| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
test: move ST cases to pipeline and remove duplicate UT cases Co-authored-by: Rostellaria<xiaoshaoan@huawei.com> # message auto-generated for no-merge-commit merge: !4867 merge master into master test: move ST cases to pipeline and remove duplicate UT cases Created-by: Rostellaria Commit-by: Rostellaria Merged-by: ascend-robot Description: [#1678](https://gitcode.com/Ascend/MindSpeed-LLM/issues/1678) ## What this PR does / why we need it? This PR migrates ST cases ( deepseek_v2_mcore_tp1_pp1_ep8 , mamba2_8b_tp4_pp1_cp2_recompute_4k_ptd , mamba2_2.7b_tp1_pp1 , deepseek_500b_tp1_pp2_ep2_cp2_overlap to legacy; deepseek32_tp1_pp2_vpp1_ep4 , qwen3_next_80b_4K_A3_ptd , qwen3_next_ascendc-GDN_pack_cp2 to pipeline/st) with their baselines from tests/st to tests/pipeline , and removes UT cases that fully overlap with tests/pipeline/ut (checkpoint dsv2, evaluation llama2, inference llama2/llama3, process_data merge/instruction). README tables are updated accordingly. ## Does this PR introduce any user-facing change? No. It only relocates test files and removes redundant duplicates; no public API, CLI, or user-facing behavior is changed. ## How was this patch tested? Verified moved files are auto-discovered by tests/pipeline/st/test_pipeline_st.py (recursive .sh scan + flat baseline matching), confirmed remaining UT cases have no identical counterpart in pipeline, and validated all Python/JSON syntax and README link paths. See merge request: Ascend/MindSpeed-LLM!4867 | 1 个月前 | |
test: move ST cases to pipeline and remove duplicate UT cases Co-authored-by: Rostellaria<xiaoshaoan@huawei.com> # message auto-generated for no-merge-commit merge: !4867 merge master into master test: move ST cases to pipeline and remove duplicate UT cases Created-by: Rostellaria Commit-by: Rostellaria Merged-by: ascend-robot Description: [#1678](https://gitcode.com/Ascend/MindSpeed-LLM/issues/1678) ## What this PR does / why we need it? This PR migrates ST cases ( deepseek_v2_mcore_tp1_pp1_ep8 , mamba2_8b_tp4_pp1_cp2_recompute_4k_ptd , mamba2_2.7b_tp1_pp1 , deepseek_500b_tp1_pp2_ep2_cp2_overlap to legacy; deepseek32_tp1_pp2_vpp1_ep4 , qwen3_next_80b_4K_A3_ptd , qwen3_next_ascendc-GDN_pack_cp2 to pipeline/st) with their baselines from tests/st to tests/pipeline , and removes UT cases that fully overlap with tests/pipeline/ut (checkpoint dsv2, evaluation llama2, inference llama2/llama3, process_data merge/instruction). README tables are updated accordingly. ## Does this PR introduce any user-facing change? No. It only relocates test files and removes redundant duplicates; no public API, CLI, or user-facing behavior is changed. ## How was this patch tested? Verified moved files are auto-discovered by tests/pipeline/st/test_pipeline_st.py (recursive .sh scan + flat baseline matching), confirmed remaining UT cases have no identical counterpart in pipeline, and validated all Python/JSON syntax and README link paths. See merge request: Ascend/MindSpeed-LLM!4867 | 1 个月前 | |
test: move ST cases to pipeline and remove duplicate UT cases Co-authored-by: Rostellaria<xiaoshaoan@huawei.com> # message auto-generated for no-merge-commit merge: !4867 merge master into master test: move ST cases to pipeline and remove duplicate UT cases Created-by: Rostellaria Commit-by: Rostellaria Merged-by: ascend-robot Description: [#1678](https://gitcode.com/Ascend/MindSpeed-LLM/issues/1678) ## What this PR does / why we need it? This PR migrates ST cases ( deepseek_v2_mcore_tp1_pp1_ep8 , mamba2_8b_tp4_pp1_cp2_recompute_4k_ptd , mamba2_2.7b_tp1_pp1 , deepseek_500b_tp1_pp2_ep2_cp2_overlap to legacy; deepseek32_tp1_pp2_vpp1_ep4 , qwen3_next_80b_4K_A3_ptd , qwen3_next_ascendc-GDN_pack_cp2 to pipeline/st) with their baselines from tests/st to tests/pipeline , and removes UT cases that fully overlap with tests/pipeline/ut (checkpoint dsv2, evaluation llama2, inference llama2/llama3, process_data merge/instruction). README tables are updated accordingly. ## Does this PR introduce any user-facing change? No. It only relocates test files and removes redundant duplicates; no public API, CLI, or user-facing behavior is changed. ## How was this patch tested? Verified moved files are auto-discovered by tests/pipeline/st/test_pipeline_st.py (recursive .sh scan + flat baseline matching), confirmed remaining UT cases have no identical counterpart in pipeline, and validated all Python/JSON syntax and README link paths. See merge request: Ascend/MindSpeed-LLM!4867 | 1 个月前 | |
test: move ST cases to pipeline and remove duplicate UT cases Co-authored-by: Rostellaria<xiaoshaoan@huawei.com> # message auto-generated for no-merge-commit merge: !4867 merge master into master test: move ST cases to pipeline and remove duplicate UT cases Created-by: Rostellaria Commit-by: Rostellaria Merged-by: ascend-robot Description: [#1678](https://gitcode.com/Ascend/MindSpeed-LLM/issues/1678) ## What this PR does / why we need it? This PR migrates ST cases ( deepseek_v2_mcore_tp1_pp1_ep8 , mamba2_8b_tp4_pp1_cp2_recompute_4k_ptd , mamba2_2.7b_tp1_pp1 , deepseek_500b_tp1_pp2_ep2_cp2_overlap to legacy; deepseek32_tp1_pp2_vpp1_ep4 , qwen3_next_80b_4K_A3_ptd , qwen3_next_ascendc-GDN_pack_cp2 to pipeline/st) with their baselines from tests/st to tests/pipeline , and removes UT cases that fully overlap with tests/pipeline/ut (checkpoint dsv2, evaluation llama2, inference llama2/llama3, process_data merge/instruction). README tables are updated accordingly. ## Does this PR introduce any user-facing change? No. It only relocates test files and removes redundant duplicates; no public API, CLI, or user-facing behavior is changed. ## How was this patch tested? Verified moved files are auto-discovered by tests/pipeline/st/test_pipeline_st.py (recursive .sh scan + flat baseline matching), confirmed remaining UT cases have no identical counterpart in pipeline, and validated all Python/JSON syntax and README link paths. See merge request: Ascend/MindSpeed-LLM!4867 | 1 个月前 | |
