[bong-water-water-bong]

Docs only. Two additions to docs/vulkan.md.

The two legacy-format ZAYA base models, now convertible (llama.cpp #14). Checked on Strix Halo, Q4_K_M on Vulkan0, wikitext-2 test (60 × 512):

Model	Perplexity	Decode
ZAYA1-base	8.55 ± 0.17
ZAYA1-reasoning-base	10.15 ± 0.22	89–91 tok/s in chat, with its own template
ZAYA1-8B, for comparison	32.13 ± 1.08

The multi-sequence fix already pinned on main (llama.cpp #15). ZAYA returned wrong results whenever a ubatch held more than one sequence, which includes concurrent requests to 1bit serve.

🤖 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 a7bcf6b

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 14 - Partially compliant Compliant requirements: Pin ZINC at upstream main c50b4add Build ZINC with scripts/build-zinc.sh for Vulkan, ROCm, and CUDA Verified ZINC on Vulkan and ROCm Verified ZINC CUDA backend on RTX 5090 with Qwen3.5-9B Updated docs/zinc.md and docs/PORTING.md Added .github/workflows/bump-zinc.yml Non-compliant requirements: None Requires further human verification: CUDA verification on RTX 5090 (already mentioned in PR description)

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Documentation Clarity The documentation mentions that ZAYA1-reasoning-base answers chat prompts with its own template, thinking first. However, it does not specify what the actual prompt or template is, which could be important for reproducibility or understanding the behavior. Clarifying this would improve the documentation's utility. | ZAYA1-reasoning-base | 10.15 ± 0.22 | 89-91 tok/s (chat) |