Use expert gradients, not expert weights, in the FP16_Optimizer MoE grad norm (#8635)
`FP16_Optimizer.step` builds two gradient-norm inputs, and the two
consumers expect different things:
- the non-expert list goes to `get_flattened_grad_norm`, which filters
on `p.grad` and reads `p.grad.data`, so it expects **parameters**;
- the expert list goes to `get_norm_with_moe_layers` →
`get_global_norm_of_tensors`, which reads each tensor's own data, and
whose docstring asks for **gradient tensors**.
Both lists were filled with the FP32 master parameter
`self.fp32_groups_flat[i]`. That is correct for the first consumer and
wrong for the second, so the expert contribution to the global norm was
the expert **weight** norm rather than the expert gradient norm. With
the default `gradient_clipping=1.0`, that inflates the norm and shrinks
the clip scale applied to every gradient, including non-expert ones.
The fix passes `grads_groups_flat[i]` to the expert list.
`BF16_Optimizer` (`self.fp32_groups_gradients[i][j]`) and
`FP16_UnfusedOptimizer`
(`split_params_grads_into_shared_and_expert_params`) already take the
expert norm from gradients and are not affected.
### Test
`TestMoEFP16GradientNorm::test_expert_weight_values_do_not_affect_gradient_norm`
trains a small MoE model with fp16 and ZeRO stage 0, which selects
`FP16_Optimizer`. It sets every gradient to the same fixed value, scales
only the expert weights by 100, and requires the norm reported by
`engine.get_global_grad_norm()` to be exactly unchanged. It also checks
that the norm is above the default clipping threshold of 1.0, so that
clipping acts on it.
H100 80GB, torch 2.10 (CUDA 13.0):
| Source | Result |
|---|---|
| test only, unmodified master | fails: `Scalars are not equal!` |
| test + fix | passes |
These runs predate the rebase onto current master, which does not touch
either file. The test needs fp16, so CPU runners skip it.
Signed-off-by: yh0903 <helloyu0903@gmail.com>