fix: isolate DSA ShareIndex cache during full recomputation Co-authored-by: LinShua<707894133@qq.com> # message auto-generated for no-merge-commit merge: !4906 merge master_glm52_share_index into master fix: isolate DSA ShareIndex cache during full recomputation Created-by: LinShua Commit-by: LinShua Merged-by: ascend-robot Description: Fixes #1701 ## What this PR does / why we need it? DSA ShareIndex previously stored only the most recently computed TopK tensor in a process-global variable. With multiple share groups, such as L3 -> L4-L6 and L7 -> L8-L10, the later Compute layer could overwrite the source needed by an earlier group during reverse-order activation recomputation. 1. Replaces the single global tensor with a Megatron-style per-forward holder keyed by the global Compute layer: dict[source_layer, topk_indices]. 2. Carries the holder on packed_seq_params, the attention mask, or the RoPE tensor for microbatch isolation, with the model config used only as a final fallback. 3. Resolves each Share layer's exact source from index_topk_pattern or the existing frequency/offset settings, rejects share groups split across PP/VPP boundaries, and validates batch/sequence shape before reuse. 4. Stores tensor references without cloning, so all Share layers in one group reuse the Compute layer's tensor storage while different Compute layers retain separate dictionary entries. 5. Restores the same holder reference in MindSpeed's existing checkpoint argument wrapper after checkpoint backward detaches tensor inputs. This keeps the implementation compatible with reverse-order recomputation without replacing TransformerBlock._checkpointed_forward, ContextVar state, or checkpoint-closure snapshots. 6. Adds focused unit coverage for storage reuse, L3/L7 reverse replay, per-layer overwrite behavior, missing sources, dynamic shape mismatches, attention-mask sequences, and packed/mask-free microbatch isolation. ## Does this PR introduce any user-facing change? No new CLI options are introduced. DSA ShareIndex now reuses TopK indices from the configured source layer during full activation recomputation. Invalid PP/VPP placement, a missing source layer, or a batch/sequence shape mismatch fails with an explicit error instead of silently reusing the latest global cache entry. ## How was this patch tested? 1. Ran repository pre-commit hooks on all four files changed by this PR. Trailing-whitespace, end-of-file, added-large-file, merge-conflict, private-key, Ruff check/format, codespell, Pylint, Bandit, and typos hooks passed. The configured local Gitleaks hook could not run because this checkout does not contain ./gitleaks. 2. Ran Python py_compile successfully on all four changed Python files. 3. Ran git diff --check successfully. 4. The focused pytest file and NPU training scenarios were not executed in this Windows validation environment because it does not provide torch/pytest or an NPU runtime. 5. GitCode push hooks reported PASSED for commit d15fbf19. See merge request: Ascend/MindSpeed-LLM!4906 | 25 天前 | |
test: move ST cases to pipeline and remove duplicate UT cases Co-authored-by: Rostellaria<xiaoshaoan@huawei.com> # message auto-generated for no-merge-commit merge: !4867 merge master into master test: move ST cases to pipeline and remove duplicate UT cases Created-by: Rostellaria Commit-by: Rostellaria Merged-by: ascend-robot Description: [#1678](https://gitcode.com/Ascend/MindSpeed-LLM/issues/1678) ## What this PR does / why we need it? This PR migrates ST cases ( deepseek_v2_mcore_tp1_pp1_ep8 , mamba2_8b_tp4_pp1_cp2_recompute_4k_ptd , mamba2_2.7b_tp1_pp1 , deepseek_500b_tp1_pp2_ep2_cp2_overlap to legacy; deepseek32_tp1_pp2_vpp1_ep4 , qwen3_next_80b_4K_A3_ptd , qwen3_next_ascendc-GDN_pack_cp2 to pipeline/st) with their baselines from tests/st to tests/pipeline , and removes UT cases that fully overlap with tests/pipeline/ut (checkpoint dsv2, evaluation llama2, inference llama2/llama3, process_data merge/instruction). README tables are updated accordingly. ## Does this PR introduce any user-facing change? No. It only relocates test files and removes redundant duplicates; no public API, CLI, or user-facing behavior is changed. ## How was this patch tested? Verified moved files are auto-discovered by tests/pipeline/st/test_pipeline_st.py (recursive .sh scan + flat baseline matching), confirmed remaining UT cases have no identical counterpart in pipeline, and validated all Python/JSON syntax and README link paths. See merge request: Ascend/MindSpeed-LLM!4867 | 1 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 25 天前 | ||
| 1 个月前 |