| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
[mindspore] support qwen3vl finetune for megatron backend Co-authored-by: ffmh<fengminghao2@huawei.com> # message auto-generated for no-merge-commit merge: !1785 merge qwen3vl_megatron into master [mindspore] support qwen3vl finetune for megatron backend Created-by: ffmh Commit-by: ffmh Merged-by: ascend-robot Description: ## Motivation Please describe the motivation of this PR and the goal you want to achieve through this PR. 适配 qwen3vl 8B/30B megatron后端 ## Modification Please briefly describe what modification is made in this PR. 添加 layer_spec修改、text模块transformer block/layer添加deepstack相关适配 ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1785 | 7 个月前 | |
[Feature] MM_FSDP2: Add attention patch for transformers. And add ring & usp for qwen3vl Co-authored-by: zs-derrick1<1434012475@qq.com> # message auto-generated for no-merge-commit merge: !2120 merge fsdp2_dev2 into fsdp2_dev [Feature] MM_FSDP2: Add attention patch for transformers. And add ring & usp for qwen3vl Created-by: zs-derrick1 Commit-by: zs-derrick1 Merged-by: ascend-robot Description: ## Motivation Add attention patch for transformers. And add ring & usp for qwen3vl ## Modification 1.Add transformers attention patch to support cp（ulysses/ring/usp) 2.add ring&usp for qwen3vl ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [x] CLA has been signed and all committers have signed the CLA in this PR. - [x] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!2120 | 4 个月前 | |
[Feature] keep profiling of fsdp2 dev the same with the master branch Co-authored-by: weixin_44031810<gaojie75@huawei.com> # message auto-generated for no-merge-commit merge: !2196 merge fsdp2_dev into fsdp2_dev [Feature] keep profiling of fsdp2 dev the same with the master branch Created-by: gaojie_ Commit-by: weixin_44031810 Merged-by: ascend-robot Description: ## Motivation keep profiling same with the master branch ## Modification Please briefly describe what modification is made in this PR. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!2196 | 4 个月前 | |
[Modify][MindSpore] Optimize data preprocessing performance for Qwen3vl Co-authored-by: weixin_47897441<wuyouqi1@h-partners.com> # message auto-generated for no-merge-commit merge: !1926 merge master-qwen3vl into master [Modify][MindSpore] Optimize data preprocessing performance for Qwen3vl Created-by: weixin_47897441 Commit-by: weixin_47897441 Merged-by: ascend-robot Description: ## Motivation This PR aims to optimize the performance of qwenvl3 data preprocessing by replacing partial tensor operations with tensor-to-numpy conversions, reducing unnecessary tensor memory copies and improving the overall efficiency of the preprocessing pipeline. ## Modification Replace tensor-based data operations that cause redundant memory copies with tensor-to-numpy conversion operations. ## Self-test (Optional) 1.After modification, the output of qwenvl3 data preprocessing is consistent with the original version. 2.The preprocessing time is reduced from 10xPTA to 5.2xPTA.  ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [x] CLA has been signed and all committers have signed the CLA in this PR. - [x] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1926 | 6 个月前 | |
[Bugfix]fix fpdt nonblocking data dependency bug Co-authored-by: Bian Zheng<bianzheng8@huawei.com> # message auto-generated for no-merge-commit merge: !1951 merge fix_fpdt_nonblocking into master [Bugfix]fix fpdt nonblocking data dependency bug Created-by: rewindz Commit-by: Bian Zheng Merged-by: ascend-robot Description: ## Motivation Fix the non-blocking data dependency bug in the FPDT (Fully Pipelined Distributed Transformer, aka ulysses offload) mechanism. In the current implementation, data transfer between CPU and NPU uses non-blocking mode, but data dependencies are not properly handled, which may lead to data inconsistency or calculation errors. ## Modification 1. Optimize data transfer mode : Change data transfer between CPU and NPU from non_blocking=False to non_blocking=True to improve data transfer efficiency. 