docs(changelog): record Qwen3-14B NaN inference fix (#5)
The Qwen3-14B "Invalid sampling probability ... NaN / NaN/Inf logits" failure is a half-precision (F16) overflow in the onde inference engine on GPUs without bfloat16 support (notably Intel Macs): the 14B model's attention scores exceed F16's range and become NaN. Fixed in onde by preferring F32 over the F16 fallback (see onde/patches/), and siGit now shows an actionable hint instead of the raw engine error. Reaches siGit when the onde dependency is bumped. Refs #5 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0162qAWnHPNWsoD11y3TKPVC
Claude committed
Jun 25, 2026 at 15:36 UTC
a0b2ed27d7910e04981c10a51fa840d04b54005c
1 file changed
+12
CHANGELOG.md
+12
@@ -1,5 +1,17 @@
1
# Changelog
2
3
+## Unreleased
4
+
5
+### Fixed
6
+
7
+* **Qwen3-14B on-device inference** (#5): selecting Qwen3-14B and prompting it failed with
8
+ `Inference failed: inference error: Invalid sampling probability at index 0: NaN. The model
9
+ likely produced NaN/Inf logits.` This is a half-precision (F16) overflow in the `onde`
10
+ inference engine on GPUs without bfloat16 support (notably Intel Macs); the 14B model's
11
+ attention scores exceed F16's range and become `NaN`. Fixed in the engine by preferring F32
12
+ over the F16 fallback (see `onde/patches/`), surfaced here once the `onde` dependency is
13
+ bumped. siGit also now shows an actionable hint instead of the raw engine error.
14
+
15
## 1.2.1
16
17
Stabilizes the Zed/ACP integration and finishes the cloud-tier wiring on top of 1.2.0.