docs/hrx: root cause of the #123/#140 NaNs (decode-split q8 pack barrier)

[bong-water-water-bong]

Docs follow-up to #176. docs/hrx.md still called the NaN source open and pointed at page migration. It now states the root cause (the LDS-only barrier before pack_completed_q8, fixed in llama.cpp 00adc2b via 1bit-MONSTER/llama.cpp#28 and #29) and the measured fix. The decode-split note now says the kernel stays off by default for speed, not correctness: per the existing table it is slower on Qwen3-0.6B and ZAYA1-8B and faster only on Coder-30B. Docs only.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 7d3e911

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 176 - Partially compliant Compliant requirements: third_party/llama.cpp moved from fa226f9 to 00adc2b Root cause fix implemented with global barriers in llama.cpp Measurements provided showing reduction in divergent results and NaN occurrences Non-compliant requirements: None Requires further human verification: None 28 - Partially compliant Compliant requirements: third_party/llama.cpp-vulkan pinned to v0.5.0 ONEBIT_VULKAN flag support added Documentation updated with vulkan.md and other relevant docs bump-llama-vulkan.yml workflow added Non-compliant requirements: None Requires further human verification: None 29 - Partially compliant Compliant requirements: third_party/llama.cpp pinned to 1bit/hrx-vulkan-patched branch IQ3_XXS matmul kernel and matcher added Honest op claims implemented to prevent unsupported node dispatching Rebase logic implemented in bump-hrx.yml Previous tip tagged for reachability Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation update correctly identifies and describes the root cause of the NaNs in decode-split flash attention. However, it could be more explicit about how the fix impacts performance trade-offs, especially since the decode-split kernel is slower on some models (like Qwen3-0.6B and ZAYA1-8B) but faster on others (like Qwen3-Coder-30B). Clarifying this in the documentation would help users make informed decisions about enabling or disabling the feature. - **Decode-split flash attention is off by default.** It used to give nondeterministic attention on HRX0 ([#140](https://github.com/1bit-MONSTER/engine/issues/140), up to 3.66 nats apart between identical requests on Qwen3-0.6B) and to fault Qwen3-Coder-30B-A3B. That is fixed since llama.cpp `00adc2b` (see the #123 section below), but the kernel is only faster on some models, so `1bit serve --device hrx` still starts llama-server with `GGML_HRX_DISABLE_DISPATCH=decode_split`. The table was measured before the fix: