Issues

• Qwen3-VL-8B-Instruct fails to execute graph with Error 1007 (DSP CONN Reset / SSR) on Snapdragon X Plus Surface Pro 11

  qualcomm/GenieX#1596 · Filter by author FuzzyIkigai opened on Oct 24, 2026

• llama_cpp plugin on Windows: CPU threadpool spin-waits (ggml poll 50) through every token while the model runs on the NPU

  qualcomm/GenieX#1590 · Filter by author Jcollier0120 opened on Oct 22, 2026
  

• llama_cpp plugin: draft-mtp with nextn.* drafters (Qwen3.5) crashes — load_mtp not set, and the drafter KV is not kept in step with the target (reset / prefix rollback / draft region)

  qualcomm/GenieX#1584 · Filter by author skenaid opened on Oct 22, 2026

• Consecutive request failures in Qwen 3.5

  qualcomm/GenieX#1558 · Filter by author zhuxiaodongzxd opened on Oct 18, 2026
  1

• llama_cpp plugin (Windows, NPU + OpenCL vision): ~1.4 GB more memory than needed for Qwen3-VL-2B

  qualcomm/GenieX#1557 · Filter by author Jcollier0120 opened on Oct 17, 2026

• [Feature Request] qairt VLMs: load the embedding table quantized and memory-mapped, as text LLMs and Gemma4 already do

  qualcomm/GenieX#1547 · Filter by author Jcollier0120 opened on Oct 15, 2026
  2

  

• How could we sign libggml-htp.cat with Microsoft Windows Hardware Compatibility Publisher

  qualcomm/GenieX#1542 · Filter by author zhang261007 opened on Oct 15, 2026
  1

• bug(bench): Gemma-4-E4B HTP benchmark fails FastRPC mapping with default settings on SM8650

  qualcomm/GenieX#1531 · Filter by author mansiverma897993 opened on Oct 13, 2026
  3

  

• Ternary (1.58-bit) GGUF from Hugging Face: which files the pinned llama.cpp refuses, and an offer to help

  qualcomm/GenieX#1515 · Filter by author dmitrii-f-t27 opened on Oct 6, 2026
  2

  

• QAIRT Qwen3-4B-Instruct-2507: decode ~30x slower when the context binary holds 10 graphs per part; same graphs relinked as ≤8 per part run at full speed (Snapdragon X X1-26-100)

  qualcomm/GenieX#1509 · Filter by author ylu861 opened on Oct 4, 2026
  4

• [Bug] ggml-hex (HTP): Silent corruption + dspqueue Error 0xc timeouts on Hexagon v73 (SM7635)

  qualcomm/GenieX#1508 · Filter by author algo1127 opened on Oct 3, 2026
  3

  

• v0.8.0 Android/SM8650: HTP session fails to open with error 0x80000406

  qualcomm/GenieX#1507 · Filter by author crab3e opened on Oct 3, 2026
  3

  

• [Feature Request]: received error while streaming: SDKError(Input prompt too long)

  qualcomm/GenieX#1498 · Filter by author josergc opened on Sep 30, 2026
  2

  

• Phi-3.5-Mini-Instruct doesn't load in Elite X (12 cores) X1E80100 Hexagon NPU

  qualcomm/GenieX#1494 · Filter by author josergc opened on Sep 28, 2026
  1

• Poor performance of Qwen models in the X Elite (12 cores) X1E80100 chip

  qualcomm/GenieX#1493 · Filter by author josergc opened on Sep 28, 2026
  

• Gemma-4-26B (GGUF, GenieX-served) does not reliably conform to provided tool-call JSON schema — flattens nested arguments and guesses tool/parameter names instead of reading the schema

  qualcomm/GenieX#1479 · Filter by author EurusZhang opened on Sep 22, 2026
  1

  

• GenieX: chat template mishandles empty-thought boundary tokens and occasionally drops tool_calls entirely after a tool-error correction turn (Gemma-4-26B GGUF, NPU backend)

  qualcomm/GenieX#1478 · Filter by author EurusZhang opened on Sep 22, 2026
  2

  

• llama_cpp backend: VLM vision output silently wrong on compute_unit=npu (HTP0), crashes on compute_unit=gpu

  qualcomm/GenieX#1476 · Filter by author ppl520zfl opened on Sep 22, 2026
  3

  

• [FEATURE] Allow prompt_utf8 input in geniex_llm_forward_logits

  qualcomm/GenieX#1473 · Filter by author nos1609 opened on Sep 20, 2026
  3

• QAIRT Qwen3-4B-Instruct-2507 crashes with stack overflow on Snapdragon X Plus X1P42100

  Stale
  qualcomm/GenieX#1471 · Filter by author dawalmalik0405-spec opened on Sep 18, 2026
  3

  

• Hunyuan Model Support

  Stale
  qualcomm/GenieX#1465 · Filter by author RichardDoong opened on Sep 16, 2026
  1

  

• design(vlm): support multimodal KV prefix reuse

  Stale
  qualcomm/GenieX#1455 · Filter by author mansiverma897993 opened on Sep 11, 2026
  2

• ggml-hexagon on Hexagon v73 (QCM6690): NPU output is byte-identical regardless of the prompt — silent corruption, no crash, graph splits = 3

  Bug

  qualcomm/GenieX#1429 · Filter by author sGeeK44 opened on Sep 4, 2026
  3

• Add transaction-safe lineage validation for retained chat caches

  Stale
  Feature

  qualcomm/GenieX#1413 · Filter by author samkwak188 opened on Aug 31, 2026
  2

  

• Add fidelity-to-teacher KLD and long-context retrieval eval types

  Stale
  Feature

  qualcomm/GenieX#1411 · Filter by author samuelsalfati opened on Aug 30, 2026
  3