0
代码
代码
Issues
Pull Requests
讨论
Wiki
分析
0
ascend-robotascend-robottest: add UT for processor/quant/autoround_utils and processor/quarot/common
12c0e6d4创建于 13 小时前历史提交
文件最后提交记录最后更新时间
__init__.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_attention_processor.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_autoround_processor.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_autoround_utils.py
test: add UT for processor/quant/autoround_utils and processor/quarot/common Co-authored-by: tanxiangyuu<tanxiangyu2@huawei.com> # message auto-generated for no-merge-commit merge: !511 merge feature/add-ut-quant-quarot into master test: add UT for processor/quant/autoround_utils and processor/quarot/common Created-by: tanxiangyuu Commit-by: tanxiangyuu Merged-by: ascend-robot Description: 覆盖率补充请看：https://gitcode.com/tanxiangyuu/msmodelslim/issues/1 See merge request: Ascend/msmodelslim!51113 小时前
test_autoround_validation.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_fa3_processor.py
【feature】 支持FA3的混合量化配置及保存 Co-authored-by: wangsihao<wangsihao5@h-partners.com> # message auto-generated for no-merge-commit merge: !483 merge master into master 【feature】 支持FA3的混合量化配置及保存 Created-by: wangsihao Commit-by: wangsihao Merged-by: ascend-robot Description: 感谢您贡献的Pull Request！ 在提交之前，请务必阅读 [CONTRIBUTING.md](https://gitcode.com/Ascend/msmodelslim/blob/master/CONTRIBUTING.md)。 ## PR描述 (What this PR does / why we need it?) 支持FA3量化中的不同策略和粒度的混合配置 如：fa_q 动态量化，fa_k/fa_v 静态量化 量化粒度支持FP8/INT8，其中INT8仅支持静态量化 ## 功能验证 (How was this patch tested?) - [_] 功能自验 - [_] 本地自验截图(涉及个人标识符等敏感信息请注意脱敏) - [_] 新增/变更内容是否已新增/适配UT测试用例看护 See merge request: Ascend/msmodelslim!48320 小时前
test_linear_quant_processor.py
feat: LinearQuantProcessor DTS integration Co-authored-by: rookie_hongchuan<hongchuan6@h-partners.com> # message auto-generated for no-merge-commit merge: !395 merge feature/linearquant-dts-v2 into master feat: LinearQuantProcessor DTS integration Created-by: rookie_hongchuan Commit-by: rookie_hongchuan Merged-by: ascend-robot Description: ## 概述 将 LinearQuantProcessor 与 DistributedTaskScheduler (DTS) 集成，在 DP 逐层量化管线中并行化 data-free 权重量化器的校准计算，减少 MoE 模型 W4A8 SSZ 量化的端到端耗时。 ## 动机 Qwen3-Next-80B（48层，每层 MoE 100 个专家）的 W4A8 SSZ per_channel 量化计算量极大——单卡每层约 4 分钟。在 DP 逐层量化管线中，每层的 LinearQuantProcessor 处理 3 个子图（self_attn、mlp.experts、linear_attn），内部包含大量线性子模块。原始的 SSZ forward 调用是串行的，成为量化瓶颈。 关键洞察：SSZ 权重量化器是 **data-free** 的（确定性算法，不需要激活数据），其 forward 计算可以在各 rank 间安全并行，无需校准数据交换。 ## 设计方案 ### 架构 ``` AutoWeightQuantizer（基类，继承 DTSMixin） ├── distributed_sync() 默认 → self.forward(None) # 各 rank 独立重算 └── WeightPerChannelSsz 覆盖 distributed_sync() → _broadcast_quantizer_state: owner 计算，广播到所有 rank ``` ### LinearQuantProcessor 中的 DTS 调度 ``` preprocess() └── _calibrate_shared_data_free_with_dts() ├── 遍历 LinearQuantizer 子模块 ├── 筛选: weight_quantizer.is_data_free() && dist_helper.is_shared(name) └── 提交 DTS 任务 → 4 rank 并行执行 ``` ### Broadcast 协议（SSZ 专用） ``` owner_rank → broadcast_object_list（元数据: shape/dtype） → broadcast_tensor_process_group_safe（3 