biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Clean up kernel unit tests (#938)

Browse Source
This commit is contained in:
Woosuk Kwon
2023-09-06 08:57:38 +09:00
committed by GitHub
parent 22379d5513
commit fbd80ad409
6 changed files with 368 additions and 403 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
tests/kernels/test_activation.py
View File

@@ -1,20 +1,34 @@
import pytest
import torch
import torch.nn.functional as F
from transformers.activations import get_activation
from vllm import activation_ops
DTYPES = [torch.half, torch.bfloat16, torch.float]
NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing
D = [512, 4096, 5120, 13824] # Arbitrary values for testing
SEEDS = [0]
def ref_silu_and_mul(x: torch.Tensor) -> torch.Tensor:
x1, x2 = x.chunk(chunks=2, dim=1)
return F.silu(x1) * x2
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("d", D)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@torch.inference_mode()
def run_silu_and_mul(
def test_silu_and_mul(
num_tokens: int,
d: int,
dtype: torch.dtype,
seed: int,
) -> None:
torch.random.manual_seed(seed)
torch.cuda.manual_seed(seed)
x = torch.randn(num_tokens, 2 * d, dtype=dtype, device='cuda')
out = torch.empty(num_tokens, d, dtype=dtype, device='cuda')
activation_ops.silu_and_mul(out, x)
@@ -22,20 +36,19 @@ def run_silu_and_mul(
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
def test_silu_and_mul() -> None:
for dtype in [torch.half, torch.bfloat16, torch.float]:
for num_tokens in [7, 83, 2048]:
for d in [512, 4096, 5120, 13824]:
print(f'Testing dtype={dtype}, num_tokens={num_tokens}, d={d}')
run_silu_and_mul(num_tokens, d, dtype)
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("d", D)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
@torch.inference_mode()
def run_gelu_new(
def test_gelu_new(
num_tokens: int,
d: int,
dtype: torch.dtype,
seed: int,
) -> None:
torch.random.manual_seed(seed)
torch.cuda.manual_seed(seed)
x = torch.randn(num_tokens, d, dtype=dtype, device='cuda')
out = torch.empty(num_tokens, d, dtype=dtype, device='cuda')
activation_ops.gelu_new(out, x)
@@ -43,30 +56,20 @@ def run_gelu_new(
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
def test_gelu_new() -> None:
for dtype in [torch.half, torch.bfloat16, torch.float]:
for num_tokens in [7, 83, 2048]:
for d in [512, 4096, 5120, 13824]:
print(f'Testing dtype={dtype}, num_tokens={num_tokens}, d={d}')
run_gelu_new(num_tokens, d, dtype)
@torch.inference_mode()
def run_gelu_fast(
@pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@pytest.mark.parametrize("d", D)
@pytest.mark.parametrize("dtype", DTYPES)
@pytest.mark.parametrize("seed", SEEDS)
def test_gelu_fast(
num_tokens: int,
d: int,
dtype: torch.dtype,
seed: int,
) -> None:
torch.random.manual_seed(seed)
torch.cuda.manual_seed(seed)
x = torch.randn(num_tokens, d, dtype=dtype, device='cuda')
out = torch.empty(num_tokens, d, dtype=dtype, device='cuda')
activation_ops.gelu_fast(out, x)
ref_out = get_activation("gelu_fast")(x)
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
def test_gelu_fast() -> None:
for dtype in [torch.half, torch.bfloat16, torch.float]:
for num_tokens in [7, 83, 2048]:
for d in [512, 4096, 5120, 13824]:
print(f'Testing dtype={dtype}, num_tokens={num_tokens}, d={d}')
run_gelu_fast(num_tokens, d, dtype)