The shard filename holding name (for releaseShard scheduling).
Dispose + evict one cached tensor so its materialized bytes are freed mid-walk. Streaming whole-model transforms (fold/quantize at 27B scale) MUST release each source tensor after its consumer evaluates, or the cache accumulates the entire model in RAM. Safe to call for names never accessed; the next tensor(name) call re-opens it from the map.
Free one shard's native map AND every cached tensor from it, releasing all bytes materialized while consuming that shard. The shard transparently re-opens on the next tensor() touching it (header parse only — lazy). Required by streaming whole-model transforms at 27B scale: the native map retains every evaluated source array, so per-tensor dispose alone frees nothing — the whole model accumulates in RAM (2026-08-18 fold OOM).
Install the checkpoint normalization. Must precede the first tensor() call: the array cache and every shard lookup below are keyed through it.
Staticopen
Canonical names — with a view installed, exactly the tensors the graph can ask for (renamed, drops excluded).