Block a user
TP auto-recovery livelocks after watchdog poison: unload "in flight" + reload "already loaded" every 2s, never converges
TP load-on-demand is not disconnect-safe: client timeout kills rank-1 mid-load and leaks the leader shard
Two more facets of load-on-demand fragility observed while retrying the swap (build 1773df3):
- Load against occupied VRAM has no admission check:
POST /models/loadfor Coder-Next while…
fix(neuron): qwen3_next models get snapshots, prefix cache + batch engine (#98)
fix(neuron): qwen3_next models get snapshots, prefix cache + batch engine (#98)
TP load-on-demand is not disconnect-safe: client timeout kills rank-1 mid-load and leaks the leader shard
feat(neuron): continuous batching — TP batched decode + engine backend seam (#98)
feat(neuron): continuous batching — TP batched decode + engine backend seam (#98)
F4d — perf(neuron): continuous batching for agentic fan-out
Slices 1–3b merged in #121 (single-GPU path complete, default-off):
- Design: gang batching with lockstep decode + cat-based rebatch. Prefill stays B=1 (chunked prefill + prefix cache reused…
feat(neuron): continuous batching — batched decode + engine (#98, slices 1–3b)
feat(neuron): continuous batching — batched decode + engine (#98, slices 1–3b)