Qwen3-0.6B: QNN HTP produces degenerate output at all precisions (XNNPACK correct on same device)

🐛 Describe the bug

Qwen3-0.6B produces token id 0 every decode step through QNN HTP, at both 16a4w and 8a8w. Same weights/device/binary via XNNPACK produces correct output.

Environment

Snapdragon XR2 Gen 2 (SXR2230P), HTP v69, QNN SDK 2.37, Quest 3, ExecuTorch main.

Same-device A/B

XNNPACK and QNN .pte exported from the same edge program. Weights, tokenizer,
runner binary all identical.

Ruled out

• Quantization — 8a8w (near-lossless) equally broken
• Model def / weights / export — shared with XNNPACK path, which is correct
• Softmax — forced all 28 softmax ops to CPU, still degenerate

Qwen3 features that differ from Llama (suspects)

• XNNPACK (CPU, same device): ✅ Correct tool call, stops on EOS
• QNN HTP (qnn_8a8w): ❌ Token id 0 every step, never stops
• QNN HTP (qnn_16a4w): ❌ Same

Gemma 3 (1B) shares qk-norm and head_dim mismatch — testing on HTP to narrow urther.

Ask

Per-layer intermediate tensor diff (QNN vs CPU reference, fixed input) to find the first diverging op. Op-skipping can't isolate further — skipping RMSNorm/RoPE

Versions

attach logs and repro configs in comments.

cc @cccclai @cbilgin @abhinaykukkadapu

Start with the QNN HTP and XNNPACK paths for the shared Qwen3-0.6B edge program, using a fixed input and the CPU reference. Compare per-layer intermediate tensors to identify the first divergence, then document the diverging op and confirm whether it explains the token-id-0 output.