First diagnostic split
Did the runtime recognize the architecture?
Investigate build provenance and whether the runtime includes the merged Muse Glimmer support.
Investigate selected artifacts, context, cache allocation, offload, buffers, and available memory.
Investigate exact byte counts, incomplete downloads, file roles, and expected artifact pairing.
Symptom-led checks
Follow one branch at a time
“Unknown architecture” or unsupported model type
This message occurs before a meaningful capacity verdict. Confirm the binary's revision, package date, and embedded runtime version. Upstream Muse Glimmer support was merged at commit62bf73d25c53b8161f8a22894d4f90c4aebbd7d0 on .
- Record the actual runtime revision rather than the wrapper application's marketing version.
- Check whether that revision contains the merged support or a later retained change.
- If it predates the merge, replace or rebuild through the runtime's documented process before testing memory.
The model is recognized, then fails with out of memory
Recognition means the architecture path was reached. Now account for the main weights, selected companions, KV cache, runtime buffers, operating-system use, and whether the runtime reserves the full configured context.
- Test with a lower context allocation before changing the model file.
- Remove optional companions that are not needed for the workload.
- Compare both active and conservative capacity boundaries instead of relying on filename size alone.
- Record offload, cache dtype, batch settings, driver, and build revision for any reproducible report.
Artifact integrity, truncation, or unexpected file size
Compare the local file's exact byte count and filename with the official repository tree. An interrupted, renamed, or unofficially repacked artifact can fail independently of memory capacity.
- Identify whether the file came from Meta's official model repository.
- Compare exact bytes, not a rounded “GB” label.
- Re-download through a trusted method if the byte count differs or the transfer was interrupted.
Text loads, but an image or speculative path fails
Treat this as a companion artifact branch. Image input uses the officialmmproj-kquant.gguf companion, while DFlash usesdflash-kquant.gguf. Confirm that the chosen runtime build documents the relevant integration; this page does not invent version-sensitive flags.
- Prove that the main model works before adding a companion.
- Add one capability at a time and record which artifact introduced the failure.
- Verify the companion file's exact bytes against the official tree.
The model runs, but performance is unexpectedly slow
Capacity fit and speed are different questions. Record prompt length, generated length, warm-up, runtime revision, CPU/GPU offload, cache settings, and whether a DFlash path is actually active before comparing results.
Meta's launch speed bars are configuration-specific claims, not a universal throughput expectation for every device or build.
Reproducible evidence
Capture enough context for someone else to diagnose it
Useful fields
- Exact runtime and wrapper versions, build revision, operating system, and driver.
- Main and companion filenames, exact byte counts, and source repository.
- Available memory, offload plan, configured context, cache type, and batch settings.
- Complete error text plus the last successful stage, with secrets and private paths removed.
Limitations
- Error wording varies across llama.cpp revisions, wrappers, and downstream integrations.
- These branches identify the next check; they are not a substitute for a build-specific issue report.
- No unverified command flags are prescribed because companion and runtime interfaces can change.
- A successful load does not validate output quality, speed, or full-context stability.
Continue the diagnosis
- Re-check the hardware estimate
Compare selected files and context against active and conservative totals.
- Verify llama.cpp build support
Pin architecture support to the exact upstream merge.
- Verify the official files
Compare filenames, byte counts, and companion roles.
- Review memory boundaries
Separate main weights, companions, KV allocation, and system overhead.