Speed demo: Glm5.3 flash on 2x GB10s
https://relay.utxo.one/8125e248f9192acc5eb8cfcb812e6708c4eefdcedde6bdb1f885626180ecd779.mp4
Is this with the fallback `--moe-backend marlin` because native `NVFP4` is not supported yet on the `sm_121`? Meaning we can expect to see speed gains as the software catches up?
1