| 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 | |
fix(pytorch): fix bug of torch.mm Co-authored-by: qyzqyz<quyueze@h-partners.com> # message auto-generated for no-merge-commit merge: !5011 merge master into master fix(pytorch): fix bug of torch.mm Created-by: qyzqyz Commit-by: qyzqyz Merged-by: ascend-robot Description: ## What this PR does / why we need it? This PR fixes the logits computation by transposing head_weight before matrix multiplication. python logits = torch.mm(hidden_states, head_weight.T).float() Without the transpose, the matrix dimensions are incompatible with the layout of the LM head weight. ## Does this PR introduce any user-facing change? No. This is an internal bug fix and does not change any user-facing APIs or configurations. ## How was this patch tested? Verified that logits can be computed successfully and that the output shape is correct. See merge request: Ascend/MindSpeed-LLM!5011 | 15 天前 | |
refactor: make post_norm config-driven, fix fc_type ref, clean qkv_split Co-authored-by: wanggangguo<wanggangguo@huawei.com> # message auto-generated for no-merge-commit merge: !4738 merge refactor into master refactor: make post_norm config-driven, fix fc_type ref, clean qkv_split Created-by: isfrapples Commit-by: wanggangguo Merged-by: ascend-robot Description: ## What this PR does / why we need it? 1. post_norm 由运行时探测改为配置驱动，model_cfg.json 显式声明 2. fc_type 修复为从 load_model 读取，up_down 分支无 gate 3. 移除 qkv_split 冗余配置，ng>1 时始终 per-group 拆分 4. 旧 checkpoint 缺少 post_attn_norm.weight 时容错不崩溃 5. 移除 extra_norm_keys 遍历和 hasattr 死代码 ## Does this PR introduce any user-facing change? 1. 新增 gemma2 配置项 post_norm: true，需同步 model_cfg.json 2. 移除 qwen2-moe、qwen25、gemma2、seed-oss 的 qkv_split 配置 3. 无 API 或 CLI 接口变更，向下兼容旧 checkpoint ## How was this patch tested? 1. 13 个测试用例全部通过，覆盖 6 个模型和 2 个转换方向 2. 每个用例包含完整权重精确对比，gate_proj 参与校验 3. 旧 checkpoint 兼容性验证：gemma2 缺少 post_attn_norm 不崩溃 4. 存量 bug 修复验证：llama2 k_proj 形状错误已修复 fixes#1604 See merge request: Ascend/MindSpeed-LLM!4738 | 1 个月前 |