个张量: scale, offset, storage） 非 owner → load_quantized_from_broadcast_tensors() → 重建 QParam + QStorage ``` ### 修改文件 | 文件 | 变更 | |------|------| | msmodelslim/core/quantizer/base.py | AutoWeightQuantizer 继承 DTSMixin；默认 distributed_sync 触发 forward(None) | | msmodelslim/core/quantizer/impl/ssz.py | WeightPerChannelSsz.distributed_sync 实现 broadcast；添加 load_quantized_from_broadcast_tensors | | msmodelslim/processor/quant/linear.py | _calibrate_shared_data_free_with_dts 方法；DTS 任务提交与调度 | ## 测试结果 ### 环境 | 项目 | 内容 | |------|------| | 模型 | Qwen3-Next-80B-A3B-Instruct | | 量化方式 | W4A8 SSZ per_channel | | 设备 | 4x Ascend 910B2 (64GB HBM) | | 分布式 | DP LayerWise（4 rank）+ DTS 内层并行 | ### 性能 | 指标 | 单卡基线 | 4卡+DTS | 加速比 | |------|---------|---------|--------| | 端到端耗时 | ~2小时 | ~82分钟 | ~1.5x | | SSZ 计算（每层平均） | ~155s 串行 | ~67s 墙钟 | ~2.3x | DTS 效率：T_run / sum(task_exec) ≈ 0.43（4 rank，每层 1536 个任务） ### 一致性验证 ``` mstool cmp baseline/ 4card/ → 所有 tensor abs diff = 0 ✅ ``` 单卡与 4 卡输出对全部 30,000+ 个 tensor（权重、scales、offsets、scale_biases）逐字节一致。 ### 配套工具 新增 tools/npu-affinity-keeper.sh：通过 npu-smi info 监控 NPU 进程，自动恢复 CPU 亲和性，解决 run_in_background 将子进程锁定到单核的问题。 See merge request: Ascend/msmodelslim!3958 天前
test_sign_sgd.py
[test] add processor/quant unit test Co-authored-by: tanxiangyuu<tanxiangyu2@huawei.com> # message auto-generated for no-merge-commit merge: !509 merge dev_2 into master [test] add processor/quant unit test Created-by: tanxiangyuu Commit-by: tanxiangyuu Merged-by: ascend-robot Description: 覆盖率补充请看：https://gitcode.com/tanxiangyuu/msmodelslim/issues/1 See merge request: Ascend/msmodelslim!50913 小时前
test_trainer.py
[test] add processor/quant unit test Co-authored-by: tanxiangyuu<tanxiangyu2@huawei.com> # message auto-generated for no-merge-commit merge: !509 merge dev_2 into master [test] add processor/quant unit test Created-by: tanxiangyuu Commit-by: tanxiangyuu Merged-by: ascend-robot Description: 覆盖率补充请看：https://gitcode.com/tanxiangyuu/msmodelslim/issues/1 See merge request: Ascend/msmodelslim!50913 小时前
test_w4a4_mx_linear_quant_processor.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_w4a8_mx_linear_quant_processor.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_w8a8_mx_linear_quant_processor.py
【msmodelslim】刷新文件头部声明和版权 Co-authored-by: caishengcheng<caishengcheng@huawei.com> # message auto-generated for no-merge-commit merge: !28 merge license into master 【msmodelslim】刷新文件头部声明和版权 Created-by: caishengcheng Commit-by: caishengcheng Merged-by: ascend-robot Description: 【msmodelslim】刷新文件头部声明和版权 See merge request: Ascend/msmodelslim!284 个月前
test_wrapper.py
[test] add processor/quant unit test Co-authored-by: tanxiangyuu<tanxiangyu2@huawei.com> # message auto-generated for no-merge-commit merge: !509 merge dev_2 into master [test] add processor/quant unit test Created-by: tanxiangyuu Commit-by: tanxiangyuu Merged-by: ascend-robot Description: 覆盖率补充请看：https://gitcode.com/tanxiangyuu/msmodelslim/issues/1 See merge request: Ascend/msmodelslim!50913 小时前