Add quickreduce as alternative to custom allreduce#16804

Closed

ilmarkov wants to merge 28 commits intovllm-project:mainfrom

neuralmagic:experimental/quick_reduce

Commits on Jun 17, 2025

Add quickreduce as alternative to custom allreduce
ilmarkov
committed
WIP
ilmarkov
committed
Add bf16 support
ilmarkov
committed
WIP
ilmarkov
committed
Refactor QuickReduce
ilmarkov
committed
Cleanup
ilmarkov
committed
Remove config param. Add faster low latency OneShot algo
ilmarkov
committed
Some fixes
ilmarkov
committed
fix bfloat16 recv

lihaoyang-amd
authored and
ilmarkov
committed
fix env

lihaoyang-amd
authored and
ilmarkov
committed
fix log info

lihaoyang-amd
authored and
ilmarkov
committed
for env

lihaoyang-amd
authored and
ilmarkov
committed
Add int8 quantization. Remove changes to custom_allreduce
ilmarkov
committed
Update after review comments
ilmarkov
committed
add Q6 support

lihaoyang-amd
authored and
ilmarkov
committed
Adjusted to static constexpr int

lihaoyang-amd
authored and
ilmarkov
committed
Remove useless functions

lihaoyang-amd
authored and
ilmarkov
committed
fix max size err

lihaoyang-amd
authored and
ilmarkov
committed
adjust for comments

lihaoyang-amd
authored and
ilmarkov
committed
integrate_qr2cr

lihaoyang-amd
authored and
ilmarkov
committed
fix message size

lihaoyang-amd
authored and
ilmarkov
committed
Fix fp 2GB bug
ilmarkov
committed
adjust condition

lihaoyang-amd
authored and
ilmarkov
committed
fix vll_config

lihaoyang-amd
authored and
ilmarkov
committed
change comment

lihaoyang-amd
authored and
ilmarkov
committed
Update test. Disable QR by default. Set fp16 ovfl flag.
ilmarkov
committed
Fix CodecQ4
ilmarkov
committed
Update min sizes
ilmarkov
committed