per_token_quant_fp8

from sglang.srt.utils import is_hip

_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn


def torch_per_token_quant_fp8(tensor, inv_scale):
    # The reference implementation that fully aligns to
    # the kernel being tested.
    finfo = torch.finfo(torch.float8_e4m3fn)
    inv_scale = inv_scale.view(-1, 1)
    scale = inv_scale.reciprocal()
    qweight = (tensor.to(torch.float32) * scale).clamp(min=finfo.min, max=finfo.max)
    qweight = qweight.to(torch.float8_e4m3fn)
    return qweight


def sglang_per_token_quant_fp8(
    input: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
    scale = torch.zeros(input.size(0), device=input.device, dtype=torch.float32)
    output = torch.empty_like(input, device=input.device, dtype=fp8_type_)

    sgl_per_token_quant_fp8(input, output, scale)
    scale = scale.reshape(-1, 1)

    return output, scale


PER_TOKEN_QUANT_CASES = list(
    itertools.product([128, 256, 512], [512, 1076, 1368, 2048, 4096])
) + [
    (39, 1536),
    (1392, 1536),
    (7807, 1536),
]


@pytest.mark.parametrize("num_tokens,hidden_dim", PER_TOKEN_QUANT_CASES)
def test_per_token_quant_compare_implementations(
    num_tokens: int,
    hidden_dim: int,
):
    device = torch.device("cuda")
    x = torch.rand((num_tokens, hidden_dim), dtype=torch.float16, device=device)

    sglang_out, sglang_scale = sglang_per_token_quant_fp8(x)
    torch_out = torch_per_token_quant_fp8(x, sglang_scale)

    torch.testing.assert_close(
        sglang_out.float(), torch_out.float(), rtol=1e-3, atol=1e-3
    )


if __name__ == "__main__":
    sys.exit(pytest.main([__file__]))