Pin llama.cpp 4380daf: ZAYA1-74B (sliding-window layers) on Vulkan

[bong-water-water-bong]

Moves third_party/llama.cpp from 1e775cd (#119) to 4380daf, the merge of 1bit-MONSTER/llama.cpp#11. That adds ZAYA1-74B-preview's sliding-window layers (4,097-token window, rope theta 1e4) and its second rope base (1e7). The pin is the head of 1bit/hrx-vulkan-patched.

Checked on Strix Halo:

• ZAYA1-8B unchanged. Q4_K_M perplexity on Vulkan is 21.5731 before and after, and test-llama-archs -a zaya passes on Vulkan and CPU.
• ZAYA1-74B-preview. Q4_K_M (45.7 GB) passes tests/serve_e2e.sh through 1bit serve --device vulkan --ctx-size 8192, run under mem-guard.sh. There's no reference check: a 74B FP32 transformers run doesn't fit on this machine, and docs/vulkan.md says so.

registry/architectures.json is regenerated. It had not been updated when #119 moved the pin, and still named 5556bf2; it now records 4380daf.

🤖 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 bed4e9f

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit bed4e9f)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 119 - Partially compliant Compliant requirements: Bump third_party/llama.cpp to pick up fixes for HRX issues HRX: dense on HRX0 with experts on another device fails for Qwen MoE models (fused router) #108 and HRX: flash-attention decode-split template resolution fails ("all_rejected") at longer context (~3800+ tokens) #115 Fix route-ids binding-length crash in llama.cpp#7 Fix fused router's launch geometry issue in llama.cpp#12 Fix capacity-cap fix for HRX flash-attention decode-split in llama.cpp#9 Verify fixes with exact repro and end-to-end testing Non-compliant requirements: None Requires further human verification: None 11 - Partially compliant Compliant requirements: Move engine's llama.cpp to AMD's maintained HRX line Pin to AMD's tested pair and keep current with daily workflow Update submodules and build system Update device name from HRX20 to HRX0 Update documentation and tests Verify on Strix Halo with speed improvements 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 mentions ZAYA1-74B-preview's sliding-window layers and rope theta values, but does not specify the exact impact of these changes on performance or memory usage beyond the context of the vulkan device. Clarification on how these parameters affect inference behavior would be beneficial for users. ZAYA1-74B-preview (1bit-MONSTER/llama.cpp #11) alternates sliding-window layers (a 4,097-token window, rope theta 1e4) with full-attention layers (theta 1e7); the converter writes the window, the per-layer pattern and the second rope base, and ZAYA1-8B, which has no sliding layers, is unchanged (same perplexity to four decimals). Its Q4_K_M (45.7 GB) passes `tests/serve_e2e.sh` on `--device vulkan` with `--ctx-size 8192`; there is no reference check for it here, since a 74B FP32 transformers run does not fit on Strix Halo.

[github-actions]

Persistent review updated to latest commit bed4e9f