7e7e9a4b69
Remove standalone worker.py daemon. Training now runs inside vLLM: - train_router.py: FastAPI router patched into vLLM's build_app() - /train served on same port as /completions, /score - Lazy-loads HF model with vLLM weight views on first request - HOGWILD training: no pause, weights updated in-place The previous architecture had a separate daemon on port 8080 that communicated with vLLM via pause/resume endpoints. This was wrong - training should run in-process, sharing GPU memory directly. Co-Authored-By: Proof of Concept <poc@bcachefs.org>
71 lines
2.2 KiB
Python
71 lines
2.2 KiB
Python
"""Monkey-patch vLLM to export weight IPC handles on startup.
|
|
|
|
Usage — install the apollo_plugin package:
|
|
|
|
pip install -e /path/to/training
|
|
|
|
Then vLLM auto-discovers and loads via entry point. Or filter:
|
|
|
|
VLLM_PLUGINS=apollo vllm serve Qwen/Qwen3.5-27B ...
|
|
|
|
The hook patches vLLM's model runner to export IPC handles after
|
|
model loading completes. The handles are saved to a file that the
|
|
Apollo training process reads.
|
|
"""
|
|
|
|
import atexit
|
|
import torch
|
|
from pathlib import Path
|
|
|
|
HANDLE_PATH = "/tmp/vllm_weight_handles.pt"
|
|
|
|
|
|
def export_model_weights(model):
|
|
"""Export CUDA IPC handles for all model parameters."""
|
|
from torch.multiprocessing.reductions import reduce_tensor
|
|
|
|
handles = {}
|
|
total_bytes = 0
|
|
|
|
for name, param in model.named_parameters():
|
|
if param.device.type != 'cuda':
|
|
continue
|
|
handle = reduce_tensor(param.data)
|
|
handles[name] = {
|
|
'handle': handle,
|
|
'shape': list(param.shape),
|
|
'dtype': str(param.dtype),
|
|
}
|
|
total_bytes += param.nelement() * param.element_size()
|
|
|
|
torch.save(handles, HANDLE_PATH)
|
|
print(f"[apollo] Exported {len(handles)} weight handles "
|
|
f"({total_bytes / 1e9:.1f} GB) to {HANDLE_PATH}")
|
|
|
|
|
|
def _patch_model_runner():
|
|
"""Patch gpu_worker to export handles after model loading.
|
|
|
|
vLLM loads the model in a subprocess (EngineCore_DP0), so we
|
|
can't patch from the parent. Instead, patch the worker's
|
|
init_device or load_model at the module level — the subprocess
|
|
imports the same modules.
|
|
"""
|
|
from vllm.v1.worker import gpu_worker
|
|
|
|
original_load = gpu_worker.Worker.load_model
|
|
|
|
def patched_load(self, *args, **kwargs):
|
|
result = original_load(self, *args, **kwargs)
|
|
try:
|
|
export_model_weights(self.model_runner.model)
|
|
# Set model path for training router
|
|
model_path = self.vllm_config.model_config.model
|
|
from .train_router import set_model_path
|
|
set_model_path(model_path)
|
|
except Exception as e:
|
|
print(f"[apollo] Failed to export weights: {e}")
|
|
return result
|
|
|
|
gpu_worker.Worker.load_model = patched_load
|
|
print("[apollo] Weight export hook installed")
|