If your local inference workload is already known to require ROLL and you are considering Hexagon execution, llama.cpp b11045 gives you a specific reason to revisit compatibility. It does not establish that your model is ready to deploy.
The upstream b11045 release page labels the build a prerelease and announces “hexagon: add ROLL op support”, linking PR #29105. That is an operator-support announcement. The release page does not identify compatible models or devices, supported tensor configurations, numerical validation results, full-graph offload, or performance gains. The linked PR’s implementation and tests are not assessed here.
Start with the operator, not the whole model
My recommendation is a bounded compatibility trial before investing in full-model validation. These are proposed checks, not results reported by the release:
- Confirm that ROLL is relevant to your workload. Identify the operation and input configurations you actually need to exercise. Without that connection, this release note alone is not a reason to prioritize an upgrade.
- Establish the build and backend being tested. Record the exact runtime revision, build configuration and hardware. Verify that the intended Hexagon path is available before interpreting any successful run as evidence for it.
- Check actual ROLL execution on Hexagon. Use execution evidence sufficient to distinguish the intended backend from another execution path. A completed request without that attribution leaves the operator compatibility question open.
- Compare outputs with a trusted reference. Use the same inputs and operation configuration, with an explicit correctness criterion appropriate to the workload. The release note supplies neither a tolerance nor numerical results, so do not treat either as an upstream guarantee.
There is a download-selection trap here: the page lists Android arm64 (CPU) and Windows arm64 (OpenCL Adreno) assets. Neither label establishes Hexagon compatibility. Confirm the artifact’s backend support separately; an arm64 label is not enough to select a binary for this trial.
Keep the result narrower than the announcement
Use the trial to make the next investment decision:
| Trial observation | Next step |
|---|---|
| You cannot establish that ROLL executed on Hexagon | Resolve build or execution attribution before calling the trial a compatibility pass. |
| Execution is established, but output comparison fails | Investigate that configuration before advancing to full-model validation. |
| Execution and reference-output checks pass | Proceed to full-model correctness and graph-coverage checks, retaining the exact tested configuration. |
A passing operator trial should be recorded as evidence for that build, hardware, operation configuration and input set, not for every model that might use ROLL. Full-model correctness, which parts of the graph execute on the intended backend, and end-to-end performance remain separate questions to measure.
The useful decision from b11045 is therefore whether to reopen a specific ROLL compatibility test. It is not yet a deployment recommendation or a speedup claim.