MoE fp8 backward: in-kernel drhs operand quantization + configurable gradient dtype - #4971

Open

ultrons wants to merge 4 commits into

AI-Hypercomputer:test_964807349from

ultrons:sivaibhav-moe-bwd-inkernel-quant

Commits on Aug 22, 2026

MoE ring-of-experts: TC Pallas ring all-gather for the combine backward + x_sorted remat save
ultrons
committed
82832bd
review: single descriptor for the local shard copy (start/wait on one object)
ultrons
committed
63cdc7c
MoE fp8 backward: in-kernel drhs operand quantization + configurable gradient dtype
ultrons
committed
dfa279d
review: quantize only the dlhs cotangent on the in-kernel path; Literal type for bwd_quantization_dtype
ultrons
committed
a7bf842