Turjman-RL-merged

Turjman-RL Merged

This is the merged version of Yousefbahr/Turjman-RL, a LoRA adapter fused into Yousefbahr/Turjman-Cold-Start using peft.merge_and_unload().

The merged weights are mathematically identical to running the original LoRA with PEFT. This version is compatible with vLLM for fast batched inference.

Usage with vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="0x0515/Turjman-RL-merged",
    dtype="float16",
    tensor_parallel_size=2,   # Kaggle 2xT4
    gpu_memory_utilization=0.65,
    max_model_len=2048,
)