Fix safetensors bank validation ordering

training/fable_router_common.py

@@ -19,6 +19,8 @@ EXPECTED_PROJECTIONS = {
     "down_proj.weight": (2048, 512),
 }
 
+DTYPE_BYTES = {"BF16": 2, "F16": 2}
+
 
 def sha256(path: Path, chunk_size: int = 8 * 1024 * 1024) -> str:
     digest = hashlib.sha256()
@@ -109,7 +111,7 @@ def validate_bank_header_entries(
 
     dtypes: set[str] = set()
     payload_bytes = 0
-    previous_end = 0
+    intervals: list[tuple[int, int, str]] = []
     for layer, expert, name, expected_shape in expected:
         entry = tensor_entries[name]
         shape = tuple(int(item) for item in entry.get("shape", []))
@@ -119,18 +121,34 @@ def validate_bank_header_entries(
                 f"(layer={layer}, expert={expert})"
             )
         dtype = str(entry.get("dtype"))
-        if dtype not in {"BF16", "F16"}:
+        if dtype not in DTYPE_BYTES:
             raise ValueError(f"{name} has unsupported training dtype {dtype}")
         offsets = entry.get("data_offsets")
         if not isinstance(offsets, list) or len(offsets) != 2:
             raise ValueError(f"{name} has malformed data offsets")
         start, end = map(int, offsets)
-        if start < 0 or end <= start or start < previous_end:
-            raise ValueError(f"{name} has invalid or overlapping data offsets {offsets}")
-        previous_end = end
+        expected_bytes = expected_shape[0] * expected_shape[1] * DTYPE_BYTES[dtype]
+        if start < 0 or end <= start or end - start != expected_bytes:
+            raise ValueError(
+                f"{name} has invalid data offsets {offsets}; expected {expected_bytes} bytes"
+            )
+        intervals.append((start, end, name))
         payload_bytes += end - start
         dtypes.add(dtype)
 
+    # Safetensors does not require tensor entries (or our selected-expert
+    # manifest) to be ordered by their payload offsets.  Validate the physical
+    # layout independently from semantic tensor identity.
+    previous_end = 0
+    for start, end, name in sorted(intervals):
+        if start != previous_end:
+            relation = "overlap" if start < previous_end else "gap"
+            raise ValueError(
+                f"bank payload has a {relation} before {name}: "
+                f"expected offset {previous_end}, found {start}"
+            )
+        previous_end = end
+
     if previous_end != file_payload or payload_bytes != file_payload:
         raise ValueError(
             f"bank payload coverage mismatch: last={previous_end}, summed={payload_bytes}, "