跳转至

功能矩阵

下表展示了互斥功能以及在 Ascend 硬件上的支持情况,扩展自 vLLM 表格.

所使用的符号含义如下:

  • ✅ = 完全兼容
  • 🟠 = 部分兼容
  • ❌ = 不兼容
  • ❔ = 未知或待定
功能 ACLGraph Full_Decode_Only ACLGraph Piecewise Async Scheduling APC Chunked Prefill Decode Context Parallel CPU Binding DP Disaggregated Prefill Eagle3 EPLB EP KV Cache Pool Lmhead TP MLAPO mm Multistream MoE Shared Expert DP Quantization W4A4 Quantization W4A8 Quantization W8A8 TP Weight nz KVPP (Experimental)
ACLGraph Full_Decode_Only ✅
ACLGraph Piecewise ❌ ✅
Async Scheduling ✅ ✅ ✅
APC ✅ ✅ ✅ ✅
Chunked Prefill ✅ ✅ ✅ ✅ ✅
Decode Context Parallel ✅ ✅ ✅ ✅ ✅ ✅
CPU Binding ✅ ✅ ✅ ✅ ✅ ✅ ✅
DP ✅ ✅ ✅ ✅ ✅ 🟠1 ✅ ✅
Disaggregated Prefill ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
Eagle3 ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
EPLB ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
EP ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
KV Cache Pool ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
Lmhead TP ✅ ✅ ✅ ✅ ✅ ❔ ✅ 🟠2 ✅ ✅ ✅ ✅ ❔ ✅
MLAPO ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ 🟠3 ✅ ✅ ✅ ❔ ✅ ✅
mm ✅ ✅ ✅ ✅ ✅ 🟠 ✅ ✅ ✅ ✅ ✅ ✅ ❌ ✅ ✅ ✅
Multistream MoE ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
Shared Expert DP ✅ ✅ ✅ ✅ ✅ 🟠1 ✅ ✅ ✅ ✅ ✅ ✅ ✅ ❔ ✅ ✅ ❔ ✅
Quantization W4A4 ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ❌ ❔ ✅ ✅ ❔ ❌ ❔ ❔ ✅
Quantization W4A8 ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ❌ ✅ ✅ ✅ ❔ ❌ ✅ ✅ ❔ ✅
Quantization W8A8 ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ❔ ✅ ✅
功能 ACLGraph Full_Decode_Only ACLGraph Piecewise Async Scheduling APC Chunked Prefill Decode Context Parallel CPU Binding DP Disaggregated Prefill Eagle3 EPLB EP KV Cache Pool Lmhead TP MLAPO mm Multistream MoE Shared Expert DP Quantization W4A4 Quantization W4A8 Quantization W8A8 TP Weight nz KVPP (Experimental)
TP ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅
Weight nz ✅ ✅ ✅ ✅ ✅ ❔ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ✅ ❌ 🟠 ✅ ✅ ✅
KVPP (Experimental) ❌ ❌ ✅ ✅ ✅ ❌ ❔ ❔ 🟠 ❌ ❔ ✅ 🟠 ❔ ❔ ❔ ❔ ❔ ❔ ❔ ❔ ✅ ❔ ✅
  • 1 仅 dcp 支持 dp,而 pcp 不支持 dp。
  • 2 Lmhead TP 仅在纯 dp 场景中启用。
  • 3 MLAPO 仅在解码阶段支持。

本页原文 Markdown:在 AtomGit 查看·内容源自开源项目 vl/vllm-ascend