HollowMan6/GLM-5-NOOP-LoRA

GLM-5 Empty LoRA Adapter (All-Linear + MoE Experts)

Model Summary

This repository contains an empty-initialized PEFT LoRA adapter for zai-org/GLM-5.

It is intended for:

• LoRA loading/integration tests
• Runtime compatibility checks (PEFT / vLLM)
• A clean initialization starting point before actual LoRA training

This adapter is initialized as a near no-op:

• lora_A: Kaiming-uniform
• lora_B: zeros

So generation quality should be close to the base model before any fine-tuning.

Model Details

• Developed by: This script
• Model type: PEFT LoRA adapter checkpoint
• Base model: zai-org/GLM-5
• Language(s): Same as base model
• License: Same as base model license
• Framework: PEFT

Adapter Construction

This checkpoint was generated programmatically (not fine-tuned from data), targeting:

• all linear-like modules (excluding lm_head)
• detected MoE expert projections (gate_proj, up_proj, down_proj) and gate when available

Intended Use

• Verifying LoRA checkpoint loading
• Testing MoE LoRA plumbing
• Serving/inference pipeline validation

Out-of-Scope Use

• Task performance improvement without training
• Benchmark comparisons against fine-tuned adapters

Training Details

No training was performed.
This is an initialization-only adapter checkpoint.

Evaluation

No task evaluation metrics are reported for this adapter.
Expected behavior is close to the base model due to zero-initialized lora_B.

Risks and Limitations

• Inherits all limitations and biases of the base model.
• Not suitable as a production task adapter without fine-tuning.
• Minor output differences may still appear due to runtime/kernel nondeterminism.

Usage

Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "zai-org/GLM-5"
adapter = "/path/to/this/adapter"

tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)

vLLM

from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

llm = LLM(
    model="zai-org/GLM-5",
    trust_remote_code=True,
    enable_lora=True,
    max_loras=1,
    max_lora_rank=8,  # set >= adapter rank
)

outputs = llm.generate(
    ["Hello!"],
    SamplingParams(temperature=0.0, max_tokens=32),
    lora_request=LoRARequest("empty-lora", 1, "/path/to/this/adapter"),
)
print(outputs[0].outputs[0].text)

Framework Versions

• PEFT 0.18.1