zyr-AGENT / zyr3-v1

ZYR3 v1

ZYR3 v1 is an agentic coding assistant. It behaves as an AI coding assistant rather than only a code generator: it explains code, edits, refactors, debugs, and works through multi-step coding tasks iteratively.

This repository ships two ways to use it:

1. As a servable transformers model (custom code) — loads like the ZYR3 classic repo with trust_remote_code=True . No weights to download; the model delegates to the ZYR3-V1 API configured through environment variables. This is what makes the model appear as a deployable transformers checkpoint on Hugging Face (Amazon / Google Cloud / Microsoft Foundry options).
2. As a PEFT LoRA adapter — merge or load the adapter onto a compatible base model you supply through ZYR3_BASE_MODEL . The underlying base model is not disclosed or distributed; nothing about it is hardcoded here.

Capabilities

• Coding
• Code explanation
• Code editing
• Debugging
• Bug finding
• Refactoring
• Iterative problem solving
• Programming reasoning
• Multi-step coding tasks
• Edge-case handling

Use as a full transformers model (custom code)

Identical to the ZYR3 classic repo: HF sees a servable transformers model with custom_code . The wrapper needs no GPU and no local weights — generation is delegated to the ZYR3-V1 API, configured entirely through the environment:

export ZYR3_API_BASE_URL="https://your-zyr3-v1-endpoint/v1"
export ZYR3_API_KEY="your-api-key"

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "zyr-AGENT/zyr3-v1", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("zyr-AGENT/zyr3-v1")

reply = model.chat("Write a Flask web server")
print(reply)

for chunk in model.chat("Stream this reply", stream=True):
    print(chunk, end="", flush=True)

The endpoint and key are never baked into the repo — they come from the environment or from zyr3_config in config.json .

Environment-driven configuration

Everything is read from the environment at runtime:

Nothing in the files references a concrete base model; set ZYR3_BASE_MODEL to any compatible checkpoint your runtime can access.

Download and use with PEFT + transformers

The adapter downloads directly from this Hugging Face repo (public, no auth needed) with standard transformers + peft . The repo ships adapter_config.json and adapter_model.safetensors , so PeftConfig.from_pretrained("zyr-AGENT/zyr3-v1") resolves out of the box.

import os
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftConfig, PeftModel

base_model = os.environ["ZYR3_BASE_MODEL"]
adapter = os.environ.get("ZYR3_ADAPTER", "zyr-AGENT/zyr3-v1")

config = PeftConfig.from_pretrained(adapter)
assert config.peft_type.name == "LORA"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    torch_dtype="auto",
)

model = PeftModel.from_pretrained(model, adapter)
model.eval()

Or from the shell:

# Set your base model in the environment (or .env / .env.local)
export ZYR3_BASE_MODEL=<your compatible base model>

python infer.py

Example one-shot generation:

import os
import torch

prompt = "def add(a, b):"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Launch in Ollama

Ollama cannot load a PEFT adapter directly, so merge the adapter into the base first, convert to GGUF, then create an Ollama model. All steps are env-driven; run them on a machine with ~20 GB RAM (or a 16 GB GPU).

# 1. Merge adapter into the base model
export ZYR3_BASE_MODEL=<your compatible base model>
python merge_for_ollama.py          # writes ./merged_zyr3

# 2. Convert to GGUF (llama.cpp: https://github.com/ggerganov/llama.cpp)
python llama.cpp/convert_hf_to_gguf.py merged_zyr3 \
    --outfile zyr3-v1-f16.gguf --outtype f16

# 3. Optional: quantize to 4-bit for smaller size
llama.cpp/build/bin/llama-quantize zyr3-v1-f16.gguf \
    zyr3-v1-q4_k_m.gguf q4_k_m

# 4. Create and run the Ollama model
ollama create zyr3-v1 -f ollama/Modelfile
ollama run zyr3-v1 "Write a Flask web server"

ollama/Modelfile uses ChatML formatting ( <|im_start|> / <|im_end|> ) and a ZYR3 system prompt. The FROM line points at zyr3-v1-q4_k_m.gguf ; adjust it if you skip quantization (use ./zyr3-v1-f16.gguf ).

Project

ZYR3

Version: v1