| from __future__ import annotations |
|
|
| import torch |
| import torch.distributed as dist |
| from torch import Tensor |
|
|
| _FP8_E4M3_MAX = 448.0 |
|
|
|
|
| @torch.no_grad() |
| def _update_amax_history(amax_history: Tensor, cur_abs_max: Tensor) -> Tensor: |
| out = torch.roll(amax_history, shifts=-1, dims=0) |
| out[-1] = cur_abs_max.to(dtype=out.dtype) |
| return out |
|
|
|
|
| @torch.no_grad() |
| def _fp8_round_trip_bf16(x: Tensor, scale: Tensor) -> Tensor: |
| xf = x.float() |
| qs = xf / scale |
| q = qs.to(torch.float8_e4m3fn) |
| return (q.float() * scale).to(dtype=x.dtype) |
|
|
|
|
| @torch.no_grad() |
| def solution(flat_grads: Tensor, amax_history: Tensor) -> tuple[Tensor, Tensor]: |
| world_size = dist.get_world_size() |
| n = flat_grads.numel() |
| shard_elems = n // world_size |
|
|
| cur_abs_max = flat_grads.abs().max().to(torch.float32) |
| updated_hist = _update_amax_history(amax_history, cur_abs_max) |
|
|
| scale = updated_hist.max().clamp(min=1e-12).to(torch.float32) / _FP8_E4M3_MAX |
| recon = _fp8_round_trip_bf16(flat_grads, scale) |
|
|
| out_shard = torch.empty(shard_elems, dtype=flat_grads.dtype, device=flat_grads.device) |
| dist.reduce_scatter_tensor(out_shard, recon.contiguous(), op=dist.ReduceOp.SUM) |
| out_shard.div_(world_size) |
|
|
| return out_shard, updated_hist |
|
|
