A GGUF build of the BOOMI Stage-A model (text → floor-plan SPEC JSON), made by merging the LoRA adapter BDivyesh/boomi-stage-a-1.5b into its Apache-2.0 base Qwen2.5-Coder-1.5B and quantizing to Q8_0 (~1.65 GB).

This is the file that powers the live BOOMI demo Space: llama-cpp-python runs it on a free 2-vCPU CPU much faster than full-precision transformers, and a JSON grammar (GBNF) guarantees the output is always valid SPEC JSON.

Use (llama-cpp-python)

python

from huggingface_hub import hf_hub_download
from llama_cpp import Llama, LlamaGrammar
path = hf_hub_download("BDivyesh/boomi-stage-a-1.5b-gguf", "boomi-stage-a-1.5b-q8_0.gguf")
llm = Llama(model_path=path, n_ctx=2048, n_threads=2)
# Prompt: BOOMI ChatML (system = stage_a_common.SYSTEM_PROMPT, user = the description).
# Constrain to a single JSON object so the output is always a valid SPEC, then feed it
# to the CP-SAT solver -> editable DXF.

The SPEC → geometry step is a separate OR-Tools CP-SAT solver (valid-by-construction); see the demo Space for the full flow.

Provenance

Base Qwen/Qwen2.5-Coder-1.5B (Apache-2.0). LoRA SFT on ResPlan (MIT) + RERA-derived Indian per-dwelling rows (PII-stripped, geometry-only) + grammar captions. Fully permissive. Drafts only — not for sanction/construction without licensed review.