| Files | Last commit | Last update |
|---|---|---|
fix(pytorch): checkpoint_convert_hf_mcore.md sunset Co-authored-by: cjy840282<chenjingyi9@huawei.com> # message auto-generated for no-merge-commit merge: !4697 merge ckpt into master fix(pytorch): checkpoint_convert_hf_mcore.md sunset Created-by: cjy840282 Commit-by: cjy840282 Merged-by: ascend-robot Description: ## What this PR does / why we need it? Due to an upgrade in the transformers version, support for convert_ckpt_V1 has been discontinued. Running the example code provided in the checkpoint_convert_hf_mcore.md document for convert_ckpt.py now results in errors. Consequently, this PR has removed the documentation related to convert_ckpt_V1. ## Does this PR introduce any user-facing change? Guide users to use the convert_ckpt_V2, which offers better usability. ## How was this patch tested? This PR primarily involves documentation updates, with no code modifications. It has been tested and verified through manual review. See merge request: Ascend/MindSpeed-LLM!4697 | 6 days ago | |
feat: add MTP loss support with mtp_loss_scaling_factor parameter Co-authored-by: wxy1105<wangxinyu136@huawei.com> # message auto-generated for no-merge-commit merge: !4633 merge master_train into master feat: add MTP loss support with mtp_loss_scaling_factor parameter Created-by: wxy1105 Commit-by: wxy1105 Merged-by: ascend-robot Description: ## What this PR does / why we need it? Background : This PR adds MTP (Multi Token Prediction) loss support for DeepSeek-V4 model training in FSDP2 backend, enabling multi-step prediction during training to improve model performance. Detailed changes : -Added mtp_loss_scaling_factor parameter ( default=0.3 ) in arguments.py for controlling MTP loss weight -Created LossOutput dataclass in trainer.py to encapsulate total loss, LM loss, and MTP loss -Modified _compute_loss method to calculate and accumulate MTP loss from model outputs -Updated training_step to return separate LM loss and MTP loss values for logging -Enhanced train_monitor.py to log MTP loss metrics alongside LM loss -Added documentation for mtp_loss_scaling_factor in arguments.md ## Does this PR introduce any user-facing change? Yes : -New mtp_loss_scaling_factor parameter available in FSDP2 training configuration -Training logs now include separate lm loss and mtp loss metrics -Related docs: arguments.md ## How was this patch tested? -Compilation and unit tests for FSDP2 trainer with MTP loss integration -Training monitor logging validation with MTP metrics -End-to-end training tests with DeepSeek-V4 models See merge request: Ascend/MindSpeed-LLM!4633 | 4 days ago | |
docs(pytorch): update FAQ based on frequently asked questions Co-authored-by: HanhuiChen<chenhanhui1@h-partners.com> # message auto-generated for no-merge-commit merge: !4613 merge master into master docs(pytorch): update FAQ based on frequently asked questions Created-by: HANHU1CHEN Commit-by: HanhuiChen Merged-by: ascend-robot Description: ## What this PR does / why we need it? Add a MindSpeed-LLM FAQ document based on frequently asked questions collected from resolved community issues. The FAQ covers common problems related to: - Dataset preparation and loading - Checkpoint conversion - NPU environment configuration - Training startup failures - Out-of-memory (OOM) errors This helps users quickly locate solutions to common issues and improves the onboarding experience. ## Does this PR introduce any user-facing change? Yes. This PR adds a new FAQ section/document for users. No code logic or training behavior is changed. Documentation: - docs/zh/FAQ.md (or corresponding FAQ document path) ## How was this patch tested? - Reviewed resolved issues from the MindSpeed-LLM community. - Consolidated duplicated questions and extracted representative solutions. - Verified that the documented FAQs are consistent with existing documentation and issue resolutions. No functional code changes are included in this PR. See merge request: Ascend/MindSpeed-LLM!4613 | 24 days ago | |
[pytorch][docs] add docs license Co-authored-by: leizhenzhen23<leizhenzhen4@huawei.com> # message auto-generated for no-merge-commit merge: !4240 merge llm_0225 into master [pytorch][docs] add docs license Created-by: leizhenzhen23 Commit-by: leizhenzhen23 Merged-by: ascend-robot Description: [pytorch][docs] add docs license See merge request: Ascend/MindSpeed-LLM!4240 | 4 months ago | |
docs: add torch.load security risk statement Co-authored-by: zhyebin01<zhangyebin@h-partners.com> # message auto-generated for no-merge-commit merge: !4589 merge h into master docs: add torch.load security risk statement Created-by: zhyebin01 Commit-by: zhyebin01 Merged-by: ascend-robot Description: ## What this PR does / why we need it? add torch.load security risk statement ## Does this PR introduce any user-facing change? NA. ## How was this patch tested? doc-ci test pass See merge request: Ascend/MindSpeed-LLM!4589 | 18 days ago | |
fix(pytorch): checkpoint_convert_hf_mcore.md sunset Co-authored-by: cjy840282<chenjingyi9@huawei.com> # message auto-generated for no-merge-commit merge: !4697 merge ckpt into master fix(pytorch): checkpoint_convert_hf_mcore.md sunset Created-by: cjy840282 Commit-by: cjy840282 Merged-by: ascend-robot Description: ## What this PR does / why we need it? Due to an upgrade in the transformers version, support for convert_ckpt_V1 has been discontinued. Running the example code provided in the checkpoint_convert_hf_mcore.md document for convert_ckpt.py now results in errors. Consequently, this PR has removed the documentation related to convert_ckpt_V1. ## Does this PR introduce any user-facing change? Guide users to use the convert_ckpt_V2, which offers better usability. ## How was this patch tested? This PR primarily involves documentation updates, with no code modifications. It has been tested and verified through manual review. See merge request: Ascend/MindSpeed-LLM!4697 | 6 days ago | |
