| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
feat: profiling-based empirical performance model with CSV data source Co-authored-by: Horacehxw<horacehxw@gmail.com> # message auto-generated for no-merge-commit merge: !123 merge pr/perf-db-a into develop feat: profiling-based empirical performance model with CSV data source Created-by: Horacehxw Commit-by: Horacehxw Merged-by: ascend-robot Description: **PR Type / PR类型** - [x] Feature(功能新增) - [ ] Bugfix(Bug 修复) - [ ] Docs(文档更新) - [ ] CI/CD(持续集成/持续部署) - [x] Refactor(代码重构) - [ ] Perf(性能优化) - [x] Test-Cases(测试用例更新) - [ ] Other(其他) ## 🔍 Motivation / 变更动机 TensorCast 现有的 Roofline 解析模型( AnalyticPerformanceModel)对昇腾 NPU 的性能预测精度有限:融合算子(SwiGlu、AddRmsNorm、DispatchFFNCombine)无法建模,HCCL 集合通信与理论带宽差距显著,FRACTAL_NZ 格式等硬件特性无法通过 Roofline 捕获。 本 PR 实现了基于真实 NPU Profiling 数据的**实测算子性能估算系统**,将 kernel 实测耗时接入 TensorCast 仿真框架。 **与 PR#96 的关系**:PR#96 已合入 develop,定义了 DataSourcePerformanceModel 接口骨架(stub)和 CLI 集成。本 PR 提供完整的功能实现:CSV 查询引擎(9 种 TC-vs-NPU shape matching 规则)、op_mapping 映射(60+ 算子)、插值、M1-M6 指标体系、以及 DFC/FlashComm 编译 Pass。接口完全兼容。 > 📌 配套的离线数据采集工具链将在后续 PR 中提交(tools/perf_data_collection/,与本 PR 无代码依赖)。 ------ ## 📝 Modification / 修改内容 ### 1. Profiling Data Source 核心实现(替换 PR#96 stub) | 文件 | 说明 | |------|------| | profiling_database/profiling_data_source.py (+1,885) | ProfilingDataSource:op_mapping.yaml 驱动的 CSV 查询引擎,支持 9 种 TC-vs-NPU shape 差异处理(batch dim stripping、seq padding、FRACTAL_NZ、ND transpose、SwiGlu concat、RoPE layout/kernel、composite 分解、flatten batch) | | profiling_database/interpolating_data_source.py (+702) | InterpolatingDataSource:nearest-neighbor + 线性插值包装器 | | profiling_database/data_source.py (修改) | DataSourcePerformanceModel ABC 扩展(新增 EXTRAPOLATED enum、details 字段) | ### 2. EmpiricalPerformanceModel 增强 (+436) 在 PR#96 基础上增加 **M1-M6 指标追踪**: - M1-M4:覆盖率指标(raw count → fused → compute-only → per-shape) - M5:延迟加权覆盖率 - M6 input:empirical hit total(用于离线 E2E ratio 计算) - log_stats():结构化 HIT/MISS 日志 - export_hit_miss_report():JSON 格式指标导出 ### 3. 编译 Passes (+875) | Pass | 说明 | |------|------| | dispatch_ffn_combine_pass.py | DispatchFFNCombine 超级融合(init_routing_v2 + GroupedMatmul + unpermute_tokens → 单 op),支持 5 种量化变体 | | flashcomm_v1_pass.py | FlashComm V1 图重写(matmul_all_reduce → 通信隐藏),对标 vLLM-ascend ENABLE_FLASHCOMM1=1 | ### 4. op_mapping.yaml(3 个版本,共 ~3,600 行) | 版本 | 算子数 | |------|:------:| | vllm0.13.0_torch2.8.0_cann8.3 | ~45 | | vllm0.15.0_torch2.9.0_cann8.5 | ~55 | | vllm0.18.0_torch2.9.0_cann8.5 | ~60 | ### 5. CSV Profiling Data(~250 files,Git LFS) ATLAS 800 A3 752T 128G 设备数据:HCCL 通信基准 + 3 个 vLLM 版本的 kernel 数据 + 微基准补充数据。 ### 6. 