| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
fix(pytorch):fix bug of chunkloss Co-authored-by: qyzqyz<quyueze@h-partners.com> # message auto-generated for no-merge-commit merge: !5009 merge master into master fix(pytorch):fix bug of chunkloss Created-by: qyzqyz Commit-by: qyzqyz Merged-by: ascend-robot Description: ## What this PR does / why we need it? The compatibility environment variable was previously enabled globally. However, this caused some test cases to fail when Expert Parallelism (EP) was enabled. This PR removes the global environment variable configuration and implements the required compatibility handling directly in the code. This avoids affecting unrelated execution paths while preserving the intended behavior. ## Does this PR introduce any user-facing change? No. Users no longer need to configure the compatibility environment variable manually. The compatibility handling is now applied internally where needed. Existing usage and configuration remain unchanged, including scenarios with EP enabled. No documentation changes are required. ## How was this patch tested? The patch was tested by: * Running the affected test cases with EP enabled and verifying that they pass. * Verifying the relevant compatibility behavior after removing the global environment variable. * Running tests with EP disabled to ensure existing behavior is preserved. * Confirming that no global environment variable configuration is required. The code-based compatibility handling only applies to the relevant execution path and does not globally change operator behavior. See merge request: Ascend/MindSpeed-LLM!5009 | 2 天前 | |
feat(mcore): add GLM-5.2 A5 POC scripts Co-authored-by: LinShua<707894133@qq.com> # message auto-generated for no-merge-commit merge: !5013 merge master_glm52_A5_sh into master feat(mcore): add GLM-5.2 A5 POC scripts Created-by: LinShua Commit-by: LinShua Merged-by: ascend-robot Description: Fixes #1825 ## What this PR does / why we need it? 1. Replaces the existing GLM-5.2 32B A5 POC entry with a 51B 4K FP8 configuration using TP1/PP2/EP4 parallelism, Transformer Engine MXFP8, index-topk=1024, fixed routing, the pipeline layout, and a matching log name. 2. Adds a GLM-5.2 210B 4K A5 FP8 POC entry with an eight-node TP1/PP2/EP32 configuration, Transformer Engine MXFP8 options, and portable data/tokenizer/checkpoint path placeholders. 3. Aligns --lr-warmup-iters to 500 in the GLM-5.2 744B A3 example, 744B A3 POC, and 210B A3 POC scripts. 4. This change is limited to five shell launch configurations and does not modify training runtime code or public APIs. ## Does this PR introduce any user-facing change? Yes. Users can run the new tests/poc/glm52/pretrain_glm52_210b_4k_A5_fp8_ptd.sh entry and the renamed tests/poc/glm52/pretrain_glm52_51b_4k_A5_ptd.sh entry. The 51B entry now uses the A5 MXFP8 and TP1/PP2/EP4 configuration, and the related A3 scripts now default to 500 learning-rate warmup iterations. ## How was this patch tested? 1. bash -n passed for all five changed shell scripts. 2. Changed-file pre-commit passed trailing-whitespace, end-of-file, added-large-file, merge-conflict, private-key, codespell, and typos checks. YAML, JSON, Python, and C/C++ hooks had no matching files. The repository-local gitleaks-offline-scan hook could not run because the required ./gitleaks executable is not present in the repository or local environment. 3. git diff --cached --check passed before amend. 4. The remote Git hook passed during git push --force-with-lease. 5. Functional, accuracy, performance, and end-to-end NPU training tests were not run. See merge request: Ascend/MindSpeed-LLM!5013 | 1 天前 | |
docs(pytorch): rename install_guide and updated the latest link Co-authored-by: LQ1206<liuqian164@h-partners.com> # message auto-generated for no-merge-commit merge: !4910 merge master into master docs(pytorch): rename install_guide and updated the latest link Created-by: LQ1206 Commit-by: LQ1206 Merged-by: ascend-robot Description: ## What this PR does / why we need it? rename install_guide and updated the latest link ## Does this PR introduce any user-facing change? rename install_guide and updated the latest link ## How was this patch tested? No test See merge request: Ascend/MindSpeed-LLM!4910 | 30 天前 | |
fix(pytorch):add ckpt-format argument to scripts Co-authored-by: z__y<z4t155664@163.com> # message auto-generated for no-merge-commit merge: !4371 merge add_ckpt_torch_dist_argument_for_shells into master fix(pytorch):add ckpt-format argument to scripts Created-by: z__y Commit-by: z__y Merged-by: ascend-robot Description: ## What this PR does / why we need it? This PR explicitly adds ckpt-format torch to all repository scripts to support the asynchronous checkpoint saving feature. ## Does this PR introduce any user-facing change? No. This change only adjusts internal script parameters to maintain existing behavior. There are no user-facing API or usage changes. ## How was this patch tested? Tests confirm that asynchronous checkpoint saving works correctly and that the original torch format checkpoint behavior is preserved. See merge request: Ascend/MindSpeed-LLM!4371 | 5 个月前 |