third_party: bump llama.cpp pin for engine#108 and engine#115 HRX fixes

[bong-water-water-bong]

Bumps third_party/llama.cpp to pick up every landed fix for both currently-open HRX issues. The pin was 4 commits behind 1bit/hrx-vulkan-patched.

Closes #108, closes #115.

engine#108 — dense on HRX0, MoE experts on another device (fused router)

Two independent bugs, both now fixed on the upstream fork:

• Route-ids binding-length crash — llama.cpp#7.
• Wrong output from the fused router's launch geometry (half the experts' logits never computed on gfx1151) — llama.cpp#12. PR Step 2: HRX + Vulkan in one llama.cpp build, wired into Lemonade #7's description claimed this was included; it wasn't (verified by reading the merged diff) — llama.cpp#12 is the real fix, found and applied this session.

Verified against the exact repro (-dev HRX0,Vulkan0 -ts 1,0 -ot exps=Vulkan0, Qwen3-Coder-30B-A3B-Instruct-Q4_K_M): correct and deterministic across a variety of prompts (arithmetic, factual, multi-turn) with cache_prompt:false, and GGML_HRX_DISABLE_DISPATCH=moe_router reproduces the issue's documented fallback error ("Compute error.") exactly, confirming the fix targets the dispatch actually being exercised.

engine#115 — HRX flash-attention decode-split all_rejected above 2048 KV tokens

Capacity-cap fix — llama.cpp#9. The decode-split matcher now declines above the kernel corpus's real 2048-token capacity and falls through to the general (non-split) flash_attention_f32_f16_wmma dispatch instead of crashing. There is a measured ~32% decode speed cost above that boundary (further declining with depth); a follow-up attempt to recover it with a two-dispatch producer/reducer design was found to have a race/aliasing bug during rigorous re-testing and was retracted (not part of this PR).

Verification

Built onebit (-DONEBIT_HRX=ON) against the bumped pin and re-ran both fixes end to end before opening this PR.

🤖 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 a4fc66a

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 108 - Partially compliant Compliant requirements: Fix for binding length issue in MoE router dispatch Fix for wrong output due to fused router launch geometry Verified with exact repro case Confirmed fallback behavior with GGML_HRX_DISABLE_DISPATCH=moe_router Non-compliant requirements: None Requires further human verification: End-to-end verification of the fix with the exact repro case 115 - Partially compliant Compliant requirements: Fix for flash-attention decode-split template resolution failure Prevents crashes by declining above kernel corpus capacity and falling back to general dispatch Non-compliant requirements: None Requires further human verification: Verification that the fix works at the exact token threshold (~3800 tokens) Performance testing to confirm ~32% decode speed cost is acceptable 7 - Partially compliant Compliant requirements: Bumped llama.cpp to include HRX fixes Build configuration with -DONEBIT_HRX=ON is supported Kernel artifacts are byte-identical Speed parity verified Lemonade wiring with hrx_device option implemented Non-compliant requirements: None Requires further human verification: End-to-end testing with Lemonade to confirm device routing works correctly

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Subproject commit update The PR updates the third_party/llama.cpp subproject commit to include fixes for HRX issues #108 and #115. While this is a necessary change to incorporate the upstream fixes, it's important to verify that all changes in the new commit are compatible with the current engine's usage and that no regressions are introduced. Specifically, ensure that the MoE router fixes and flash attention fixes are correctly applied and tested. Subproject commit 1e775cdf6483140ad2d852c27413aafaee32ccf3