IX2-0455
① SA Source
- Source: 開啟完整 SA 文章
- Section:
Disaggregated Prefill - Line hint:
670
Context Before
Source: DistServe ↗
Evidence
different memory allocation ratios
Context After
The key cost of disaggregation is KV cache transfer. After prefill completes, the full KV cache for that request must be transmitted from the prefill node to the decode node before the first decode token can be generated. For a model like DeepSeek R1 with 61 layers and FP8 KV cache, an 8192-token prefill produces roughly 500MB of KV data that must cross the network, adding directly to TTFT. This transfer is performed over RDMA (typically RoCE or InfiniBand) using zero-copy GPU-to-GPU data movement without CPU involvement. Libraries like NIXL (NVIDIA Inference Transfer Library) abstract the data movement layer behind a unified asynchronous API with pluggable backends for UCX, GPUDirect Storage, and other transports. This decouples the inference engine from any specific transfer protocol and enables disaggregation across heterogeneous hardware where prefill and decode instances may span different device types or interconnects.
② Atomic Claim
在 disaggregated serving 中，prefill 與 decode 可以使用不同 memory-allocation ratios。
- Epistemic Mode:
ASSERTED - Mapping Status:
COMPLETE
③ Semantic Frame
{
"frame_type": "NARY_RELATION",
"participants": [
{
"node": {
"id": "04_knowledge_base/Disaggregated serving",
"label": "disaggregated serving"
},
"role": "user_or_subject"
},
{
"node": {
"id": "04_knowledge_base/Prefill",
"label": "prefill"
},
"role": "used_entity"
},
{
"node": {
"id": "04_knowledge_base/Decode",
"label": "decode"
},
"role": "used_entity"
}
],
"qualifiers": {
"condition_text": null,
"numeric_mentions": [],
"temporal_mentions": []
},
"relation_type": "USES"
}④ Canonical Entity Mapping
| Role | Surface Label | Canonical Target |
|---|---|---|
| user_or_subject | disaggregated serving | 04_knowledge_base/Disaggregated serving |
| used_entity | prefill | Prefill |
| used_entity | decode | Decode |
⑤ Human Review
請在 Properties 逐項確認：
- 原文 → Atomic Claim 是否忠實
- Atomic Claim → Semantic Frame 是否忠實
- Canonical Entity mapping 是否正確
- Epistemic mode 是否保留原文語氣
- 最後選擇
review_action
Review state
Markdown 內文不是正式 approval。只有 Apply bridge 寫入的 Decision Ledger event 才是正式決策。