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

TP/quantization/weight loading refactor part 1 - Simplify parallel linear logic (#1181)

Browse Source
This commit is contained in:
Zhuohan Li
2023-10-02 15:36:09 -07:00
committed by GitHub
parent 84e4e37d14
commit ba0bfd40e2
42 changed files with 819 additions and 1547 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/kernels/test_activation.py
View File

@@ -29,8 +29,8 @@ def test_silu_and_mul(
) -> 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')
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)
ref_out = ref_silu_and_mul(x)
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
@@ -49,8 +49,8 @@ def test_gelu_new(
) -> 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')
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)
ref_out = get_activation("gelu_new")(x)
assert torch.allclose(out, ref_out, atol=1e-5, rtol=1e-5)
@@ -68,8 +68,8 @@ def test_gelu_fast(
) -> 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')
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)