fix: wire --metal flag into remote FFN/MoE paths, add post-FFN norms,…#115
Open
MELDApps wants to merge 1 commit into
Open
fix: wire --metal flag into remote FFN/MoE paths, add post-FFN norms,…#115MELDApps wants to merge 1 commit into
MELDApps wants to merge 1 commit into
Conversation
… fix stdout flush - Bug 1: Add stdout flush after token print loop in run_with_remote_ffn and run_with_moe_shards - Bug 2: Apply pre-FFN norm (apply_norm_for_ffn) before sending to remote server in prefill - Bug 3: Apply post-FFN norm (apply_post_ffn_norm) to server response in both prefill and decode - Bug 4: Wire --metal flag into run_with_remote_ffn (was always using CPU backend) - Bug 4b: Wire --metal flag into run_with_moe_shards (was always using CPU backend) Note: multi-token decode still produces zero hidden states in Metal backend decode_token_with_moe - root cause unresolved, see GitHub issue
Owner
|
Hey @MELDApps — thanks for digging into these in #114. Reviewing the diff I found three things blocking a clean merge:
The other four fixes are great and I wanted to land them, so I opened #122 as a cherry-pick that keeps:
Your commit attribution is preserved on the first commit of #122. Will close this in favour of that. Happy to revisit the decode-loop norm gap (keeping the Q8K fast path) as a follow-up — that's the real fix for the multi-token zero issue. Note for the type-ascription change in |
chrishayuk
pushed a commit
that referenced
this pull request
May 22, 2026
…h stdout Cherry-pick of the four legitimate fixes from #115 against current main. Drops the decode-loop changes (`break;`, debug eprintlns, Q8K-fast-path removal) which were author-acknowledged work-in-progress and would have regressed performance + correctness. Issues found during distributed inference testing across M1 Mac Mini + HP ProDesk: * `run_with_remote_ffn` and `run_with_moe_shards` print tokens via `print!` without flushing. Add `std::io::Write::flush(&mut stdout())` after each token loop so users see output immediately rather than only at process exit. * `run_with_remote_ffn` always called `larql_compute::default_backend()` which is CPU-only, ignoring `--metal`. Wire the flag through: `metal_backend()` when set (with CPU fallback if Metal init fails), `cpu_backend()` otherwise. * Remote FFN servers return the raw FFN result without applying the pre/post FFN norms a local forward would. On post-norm archs (Gemma 3 / 4) this causes the remote path's residual stream to diverge from the local one. Add `apply_post_ffn_norm` next to the existing `apply_norm_for_ffn`, and call both in the prefill `moe_fn` closure of `generate_with_remote_ffn`. The matching changes inside `generate_with_remote_ffn`'s decode loop (also missing pre/post-FFN norms) are intentionally left for a follow-up so this PR can land cleanly without the broken Q8K rewrite. Supersedes #115.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes for bugs found during distributed inference testing across M1 Mac Mini + HP ProDesk.
Closes #114
Changes
Note: multi-token decode zero hidden state issue remains unresolved - see linked issue for full details.