集成改动 | 文件 | 改动 | |------|------| | model_runner.py | profiling 模式集成(perf_models[] + log_stats + ProfilingDataSource 创建) | | user_config.py | --profiling-database 参数 | | scripts/text_generate.py | --export-metrics CLI + FlashComm 配置 | | ops/fused_moe.py | 新增 dispatch_ffn_combine op | | compile_backend.py | 注册 DFC + FlashComm passes | ------ ## 📐 Associated Test Results / 关联测试结果 ### 单元测试 $ pytest tests/perf_database/ -q 266 passed, 3 warnings in 1.94s $ pytest tests/test_tensor_cast/test_empirical.py tests/test_tensor_cast/test_dfc_pass.py -q 8 passed, 1 skipped in 120.75s $ lintrunner -a ok No lint issues. ### 功能验证 bash # Analytic 模式(行为不变) $ python -m tensor_cast.scripts.text_generate Qwen/Qwen3-32B \ --num-queries 2 --query-length 3500 --device TEST_DEVICE → [analytic] Execution time: 1.744s, TPS/Device: 4013 token/s ✅ # Profiling 模式(新功能) $ python -m tensor_cast.scripts.text_generate Qwen/Qwen3-32B \ --num-queries 1 --query-length 4112 --word-embedding-tp row \ --device ATLAS_800_A3_752T_128G_DIE --world-size 16 --tp-size 16 \ --quantize-linear-action DISABLED \ --performance-model profiling --compile \ --profiling-database tensor_cast/performance_model/profiling_database/data/ATLAS_800_A3_752T_128G_DIE/vllm_ascend/vllm0.18.0_torch2.9.0_cann8.5 → [empirical] Execution time: 0.156s, TPS/Device: 1651 token/s ✅ ### M1-M5 指标 | 场景 | M3 (计算算子 HR) | M5 (延迟覆盖) | |------|:---------------:|:------------:| | Qwen3-32B Prefill (BF16) | **61.5%** ✅ (>50%) | **89.0%** ✅ (>80%) | | Qwen3-32B Decode (BF16) | 38.5% | **80.1%** ✅ (>80%) | | DeepSeek-V3 Prefill (W8A8) | **52.6%** ✅ (>50%) | 68.9% | | DeepSeek-V3 Decode (W8A8) | 15.8% | 54.3% | ------ ## 🌟 Use cases (Optional) / 使用案例(可选) bash # 1. 使用实测数据替代 Roofline 估算 python -m tensor_cast.scripts.text_generate <model_id> \ --performance-model profiling --compile \ --profiling-database <path_to_data_dir> # 2. 导出 M1-M5 指标 JSON(用于离线 M6 计算) python -m tensor_cast.scripts.text_generate <model_id> \ --performance-model profiling --compile \ --profiling-database <path_to_data_dir> \ --export-metrics results/metrics.json # 3. 同时运行 analytic + profiling 对比 python -m tensor_cast.scripts.text_generate <model_id> \ --performance-model analytic --performance-model profiling --compile \ --profiling-database <path_to_data_dir> ------ ## ✅ Checklist / 检查列表 **Before PR**: - [x] [Linting tools](https://gitcode.com/Ascend/msmodeling/blob/develop/tensor_cast/README.md#coding-style) are used to fix the potential lint issues. - [x] Bug fixes are fully covered by unit tests, the case that causes the bug should be added in the unit tests. - [x] The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness. - [ ] All relevant documentation (API docs, docstrings, example tutorials) has been updated to reflect these changes. - [x] Please ensure code files contain no Chinese comments. ``` See merge request: Ascend/msmodeling!123 | 1 个月前 | |
