vLLM's TransformersForCausalLM backend registers FP8 parameters as input_scale/weight_scale and errors on other names. Transformers 5 and Unsloth expect activation_scale/weight_scale_inv. Neither tolerates the other's names.

Using this repo for LoRA training ensures the adapter trains against the true FP8 ground truth weights — the same values used at serving time. No dequant/re-quant mismatch.

Usage with Unsloth

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "levara/Devstral-Small-2-24B-TextOnly-FP8-Training",
    max_seq_length=8192,
    load_in_4bit=False,
)

model = FastLanguageModel.get_peft_model(model, r=16, target_modules=[
    "q_proj", "k_proj", "v_proj", "o_proj",
    "gate_proj", "up_proj", "down_proj",
])

Serving

For vLLM serving, use the companion checkpoint: levara/Devstral-Small-2-24B-TextOnly-FP8

bash

vllm serve levara/Devstral-Small-2-24B-TextOnly-FP8 \
    --tensor-parallel-size 2 \
    --max-model-len 32768 \
    --enable-lora \
    --lora-modules my-adapter=path/to/adapter

Model Details

Property	Value
Architecture	`Ministral3ForCausalLM`
Parameters	23.57B
Quantization	FP8 W8A8 static (`float8_e4m3fn`)
Layers	40
Hidden size	5120
Context length	393K tokens (YaRN RoPE)