The adapter failure was a specific dispatch problem

The AMD report concerns MUL_MAT_ID with F32, F16 or BF16 inputs and small batches of at most eight tokens. Its fast path selected an operation that reads values in pairs and requires an even column count. A rank-one LoRA adapter on MoE expert weights can create an odd dimension, triggering an assertion and stopping inference. The correction checks that dimension before selecting the fast path; other shapes use the existing per-expert fallback. This changes how an adapter is applied during inference, rather than introducing a new method of fine-tuning it.

Reported test results have a defined hardware and model scope

The contributor tested a Radeon AI PRO R9700 with ROCm 10.1.0. Two added backend cases aborted before the correction, while the resulting run passed 1,042 cases. The end-to-end example used LFM2.5-8B-A1B Q4_0 and synthetic F32 adapters of ranks one and two. These are the contributor’s reported results, not an independent certification of every AMD GPU. The description explicitly excludes testing the separate 24-billion-parameter model, the original adapter file from the issue, and some other devices. Compatibility claims should preserve those boundaries.

Vulkan tie-breaking now follows a stable column order

The Vulkan radix top-k path previously gathered selected elements through a shared atomic counter. When equal values competed for fewer remaining slots, thread scheduling could affect which columns survived. The merged implementation ranks overflowing ties in column order so that lower-indexed tied columns are retained, matching the behavior described for CUDA and Metal. The discussion covers large-k radix selection and the fused TOPK_QSA operation. It does not establish that every model response is now bit-for-bit identical across devices, numerical formats or generation settings.

What this means for local evaluation and deployment

For teams applying custom adapters or comparing GPU backends, the fixes remove two identifiable sources of trouble: one invalid tensor-shape dispatch and one unstable ranking operation. Vulkan tests and performance observations in the discussion belong to specific cases and should not be generalized into a universal inference speedup. The release pages reviewed here label these builds as prereleases. A deployment assessment should therefore retain the actual model, adapter rank and layout, backend and evaluation workload. Neither change publishes new model weights, changes a model license or establishes a new universal memory requirement.