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

[Hardware][ROCM] using current_platform.is_rocm (#9642)

Browse Source 
Signed-off-by: wangshuai09 <391746016@qq.com>
This commit is contained in:
wangshuai09
2024-10-28 12:07:00 +08:00
committed by GitHub
parent 34a9941620
commit 4e2d95e372
32 changed files with 165 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/kernels/test_blocksparse_attention.py
View File

@@ -7,7 +7,8 @@ import torch
from vllm import _custom_ops as ops
from vllm.attention.ops.blocksparse_attention.interface import (
LocalStridedBlockSparseAttn)
from vllm.utils import get_max_shared_memory_bytes, is_hip, seed_everything
from vllm.platforms import current_platform
from vllm.utils import get_max_shared_memory_bytes, seed_everything
from .allclose_default import get_default_atol, get_default_rtol
@@ -316,8 +317,8 @@ def test_paged_attention(
# NOTE(woosuk): Due to the kernel-level differences in the two
# implementations, there is a small numerical difference in the two
# outputs. Thus, we use a relaxed tolerance for the test.
atol = get_default_atol(output) if is_hip() else 1e-3
rtol = get_default_rtol(output) if is_hip() else 1e-5
atol = get_default_atol(output) if current_platform.is_rocm() else 1e-3
rtol = get_default_rtol(output) if current_platform.is_rocm() else 1e-5
# NOTE(zhaoyang): FP8 KV Cache will introduce quantization error,
# so we use a relaxed tolerance for the test.