fix(fsdp): preserve internal module output dtype (#3657)
* fix(nemotron-v3): align latent projection input dtype Cast latent MoE expert output at the model-specific projection boundary so nested FSDP FP32 outputs remain compatible with BF16 projection weights. Signed-off-by: HuiyingLi <willwin.lee@gmail.com> * fix(nemotron-v3): preserve autocast projection dtype Signed-off-by: HuiyingLi <willwin.lee@gmail.com> * test(nemotron-v3): keep dtype hook compile-safe Signed-off-by: HuiyingLi <willwin.lee@gmail.com> * fix(fsdp): preserve internal module output dtype Signed-off-by: HuiyingLi <willwin.lee@gmail.com> --------- Signed-off-by: HuiyingLi <willwin.lee@gmail.com>
H
Huiying committed
3732040ad93622c7af4afbe44278b5ebcca27b66
Parent: cc6d64c
Committed by GitHub <noreply@github.com>
on 8/27/2026, 3:14:07 AM