feat: profiling-based empirical performance model with CSV data source Co-authored-by: Horacehxw<horacehxw@gmail.com> # message auto-generated for no-merge-commit merge: !123 merge pr/perf-db-a into develop feat: profiling-based empirical performance model with CSV data source Created-by: Horacehxw Commit-by: Horacehxw Merged-by: ascend-robot Description: **PR Type / PR类型** - [x] Feature(功能新增) - [ ] Bugfix(Bug 修复) - [ ] Docs(文档更新) - [ ] CI/CD(持续集成/持续部署) - [x] Refactor(代码重构) - [ ] Perf(性能优化) - [x] Test-Cases(测试用例更新) - [ ] Other(其他) ## 🔍 Motivation / 变更动机 TensorCast 现有的 Roofline 解析模型( AnalyticPerformanceModel)对昇腾 NPU 的性能预测精度有限:融合算子(SwiGlu、AddRmsNorm、DispatchFFNCombine)无法建模,HCCL 集合通信与理论带宽差距显著,FRACTAL_NZ 格式等硬件特性无法通过 Roofline 捕获。 本 PR 实现了基于真实 NPU Profiling 数据的**实测算子性能估算系统**,将 kernel 实测耗时接入 TensorCast 仿真框架。 **与 PR#96 的关系**:PR#96 已合入 develop,定义了 DataSourcePerformanceModel 接口骨架(stub)和 CLI 集成。本 PR 提供完整的功能实现:CSV 查询引擎(9 种 TC-vs-NPU shape matching 规则)、op_mapping 映射(60+ 算子)、插值、M1-M6 指标体系、以及 DFC/FlashComm 编译 Pass。接口完全兼容。 > 📌 配套的离线数据采集工具链将在后续 PR 中提交(tools/perf_data_collection/,与本 PR 无代码依赖)。 ------ ## 📝 Modification / 修改内容 ### 1. Profiling Data Source 核心实现(替换 PR#96 stub) | 文件 | 说明 | |------|------| | profiling_database/profiling_data_source.py (+1,885) | ProfilingDataSource:op_mapping.yaml 驱动的 CSV 查询引擎,支持 9 种 TC-vs-NPU shape 差异处理(batch dim stripping、seq padding、FRACTAL_NZ、ND transpose、SwiGlu concat、RoPE layout/kernel、composite 分解、flatten batch) | | profiling_database/interpolating_data_source.py (+702) | InterpolatingDataSource:nearest-neighbor + 线性插值包装器 | | profiling_database/data_source.py (修改) | DataSourcePerformanceModel ABC 扩展(新增 EXTRAPOLATED enum、details 字段) | ### 2. EmpiricalPerformanceModel 增强 (+436) 在 PR#96 基础上增加 **M1-M6 指标追踪**: - M1-M4:覆盖率指标(raw count → fused → compute-only → per-shape) - M5:延迟加权覆盖率 - M6 input:empirical hit total(用于离线 E2E ratio 计算) - log_stats():结构化 HIT/MISS 日志 - export_hit_miss_report():JSON 格式指标导出 ### 3. 编译 Passes (+875) | Pass | 说明 | |------|------| | dispatch_ffn_combine_pass.py | DispatchFFNCombine 超级融合(init_routing_v2 + GroupedMatmul + unpermute_tokens → 单 op),支持 5 种量化变体 | | flashcomm_v1_pass.py | FlashComm V1 图重写(matmul_all_reduce → 通信隐藏),对标 vLLM-ascend ENABLE_FLASHCOMM1=1 | ### 4. op_mapping.yaml(3 个版本,共 ~3,600 行) | 版本 | 算子数 | |------|:------:| | vllm0.13.0_torch2.8.0_cann8.3 | ~45 | | vllm0.15.0_torch2.9.0_cann8.5 | ~55 | | vllm0.18.0_torch2.9.0_cann8.5 | ~60 | ### 5. CSV Profiling Data(~250 files,Git LFS) ATLAS 800 A3 752T 128G 设备数据:HCCL 通信基准 + 3 个 vLLM 版本的 kernel 数据 + 微基准补充数据。 ### 6. 