You can deploy this model and start earning money today!

volkancirik/
Laguna-XS.2-spatial-grounding-lora

Name: Laguna-XS.2-spatial-grounding-lora
Availability: InStock

Views

No views yet

Laguna-XS.2 — Spatial Grounding LoRA

A LoRA adapter for poolside/Laguna-XS.2 (33B-total / 3B-active Mixture-of-Experts), trained with reinforcement learning on PrimeIntellect hosted training against a spatial-grounding environment.

Adapter configuration

Field	Value
`peft_type`	`LORA`
`task_type`	`CAUSAL_LM`
`r`	16
`lora_alpha`	32
`lora_dropout`	0.0
`target_modules`	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`, `experts`
base model	`poolside/Laguna-XS.2`

The adapter targets attention projections and the MoE expert / MLP projections.

Serving with vLLM

Laguna XS.2 has launch-day support in vLLM (>= 0.21.0). Serve the base model with this adapter loaded as a LoRA module:

bash

pip install 'vllm>=0.21.0'

vllm serve poolside/Laguna-XS.2 \
    --enable-lora \
    --lora-modules spatial=volkancirik/Laguna-XS.2-spatial-grounding-lora \
    --tool-call-parser poolside_v1 \
    --reasoning-parser poolside_v1 \
    --enable-auto-tool-choice \
    --max-lora-rank 16 \
    --served-model-name laguna

Note: LoRA over MoE expert layers is not supported by every serving stack. If vLLM rejects the expert-targeted modules, merge the adapter into the base weights first (PeftModel.merge_and_unload() on a GPU/large-RAM host) and serve the merged checkpoint.

Then query the OpenAI-compatible endpoint (http://localhost:8000/v1/chat/completions), passing spatial as the model name.

Loading with PEFT / Transformers

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "poolside/Laguna-XS.2"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "volkancirik/Laguna-XS.2-spatial-grounding-lora")

Requires transformers >= 5.7.0 (Laguna support) and peft.

Training

Platform: PrimeIntellect hosted RL (prime-rl)
Base: poolside/Laguna-XS.2
Domain: spatial grounding (relative-position reasoning over rendered scenes)
Checkpoint id: nm0otq2i6zkmwk6xxo91zcg5