[bong-water-water-bong]

Moves third_party/llama.cpp from 4380daf (#127) to 567d2dc, the merge of 1bit-MONSTER/llama.cpp#14. It changes the converter only, not the runtime:

• Legacy ZAYA checkpoints. ZAYA1-base, ZAYA1-reasoning-base and Zyphra's *-legacy repos use the original Megatron-style layout, which transformers cannot load. The converter maps them to the transformers layout.
• --remote fetches chat_template.jinja. Remote conversions (such as ZAYA1-74B-preview's) had no chat template before. Upstream has the same gap; a draft patch is in ~/upstream-drafts for the user.

Checked on Strix Halo:

• --remote Zyphra/ZAYA1-8B-legacy converts to a GGUF whose 1,283 tensors are byte-identical to the one from Zyphra/ZAYA1-8B, with identical model metadata.
• The normalized legacy configs equal the new-format ones for the 8B and 74B pairs, including the 74B's sliding window and per-layer rope bases.
• ZAYA1-74B-preview Q4_K_M, llama-bench on Vulkan0: pp512 505.8 tok/s, tg128 35.4 tok/s. Chat with the model's own template gives coherent answers.
• ZAYA1-8B F16 on the new code matches transformers FP32 at 95 of 96 teacher-forced positions, unchanged.

Docs:

• docs/vulkan.md: Pin llama.cpp 4380daf: ZAYA1-74B (sliding-window layers) on Vulkan #127's 74B paragraph becomes a section with the convert, requantize and serve commands, the bench numbers and the chat-template note, plus a section on legacy checkpoints.
• README status: names ZAYA1-74B and the Zyphra family.

registry/architectures.json is regenerated for the new pin; only the recorded sha changes.

ZAYA1-base and ZAYA1-reasoning-base GGUFs are being converted now; their checks will follow in a separate 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 d32bd46

[github-actions]

Preparing review...

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 127 - Partially compliant Compliant requirements: Move third_party/llama.cpp from 1e775cd to 4380daf Add ZAYA1-74B-preview's sliding-window layers (4,097-token window, rope theta 1e4) Add second rope base (1e7) Regenerate registry/architectures.json to record 4380daf Non-compliant requirements: None Requires further human verification: Verification of Q4_K_M perplexity on Vulkan for ZAYA1-8B and ZAYA1-74B-preview Confirmation that test-llama-archs -a zaya passes on Vulkan and CPU Confirmation that tests/serve_e2e.sh passes for ZAYA1-74B-preview on Vulkan 14 - Partially compliant Compliant requirements: Pin ZINC at upstream main c50b4add Build ZINC with Zig 0.15.2 Add documentation for ZINC build and verification Verify Qwen3.5-9B on RTX 5090 with CUDA backend Verify Qwen3.5-0.6B on Strix Halo with Vulkan backend Verify ROCm build and link Non-compliant requirements: None Requires further human verification: Confirmation that the user's fork bong-water-water-bong/zinc commits are not included Confirmation that no Lemonade backend spawns zinc yet

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Inconsistency The documentation states that ZAYA1-74B-preview's Q4_K_M perplexity matches to four decimals with ZAYA1-8B, but the PR description mentions that ZAYA1-8B F16 on the new code matches transformers FP32 at 95 of 96 teacher-forced positions. This discrepancy should be clarified to ensure consistency in the documentation. | Check | Result | |---|---| | `tests/serve_e2e.sh`, `--device vulkan --ctx-size 8192` | PASS | | Chat with the model's own template (code, prose, a one-line fact) | coherent answers | | `llama-bench`, Vulkan0 | prefill (pp512) 505.8 tok/s, decode (tg128) 35.4 tok/s |