master

ascend-robotfix npu_sparse_attn_sharedkv ops in Mindspeed

e51bc753创建于 8 天前历史提交

文件	最后提交记录	最后更新时间
csrc	fix npu_sparse_attn_sharedkv ops in Mindspeed Co-authored-by: boes129<chenqi185@huawei.com> # message auto-generated for no-merge-commit merge: !3477 merge chenqi_f1 into master fix npu_sparse_attn_sharedkv ops in Mindspeed Created-by: boes129 Commit-by: boes129 Merged-by: ascend-robot Description: What this PR does / why we need it? 1.新版npu_sparse_attn_sharedkv算子的aclnn接口变了，新增了orikv stride和cmpkv stride参数，以适配不同内存框架。 但mindspeed没感知到，代码里还是按照老接口的调用的，在deepseekv4 flash预训练时会导致调用算子报core dump，参考新算子接口示例 https://gitcode.com/cann/cann-recipes-infer/pull/387/diffs 和ops-transformer仓PR https://gitcode.com/cann/ops-transformer/commit/fc04f943c12b87c6581527bd558fbe38cee31879?ref=master ,修复此问题。 2.原先的文件不满足.clang-format的格式要求，无法通过流水线，按照项目中的.clang-format文件格式化了文件。 Does this PR introduce any user-facing change? NA How was this patch tested? 使用新版算子包， 替换了修改后的npu_sparse_attn_shared_kv.cpp，通过mindspeed对deepseekv4 flash完成预训练，验证了修改成功生效 See merge request: Ascend/MindSpeed!3477	8 天前
triton	style:Triton operator path normalization. Co-authored-by: LinShua<707894133@qq.com> # message auto-generated for no-merge-commit merge: !3432 merge master_GDN_triton_stype into master style:Triton operator path normalization. Created-by: LinShua Commit-by: LinShua Merged-by: ascend-robot Description: What this PR does / why we need it? 删除lite路径下的GDN算子，后续统一归档到ops路径下； 对应关联issue：https://gitcode.com/Ascend/MindSpeed/issues/147 Does this PR introduce any user-facing change? NA How was this patch tested? 见PR中的test文件 See merge request: Ascend/MindSpeed!3432	1 个月前
__init__.py	!371 MoE MLP not use mc2 linear and remove ops from init Merge pull request !371 from Jializheng/master	1 年前
dropout_add_layer_norm.py	!892 公开接口声明+Readme更新 Merge pull request !892 from TJJ/master	1 年前
dropout_add_rms_norm.py	!359 change ascendspeed to mindspeed Merge pull request !359 from 邓佳/master	1 年前
ffn.py	!937 add base st Merge pull request !937 from 王源升/master	1 年前
fusion_attention_v2.py	!892 公开接口声明+Readme更新 Merge pull request !892 from TJJ/master	1 年前
gmm.py	!1297 bugfix_GmmAdd_MmAdd Merge pull request !1297 from zengshu/master	1 年前
gmm_mxfp8.py	feat: add mxfp8 gmm Co-authored-by: EVA1<jingsiyu1@huawei.com> # message auto-generated for no-merge-commit merge: !3291 merge master into master feat: add mxfp8 gmm Created-by: EVA1 Commit-by: EVA1 Merged-by: ascend-robot Description: 添加MXFP8低精gmm模块 See merge request: Ascend/MindSpeed!3291	2 个月前
grouped_matmul.py	feat:mindspeed fsdp support qwen3vl fused ep Co-authored-by: ryanyeee<yezhenyuan@huawei.com> # message auto-generated for no-merge-commit merge: !3254 merge master_0205 into master feat:mindspeed fsdp support qwen3vl fused ep Created-by: ryanyeee Commit-by: ryanyeee Merged-by: ascend-robot Description: feat:mindspeed fsdp support qwen3vl fused ep See merge request: Ascend/MindSpeed!3254	3 个月前
grouped_matmul_mc2.py	feat:mindspeed fsdp Co-authored-by: ryanyeee<yezhenyuan@huawei.com> # message auto-generated for no-merge-commit merge: !3198 merge master into master feat:mindspeed fsdp Created-by: ryanyeee Commit-by: ryanyeee Merged-by: ascend-robot Description: feat:mindspeed fsdp See merge request: Ascend/MindSpeed!3198	4 个月前
lcal_functional.py	!1769 2dtp支持lcoc融合算子 Merge pull request !1769 from kirliavc/2dtp_backward	1 年前
