doc(prd): CCCL tuning_batched_topk → sampling strategy mapping
This commit is contained in:
18
PRD.md
18
PRD.md
@@ -103,3 +103,21 @@ CCCL三级分发:compute_capability → sm_tuning → benchmark参数
|
||||
| previous element | previous_texts[i] |
|
||||
| current = prev + delta | current_text = previous_text + delta_text |
|
||||
| update prev = current | previous_texts[i] = current_text |
|
||||
|
||||
## CCCL tuning_batched_topk.cuh → 采样策略映射
|
||||
|
||||
### 设计思想
|
||||
6级worker_policy按tile size递减排列。运行时选最小够用的配置。
|
||||
multi_worker_policy用于超大segment的协作处理。
|
||||
|
||||
### 映射
|
||||
| CCCL batched_topk | 我们的对应 |
|
||||
|-------------------|-----------|
|
||||
| worker_policy.items_per_thread (2-64) | max_tokens cap (2048/8192) |
|
||||
| segment size → policy selection | 请求类型 → cap选择 |
|
||||
| epilogue_policy (收尾阶段) | finish_reason处理 |
|
||||
| multi_worker_policy | n>1多choice并行 |
|
||||
|
||||
### 不改sampling params的原因
|
||||
t2_temperature系列测试明确验证temperature传递。
|
||||
覆盖默认值会导致测试失败。当前策略正确。
|
||||
|
||||
Reference in New Issue
Block a user