feat(MODEL-MM-QWEN4-EXP): stage a large pageable H2D through a bounded pinned ring on ROCm - #3180

Merged

mudler merged 10 commits into

mainfrom

row/MODEL-MM-QWEN4-EXP-ROCM-CHUNKED-H2D

Sep 13, 2026

Commits on Sep 13, 2026

spec(MODEL-MM-QWEN4-EXP): chunked H2D through a bounded pinned bounce ring
mudler
committed
45633aa
test(MODEL-MM-QWEN4-EXP): gate the pinned H2D ring's plan, reuse order and predicate
mudler
committed
054546b
feat(MODEL-MM-QWEN4-EXP): stage a large pageable H2D through a bounded pinned ring on ROCm
mudler
committed
868db7d
record(MODEL-MM-QWEN4-EXP): the ring produces the row's first token on gfx1151, and the same binary without it does not
mudler
committed
556475c
record(MODEL-MM-QWEN4-EXP): separate the CIFS confound, and give the decode number four samples
mudler
committed
d4a0388
test(MODEL-MM-QWEN4-EXP): read the pinned ring's own residency number on the managed arm
mudler
committed
d83c4b1
fix(MODEL-MM-QWEN4-EXP): allocate the pinned ring only for a copy that will use it
mudler
committed
a6621c9
record(MODEL-MM-QWEN4-EXP): say what the ROCm arm produced, and correct three figures it got wrong
mudler
committed
933039a
record(BACKEND-GATE-ROCM-VLLM): file the low-disk failure mode of the Strix oracle suite
mudler
committed
309f626
test(MODEL-MM-QWEN4-EXP): make the split-drift case convict a deleted term, and correct what it was said to prove
mudler
committed
92d3305