Unverified85% confidenceFactExact time
AITER GEMM 和 FlyDSL FusedMoE 内核同时支持 Tensor Parallelism（TP）和 Data Parallelism + Expert Parallelism（DP+EP）两种并行策略
1
Sources
85%
Confidence
Medium-term (~90 days)
Relevance
5/28/2026
First Seen
Valid until: 8/26/2026
Sources
AMD MI355X击败B200：DeepSeek-R1推理TCO低5%的全栈优化解析
twitterlmsysorg5/28/2026
Related Entities
Related Claims
Unverified张量并行（Tensor Parallelism）是将单个模型层的权重矩阵切分到多块 GPU 上并行计算，适用于模型参数超出单卡显存的场景71% similarUnverifiedGGUF 现已支持张量并行（Tensor Parallelism），可带来约 +30% 的推理吞吐提升65% similarUnverifiedNVIDIA提出了非均匀张量并行（Nonuniform Tensor Parallelism）技术，用于在GPU故障时最大化LLM训练的有效产出（Goodput）63% similarUnverifiedTensorFlow Lite、ONNX Runtime等推理框架可与底层驱动和硬件加速器（NPU/DSP）协同优化62% similarUnverifiedTensor并行是大模型推理中应对单卡显存不足的核心策略，由Megatron-LM团队系统化提出61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/621API
curl https://kongchang.com/api/v1/knowledge/claims/621MCP
get_claim(id=621)