Scheduler UMA ring buffer (+ sanitizer and fixes) - #27311

Open

pwilkin wants to merge 17 commits into

ggml-org:masterfrom

pwilkin:sched-uma-ring

Commits on Aug 21, 2026

ggml: add a scheduler sanitizer

am17an
authored and
pwilkin
committed
76870ad
ggml: add a non-fatal mode to the scheduler sanitizer
pwilkin
committed
22994a6
ggml,llama: let the sanitizer see direct writes to host tensors
pwilkin
committed
0051b45
llama: give every graph input a name
pwilkin
committed
292887a
cuda: verify the cgraph uid fast path instead of trusting it
pwilkin
committed
70192ac
ggml-backend: ring buffer graph inputs when a backend computes on host memory
pwilkin
committed
f141436
ggml-backend,llama: rotate the graph input ring on the reuse path
pwilkin
committed
9bdc6ea
ggml-backend: document the graph input ring buffer in the scheduler docs
pwilkin
committed
2dd7ff1
ggml: report the allocation root in sanitizer race reports
pwilkin
committed
c7f83de
ggml-backend: keep memory read in place by another backend out of the reuse pool
pwilkin
committed
f543f23
ggml-backend: run ops that alias their source on the source's backend
pwilkin
committed
9dea828
docs: document the backend scheduler
pwilkin
committed
20c6c5b
ggml-backend: do not ring buffer graph inputs for a CPU-only scheduler
pwilkin
committed
5db5750
ggml-backend: only ring buffer graph inputs for an asynchronous backend
pwilkin
committed
c8d05db
ggml-backend: only move an aliasing op to a backend that supports it
pwilkin
committed
508d9a2
ggml: harden scheduler ring allocation
pwilkin
committed
1486aa7
cuda: simplify graph uid verification setting
pwilkin
committed
b3823d8