fix: prevent OOM on receiver by using lazy=True with broadcast weights

When the receiver has no safetensors files, load_model's internal
nn.quantize skips quantization (class_predicate finds no .scales keys
in empty weights dict), leaving the model un-quantized as full fp16.
With lazy=False, mx.eval(model.parameters()) materializes ~72GB of
fp16 data for a 36B-param model on a 24GB machine → silent OOM kill.

Fix: use lazy=True when broadcast_weights is available. This skips the
eager eval, and our code handles quantization correctly before loading
the broadcast weights.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

src/exo/worker/engines/mlx/utils_mlx.py

@@ -229,9 +229,12 @@ def shard_and_load(
         broadcast_weights = broadcast_model_weights(model_path, group, is_source)
 
     # Create model architecture (all ranks have config.json on disk now).
-    # Use lazy=False when receiver has no local weights: lazy=True would create
-    # dangling references to nonexistent safetensors files.
-    use_lazy = has_local_weights or not needs_transfer
+    # Always use lazy=True when we have broadcast weights: load_model's internal
+    # nn.quantize skips quantization when weights dict is empty (no safetensors),
+    # leaving the model un-quantized. lazy=False would then mx.eval() the full
+    # fp16 model (~72GB for a 36B-param model), causing OOM on the receiver.
+    # We handle quantization ourselves below before loading broadcast weights.
+    use_lazy = has_local_weights or broadcast_weights is not None
     model, _ = load_model(model_path, lazy=use_lazy, strict=False)
     logger.debug(model)
     if hasattr(model, "model") and isinstance(model.model, DeepseekV3Model):  # type: ignore