NameError in Unsloth-generated Linear_peft_forward.py: VARIANT_KWARG_KEYS not defined

[Nyovelt]

Bug Report

• Summary: The Unsloth-generated Linear_peft_forward.py references VARIANT_KWARG_KEYS when constructing variant_kwargs, but the constant is never imported or defined, so every LoRA Linear forward immediately crashes with NameError.
• Environment: Python 3.10, torch ≥2.8.0, transformers 4.56.2, latest unsloth/unsloth_zoo from Git (ROCm workstation, but this is a pure-Python failure).
• GPU: AMD RYZEN AI MAX+ 395 w/ Radeon 8060S, rocm 7.0+,
• Steps to Reproduce:
  1. Generate any LoRA-wrapped model with Unsloth (e.g., run examples/gpt_oss_(20B)_Reinforcement_Learning_2048_Game_BF16.py).
  2. When the decoder calls a LoRA projection, unsloth_forward in Linear_peft_forward.py executes.
  3. The function evaluates variant_kwargs = {k: kwargs.pop(k, None) for k in VARIANT_KWARG_KEYS} and Python raises NameError, halting the run.
• Minimal Reproduction (standalone):

import textwrap, types, torch

LINEAR_SOURCE = textwrap.dedent(
    """
    import torch
    def unsloth_forward(self, x, *args, **kwargs):
        variant_kwargs = {k: kwargs.pop(k, None) for k in VARIANT_KWARG_KEYS}
        return self.base_layer(x, *args, **kwargs)
    """
)

mod = types.ModuleType("broken_linear")
exec(LINEAR_SOURCE, mod.__dict__)

class Dummy:
    disable_adapters = False
    merged = False
    active_adapters = []
    lora_A = lora_B = lora_dropout = {}
    scaling = {}
    lora_variant = {}
    def base_layer(self, x, *args, **kwargs):
        return x

mod.unsloth_forward(Dummy(), torch.zeros(1, 1))

Running this script reproduces the NameError.

• Expected Result: variant_kwargs should be built successfully so the base layer forward completes.
• Actual Result: NameError prevents any LoRA projection from running.
• Proposed Fix: Have the generator import VARIANT_KWARG_KEYS from peft.tuners.lora.layer, fallback to peft.tuners.lora.bnb, or default to ["alora_offsets"] if neither import is available.

[enesbol]

Replace first block as following

import os, re

os.environ["UNSLOTH_COMPILE"] = "0"  # Add this line

if "COLAB_" not in "".join(os.environ.keys()):
    !pip install unsloth
else:
    import torch; v = re.match(r"[0-9]{1,}\.[0-9]{1,}", str(torch.__version__)).group(0)
    xformers = "xformers==" + ("0.0.33.post1" if v=="2.9" else "0.0.32.post2" if v=="2.8" else "0.0.29.post3")
    !pip install --no-deps bitsandbytes accelerate {xformers} peft==0.13.2 trl triton cut_cross_entropy unsloth_zoo
    !pip install sentencepiece protobuf "datasets==4.3.0" "huggingface_hub>=0.34.0" hf_transfer
    !pip install --no-deps unsloth

!pip install transformers==4.56.2
!pip install --no-deps trl==0.22.2
!pip install librosa soundfile evaluate jiwer torchcodec "datasets>=3.4.1,<4.0.0"

[shimmyshimmer]

Apologies @Nyovelt it should be fixed now! Could you please update unsloth? Thanks!

[Nyovelt]

Apologies @Nyovelt it should be fixed now! Could you please update unsloth? Thanks!

Thanks! I will do a quick try these days.