linuxid10t/Laguna-XS.2-GGUF

Laguna-XS.2 GGUF

GGUF quantizations of poolside/Laguna-XS.2. Converted and tested with a patched fork of llama.cpp that implements LLM_ARCH_LAGUNA from scratch.

Requires the custom fork. Upstream llama.cpp does not support this architecture. Fork: https://github.com/linuxid10t/llama.cpp-add-laguna

Files

File	Quant	Size	Notes
Laguna-XS.2-f16.gguf	f16	~67 GB	Full precision, reference
Laguna-XS.2-Q4_K_M.gguf	Q4_K_M	~20 GB	Recommended for most users

Usage

git clone https://github.com/linuxid10t/llama.cpp-add-laguna
cd llama.cpp-add-laguna && cmake -B build && cmake --build build -j$(nproc)

./build/bin/llama-cli \
  -m Laguna-XS.2-Q4_K_M.gguf \
  --ctx-size 131072 \
  --temp 0 \
  -p "The capital of France is"

Chat / thinking mode

# Thinking on (default) — model prefills <think> and generates a reasoning trace
./build/bin/llama-cli -m Laguna-XS.2-Q4_K_M.gguf -cnv --ctx-size 32768

# Thinking off — direct answer, no reasoning trace
./build/bin/llama-cli -m Laguna-XS.2-Q4_K_M.gguf -cnv --ctx-size 32768 --reasoning off

The chat template, thinking-mode prefix, EOT token (</assistant>, token 24), and streaming parser are all handled automatically — no manual prompt engineering needed.

Architecture

Property	Value
Parameters	33B total, ~3B active per token
Layers	40 (10 global + 30 SWA, 3:1 pattern)
Attention heads	64 (global layers) / 48 (SWA layers)
KV heads	8 (GQA)
Q/K norm	RMSNorm per head
Attention gate	Head-wise softplus gate on o_proj input, SWA layers only
SWA window	512 tokens
Experts	256 routed (top-8) + 1 shared
Dense layers	Layer 0 only
MoE router	Sigmoid + e_score_correction_bias (added at selection, not routing)
Routing scale	moe_routed_scaling_factor with weight normalization
Global RoPE	YaRN: base 500K, factor 32, original_max 4096, β_fast 64, β_slow 1, partial_rotary 0.5
SWA RoPE	Default: base 10K, full rotary
Context length	131,072 tokens

Implementation Notes

Attention gate: SWA layers have an extra self_attn.g_proj that produces a per-head scalar. This is applied as a softplus gate to the SDPA output before o_proj. Global layers have no gate.

Partial rotary: Global layers rotate only the first half of head dimensions (matching partial_rotary_factor=0.5). SWA layers use full rotary. This is handled via llama.cpp's existing n_rot_full / 2 mechanism.

Dual RoPE: Global and SWA layers use entirely different RoPE configs. The fork adds per-layer-type YaRN siblings to llama_hparams / llama_cparams and selects at graph build time via hparams.is_swa(il).

MoE routing: Router logits pass through sigmoid (not softmax), biased by a per-expert e_score_correction_bias that is added only during top-k selection, not during weight computation. Routing weights are L1-normalized before scaling by moe_routed_scaling_factor.

Thinking mode prefill: The chat template injects <think> as the generation prompt prefix when thinking is enabled. The streaming parser is patched to treat this as a delimiter-style boundary (not a generated token), so the <think> tag never appears in streaming output.

Stop token: </assistant> is token 24, a regular vocabulary token (not a special token). The fork adds it to antiprompt so the stop-word erase logic strips it before output.

Tested

• Greedy decode on "The capital of France is" → " Paris.\nThe capital of Germany is" ✓
• 10K-token prompt (exercises SWA layers) — no corruption ✓
• 35K needle-in-haystack (SWORDFISH99) — correctly retrieved (exercises global YaRN layers) ✓

Known Limitations

• Numerical validation against HF Transformers not yet done (requires CUDA/ROCm).
• Q4_K_M vs f16 top-1 token agreement not formally checked (f16 GGUF exceeds available RAM for full eval).