2. Add stream synchronization mechanism : - When loading data to NPU, use offload_stream for asynchronous loading and wait for completion in compute_stream - Add compute_stream.synchronize() and general_offload_stream.synchronize() to ensure data transfer is complete before computation - Add torch_npu.npu.synchronize() before backpropagation to ensure all operations are complete 3. Refactor data loading logic : Encapsulate data loading operations in a stream context manager to clarify the asynchronous nature of data transfer. ## Self-test (Optional) The correctness of the modification has been verified through the following tests: - Functional test: Ensure the FPDT mechanism works normally in non-blocking mode - Performance test: Verify the speedup of step_time with 10~15% (baseline: blocking FPDT) - Consistency test: Ensure the calculation results are consistent with those before modification - Precision alignment: Precision has been aligned with PR #1886 ## BC-breaking (Optional) No compatibility issues. The modification only involves internal implementation details and does not affect external interfaces. The required cann/torch_npu version remains consistent with before modification. ## Checklist **Before PR**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1951 | 6 个月前 | |
[Feature] Add the muon optimizer and adapt it for FSDP2. Co-authored-by: hanyyy<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !1964 merge master into master [Feature] Add the muon optimizer and adapt it for FSDP2. Created-by: vasileone Commit-by: hanyyy Merged-by: ascend-robot Description: ## Motivation Added the Muon optimizer and enabled its compatibility with FSDP2. ## Modification - Added Muon optimizer implementation with FSDP2 compatibility. - Added patches for Megatron-core to enable zero-code-change injection. - Fixed the issue of Muon optimizer failing under FSDP2 context: sharded parameter handling, gradient sync, and step behavior. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - The new code needs to comply with the Clean Code specification. - The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - CLA has been signed and all committers have signed the CLA in this PR. - The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1964 | 6 个月前 | |
[Feature] Add the muon optimizer and adapt it for FSDP2. Co-authored-by: hanyyy<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !1964 merge master into master [Feature] Add the muon optimizer and adapt it for FSDP2. Created-by: vasileone Commit-by: hanyyy Merged-by: ascend-robot Description: ## Motivation Added the Muon optimizer and enabled its compatibility with FSDP2. ## Modification - Added Muon optimizer implementation with FSDP2 compatibility. - Added patches for Megatron-core to enable zero-code-change injection. - Fixed the issue of Muon optimizer failing under FSDP2 context: sharded parameter handling, gradient sync, and step behavior. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - The new code needs to comply with the Clean Code specification. - The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - CLA has been signed and all committers have signed the CLA in this PR. - The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1964 | 6 个月前 | |
[Modify] delete qwen2.5vl grpo from master Co-authored-by: lu-jinfu1999<lujinfu1@h-partners.com> # message auto-generated for no-merge-commit merge: !1896 merge master into master [Modify] delete qwen2.5vl grpo from master Created-by: lu-jinfu1999 Commit-by: lu-jinfu1999 Merged-by: ascend-robot Description: ## Motivation delete qwen2.5vl grpo from master. ## Modification delete qwen2.5vl grpo from master. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [x] CLA has been signed and all committers have signed the CLA in this PR. - [x] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1896 | 6 个月前 | |
[Feature]VACE Model and Data Process Co-authored-by: feng0w0<houyufeng4@huawei.com> # message auto-generated for no-merge-commit merge: !1669 merge master into master [Feature]VACE Model and Data Process Created-by: feng0w0 Commit-by: feng0w0 Merged-by: ascend-robot Description: ## Motivation Support VACE Model ## Modification Added VACE data processing pipeline and model components. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1669 | 7 个月前 | |
