Block a user
R6 — packaging / deploy / CI for the ROCm engine + neuron AMD mode
R5 — observability integration for the AMD path (bridge engine metrics into cortex)
R4 — model bring-up on MI355X via the harness, validated end-to-end through cortex
R3 — neuron AMD-node mode: ROCm discovery + health + harness-wrapper operation
R2 — external-engine Harness impl (vLLM/SGLang-ROCm behind the retained trait)
R1 — MI355X cloud spike: prove GLM 5.2 / Hy3 on vLLM-ROCm (decision gate)
[epic] neuron on AMD/ROCm: external-engine Harness for MI-class (Path C); candle-native deferred
HC — pipeline / expert parallelism (contingency: PCIe-bound high-TP)
H6 — quantized weight load + fleet validation
H5 — batched decode engine + snapshot wiring
H4 — tensor parallelism (degree-parameterized: TP=2 and TP=4)
H3 — decoder assembly + hy_v3 config + registration
H2 — noaux_tc MoE block (sigmoid + expert-bias) — shared with #131
H1 — hy_v3 arch: decoupled-head GQA + QK-norm (single-GPU)