106 lines
5.9 KiB
Markdown
106 lines
5.9 KiB
Markdown
# PROJECT_SUMMARY — project_6
|
||
|
||
## 项目背景
|
||
天垓100 (BI-V100) 推理引擎竞赛，在 4×BI-V100 上运行 Qwen3.5-27B 推理服务。
|
||
竞赛目标：Token吞吐加权值 ≥ 8000（Output TPS × 83% + Input TPS × 14% + Cache TPS × 3%）
|
||
|
||
## 技术栈
|
||
- Base image: bi100-3.2.3-x86-ubuntu20.04-py3.10-poc-llm-infer:v1.2.3
|
||
- vLLM 0.6.3 (base) + serving层patch
|
||
- ixformer (CoreX SDK, 含 flash_attn / paged_attention / silu_and_mul 等)
|
||
- Tensor Parallel = 4, enforce_eager=True
|
||
|
||
## 文件结构
|
||
|
||
```
|
||
project_6/
|
||
├── PRD.md # 竞赛需求 + CCCL→base映射
|
||
├── SYSTEM_DESIGN.md # 架构设计: Docker/Build/Runtime/GDN dispatch
|
||
├── Dockerfile # Docker构建
|
||
├── computility-run.yaml # vLLM启动参数
|
||
├── qwen3_6_scripts/ # serving层 + model patches (部署到vllm)
|
||
│ ├── qwen3_5.py (2040行) 模型代码: GDN + MoE + Attention
|
||
│ ├── serving_chat.py OpenAI API处理核心
|
||
│ ├── protocol.py 请求/响应模型
|
||
│ ├── api_server.py FastAPI入口
|
||
│ ├── patch_ops.sh 部署脚本 (全部patch的安装器)
|
||
│ ├── flash_qla_sm70/ GDN CUDA kernel (gdn_forward.cu 1919行)
|
||
│ └── ... 其他patches
|
||
├── ex_engine/ # EX引擎: 算法因子置换层
|
||
│ ├── csrc/
|
||
│ │ ├── ix_full_bridge.cpp (331行) pybind11桥接→ixformer::infer 14个C++函数
|
||
│ │ ├── ix_moe_bridge.cpp (258行) MoE-only子集桥接
|
||
│ │ └── moe_topk_softmax_v3.cu (148行) 独立CUDA topk kernel
|
||
│ ├── python/
|
||
│ │ ├── corex_moe.py (196行) MoE分发: ix_bridge→ixformer::infer 7步pipeline
|
||
│ │ ├── corex_gdn.py (217行) GDN分发: chunked delta rule + decode
|
||
│ │ ├── corex_fa2.py (228行) FA2分发: packed/paged/chunked三模式
|
||
│ │ ├── ix_bridge.py (162行) ix_full_bridge.so加载器
|
||
│ │ └── moe_topk.py CUDA topk Python wrapper
|
||
│ ├── build.sh 编译脚本 (corex clang/16)
|
||
│ └── include/ C++ headers
|
||
├── cccl_upstream/ (8900文件) NVIDIA CCCL strategic subset
|
||
│ ├── cub/ tuning headers + benchmarks + tests
|
||
│ ├── thrust/ examples + tests
|
||
│ └── libcudacxx/ C++ STL headers
|
||
├── muh/ muh工具链: BI-V100 tuning parameter生成
|
||
│ ├── include/muh/tuning/ 27个BI-V100 policy_selector headers
|
||
│ └── gen_patch.py C++ header → vllm unified diff
|
||
├── upstream_ref/ 上游参考代码
|
||
│ ├── ds_vllm/ ds-vllm (vllm fork, 含topk_softmax_kernels.cu)
|
||
│ └── xllm/ xllm (ILU backend: kernels/ilu + layers/ilu)
|
||
├── vllm/ vllm源码副本 (参考用)
|
||
└── docs/ 分析文档
|
||
```
|
||
|
||
## 关键文件说明
|
||
|
||
### ex_engine/csrc/ix_full_bridge.cpp
|
||
- `ix_topk_softmax()` → `ixformer::infer::topk_softmax`
|
||
- `ix_moe_gen_idx()` → `ixformer::infer::moe_compute_token_index_api`
