transformers
baad21b7
- Let FSDP read the resolved EP plan so Mega MoE experts stay local
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Commit
View On
GitHub
Commit
27 days ago
Let FSDP read the resolved EP plan so Mega MoE experts stay local Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
References
#49775 - Experts and TP: per-instance experts flags, load-time-only experts, scalars and per-expert scales in TP and EP
#49776 - Fine-grained block-scaled quantization: the shared machinery
#49777 - FP8 on the fine-grained machinery (`Fp8Config`)
#49778 - 🚨 MXFP4 on the fine-grained machinery, and DeepSeek-V4's FP8 + MXFP4 plan
#49779 - NVFP4 on the fine-grained machinery
#49780 - MXFP8 on the fine-grained machinery (`Mxfp8Config`)
#49818 - Fine-grained quantization: a host quantizer that mixes methods per module
#49823 - Load ModelOpt and compressed-tensors exports on the fine-grained formats
Author
IlyasMoutawwakil
Parents
1ab4d06a
Loading