vllm.distributed.engram_pool.rows
¶
How kernels find rows: which GPU owns an id (transports), and where a
fetched row is once it has landed (landed_row, for models' kernels).
Functions:
-
landed_row–Address of fetched row
rowoftable(args fromEngramFetch.landed), -
locate–Item
itemof[..., TABLES, K]ids: its table index andmetarow, -
wait_landed–Spin until every QP that gathered rows of
tablein the fetchready
landed_row(rows, slots, ready, table, row, valid, ROW_STRIDE, PLACED)
¶
Address of fetched row row of table (args from EngramFetch.landed),
and whether it holds data (else load zeros). Call from every program.
Source code in vllm/distributed/engram_pool/rows.py
locate(ids, meta, item, valid, TABLES, K)
¶
Item item of [..., TABLES, K] ids: its table index and meta row,
the owner GPU of its row and the row's index there, and whether it is live.
Source code in vllm/distributed/engram_pool/rows.py
wait_landed(ready, table)
¶
Spin until every QP that gathered rows of table in the fetch ready
describes has flagged them landed; trap after the timeout.