Issues

• [Performance] z-image-turbo is ~16% slower than native on GB10

  NVIDIA/TensorRT-Model-Connect#2226 · Filter by author chaofengw-nv opened on Nov 11, 2026

• encoder-parity: camembert-base, roberta-large and fnet-base miss the per-sentence tolerance on 1-2 of 400 sentences on GB10 and Thor

  bug
  NVIDIA/TensorRT-Model-Connect#2225 · Filter by author chaofengw-nv opened on Nov 11, 2026

• aiperf qual: tts-validity fails bark-small for an utterance the native model also renders silent

  question
  NVIDIA/TensorRT-Model-Connect#2224 · Filter by author chaofengw-nv opened on Nov 11, 2026

• aiperf qual: the code sandbox needs root to switch to nobody (setpriv), so codegen-350m and starcoder2-3b get no result on non-root hosts

  bug
  NVIDIA/TensorRT-Model-Connect#2219 · Filter by author chaofengw-nv opened on Nov 11, 2026
  1

• internvl3: the InternViT vision engine is always FP32 (55 ms per tile on Thor), OCRBench 1.4-1.9x slower than PyTorch

  bug
  NVIDIA/TensorRT-Model-Connect#2206 · Filter by author chaofengw-nv opened on Nov 10, 2026

• nemotron-speech-streaming-en-0.6b-fp8: LibriSpeech WER 3.49 vs 2.49, the FP8 attention projections cost a point

  bug
  NVIDIA/TensorRT-Model-Connect#2203 · Filter by author chaofengw-nv opened on Nov 10, 2026

• phi4-multimodal: prefill and decode attend over the whole 10368-row KV allocation (2-2.3x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2201 · Filter by author chaofengw-nv opened on Nov 10, 2026

• dinov3-vits16: FP16 features on GB10 are 4-9% off (0/50 parity), the attention softmax loses precision

  bug
  NVIDIA/TensorRT-Model-Connect#2199 · Filter by author chaofengw-nv opened on Nov 10, 2026

• qwen3-vl-8b: OCRBench fails (78.7 vs 88.2), reference sees full-resolution documents; 2.5-point gap remains at equal budget

  bug
  NVIDIA/TensorRT-Model-Connect#2197 · Filter by author chaofengw-nv opened on Nov 10, 2026
  2

• fast-foundation-stereo: GB10 engine build fails, no FP16 tactic for a kernel-4 stride-2 deconvolution

  bug
  NVIDIA/TensorRT-Model-Connect#2195 · Filter by author chaofengw-nv opened on Nov 10, 2026

• stable-diffusion-v1-5: UNet engine built on Thor fails createExecutionContext (Myelin cask6 initialize)

  bug
  NVIDIA/TensorRT-Model-Connect#2190 · Filter by author chaofengw-nv opened on Nov 10, 2026
  1

• 27B BF16 bundles (gemma-3-27b, qwen36-27b, qwen38-27b) run out of memory while building on GB10 and Thor

  bug
  NVIDIA/TensorRT-Model-Connect#2189 · Filter by author chaofengw-nv opened on Nov 10, 2026
  1

• rwkv: prompts are prefilled one token per decoder launch (1.2-2x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2188 · Filter by author chaofengw-nv opened on Nov 10, 2026

• aiperf qual: dpr-ctx-encoder SciFact documents cut to 256 tokens overflow its 256-token bundle with [CLS]/[SEP]

  bug
  NVIDIA/TensorRT-Model-Connect#2186 · Filter by author chaofengw-nv opened on Nov 10, 2026

• glmasr: the audio prompt is prefilled one position per decoder launch (2-3.7x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2185 · Filter by author chaofengw-nv opened on Nov 10, 2026

• aiperf qual: InternLM2 remote-code tokenizer counts one token per character on transformers 5.x, so internlm2-1.8b plans no GSM8K problem

  bug
  NVIDIA/TensorRT-Model-Connect#2183 · Filter by author chaofengw-nv opened on Nov 10, 2026
  1

• k2_horizon_uno: prompts are prefilled in 8-row blocks (about 2x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2182 · Filter by author chaofengw-nv opened on Nov 10, 2026

• mamba: prompts are prefilled one token per decoder launch (1.6-3.2x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2181 · Filter by author chaofengw-nv opened on Nov 10, 2026

• nemotron_h: single-GPU bundles prefill one token per decoder launch (5-10x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2180 · Filter by author chaofengw-nv opened on Nov 10, 2026

• qwen3_5: text prompts are prefilled one token per decoder launch (5-18x slower than PyTorch on GB10 and Thor)

  bug
  NVIDIA/TensorRT-Model-Connect#2179 · Filter by author chaofengw-nv opened on Nov 10, 2026

• Single-GPU Mixtral prefills the prompt one token per engine launch

  NVIDIA/TensorRT-Model-Connect#2177 · Filter by author chaofengw-nv opened on Nov 10, 2026

• Qwen3-Omni Thinker prefill copies every row's top-k expert matrices

  NVIDIA/TensorRT-Model-Connect#2174 · Filter by author chaofengw-nv opened on Nov 10, 2026

• Muse-Glimmer native target prefills the prompt one token per engine launch

  NVIDIA/TensorRT-Model-Connect#2172 · Filter by author chaofengw-nv opened on Nov 10, 2026

• LFM2 prefills the prompt one token per engine launch

  NVIDIA/TensorRT-Model-Connect#2163 · Filter by author chaofengw-nv opened on Nov 10, 2026

• Feature: nemotron_omni decode is bound by 2944 conditional expert subgraphs per token (about 50 ms/token for a 3B-active model)

  enhancement
  NVIDIA/TensorRT-Model-Connect#2060 · Filter by author zhenshanx-nv opened on Nov 7, 2026