[bong-water-water-bong]

Next step of Zyphra-family support. third_party/llama.cpp-vulkan → 1bit/vulkan-upstream f6261222 (the current ee6e4e0d + one commit of ours): model: BlackMamba (Zyphra/BlackMamba-1.5B, -2.8B), our code in src/models/blackmamba.cpp / conversion/blackmamba.py.

• Mamba-1 layers alternate with a Switch MoE: router (with bias) → sigmoid → top-1 expert, scaled by its probability; experts are gated erf-GELU MLPs (ggml_geglu_erf_split; the config says swiglu but the activation is F.gelu).
• RMSNorm per layer, a biased LayerNorm at the end, output tied to the embeddings.
• No transformers implementation exists: the checkpoint is Megatron-style ({"model": state_dict}, config.json without architectures), so conversion/base.py routes on mamba_moe_layers (next to the existing ssm_cfg case) and the converter uses the built-in GPT-NeoX vocab.

Check (Strix Halo)	Result
Teacher-forced top-1 vs Zyphra's own PyTorch code (FP32, CPU; Triton/CUDA ops swapped for plain PyTorch)	96/96
Greedy text and prompt tokens vs the reference	identical
BlackMamba-1.5B F16	pp256 7180, tg64 250 tok/s
BlackMamba-1.5B Q8_0	pp256 10136, tg64 369 tok/s

Registry: BlackMambaForCausalLM → vulkan. It is a base model, so there is no registry_check row (that check asks a chat question).

🤖 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 b66994a

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

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

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Missing Performance Benchmark Details The PR description includes performance benchmarks for BlackMamba on Vulkan, but the specific hardware and conditions under which these benchmarks were measured are not clearly stated. This makes it difficult to assess the validity and reproducibility of the claimed performance numbers (e.g., pp256 7180, tg64 250 tok/s). BlackMamba (`f6261222`), `Zyphra/BlackMamba-1.5B`: teacher-forced top-1 vs Zyphra's own PyTorch code (FP32, CPU) **96/96** on three prompts, identical greedy text and prompt tokens; F16 pp256 7180 / tg64 250 tok/s, Q8_0 pp256 10136 / tg64 **369** tok/s. No HF `architectures` in its config.json: the converter routes on `mamba_moe_layers`. Architecture Registry Update The registry update adds BlackMambaForCausalLM to the Vulkan backend, but there is no corresponding entry in the architectures section for BlackMambaForCausalLM in the src/models/blackmamba.cpp file or any other model-specific code. This could lead to inconsistencies in how the model is handled across different parts of the system. "BlackMambaForCausalLM": { "gguf": "blackmamba", "backends": [ "vulkan" ] },