Block a user
HeartMuLa H5: memory discipline — LM/codec placement and the 19 GiB peak
HeartMuLa H4: candle port of HeartCodec (token-to-waveform decoder)
HeartMuLa H3: candle port of the MuLa LM (multi-codebook autoregressive decoder)
HeartMuLa H2: reference fixtures from heartlib at a pinned revision
feat(neuron): HeartMuLa serving — fixtures, candle port, audio-path integration (#390–#394)
feat(neuron): HeartMuLa serving — fixtures, candle port, audio-path integration (#390–#394)
rust runner on anjie exits mid-job with no log; Gitea marks the job failed
There's a second, clearer way a job gets killed, this time on beast (2026-10-17, helexa run 9602, CUDA type-check, runner gf-05b8eb7e):
feat(deploy): serve ACE-Step on benjy; validate-audio.sh; bench audio scenario (#387)
feat(deploy): serve ACE-Step on benjy; validate-audio.sh; bench audio scenario (#387)
ACE-Step A1: reference fixtures from the upstream pipeline at a pinned revision
feat(neuron): ACE-Step 1.5 music engine, with upstream parity (#381, #383)
feat(neuron): ACE-Step 1.5 music engine, with upstream parity (#381, #383)
HeartMuLa H4: candle port of HeartCodec (token-to-waveform decoder)
Codec port: branch
Codec port: branch feat/391-heartmula-port @ ed3a48f4
The codec lives in crates/neuron/src/harness/arch/heartmula/codec.rs. It covers:
- RVQ lookup: 8 × 8192 × 32 codebooks, then…
HeartMuLa H3: candle port of the MuLa LM (multi-codebook autoregressive decoder)
LM port: branch
LM port: branch feat/391-heartmula-port @ ed3a48f4
The LM lives in crates/neuron/src/harness/arch/heartmula/lm.rs. It covers:
- the torchtune
llama3_23B backbone and the 300M local…