Upload serving/serve_example.py with huggingface_hub
Browse files- serving/serve_example.py +87 -0
serving/serve_example.py
ADDED
|
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Minimal example: load Clover 2's base model, attach the 7 trained LoRA experts and the trained router for
|
| 2 |
+
real per-token dynamic routing, then generate.
|
| 3 |
+
|
| 4 |
+
Requires: vllm, torch. Tested against vllm==0.31.0.
|
| 5 |
+
|
| 6 |
+
IMPORTANT -- compilation_config: torch.compile silently drops every LoRA delta under CUDA graph capture
|
| 7 |
+
unless it's disabled this way. Stubbing GPUModelRunner.capture_model only defers the CUDA-graph-REPLAY-buffer
|
| 8 |
+
step; it does NOT defer torch.compile's own Dynamo tracing, which vLLM triggers during LLM()'s own internal
|
| 9 |
+
profiling/warmup -- i.e. BEFORE attach() ever runs, against the ORIGINAL unpatched mlp. That compiled graph
|
| 10 |
+
then gets reused by compile_or_warm_up_model() regardless of the live monkey-patch, silently dropping every
|
| 11 |
+
LoRA delta. Disabling torch.compile (mode=0) and relying on vLLM's CUDA graph capture directly sidesteps this
|
| 12 |
+
entirely: the capture-time warmup call then executes the patched Python forward directly.
|
| 13 |
+
"""
|
| 14 |
+
import os
|
| 15 |
+
os.environ['VLLM_ENABLE_V1_MULTIPROCESSING'] = '0'
|
| 16 |
+
import torch
|
| 17 |
+
import vllm.v1.worker.gpu.model_runner as MR
|
| 18 |
+
|
| 19 |
+
# capture_model must be stubbed out BEFORE LLM() builds the model, so graph capture doesn't run against the
|
| 20 |
+
# original (unpatched) mlp. Restored and called manually after attach()+load_adapter(), below.
|
| 21 |
+
_orig_capture_model = MR.GPUModelRunner.capture_model
|
| 22 |
+
MR.GPUModelRunner.capture_model = lambda self, *a, **kw: 0
|
| 23 |
+
|
| 24 |
+
from vllm import LLM, SamplingParams
|
| 25 |
+
import vllm_moelora as V
|
| 26 |
+
|
| 27 |
+
BASE_MODEL = "./base" # path to the base checkpoint (this repo's base/ subfolder)
|
| 28 |
+
ADAPTERS = {
|
| 29 |
+
1: "./adapters/roleplay/adapter_model.safetensors",
|
| 30 |
+
2: "./adapters/clover/adapter_model.safetensors",
|
| 31 |
+
3: "./adapters/security/adapter_model.safetensors",
|
| 32 |
+
4: "./adapters/math/adapter_model.safetensors",
|
| 33 |
+
5: "./adapters/longdoc2/adapter_model.safetensors",
|
| 34 |
+
6: "./adapters/efficiency/adapter_model.safetensors",
|
| 35 |
+
7: "./adapters/knowledge/adapter_model.safetensors",
|
| 36 |
+
}
|
| 37 |
+
ROUTER_PATH = "./router/trained_router_v2.pt"
|
| 38 |
+
|
| 39 |
+
# Security (slot 3) is masked off by default -- enabling it measurably increases compliance with offensive
|
| 40 |
+
# (dual-use) security-research requests. Remove 3 from this list to re-enable it.
|
| 41 |
+
DEFAULT_MASKED_SLOTS = [3]
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def load_safetensors_adapter(path):
|
| 45 |
+
from safetensors.torch import load_file
|
| 46 |
+
return load_file(path)
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def main():
|
| 50 |
+
llm = LLM(
|
| 51 |
+
model=BASE_MODEL,
|
| 52 |
+
dtype="bfloat16",
|
| 53 |
+
gpu_memory_utilization=0.85,
|
| 54 |
+
trust_remote_code=False,
|
| 55 |
+
additional_config={"gdn_prefill_backend": "triton"},
|
| 56 |
+
compilation_config={"mode": 0, "cudagraph_mode": "FULL"},
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
worker = llm.llm_engine.model_executor.driver_worker.worker
|
| 60 |
+
model = worker.model_runner.model
|
| 61 |
+
moles = V.attach(model, n_experts=7, rank=128, scaling=2.0, topk=2)
|
| 62 |
+
|
| 63 |
+
for slot, path in ADAPTERS.items():
|
| 64 |
+
V.load_adapter(moles, slot, load_safetensors_adapter(path))
|
| 65 |
+
|
| 66 |
+
router_state = torch.load(ROUTER_PATH, map_location="cpu", weights_only=False)
|
| 67 |
+
for layer, m in moles.items():
|
| 68 |
+
sd = router_state[layer]
|
| 69 |
+
m.router.weight.data.copy_(sd["weight"].to(m.router.weight.device, m.router.weight.dtype))
|
| 70 |
+
m.router.bias.data.copy_(sd["bias"].to(m.router.bias.device, m.router.bias.dtype))
|
| 71 |
+
|
| 72 |
+
V.set_mask(moles, DEFAULT_MASKED_SLOTS)
|
| 73 |
+
|
| 74 |
+
MR.GPUModelRunner.capture_model = _orig_capture_model
|
| 75 |
+
worker.compile_or_warm_up_model()
|
| 76 |
+
|
| 77 |
+
sp = SamplingParams(temperature=0.8, max_tokens=300)
|
| 78 |
+
outs = llm.chat(
|
| 79 |
+
[[{"role": "user", "content": "Write a short scene where a detective questions a nervous witness."}]],
|
| 80 |
+
sp,
|
| 81 |
+
chat_template_kwargs={"enable_thinking": False},
|
| 82 |
+
)
|
| 83 |
+
print(outs[0].outputs[0].text)
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
if __name__ == "__main__":
|
| 87 |
+
main()
|