[bong-water-water-bong]

First step of Zyphra-family support. Zamba2 (Mamba-2 layers plus shared transformer blocks) is the family's most downloaded line (Zamba2-1.2B-instruct alone: 213k).

Vulkan pin third_party/llama.cpp-vulkan → 1bit/vulkan-upstream a32f0d6f:

Commit	What
62484fba	ggml-org#28243, Flash-Next MTP (as before)
5ce79a56	ggml-org#21412 Zamba2 + converter. Reviewed line by line: model code self-contained; the one shared change (is_n_embd_{k,v}_gqa_variable skipping KV-less layers) only affects a warning in v0.5.0; no network/file/process/build changes. Changed lines identical to the PR (530/530).
a32f0d6f	ours: Vulkan SSM_SCAN had pipelines only for d_state 128/256, so Zamba2-2.7B/7B (d_state 64) ran the scan on the CPU in every layer (110 graph splits). Adds the 64 pipeline where it is a whole number of subgroups, plus two test-backend-ops cases.

Measured on Strix Halo (Q8_0, llama-bench):

Zamba2	tg64 before → after	pp256 before → after
1.2B-instruct	88.7 → 89.8	3227 → 3166
2.7B-instruct	4.3 → 45.2	614 → 1541
7B-Instruct	2.1 → 17.1	29 → 523

test-backend-ops SSM_SCAN: 14/14 on Vulkan0. Engine ctest 8/8 (-DONEBIT_VULKAN=ON). 1bit serve --device vulkan + Zamba2-2.7B: "The capital of France is Paris." at 45.6 tok/s. registry_check: PASS for Zamba2 1.2B / 2.7B / 7B (recorded in registry/checked.json).

Registry regenerated from the pins: Zamba2ForCausalLM → vulkan; it also brings the recorded HRX pin up to 5556bf2 (it still said c358334). docs/vulkan.md lists the carried commits and the numbers; NOTICE credits #21412.

Next for the family: ZAYA1-74B / base / reasoning (same zaya architecture), Zamba v1 and BlackMamba (our own implementations), the VL models.

🤖 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 9114757

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

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

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Performance Claim Verification The PR claims significant performance improvements for Zamba2 models on Vulkan (e.g., 2.1 → 17.1 tok/s for 7B-Instruct), but these claims are based on measurements from a specific hardware (Strix Halo) and may not be universally applicable. The PR should clarify if these performance gains are consistent across different Vulkan implementations or if they are specific to the Strix Halo hardware. v0.5.0 + #28243 + #21412 + the d_state-64 scan (`a32f0d6f`); Zamba2 converted with the carried converter (`convert_hf_to_gguf.py --outtype q8_0`), test-backend-ops SSM_SCAN 14/14 on Vulkan0: | Zamba2 (Q8_0, `llama-bench -p 256 -n 64`) | pp256 | tg64 | chat | |---|---|---|---| | 1.2B-instruct | 3166 | 89.8 | "The capital of France is Paris." | | 2.7B-instruct | 1541 (was 614) | 45.2 (was 4.3: scan on the CPU) | correct | | 7B-Instruct (2 shared blocks, RoPE) | 523 (was 29) | 17.1 (was 2.1) | correct, 16.7 tok/s | Architecture Registration The PR registers 'Zamba2ForCausalLM' architecture for the 'vulkan' backend. It's important to verify that this registration is complete and accurate, and that all necessary backend-specific code paths are properly implemented to support this architecture on Vulkan. "Zamba2ForCausalLM": { "gguf": "zamba2", "backends": [ "vulkan" ] },