torchair/tests · Ascend/torchair

文件	最后提交记录	最后更新时间
aoe	clean code Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3171 merge clean_code_clang-format into master clean code Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: clean_code_clang-format c++ clean code See merge request: Ascend/torchair!3171	3 个月前
privateuse1	clean code Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3171 merge clean_code_clang-format into master clean code Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: clean_code_clang-format c++ clean code See merge request: Ascend/torchair!3171	3 个月前
reg_op_stub	clean code Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3171 merge clean_code_clang-format into master clean code Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: clean_code_clang-format c++ clean code See merge request: Ascend/torchair!3171	3 个月前
smoke	bug fix all_to_all_single support mulit dim in eager model Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3450 merge alltoallsignl into master bug fix all_to_all_single support mulit dim in eager model Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: bug fix all_to_all_single support mulit dim in eager model：原： out_size[0] = sum(recv_counts) out_tensor = input_tensor.new_empty(out_size) 其中recv_counts为接受的元素个数而实际：c10d.all_to_all_single接受的是output_split_sizes是对input[0]的切分比例，用来表示每个卡返回的元素比例是多少。而不是要返回的元素个数就是下面举例的，入参应该是[2, 5, 6, 3]，而不是[24, 60, 72, 36] 元素个数。这样入参1、会导致错误，output_split_sizes和input[0] 个数不匹配报错，input[0] 总共是16个。 c10d.all_to_all_single( out_tensor, input_tensor, output_split_sizes=recv_counts, input_split_sizes=send_counts, group=pg, async_op=False, ) 示例条件 plaintext input_tensor shape = [16, 4, 3] group_size = 4 # 4张卡rank0~rank3 output_split_sizes = [2, 5, 6, 3] 步骤 1：计算 input_non_0_dim_size plaintext out_size[0] = 16 total_elem = 16 * 4 * 3 = 192 input_non_0_dim_size = 192 // 16 = 12 等价 D1D2 = 43 = 12 含义：第 0 维取 1 行时，内部固定 12 个元素。步骤 2：逐 rank 计算 recv_counts 公式：recv_counts[i] = output_split_sizes[i] * 12 rank0：2 × 12 = 24 rank1：5 × 12 = 60 rank2：6 × 12 = 72 rank3：3 × 12 = 36 最终 plaintext recv_counts = [24, 60, 72, 36] 含义：all_to_all 通信时，每张卡接收的元素总量。 See merge request: Ascend/torchair!3450	2 个月前
st	更新readme文件，方便ai进行读取并部署环境 Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3467 merge master into master 更新readme文件，方便ai进行读取并部署环境 Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: 更新readme文件，方便ai进行读取并部署环境 See merge request: Ascend/torchair!3467	2 个月前
ut	clean code Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3171 merge clean_code_clang-format into master clean code Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: clean_code_clang-format c++ clean code See merge request: Ascend/torchair!3171	3 个月前
utils_tools	clean code Co-authored-by: guanlongfeng<guanlongfeng1@huawei.com> # message auto-generated for no-merge-commit merge: !3171 merge clean_code_clang-format into master clean code Created-by: glf2016 Commit-by: guanlongfeng Merged-by: ascend-robot Description: clean_code_clang-format c++ clean code See merge request: Ascend/torchair!3171	3 个月前
CMakeLists.txt	!1718 add func symbol get Merge pull request !1718 from huangziming3/master	1 年前