Files
sglang/python/sglang/jit_kernel/tests/test_store_cache.py
2026-01-10 17:34:09 -08:00

36 lines
1.1 KiB
Python

import itertools
import pytest
import torch
from sglang.jit_kernel.kvcache import store_cache
BS_LIST = [2**n for n in range(0, 15)]
BS_LIST += [x + 1 + i for i, x in enumerate(BS_LIST)]
HIDDEN_DIMS = [64, 128, 256, 512, 1024, 96, 98, 100]
CACHE_SIZE = 1024 * 1024
DTYPE = torch.bfloat16
DEVICE = "cuda"
@pytest.mark.parametrize(
"batch_size,element_dim",
list(itertools.product(BS_LIST, HIDDEN_DIMS)),
)
def test_store_cache(batch_size: int, element_dim: int) -> None:
k = torch.randn((batch_size, element_dim), dtype=DTYPE, device=DEVICE)
v = torch.randn((batch_size, element_dim), dtype=DTYPE, device=DEVICE)
k_cache = torch.randn((CACHE_SIZE, element_dim), dtype=DTYPE, device=DEVICE)
v_cache = torch.randn((CACHE_SIZE, element_dim), dtype=DTYPE, device=DEVICE)
indices = torch.randperm(CACHE_SIZE, device=DEVICE)[:batch_size]
# AOT store cache
store_cache(k, v, k_cache, v_cache, indices)
assert torch.all(k_cache[indices] == k)
assert torch.all(v_cache[indices] == v)
if __name__ == "__main__":
pytest.main([__file__])