Overlay: drop a checkpoint's own table tensors when a sidecar is attached, including the monolithic weight/weight_scale form — lets FP8-table checkpoints (nvidia, official FP8) serve on one card

worker_image_quant.py

@@ -673,6 +673,15 @@ class PleOffloadRunner:
                     matched_checkpoint_tensors += 1
                     if disk_complete_tables:
                         table = _ple_disk_shard_of(mapped_name)
+                        if table is None and mapped_name.endswith(
+                            (".weight", ".weight_scale", ".weight_scale_2", ".input_scale")
+                        ):
+                            # Checkpoints that ship the table as one tensor plus a
+                            # global scale (the FP8 revision and the NVFP4 builds
+                            # derived from it) instead of shard_N.weight parts. The
+                            # sidecar owns those rows, so drop them before the loader
+                            # tries to write a 51.2B tensor into the stub.
+                            table = mapped_name.rsplit(".", 1)[0]
                         if table is not None and table.startswith(disk_complete_tables):
                             continue
                     yield weight_name, tensor