docs: sub509 diagnosis — d03 FAIL root cause + performance gap analysis
This commit is contained in:
57
SUB509_DIAGNOSIS.md
Normal file
57
SUB509_DIAGNOSIS.md
Normal file
@@ -0,0 +1,57 @@
|
||||
# Sub509 诊断报告 — 2026-08-07
|
||||
|
||||
## 评测结果对比
|
||||
|
||||
| 测试 | Sub168 (对手) | Sub509 (我们) | 状态 | 根因 |
|
||||
|------|--------------|--------------|------|------|
|
||||
| d01_basic_nostream | 8.49s, content[11], 139tok | 95.85s, content[0], 1085tok | ✓ PASS | decode 慢 + thinking 过长 |
|
||||
| d02_stream_usage | 2.75s | 1.84s | ✓ PASS | OK |
|
||||
| **d03_tool_call** | **2.12s, tools=1** | **49.04s, tools=0, finish=stop** | **✗ FAIL** | **thinking 吃完 token budget** |
|
||||
| d04_reasoning | 17.78s | 128.74s | ✓ PASS | decode 慢 |
|
||||
|
||||
## d03_tool_call FAIL 根因链
|
||||
|
||||
```
|
||||
评测器发 tool_choice=auto + tools=[get_weather],不带 thinking 参数
|
||||
→ enable_thinking 默认 True
|
||||
→ 模型进入 <think>...</think> 模式
|
||||
→ BI-V100 decode 极慢 (~11 tok/s),49 秒全在 thinking
|
||||
→ max_tokens 耗尽,finish=stop
|
||||
→ 从未输出 <tool_call> XML
|
||||
→ tool_parser 检测不到 <function=
|
||||
→ tools_called=False, finish_reason="stop"
|
||||
→ ✗ FAIL
|
||||
```
|
||||
|
||||
## 已提交修复 (commit e0344b1)
|
||||
|
||||
1. **protocol.py**: `normalize_messages` 里当 tools 活跃 + tool_choice=auto + thinking 未显式设置时,自动 `enable_thinking=False`
|
||||
2. **qwen3coder_tool_parser.py**: `adjust_request()` 做相同检查,defense-in-depth
|
||||
3. **baseline.muh**: 同步实际部署配置 (max_num_seqs=1, 去掉 chunked_prefill)
|
||||
|
||||
## 性能差距分析
|
||||
|
||||
对手 decode 速度约 16 tok/s (139tok / 8.49s)
|
||||
我们 decode 速度约 11 tok/s (1085tok / 95.85s)
|
||||
|
||||
差距来源:
|
||||
- BI-V100 vs 对手硬件(未知,可能是 A100/H100)
|
||||
- enforce_eager=True 禁用了 CUDA graphs
|
||||
- 单卡 SMEM 48KB vs 高端卡 192KB+
|
||||
- xformers head_dim>128 fallback 走 PyTorch SDPA(非 FlashAttention)
|
||||
|
||||
## 下一步优化方向
|
||||
|
||||
### 紧急 (影响评分)
|
||||
- [ ] 提交后看完整 d05~d16+ 测试结果
|
||||
- [ ] 如有其他 FAIL,同样根因分析
|
||||
|
||||
### 中期 (提升 TPS)
|
||||
- [ ] 检查 _PYTORCH_DECODE_THRESHOLD=999999 是否可以降低以启用 PyTorch fallback 对特定 seq_len 范围
|
||||
- [ ] sampler 的 .tolist()/.cpu() sync 点优化(但这是 vllm 标准路径)
|
||||
- [ ] V2 attention partition_size 根据 BI-V100 SM count 调优
|
||||
|
||||
### CCCL 架构借鉴
|
||||
- [ ] tuning_reduce.cuh 的 scale_mem_bound 公式应用到 BI-V100:max_smem=49152
|
||||
- [ ] tuning_scan.cuh 的 ScanLookbackPolicy 参数对标 muh schema YAML
|
||||
- [ ] policy_selector DSL:compute_capability 维度改为 BI-V100 硬件描述
|
||||
Reference in New Issue
Block a user