Pin llama.cpp 00adc2b: HRX decode-split publishes next_q8 after a global barrier (#123, #140)

[bong-water-water-bong]

Moves third_party/llama.cpp fa226f9 → 00adc2b (fast-forward), bringing 1bit-MONSTER/llama.cpp#28 and #29: the root-cause fix for #123 and #140.

Bug: in every HRX decode-split reduce_fused variant (direct, cooperative, multipass; loom-libs/ops and qwen_moe/qwen3_moe corpora), the barrier between the reduce's global output stores and pack_completed_q8's output loads was kernel.barrier<workgroup>, which orders LDS only (s_waitcnt lgkmcnt(0); s_barrier, no vscnt wait, no buffer_gl0_inv). The pack step could quantise stale output into the next layer's Q8 input. The f32 result was right and the Q8 copy was wrong. Preemption (any process's KFD queue eviction), contention and transient alignment only widened the window. The fix is kernel.barrier<global> at those 5 sites, plus the LDS barrier kept next to it (#29: a Loom barrier fences only the memory space it names). #29 also adds the missing global barrier to the undispatched standalone reduce_f32 kernels.

Evidence (llama.cpp#28, logs in strixhalo ~/wt/hrx123-logs):

• Qwen3-0.6B under another process's queue evictions: 0/354 divergent after the fix (before: 13–34 per 177).
• Same with mlockall + forced compaction: 0/118.
• Qwen3-Coder-30B at GGML_HRX_FA_PARTIAL_ALIGN=256 (the HRX: decode-split multipass path intermittently faults (HSA_STATUS_ERROR_MEMORY_FAULT) above capacity 2048 #123 case): 16/16 correct, 0 NaN (before: 1/8 correct, NaN guard fired 7×).
• Merged head 9ca9603: 0.6B 0/59, 30B 8/8.
• Final head 00adc2b: 0.6B under evictions 0/118, 30B at 256 alignment 8/8 with 0 NaN; every variant's ISA has vscnt(0), s_barrier and gl1/gl0_inv before the pack load; decode speed within noise.

The #170 router change and #175's loud-NaN guard remain as defence in depth. Checks on this PR's head (3411558, llama.cpp 00adc2b), strixhalo, built with ONEBIT_HRX=ON ONEBIT_VULKAN=ON:

• ctest: 11/11 pass (registry_pins, registry_gaps, serve_e2e_vulkan_upstream, serve_e2e_hrx, serve_e2e_vulkan_prefill_hrx, ...); tests/serve_e2e.sh on --device auto: PASS.
• HRX: decode-split multipass path intermittently faults (HSA_STATUS_ERROR_MEMORY_FAULT) above capacity 2048 #123/HRX: decode_split_next_q8 gives wrong, nondeterministic attention at short context (up to 3.66 nats between identical requests) #140 on the engine's own build/hrx/llama/bin/llama-server, with a separate process forcing KFD queue evictions: Qwen3-0.6B 0 of 59 divergent (only the warm-up request); Qwen3-Coder-30B, 2113-token prompt, GGML_HRX_FA_PARTIAL_ALIGN=256: 8/8 correct, 0 NaN.
• ZAYA1-8B unchanged: HRX0 llama-perplexity on ppl.txt = 21.5944 at both fa226f9 and 00adc2b (back to back; the older 21.5518 figure predates later HRX kernel changes). Vulkan is not affected: fa226f9...00adc2b only touches the two HRX decode-split .loom files, and third_party/llama.cpp-vulkan does not move.
• ZAYA chat tokenization: /tokenize "<|im_start|>user" with parse_special -> [105, 2364].

🤖 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 3411558

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 3411558)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 28 - Partially compliant Compliant requirements: Pin third_party/llama.cpp-vulkan to ggml-org/llama.cpp v0.5.0 (7fe450e19305) Build llama-server with ONEBIT_VULKAN flag 1bit serve --device vulkan prefers the vulkan build --device hrx keeps the HRX build bump-llama-vulkan.yml creates a daily PR when upstream publishes a release Update docs (docs/vulkan.md; serve.md, hrx.md, README and NOTICE) Non-compliant requirements: None Requires further human verification: None 29 - Partially compliant Compliant requirements: Move third_party/llama.cpp from AMD's commit to 1bit-MONSTER/llama.cpp 1bit/hrx-vulkan-patched (d2a9239f9) Add IQ3_XXS matmul on HRX with kernel kernels/hrx/mul_mat_vec_iq3xxs_f32.loom and matcher common/dispatch-mul-mat-iq3-xxs.{h,cpp} Add honest op claims to prevent unsupported node dispatches bump-hrx.yml rebases our commits onto each new AMD pair, tagging previous tip patched-<sha> Verify on Strix Halo with Qwen3-0.6B and perplexity tests Non-compliant requirements: None Requires further human verification: None 123 - Partially compliant Compliant requirements: Fix intermittent fault in HRX decode-split multipass path (HSA_STATUS_ERROR_MEMORY_FAULT) Ensure that the barrier between reduce's global output stores and pack_completed_q8's output loads is kernel.barrier<global> Fix the issue where stale output could be quantized into the next layer's Q8 input Validate fix with Qwen3-0.6B under queue evictions and other stress tests Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject Commit Update The subproject commit for third_party/llama.cpp has been updated from fa226f9377a3025a171aea3eb8839c2f9f052d37 to 00adc2b9d6676f39cb3fcca8b5ad97621e653fae. This update brings in the fixes for the HRX decode-split multipass path fault and the vulkan pinning changes. Ensure that this commit is correctly pinned and that all related changes are properly integrated. Subproject commit 00adc2b9d6676f39cb3fcca8b5ad97621e653fae Architecture Source Update The architecture source for llama.cpp (hrx) has been updated in registry/architectures.json to reflect the new pinned commit 00adc2b9d6676f39cb3fcca8b5ad97621e653fae. This ensures that the registry correctly identifies the source of the HRX backend. "llama.cpp (hrx)": "00adc2b9d6676f39cb3fcca8b5ad97621e653fae",

[github-actions]

Persistent review updated to latest commit 3411558