Modelpublic

giannisan/Laguna-S-2.1-ds4-gguf

sourceHugging Faceupdated 2mo agoView on Hugging Face

4likes170downloads

Model Card

Laguna-S-2.1 (118B) GGUF for pulsar / ds4 (SSD streaming, CUDA)

Mixed-precision GGUF of poolside/Laguna-S-2.1 (118B total / 8B active MoE, 256 experts top-10 + 1 shared) built for SSD-streaming inference engines: routed experts live on disk and stream per token, so the model runs on GPUs that cannot hold it.

Runs on pulsar (Rust + CUDA).

Measured on RTX 5060 Ti 16GB + RTX 4060 Ti 16GB, Ryzen 9900X, Gen5 NVMe (greedy, warm cache):

metric	pulsar
decode	17.3 tok/s
decode + CPU expert lane (`PULSAR_CPU=1`)	22.4 tok/s
prefill (328-token prompt)	~53 tok/s

Per token only 10 of 256 experts per layer are read; attention, the shared expert, and the router stay resident. At this quant ~15.5GB of hot experts stay VRAM-resident (85% cache hit rate on the decode path), which is why the 2-bit build is both smaller and faster than the Q4KM edition. Laguna's hybrid attention (one full-window layer every fourth, sliding-window 512 for the rest), per-head output gate, and per-layer-type RoPE (YaRN on full layers, plain on sliding) are all handled by the engine.

Files

file	provenance	recommendation
`Laguna-S-2.1-IQ2XXS-RouterF32-AttnQ8-SExpQ8-OutQ6.gguf`	routed gate/up experts IQ2_XXS with an importance matrix (36 GB)	use this one
`Laguna-S-2.1-Q2K-RouterF32-AttnQ8-SExpQ8-OutQ6.gguf`	routed experts Q2_K, no imatrix (41 GB)	kept for continuity

Both quantize straight from the BF16 checkpoint (235 GB) and keep an identical high-precision layout for everything that steers a token. They differ only in the routed-expert codec: the IQ2XXS build uses a general-purpose importance matrix so the 2-bit gate/up weights keep the dimensions that matter; the Q2K build is uniform 2-bit with no calibration. The imatrix file is smaller and closer to the source.

Recipe

The layout targets a streaming expert cache: routed experts must be uniform fixed-size slabs, and everything that steers a token stays high precision. Same design as antirez's GLM-5.2 ds4 build.

Tensors	Type	Why
routed experts, gate/up	IQ2XXS (imatrix) / Q2K	streamed from disk per token; uniform slabs
routed experts, down	Q2_K	streamed per token; down is the sensitive one
shared expert + dense layer 0 FFN	Q8_0	resident, paid once
attention q/k/v/output, all layers	Q8_0	resident (source ships F16, converted at quant time)
token embeddings, output head	Q6_K	resident
router (ffngateinp), expert bias, all norms	F32	decision makers stay exact

The router stays F32 because pulsar runs it through an F32 matmul; a quantized ffn_gate_inp is read as raw floats and destroys routing.

Usage

pulsar

git clone https://github.com/giannisanni/pulsar
cd pulsar
CXX=g++-12 cargo build --release -p engine

./target/release/pulsar-cli -m Laguna-S-2.1-IQ2XXS-RouterF32-AttnQ8-SExpQ8-OutQ6.gguf \
    -p "The capital of France is" -n 64
# or the OpenAI-compatible API + web UI: cargo build --release -p serve

Zero-config multi-GPU: pulsar measures each card's PCIe bandwidth at startup, streams experts over the fastest link, and fills spare GPUs with resident hot experts. First run is cold; a .warm sidecar makes every later run start hot. Main knob: PULSAR_CACHE_GB (host expert cache, defaults to measured free RAM minus a reserve).