fr0stbit3/laya-gguf

GGUF conversion of convaiinnovations/laya (Apache-2.0), f16, via llama.cpp convert_hf_to_gguf.py. laya-F16.gguf (and the quants) hold the ModernBERT backbone only, and load in llama.cpp as modern-bert (use --embeddings --pooling none). The decision head and config are in the sibling file laya-head.safetensors (like an mmproj file); read it with laya_head.py (load_head(path) returns config + head weights under the original PyTorch names). The head itself must run outside llama.cpp. Lightly tested: outputs matched the original on a limited set (100 AG News + 100 DAIR Emotion samples, a few examples); may need further testing. Quantized files (laya-Q8_0.gguf, -Q6_K.gguf, -Q4_K_M.gguf; laya-F16.gguf is the unquantized 16-bit original conversion) are made with llama-quantize from the f16 file and are even less tested: checked on a single example only, where Q8_0/Q6_K stayed close to f16 and Q4_K_M drifted slightly more (probabilities shifted by up to about 0.02 to 0.03, and score outputs by about 0.04). Prefer f16 or Q8_0 for anything important, and verify Q4_K_M on your own data.

Quant check (single example: the HF README ticket "Duplicate charge on invoice 4411"; department=billing probability, urgency score 0-2, churn_risk noul; head weights from the head file, backbone in llama.cpp). One example only, not a benchmark.

file	size	billing p	urgency	churn
HF original (PyTorch)	-	0.967	1.630	0.198
F16	933M	0.966	1.631	0.201
Q8_0	563M	0.967	1.637	0.198
Q6_K	485M	0.970	1.650	0.198
Q4_K_M	414M	0.967	1.601	0.219

llama.cpp quickstart

# 1. serve the backbone (per-token hidden states; -ub must cover your longest input)
llama-server -m laya-F16.gguf --embeddings --pooling none -c 2048 -ub 2048 -b 2048 --port 8080

# 2. python deps for the decision head + tokenizer
pip install laya requests torch safetensors

# quickstart.py: llama.cpp backbone + Laya head from the sibling head file
import types, requests, torch, laya
from laya_head import load_head

HEAD = "laya-head.safetensors"
agent = laya.load("convaiinnovations/laya", device="cpu")  # builds the head architecture + tokenizer/prompt logic

# swap in the decision-head weights from the head file
_, head = load_head(HEAD)
sd = agent.model.state_dict()
for k, v in head.items():
    if k in sd: sd[k].copy_(v)

# swap the PyTorch encoder for the llama.cpp server
D = agent.model.encoder.config.hidden_size
def llamacpp_encoder(input_ids, attention_mask=None, **_):
    out = []
    for i, row in enumerate(input_ids):
        n = int(attention_mask[i].sum())
        r = requests.post("http://localhost:8080/embedding", json={"content": [row[:n].tolist()]}).json()
        h = torch.zeros(input_ids.shape[1], D); h[:n] = torch.tensor(r[0]["embedding"]); out.append(h)
    return types.SimpleNamespace(last_hidden_state=torch.stack(out))
agent.model.encoder.forward = llamacpp_encoder

result = agent.predict(
    {"subject": "Duplicate charge on invoice 4411",
     "body": "We were billed twice for March. Please refund the duplicate."},
    {"department": {"type": "choice", "instructions": "Which team should handle this?",
                    "criteria": {"billing": "invoices, payments, refunds",
                                 "technical": "bugs and outages", "sales": "pricing"}}},
)
print(result["answers"]["department"]["choice"])  # billing

Notes: the laya package still downloads the original weights once (for the head architecture and tokenizer); inference runs the backbone in llama.cpp. Tokenize with the HF tokenizer and send token ids ("content": [ids]) as above. A native Go/C++ head is not provided.

Original model

Full model card, usage, benchmarks and license terms: convaiinnovations/laya.