← News
Runtime

Community report: PLaMo 3 loads and benchmarks on Snapdragon Hexagon but returns empty answers

Issue #31889 reports correct CPU translation but empty Hexagon output on the same tablet and weights. It is a single community report, with an unconfirmed cause and limited scope.

Source: llama.cpp 社区 issue #31889

On November 8, llama.cpp issue #31889 reported a PLaMo 3 translation failure on a Xiaomi Pad 7 Pro with Snapdragon 8s Gen 3 and Android 16. The CPU produces a translation while full Hexagon HTP0 offload ends generation immediately. At retrieval the issue was open with no comments. This is one community report, not an AraoAI reproduction or an official root-cause finding.

  1. Preserve the scope. The report names the official b12202 Android Snapdragon package and says b12125 is also affected. It identifies plamo-3-translate-2b-Q4_0.gguf, its source and hash, and the prompt. The CPU and HTP0 comparison uses the same device and weights. It does not establish failure across every Hexagon device, precision or model.

  2. Successful loading is not successful deployment. Logs report full layer offload and allocated model and KV buffers. The output is nevertheless empty. Variations in offloaded layers, Flash Attention and environment settings yield empty or repeated-token output. We do not publish the accompanying benchmark as a successful deployment speed: throughput with incorrect answers is not task validity.

  3. A hypothesis is not a diagnosis. The author suggests large activations and FP16 overflow as a possibility and explicitly marks it unverified. The evidence does not yet establish a faulty operator or a released fix. Preserve the CPU baseline, build, weight hash, prompt and logs, then vary one offload setting at a time.

For on-device selection, validate actual task output before comparing acceleration and memory. Device discovery, loading and printed throughput only confirm intermediate steps. This source remains a failure report to track, not a successful run record or evidence that users should replace their hardware.

Source: https://github.com/ggml-org/llama.cpp/issues/31889