npu_all_to_all_all_gather_bmm.py	!1017 Mc2 Moe Op Bug Fix Merge pull request !1017 from yayahello/master	1 年前
npu_apply_fused_adamw_v2.py	!2341 add swap optimizer to core_r0.12.0 Merge pull request !2341 from wangyuansheng8/master	11 个月前
npu_apply_fused_ema_adamw.py	!1603 add ema adamw optimizer Merge pull request !1603 from wangyuansheng8/master	1 年前
npu_bmm_reduce_scatter_all_to_all.py	!954 update moe MC2 torch api Merge pull request !954 from 陈建军/master	1 年前
npu_grouped_mat_mul_all_reduce.py	!937 add base st Merge pull request !937 from 王源升/master	1 年前
npu_groupmatmul_add.py	fix: adaptation for A5 chip renaming Co-authored-by: xu_yuanhui<xuyuanhui3@h-partners.com> # message auto-generated for no-merge-commit merge: !3277 merge master into master fix: adaptation for A5 chip renaming Created-by: xu_yuanhui Commit-by: xu_yuanhui Merged-by: ascend-robot Description: adaptation for A5 chip renaming See merge request: Ascend/MindSpeed!3277	3 个月前
npu_lightning_indexer.py	Feat: adaptor for DeepSeek V4 Co-authored-by: wuweiqiang24<wuweiqiang11@huawei.com> # message auto-generated for no-merge-commit merge: !3427 merge master into master Feat: adaptor for DeepSeek V4 Created-by: wuweiqiang24 Commit-by: wuweiqiang24 Merged-by: ascend-robot Description: What this PR does / why we need it? Adaptor for DeepSeek V4!!! Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed!3427	1 个月前
npu_matmul_add.py	fix: adaptation for A5 chip renaming Co-authored-by: xu_yuanhui<xuyuanhui3@h-partners.com> # message auto-generated for no-merge-commit merge: !3277 merge master into master fix: adaptation for A5 chip renaming Created-by: xu_yuanhui Commit-by: xu_yuanhui Merged-by: ascend-robot Description: adaptation for A5 chip renaming See merge request: Ascend/MindSpeed!3277	3 个月前
npu_mm_all_reduce_add_rms_norm.py	!937 add base st Merge pull request !937 from 王源升/master	1 年前
npu_mm_all_reduce_add_rms_norm_.py	!937 add base st Merge pull request !937 from 王源升/master	1 年前
npu_moe_token_permute.py	!892 公开接口声明+Readme更新 Merge pull request !892 from TJJ/master	1 年前
npu_moe_token_unpermute.py	!892 公开接口声明+Readme更新 Merge pull request !892 from TJJ/master	1 年前
npu_ring_attention_update.py	!892 公开接口声明+Readme更新 Merge pull request !892 from TJJ/master	1 年前
npu_rotary_position_embedding.py	!892 公开接口声明+Readme更新 Merge pull request !892 from TJJ/master	1 年前
npu_sparse_attn_shared_kv.py	Feat: adaptor for DeepSeek V4 Co-authored-by: wuweiqiang24<wuweiqiang11@huawei.com> # message auto-generated for no-merge-commit merge: !3427 merge master into master Feat: adaptor for DeepSeek V4 Created-by: wuweiqiang24 Commit-by: wuweiqiang24 Merged-by: ascend-robot Description: What this PR does / why we need it? Adaptor for DeepSeek V4!!! Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed!3427	1 个月前
npu_sparse_lightning_indexer_grad_kl_loss.py	Feat: adaptor for DeepSeek V4 Co-authored-by: wuweiqiang24<wuweiqiang11@huawei.com> # message auto-generated for no-merge-commit merge: !3427 merge master into master Feat: adaptor for DeepSeek V4 Created-by: wuweiqiang24 Commit-by: wuweiqiang24 Merged-by: ascend-robot Description: What this PR does / why we need it? Adaptor for DeepSeek V4!!! Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed!3427	1 个月前
quant_gmm.py	!968 Add weight quantize gmm ops Merge pull request !968 from 洪炜杰/hong0927	1 年前
weight_quant_gmm.py	!968 Add weight quantize gmm ops Merge pull request !968 from 洪炜杰/hong0927	1 年前