i changed path from previous v1 eagle to this, and now getting 0% of acceptance rate, however everything else keep unchanged
loading eagle like this
import subprocess

def start_vllm_server() -> subprocess.Popen[bytes]:
"""Start vLLM server in the background"""
os.environ["TRANSFORMERS_NO_TF"] = "1"
os.environ["TRANSFORMERS_NO_FLAX"] = "1"
os.environ["VLLM_ATTENTION_BACKEND"] = "TRITON_ATTN"
os.environ["TRITON_PTXAS_PATH"] = "/usr/local/cuda/bin/ptxas"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
# https://docs.vllm.ai/projects/recipes/en/latest/OpenAI/GPT-OSS.html#troubleshooting
os.environ["TIKTOKEN_ENCODINGS_BASE"] = (
"/kaggle/usr/lib/pip_install_aimo3_1/tiktoken_encodings"
)

sequence_length = 65_536

command: list[str] = [
    "python",
    "-m",
    "vllm.entrypoints.openai.api_server",
    "--model",
    "/dev/shm/model",
    "--served-model-name",
    "vllm-model",
    "--tensor-parallel-size",
    "1",
    "--max-num-seqs",
    "4",
    "--gpu-memory-utilization",
    "0.96",
    "--speculative-config",
    '{"method": "eagle3","model": "/dev/shm/eagle","num_speculative_tokens": 3,"draft_tensor_parallel_size": 1}',# any higher may not have enough for graph capture
    "--host",
    "0.0.0.0",
    "--port",
    "8000",
    "--dtype",
    "auto",
    "--max-model-len",
    f"{sequence_length}",
]

# Start the process in the background
with open("/kaggle/working/a-vllm.log", "w") as logfile:
    process: subprocess.Popen[bytes] = subprocess.Popen(
        command, stdout=logfile, stderr=subprocess.STDOUT, start_new_session=True
    )

print("Logs: /kaggle/working/a-vllm.log")
return process

Start the server

vllm_process: subprocess.Popen[bytes] = start_vllm_server()