|
||
- `ix_moe_expand_input()` → `ixformer::infer::moe_expand_input`
|
||
- `ix_group_gemm()` → `ixformer::infer::moe_w16a16_group_gemm`
|
||
- `ix_silu_and_mul()` → `ixformer::infer::silu_and_mul`
|
||
- `ix_moe_combine_result()` → `ixformer::infer::moe_output_reduce_sum`
|
||
- `ix_fused_moe_forward()` — 以上6步组合, 一次C++调用完成整个MoE
|
||
- `ix_paged_attention()` → `ixformer::infer::xllm_paged_attention`
|
||
- `ix_flash_attn_prefill()` → `ixformer::infer::ixinfer_flash_attn_unpad_with_block_tables`
|
||
- `ix_rms_norm()` / `ix_fused_add_rms_norm()` / `ix_rotary_embedding()` / `ix_reshape_and_cache()`
|
||
|
||
### ex_engine/python/corex_moe.py
|
||
- `moe_forward()` — 3级分发: ix_bridge全C++ → ix_bridge逐步 → Python loop
|
||
- `topk_softmax()` — ix_bridge优先, fallback到Python softmax+topk
|
||
- `moe_prefill()` / `moe_decode()` — 日志匹配comp 168格式
|
||
|
||
### qwen3_6_scripts/qwen3_5.py
|
||
- `GatedDeltaNet.forward()` — GDN层: corex_gdn dispatch
|
||
- `Qwen3_5MoE.forward()` — MoE层: Tier 0-3分发 (ix_fused_moe → ix_bridge → corex_moe → PyTorch)
|
||
|
||
## 当前状态
|
||
- 360+ commits
|
||
- 38 GitHub issues (open) + 72 draft issues (待转真issue)
|
||
- ix_full_bridge.cpp 已写完14个ixformer::infer函数桥接
|
||
- corex_moe/corex_gdn/corex_fa2 已重写, 使用真实ixformer::infer dispatch chain
|
||
- 需要真机编译 ix_full_bridge.cpp → .so 并验证MoE走C++ pipeline
|
||
|
||
## 本次任务完成内容
|
||
重写3个dlopen模块(corex_moe.py, corex_gdn.py, corex_fa2.py):
|
||
- corex_moe.py: 接入ix_bridge→ixformer::infer 7步MoE pipeline, 移除独立CUDA topk依赖
|
||
- corex_gdn.py: gate clamp[-5,0] + state clamp±100 稳定性修复
|
||
- corex_fa2.py: 3级tiered dispatch (ix_bridge C++ → ixformer Python → V1 fallback)
|
||
- 分析comp 168 docker日志确认真机dlopen调用链条
|
||
|
||
## 历史任务摘要
|
||
- CCCL upstream导入(8900文件) + 27/27 muh tuning headers + CCCL→vllm pattern mapping
|
||
- ix_full_bridge.cpp 14函数桥接 + ix_moe_bridge.cpp MoE子集 + moe_topk_softmax_v3.cu
|
||
- GDN dtype guard + NaN clamp修复 + corex_gdn/corex_moe初始版本
|
||
- serving层部署(protocol/serving_chat/api_server等) + Sub508/509功能修复
|
||
- 38 GitHub issues创建 + PRD/SYSTEM_DESIGN文档
|
||
|
||
## 遗留问题/下次继续
|
||
1. **真机编译ix_full_bridge.cpp** — 需要在Docker中JIT编译, 验证MoE走Tier 0 (C++ 7步)
|
||
2. **72个draft issues转真issue** — 内容已写好, 需要GitHub API批量关联到repo
|
||
3. **MoE Python loop性能** — 如果ix_bridge编译失败, Tier 2的Python expert loop是性能瓶颈(64 experts × 每token)
|
||
4. **GDN prefill精度** — FlashQLA .so在BI-V100上编译通过但abs_mean=inf, 需要fp32 accumulation fix
|
||
5. **benchmark实测** — Sub168基准 TPS=11.86, 需要在新dispatch chain下重测
|