Block a user
In-situ quantisation is redone on every load and discarded on unload — ~26 minutes per load for qwen4_exp
Landed and measured: 8m19s → 23s
d0e7f185 on main, validated on beast across three guarded windows.
Every run hits the same OOM at decoder layer 15 — VRAM is the
constraint, not time —…
qwen4_exp E: hyper-connections, PLE n-gram embeddings, interleaved M-RoPE
qwen4_exp D: QSA sparse-attention indexer
qwen4_exp E: hyper-connections, PLE n-gram embeddings, interleaved M-RoPE
Closing: all three blocks implemented, and compared to the implementation they were ported from
This issue grouped the new maths in the decoder and asked that the riskiest piece be parity-teste…
qwen4_exp D: QSA sparse-attention indexer
Closing: implemented, audited against the reference, and numerically verified
All four gates this issue accumulated are now closed, and by different kinds of evidence rather than four runs of…
qwen4_exp QSA assumes the visible set is contiguous — true today, false the moment it is batched