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

[CI/Build] Avoid CUDA initialization (#8534)

Browse Source
This commit is contained in:
Cyrus Leung
2024-09-18 18:38:11 +08:00
committed by GitHub
parent e351572900
commit 6ffa3f314c
55 changed files with 256 additions and 256 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/kernels/test_flashinfer.py
View File

@@ -4,6 +4,8 @@ import flashinfer
import pytest
import torch
from vllm.utils import seed_everything
NUM_HEADS = [(16, 16), (32, 8), (64, 8), (6, 1)]
HEAD_SIZES = [128, 256]
BLOCK_SIZES = [16, 32]
@@ -82,7 +84,7 @@ def test_flashinfer_decode_with_paged_kv(
soft_cap: Optional[float],
) -> None:
torch.set_default_device("cuda")
torch.cuda.manual_seed_all(0)
seed_everything(0)
num_seqs = len(kv_lens)
num_query_heads = num_heads[0]
num_kv_heads = num_heads[1]
@@ -168,7 +170,7 @@ def test_flashinfer_prefill_with_paged_kv(seq_lens: List[Tuple[int, int]],
block_size: int,
soft_cap: Optional[float]) -> None:
torch.set_default_device("cuda")
torch.cuda.manual_seed_all(0)
seed_everything(0)
num_seqs = len(seq_lens)
query_lens = [x[0] for x in seq_lens]
kv_lens = [x[1] for x in seq_lens]
@@ -266,7 +268,7 @@ def test_flashinfer_prefill_with_paged_fp8_kv(
head_size: int, dtype: torch.dtype, block_size: int,
soft_cap: Optional[float]) -> None:
torch.set_default_device("cuda")
torch.cuda.manual_seed_all(0)
seed_everything(0)
num_seqs = len(seq_lens)
query_lens = [x[0] for x in seq_lens]
kv_lens = [x[1] for x in seq_lens]
@@ -379,7 +381,7 @@ def test_flashinfer_decode_with_paged_fp8_kv(
) -> None:
# test doesn't work for num_heads = (16,16)
torch.set_default_device("cuda")
torch.cuda.manual_seed_all(0)
seed_everything(0)
num_seqs = len(kv_lens)
num_query_heads = num_heads[0]
num_kv_heads = num_heads[1]