Quantize activations before EP all gather - #4812

Open

Shuwen-Fang wants to merge 11 commits into

mainfrom

quantize_sort_2

Commits on Aug 11, 2026

Rebase quantize_sort_2 onto latest main

Shuwen-Fang
and
claude
committed
1fb67bf
Tighten scope: keep use_single_sparsecore, drop unrelated capacity fixes

Shuwen-Fang
and
claude
committed
eb1b787
Restore PR #4735's static-LHS-scaling mechanism, layered alongside ours

Shuwen-Fang
and
claude
committed
595d36f
Use qwix's should_quantize() for the "already quantized" check in ops.py

Shuwen-Fang
and
claude
committed
59a7acb
moe.py: dedup ring_ragged_sort call sites, drop redundant/unsafe quantize block

Shuwen-Fang
and
claude
committed
f438e64
Add quantize_before_ep_all_gather flag; use qpl.get_current_rule in roe_ag_and_route

Shuwen-Fang
and
claude
committed
09dc56e
Add test_quantize_before_ep_all_gather_equivalence

Shuwen-Fang
and
claude
committed
207b943
update test
Shuwen-Fang
committed
76f2097
validate
Shuwen-Fang
committed
5728eac
simplify
Shuwen-Fang
committed
b1ed0f0

Commits on Aug 12, 2026

fix
Shuwen-Fang
committed
1ed8bb5