Merge pull request #88 from andrea-fasoli/util_update

feat: expand detection of data types in model size estimation

Author: chichun-charlie-liu

fms_mo/fx/utils.py

@@ -21,6 +21,7 @@
 import os
 
 # Third Party
+import pandas as pd
 import torch
 
 # Local
@@ -433,7 +434,7 @@ def get_target_op_from_mod_or_str(mod_or_str, verbose=False):
 #############
 
 
-def model_size_Wb(mod, unit="MB"):
+def model_size_Wb(mod, unit="MB", print_to_file=True, show_details=False):
     """Checks model size, only count weight and bias
 
     NOTE:
@@ -447,40 +448,61 @@ def model_size_Wb(mod, unit="MB"):
     Returns:
         float: model size in desired unit
     """
+
     mem_use = 0
-    Nint8 = 0
-    Nfp32 = 0
+    if unit not in ["MB", "GB"]:
+        logger.warning(
+            f"Unrecognized unit for memory summary: {unit}. Will use MB instead."
+        )
+        unit = "MB"
+
+    summary_weights = {"layer": [], "shape": [], f"mem ({unit})": [], "dtype": []}
     for n, m in mod.named_modules():
         w = getattr(m, "weight", None)
+        w_dtype, w_shape = None, None
         if callable(w):  # see Note 1.
             w_mat, b_mat = w()[:2]
-            mem_use += (
+            mem_use = (
                 w_mat.numel() * w_mat.element_size()
                 + b_mat.numel() * b_mat.element_size()
             )
-            if w_mat.dtype in [torch.qint8, torch.quint8]:
-                Nint8 += 1
-            else:
-                logger.info(f"Parameter {n} should be int8 but is {w_mat.dtype}")
+            w_dtype = w_mat.dtype
+            w_shape = w_mat.shape
+
         elif isinstance(w, torch.Tensor):
-            mem_use += w.numel() * w.element_size()
+            mem_use = w.numel() * w.element_size()
             if hasattr(m, "bias") and m.bias is not None:
                 mem_use += m.bias.numel() * m.bias.element_size()
-            if w.dtype == torch.float32:
-                Nfp32 += 1
-            else:
-                logger.info(f"Parameter {n} should be fp32 but is {w.dtype}")
-    logger.info(
-        f"[check model size] Found {Nint8} INT8 and {Nfp32} "
-        "FP32 W/b tensors in this model."
+            w_dtype = w.dtype
+            w_shape = w.shape
+
+        if w_shape:
+            mem_use = mem_use / 1e9 if unit == "GB" else mem_use / 1e6
+
+            summary_weights["layer"].append(n)
+            summary_weights["shape"].append(w_shape)
+            summary_weights[f"mem ({unit})"].append(mem_use)
+            summary_weights["dtype"].append(w_dtype)
+
+    df_summary_weights = pd.DataFrame(summary_weights)
+    logger_or_print = logger.info if print_to_file else print
+    logger_or_print("[check model size] Summary of W/b tensors in this model:")
+    logger_or_print(
+        "\n%s",
+        str(
+            pd.pivot_table(
+                df_summary_weights,
+                index="dtype",
+                values=["layer", f"mem ({unit})"],
+                aggfunc={"layer": "count", f"mem ({unit})": "sum"},
+            )
+        ),
     )
 
-    if unit == "GB":
-        mem_use /= 1e9
-    else:
-        mem_use /= 1e6
+    if show_details:
+        logger_or_print(df_summary_weights.to_markdown())
 
-    return mem_use
+    return df_summary_weights[f"mem ({unit})"].sum().item()
 
 
 def plot_graph_module(