集成改动 | 文件 | 改动 | |------|------| | model_runner.py | profiling 模式集成(perf_models[] + log_stats + ProfilingDataSource 创建) | | user_config.py | --profiling-database 参数 | | scripts/text_generate.py | --export-metrics CLI + FlashComm 配置 | | ops/fused_moe.py | 新增 dispatch_ffn_combine op | | compile_backend.py | 注册 DFC + FlashComm passes | ------ ## 📐 Associated Test Results / 关联测试结果 ### 单元测试 $ pytest tests/perf_database/ -q 266 passed, 3 warnings in 1.94s $ pytest tests/test_tensor_cast/test_empirical.py tests/test_tensor_cast/test_dfc_pass.py -q 8 passed, 1 skipped in 120.75s $ lintrunner -a ok No lint issues. ### 功能验证 bash # Analytic 模式(行为不变) $ python -m tensor_cast.scripts.text_generate Qwen/Qwen3-32B \ --num-queries 2 --query-length 3500 --device TEST_DEVICE → [analytic] Execution time: 1.744s, TPS/Device: 4013 token/s ✅ # Profiling 模式(新功能) $ python -m tensor_cast.scripts.text_generate Qwen/Qwen3-32B \ --num-queries 1 --query-length 4112 --word-embedding-tp row \ --device ATLAS_800_A3_752T_128G_DIE --world-size 16 --tp-size 16 \ --quantize-linear-action DISABLED \ --performance-model profiling --compile \ --profiling-database tensor_cast/performance_model/profiling_database/data/ATLAS_800_A3_752T_128G_DIE/vllm_ascend/vllm0.18.0_torch2.9.0_cann8.5 → [empirical] Execution time: 0.156s, TPS/Device: 1651 token/s ✅ ### M1-M5 指标 | 场景 | M3 (计算算子 HR) | M5 (延迟覆盖) | |------|:---------------:|:------------:| | Qwen3-32B Prefill (BF16) | **61.5%** ✅ (>50%) | **89.0%** ✅ (>80%) | | Qwen3-32B Decode (BF16) | 38.5% | **80.1%** ✅ (>80%) | | DeepSeek-V3 Prefill (W8A8) | **52.6%** ✅ (>50%) | 68.9% | | DeepSeek-V3 Decode (W8A8) | 15.8% | 54.3% | ------ ## 🌟 Use cases (Optional) / 使用案例(可选) bash # 1. 使用实测数据替代 Roofline 估算 python -m tensor_cast.scripts.text_generate <model_id> \ --performance-model profiling --compile \ --profiling-database <path_to_data_dir> # 2. 导出 M1-M5 指标 JSON(用于离线 M6 计算) python -m tensor_cast.scripts.text_generate <model_id> \ --performance-model profiling --compile \ --profiling-database <path_to_data_dir> \ --export-metrics results/metrics.json # 3. 同时运行 analytic + profiling 对比 python -m tensor_cast.scripts.text_generate <model_id> \ --performance-model analytic --performance-model profiling --compile \ --profiling-database <path_to_data_dir> ------ ## ✅ Checklist / 检查列表 **Before PR**: - [x] [Linting tools](https://gitcode.com/Ascend/msmodeling/blob/develop/tensor_cast/README.md#coding-style) are used to fix the potential lint issues. - [x] Bug fixes are fully covered by unit tests, the case that causes the bug should be added in the unit tests. - [x] The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness. - [ ] All relevant documentation (API docs, docstrings, example tutorials) has been updated to reflect these changes. - [x] Please ensure code files contain no Chinese comments. ``` See merge request: Ascend/msmodeling!123 | 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 |