training-platform core: drop unused loop var and populate zero-sized input defaults

Two fixes to the inputTypes / inputOffsets build loop in both
generateTrainingNetwork.py and testMVPTraining.py (the two files share
this pattern verbatim).

1. The `name` loop variable in `for graph_idx, name in enumerate(...)`
   was dead — switched to `for graph_idx in range(len(...))` so the
   intent is explicit and the lint warning goes away.

2. The zero-sized-input branch (`np.prod(arr.shape) == 0`) previously
   `pass`ed without populating inputTypes[f"input_{idx}"] or
   inputOffsets[f"input_{idx}"].  ONNX permits optional placeholder
   inputs with shape like (0,) and the downstream deployer looks up
   every input by key, so a zero-sized input would later raise a
   confusing KeyError far from the cause.  Populate the entries with
   a trivial default (float32 pointer, offset 0) — the type does not
   matter for codegen since the buffer has no elements, but the key
   must exist.

Verified on Siracusa: simplemlp_train passes 0/4 (diff=0.000000 at every
step) in both non-tiled and tiled runs.

Author: runwangdl

DeeployTest/generateTrainingNetwork.py

@@ -84,7 +84,7 @@ def generateTrainingNetwork(args):
     inputOffsets = {}
 
     npz_idx = 0
-    for graph_idx, name in enumerate(graph_input_names):
+    for graph_idx in range(len(graph_input_names)):
         if graph_idx in grad_acc_set:
             inputTypes[f"input_{graph_idx}"] = PointerClass(float32_t)
             inputOffsets[f"input_{graph_idx}"] = 0
@@ -102,7 +102,11 @@ def generateTrainingNetwork(args):
                 inputTypes[f"input_{graph_idx}"] = PointerClass(float32_t)
                 inputOffsets[f"input_{graph_idx}"] = 0
             elif np.prod(arr.shape) == 0:
-                pass
+                # Zero-sized input (ONNX allows shape (0, ...) for optional
+                # placeholders).  No data to infer from, but downstream still
+                # looks up input_{idx} by key, so populate with a trivial default.
+                inputTypes[f"input_{graph_idx}"] = PointerClass(float32_t)
+                inputOffsets[f"input_{graph_idx}"] = 0
             else:
                 values = arr.reshape(-1).astype(np.float32)
                 _type, offset = inferTypeAndOffset(values, signProp = False)

DeeployTest/testMVPTraining.py

@@ -79,7 +79,7 @@ def generateTiledTrainingNetwork(args) -> None:
     inputOffsets = {}
 
     npz_idx = 0
-    for graph_idx, name in enumerate(graph_input_names):
+    for graph_idx in range(len(graph_input_names)):
         if graph_idx in grad_acc_set:
             inputTypes[f"input_{graph_idx}"] = PointerClass(float32_t)
             inputOffsets[f"input_{graph_idx}"] = 0
@@ -93,7 +93,11 @@ def generateTiledTrainingNetwork(args) -> None:
                 inputTypes[f"input_{graph_idx}"] = PointerClass(float32_t)
                 inputOffsets[f"input_{graph_idx}"] = 0
             elif np.prod(arr.shape) == 0:
-                pass
+                # Zero-sized input (ONNX allows shape (0, ...) for optional
+                # placeholders).  No data to infer from, but downstream still
+                # looks up input_{idx} by key, so populate with a trivial default.
+                inputTypes[f"input_{graph_idx}"] = PointerClass(float32_t)
+                inputOffsets[f"input_{graph_idx}"] = 0
             else:
                 values = arr.reshape(-1).astype(np.float32)
                 _type, offset = inferTypeAndOffset(values, signProp = False)