[Modify] async offload modify Co-authored-by: liyingxuan<liyingxuan3@huawei.com> # message auto-generated for no-merge-commit merge: !1909 merge async_offload into master [Modify] async offload modify Created-by: liyx616 Commit-by: liyingxuan Merged-by: ascend-robot Description: ## Motivation 优化async offload处理逻辑 ## Modification 1. 原来的代码中，stream和event的管理过于混乱，有一些冗余的wait关系，也缺乏一些必要的wait关系，进行重新梳理，梳理之后统一使用event进行管理 2. 修复反向H2D潜在的显存未及时释放问题 3. 添加必要的注释 4. 移除没有用到的函数 计算流和swap流的依赖关系梳理如下：  ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [ ] The new code needs to comply with the Clean Code specification. - [ ] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [ ] CLA has been signed and all committers have signed the CLA in this PR. - [ ] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1909 | 6 个月前 | |
[Feature]MM_FSDP2:place dependencies for non-FSDP directories inside the FSDP directory Co-authored-by: zs-derrick1<1434012475@qq.com> # message auto-generated for no-merge-commit merge: !2090 merge fsdp2_dev into fsdp2_dev [Feature]MM_FSDP2:place dependencies for non-FSDP directories inside the FSDP directory Created-by: zs-derrick1 Commit-by: zs-derrick1 Merged-by: ascend-robot Description: ## Motivation MM_FSDP2: Place dependencies for non-FSDP directories inside the FSDP directory ## Modification Please briefly describe what modification is made in this PR. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [x] CLA has been signed and all committers have signed the CLA in this PR. - [x] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!2090 | 5 个月前 | |
[Feature] Add the muon optimizer and adapt it for FSDP2. Co-authored-by: hanyyy<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !1964 merge master into master [Feature] Add the muon optimizer and adapt it for FSDP2. Created-by: vasileone Commit-by: hanyyy Merged-by: ascend-robot Description: ## Motivation Added the Muon optimizer and enabled its compatibility with FSDP2. ## Modification - Added Muon optimizer implementation with FSDP2 compatibility. - Added patches for Megatron-core to enable zero-code-change injection. - Fixed the issue of Muon optimizer failing under FSDP2 context: sharded parameter handling, gradient sync, and step behavior. ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - The new code needs to comply with the Clean Code specification. - The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - CLA has been signed and all committers have signed the CLA in this PR. - The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1964 | 6 个月前 | |
[Feature] Support custom optimizer lr scale and weight decay ignore Co-authored-by: suis-moi<zhushixiang@huawei.com> # message auto-generated for no-merge-commit merge: !1878 merge master into master [Feature] Support custom optimizer lr scale and weight decay ignore Created-by: suis-moi Commit-by: suis-moi Merged-by: ascend-robot Description: ## Motivation support custom optimizer lr scale and weight decay ignore ## Modification 1、Add condition function to determine if a parameter should be excluded from weight decay. 2、Add condition function to determine if a parameter should apply learning rate scaling (with --lr-mult). 使用示例： 在训练shell脚本的args添加 shell GPT_ARGS=" ... --weight-decay-exclude-modules norm bias \ # 排除指定参数的权重衰减，根据实际需要设置参数关键词 --lr-scale-modules vision \ # 对视觉模块参数缩放学习率（如mult=0.5） --lr-mult 0.5 \ " ## Self-test (Optional) If modifications to this PR may cause/fix function/accuracy/performance DTSs/issues, a self-inspection record needs to be attached. ## BC-breaking (Optional) If there are compatibility issues, such as dependencies on cann/torch_npu versions, they need to be explained in the PR. ## Checklist **Before PR**: - [x] The new code needs to comply with the Clean Code specification. - [x] The PR content is self-checked, and the expression can be clear and the writing standardized **After PR**: - [x] CLA has been signed and all committers have signed the CLA in this PR. - [x] The ci-pipeline is passed, Code Check is passed. See merge request: Ascend/MindSpeed-MM!1878 | 6 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 7 个月前 | ||
| 4 个月前 | ||
| 4 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 6 个月前 | ||
| 7 个月前 | ||
| 6 个月前 | ||
| 5 个月前 | ||
| 6 个月前 | ||
| 6 个月前 |