training: restructure as vLLM plugin package

- Convert to installable package with entry points for vLLM auto-discovery
- Add checkpoint_sync.py: Python replacement for Rust checkpoint binary
  - Block-level diffing of safetensors files (4KB blocks)
  - vLLM→HF weight name conversion built-in
  - Scheduled 10min after training jobs (batched)
- API change: /train now takes raw token IDs (context_ids + continuation_ids)
  - No tokenizer on training side, client owns tokenization
- Remove superseded code: standalone scripts, Rust binary, tokenizer helpers

Install: pip install -e ./training
Then vLLM auto-loads via entry point.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

training/apollo_plugin/export_hook.py

@@ -0,0 +1,67 @@
+"""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)
+        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")