Blog: milestone - the HRX decode race is solved (#123, #140)

[bong-water-water-bong]

Milestone post for the #123/#140 fix, plus README status and PORTING row.

• blog/2026-09-27-hrx-decode-race-solved.md: the cause (decode-split q8 pack read global output behind an LDS-only barrier), the evidence (forced foreign evictions: Coder-30B 1/8 -> 8/8, 0.6B 13-34 divergent per ~177 -> 0/354), decode with the kernel back on (Pin llama.cpp 00adc2b: HRX decode-split publishes next_q8 after a global barrier (#123, #140) #176, serve: decode-split flash attention back on for --device hrx #178, Pin llama.cpp a34a6b7: HRX decode-split multipass output pass is vectorised (engine#124) #180). Recap: architecture gaps closed (323 HF architectures mapped), GGUF on the NPU (npu: GGUF-on-NPU serving route (Qwen2.5-7B, MiniCPM4-8B, MiniCPM5-1B answer ' Paris' on the NPU) #179). Every number is from docs/hrx.md, docs/registry.md or docs/npu.md.
• README status and docs/PORTING.md step 2: the race is fixed and the kernel is on by default.
• Site built locally with tools/site.py (35 pages); the post's doc anchors resolve.

🤖 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 54fe30e

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 123 - Partially compliant Compliant requirements: The PR addresses the root cause by changing the barrier type from kernel.barrier<workgroup> to kernel.barrier<global> at the critical points where global memory stores are followed by global memory loads The fix ensures that the decode-split kernel no longer faults with HSA_STATUS_ERROR_MEMORY_FAULT The fix makes the decode deterministic, resolving the intermittent faults and wrong outputs Non-compliant requirements: The PR does not include specific measurements or performance benchmarks to validate the fix's effectiveness Requires further human verification: The effectiveness of the fix should be verified through testing on actual hardware with the specified conditions (e.g., forced KFD queue evictions) 140 - Partially compliant Compliant requirements: The PR addresses the race condition by ensuring proper memory ordering with global barriers The fix ensures deterministic outputs for identical requests The fix resolves the issue where the same run could produce different outputs Non-compliant requirements: The PR does not include specific measurements or performance benchmarks to validate the fix's effectiveness Requires further human verification: The effectiveness of the fix should be verified through testing on actual hardware with the specified conditions (e.g., forced KFD queue evictions) 176 - Partially compliant Compliant requirements: The PR pins llama.cpp to 00adc2b, bringing the necessary fixes The fix ensures that the kernel is on by default and working correctly The PR includes the necessary changes to the kernel barriers to fix the race condition Non-compliant requirements: The PR does not include specific measurements or performance benchmarks to validate the fix's effectiveness Requires further human verification: The effectiveness of the fix should be verified through testing on actual hardware with the specified conditions (e.g., forced KFD queue evictions)

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The blog post mentions that the fix involves changing barriers from kernel.barrier<workgroup> to kernel.barrier<global>, but does not clearly explain the implications of this change for the memory ordering and how it prevents the race condition. This could be confusing for readers trying to understand the technical details. those system settings. For a while it looked like a page-migration bug, but migration was only the trigger. The fix fences both memory spaces at all five places the kernel does this: `barrier<global>`, then `barrier<workgroup>` ([docs/hrx.md](../docs/hrx.md#moe-router-expert-id-fault-fix-engine123)).