docs: update README/quickstart/precision_issue/supported_models Co-authored-by: xuwenyue<xuwenyue1@huawei.com> # message auto-generated for no-merge-commit merge: !4550 merge fix_docs_0529 into master docs: update README/quickstart/precision_issue/supported_models Created-by: xuwenyue Commit-by: xuwenyue Merged-by: ascend-robot Description: ## What this PR does / why we need it? update README: 未来规划/社区会议/加入我们 move to bottom, 最新消息 add a details tag, update quickstart: add note add precision_issue.md update supported_models: remove mindspeed-mm supported models table ## 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-LLM!4550 | 1 month ago | |
fix(pytorch): ds4 LoRA fine-tuning gets stuck when PP is enabled Co-authored-by: cjy840282<chenjingyi9@huawei.com> # message auto-generated for no-merge-commit merge: !4678 merge ds4_lora into master fix(pytorch): ds4 LoRA fine-tuning gets stuck when PP is enabled Created-by: cjy840282 Commit-by: cjy840282 Merged-by: ascend-robot Description: ## What this PR does / why we need it? Fix ds4 LoRA fine-tuning gets stuck when PP is enabled. ## Does this PR introduce any user-facing change? No. Internal bugfix, no config or interface changes. ## How was this patch tested? 1. After the fix, launching PP no longer hangs and can proceed to the iteration normally. 2. When PP=1, run with the modified and unmodified code respectively after loading weights to ensure the loss can align. See merge request: Ascend/MindSpeed-LLM!4678 | 5 days ago | |
docs(pytorch): handling minor document errors and update menu Co-authored-by: LQ1206<liuqian164@h-partners.com> # message auto-generated for no-merge-commit merge: !4657 merge master into master docs(pytorch): handling minor document errors and update menu Created-by: LQ1206 Commit-by: LQ1206 Merged-by: ascend-robot Description: ## What this PR does / why we need it? handling minor document errors and update menu ## Does this PR introduce any user-facing change? handling document dormat errors and update menu ## How was this patch tested? No test See merge request: Ascend/MindSpeed-LLM!4657 | 12 days ago | |
docs: improve docs structure, readability and navigation Co-authored-by: leizhenzhen23<leizhenzhen4@huawei.com> # message auto-generated for no-merge-commit merge: !4604 merge master into master docs: improve docs structure, readability and navigation Created-by: leizhenzhen23 Commit-by: leizhenzhen23 Merged-by: ascend-robot Description: ## What this PR does / why we need it? Optimizes the structure and readability of project_guide.md and docs_guide.md, including section hierarchy unification, transitional content improvement, and directory tree collapsible display. Supplements missing directory entries and document navigation links to keep docs consistent with the actual repository. ## Does this PR introduce any user-facing change? Yes. Directory tree now defaults to concise view with collapsible details; more navigation links and categories added; img.png path changed to docs/zh/pytorch/figures/project_guide/img.png . ## How this patch tested? Verified Markdown rendering of collapsible blocks; cross-checked all entries and links against actual repository; confirmed no MindSpore content reintroduced. See merge request: Ascend/MindSpeed-LLM!4604 | 25 days ago | |
docs(pytorch): update document Co-authored-by: LQ1206<liuqian164@h-partners.com> # message auto-generated for no-merge-commit merge: !4661 merge master into master docs(pytorch): update document Created-by: LQ1206 Commit-by: LQ1206 Merged-by: ascend-robot Description: ## What this PR does / why we need it? update document ## Does this PR introduce any user-facing change? fix error for AIDD ## How was this patch tested? No test See merge request: Ascend/MindSpeed-LLM!4661 | 11 days ago | |
fix(pytorcch): fix describe of replace_ascend_path Co-authored-by: qyzqyz<quyueze@h-partners.com> # message auto-generated for no-merge-commit merge: !4626 merge master into master fix(pytorcch): fix describe of replace_ascend_path Created-by: qyzqyz Commit-by: qyzqyz Merged-by: ascend-robot Description: ## What this PR does / why we need it? This PR adds a global tool to change the HDK path. The tool allows users to easily switch or configure the HDK installation directory across the project, improving flexibility for environments where HDK is installed in non-standard locations. ## Does this PR introduce any user-facing change? This PR allows users to set/change the HDK path globally . Users can now specify a custom HDK path via this tool instead of manually modifying environment variables or configuration files. ## How was this patch tested? Manually tested the tool by setting different HDK paths and verifying that the environment variables and configurations are updated correctly. See merge request: Ascend/MindSpeed-LLM!4626 | 19 days ago |
| Files | Last commit | Last update |
|---|---|---|
| 6 days ago | ||
| 4 days ago | ||
| 24 days ago | ||
| 4 months ago | ||
| 18 days ago | ||
| 6 days ago | ||
| 1 month ago | ||
| 5 days ago | ||
| 12 days ago | ||
| 25 days ago | ||
| 11 days ago | ||
| 19 days ago |