Fix unused parameters in attention layers

[erwulff]

This PR fixes a critical bug in mlpf/model/mlpf.py where the output of the FFN in PreLnSelfAttentionLayer was computed but never added to the residual connection.

This was discovered when running model training with Ray Train, and the following error was encountered:

RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates that your module has parameters that were not used in producing loss. You can enable unused parameter detection by passing the keyword argument `find_unused_parameters=True` to `torch.nn.parallel.DistributedDataParallel`, and by 
making sure all `forward` function outputs participate in calculating loss. 
If you already have done the above, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's `forward` function. Please include the loss function and the structure of the return value of `forward` of your module when reporting this issue (e.g. list, dict, iterable).
Parameter indices which did not receive grad for rank 7: 102 103 104 105 106 107 114 115 116 117 118 119

This probably appears only when running with Ray Train, and not with bare DDP, because Ray Train enforces more strict checks by default.

The FFN block and second LayerNorm were effectively unused parameters, causing RuntimeError: Expected to have finished reduction... failures during distributed training (DDP).

TODO:

• run quick before/after validation (1 GPU, 1 hour is enough), post train and val loss.

[Copilot]

Pull request overview

Fixes a training-time bug in PreLnSelfAttentionLayer where the FFN output was computed but never applied to the residual stream, which could leave FFN/LN parameters unused and trigger stricter DDP/Ray Train unused-parameter failures.

Changes:

• Add the missing residual update x = residual + ffn_out after the FFN block in PreLnSelfAttentionLayer.forward.

Comments suppressed due to low confidence (1)

mlpf/model/mlpf.py:319

• save_attention branch can raise runtime errors for attention_type == LINEAR: att_mat is never defined in the LINEAR path, and self.mha.in_proj_weight doesn't exist on LinearAttention. Consider guarding this block with self.attention_type != AttentionType.LINEAR (and/or handling LinearAttention separately), and ensure att_mat is always defined before use.

        if not self.use_simplified_attention and self.save_attention:
            np.savez(
                open("{}/attn_{}_{}.npz".format(self.outdir, self.name, self.att_mat_idx), "wb"),
                att=att_mat,
                x=x.detach().cpu().numpy(),
                in_proj_weight=self.mha.in_proj_weight.detach().cpu().numpy(),
            )

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jpata]

oh, good catch! can you just run a quick before/after training, like 1h each, and post the losses for posterity?

[erwulff]

oh, good catch! can you just run a quick before/after training, like 1h each, and post the losses for posterity?

Sure! I ran a slightly longer test than you suggested. 4h on 8xH100. Losses are slightly lower after the fix.

Figure: Total training loss before and after the fix. After fix in purple, before fix in yellow.

[jpata]

That's awesome! Glad you spotted the issue!