Closed
[Bug-Report|缺陷反馈]: 910B2 grouped_matmul_finalize_routing算子开确定性之后卡死 #4334
yangchao888Add on Aug 14Closed at Aug 25
weihao18
Aug 14 Comments:
Aug 14 Comments:
/assign @zhangzhuoran3


Aug 14 Assign hunter_li as responsible person
Aug 14 Assign zhangzhuoran3 as responsible person, remove original responsible person: hunter_li
Aug 14 Assign zhangzhuoran3 as responsible person, remove original responsible person: hunter_li
Aug 19 Associated pull request
Aug 25 Close issue
Aug 25 Added label: resolved
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Describe the current behavior / 问题描述 (Mandatory / 必填)
910B2 grouped_matmul_finalize_routing算子开确定性之后卡死
Environment / 环境信息 (Mandatory / 必填)
芯片型号:Ascend910B2
CANN 版本:9.0.0
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
输入信息:
x_int8.shape: torch.Size([16548, 1536]), x_int8.dtype: torch.int8
weight_int8.shape: torch.Size([180, 1536, 1536]), weight_int8.dtype: torch.int8
group_list.shape: torch.Size([180]), group_list.dtype: torch.int64
scale.shape: torch.Size([180, 1536]), scale.dtype: torch.float32
pertoken_scale.shape: torch.Size([16548]), pertoken_scale.dtype: torch.float32
shared_input.shape: torch.Size([394, 1536]), shared_input.dtype: torch.bfloat16
logit.shape: torch.Size([16548]), logit.dtype: torch.float32
row_index.shape: torch.Size([16548]), row_index.dtype: torch.int32
output_bs: 394
group_list_type: 1
开确定性:
torch.use_deterministic_algorithms(True)
output = torch_npu.npu_grouped_matmul_finalize_routing(
x_int8,
weight_int8,
group_list,
scale=scale,
pertoken_scale=pertoken_scale,
shared_input=shared_input,
logit=logit,
row_index=row_index,
output_bs=output_bs,
group_list_type=group_list_type,
)
调用output = torch_npu.npu_grouped_matmul_finalize_routing()后,一直卡住无返回。
torch.use_deterministic_algorithms(True) 改为 torch.use_deterministic_algorithms(False) 后正常返回。
Describe the expected behavior / 预期结果 (Mandatory / 必填)
调用output = torch_npu.npu_grouped_matmul_finalize_routing()后,正常输出结果
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
调用output = torch_npu.npu_grouped_matmul_finalize_routing()后,一直卡住无返回
Special notes for this issue/备注 (Optional / 选填)