clover-2 / serving /serve_example.py
haha-hehe's picture
Upload serving/serve_example.py with huggingface_hub
034cc6f verified
Raw History Blame Contribute Delete
3.7 kB
"""Minimal example: load Clover 2's base model, attach the 7 trained LoRA experts and the trained router for
real per-token dynamic routing, then generate.
Requires: vllm, torch. Tested against vllm==0.31.0.
IMPORTANT -- compilation_config: torch.compile silently drops every LoRA delta under CUDA graph capture
unless it's disabled this way. Stubbing GPUModelRunner.capture_model only defers the CUDA-graph-REPLAY-buffer
step; it does NOT defer torch.compile's own Dynamo tracing, which vLLM triggers during LLM()'s own internal
profiling/warmup -- i.e. BEFORE attach() ever runs, against the ORIGINAL unpatched mlp. That compiled graph
then gets reused by compile_or_warm_up_model() regardless of the live monkey-patch, silently dropping every
LoRA delta. Disabling torch.compile (mode=0) and relying on vLLM's CUDA graph capture directly sidesteps this
entirely: the capture-time warmup call then executes the patched Python forward directly.
"""
import os
os.environ['VLLM_ENABLE_V1_MULTIPROCESSING'] = '0'
import torch
import vllm.v1.worker.gpu.model_runner as MR
# capture_model must be stubbed out BEFORE LLM() builds the model, so graph capture doesn't run against the
# original (unpatched) mlp. Restored and called manually after attach()+load_adapter(), below.
_orig_capture_model = MR.GPUModelRunner.capture_model
MR.GPUModelRunner.capture_model = lambda self, *a, **kw: 0
from vllm import LLM, SamplingParams
import vllm_moelora as V
BASE_MODEL = "./base" # path to the base checkpoint (this repo's base/ subfolder)
ADAPTERS = {
1: "./adapters/roleplay/adapter_model.safetensors",
2: "./adapters/clover/adapter_model.safetensors",
3: "./adapters/security/adapter_model.safetensors",
4: "./adapters/math/adapter_model.safetensors",
5: "./adapters/longdoc2/adapter_model.safetensors",
6: "./adapters/efficiency/adapter_model.safetensors",
7: "./adapters/knowledge/adapter_model.safetensors",
}
ROUTER_PATH = "./router/trained_router_v2.pt"
# Security (slot 3) is masked off by default -- enabling it measurably increases compliance with offensive
# (dual-use) security-research requests. Remove 3 from this list to re-enable it.
DEFAULT_MASKED_SLOTS = [3]
def load_safetensors_adapter(path):
from safetensors.torch import load_file
return load_file(path)
def main():
llm = LLM(
model=BASE_MODEL,
dtype="bfloat16",
gpu_memory_utilization=0.85,
trust_remote_code=False,
additional_config={"gdn_prefill_backend": "triton"},
compilation_config={"mode": 0, "cudagraph_mode": "FULL"},
)
worker = llm.llm_engine.model_executor.driver_worker.worker
model = worker.model_runner.model
moles = V.attach(model, n_experts=7, rank=128, scaling=2.0, topk=2)
for slot, path in ADAPTERS.items():
V.load_adapter(moles, slot, load_safetensors_adapter(path))
router_state = torch.load(ROUTER_PATH, map_location="cpu", weights_only=False)
for layer, m in moles.items():
sd = router_state[layer]
m.router.weight.data.copy_(sd["weight"].to(m.router.weight.device, m.router.weight.dtype))
m.router.bias.data.copy_(sd["bias"].to(m.router.bias.device, m.router.bias.dtype))
V.set_mask(moles, DEFAULT_MASKED_SLOTS)
MR.GPUModelRunner.capture_model = _orig_capture_model
worker.compile_or_warm_up_model()
sp = SamplingParams(temperature=0.8, max_tokens=300)
outs = llm.chat(
[[{"role": "user", "content": "Write a short scene where a detective questions a nervous witness."}]],
sp,
chat_template_kwargs={"enable_thinking": False},
)
print(outs[0].outputs[0].text)
if __name__ == "__main__":
main()