Merge pull request #3 from cyx-6/nmk

extract dynamic vars

example.py

@@ -1,7 +1,8 @@
 from models.vicuna_v1_7b_fp16 import Module as Vicuna
-from mlc_bench.extraction import extract_from_relax
+from mlc_bench.extraction import extract_from_relax, extract
 
 if __name__ == "__main__":
-    extract_from_relax(
-        mod=Vicuna, model_name="vicuna_v1_7b_fp_16", file_path="./extracted"
-    )
+    # extract_from_relax(
+    #     mod=Vicuna, model_name="vicuna_v1_7b_fp_16", file_path="./extracted"
+    # )
+    extract(mod=Vicuna)

mlc_bench/extraction.py

@@ -142,6 +142,16 @@ def extract_from_relax(mod: tvm.ir.IRModule, model_name: str, file_path: str):
         )
 
 
+def extract_shape(arg) -> List:
+    if isinstance(arg, (tuple, list, tvm.relax.Tuple)):
+        r = []
+        for a in arg:
+            r.extend(extract_shape(a))
+        return r
+    else:
+        return [arg.struct_info]
+
+
 def prim_func_usage_gen(mod: tvm.ir.IRModule):
     for gv, func in mod.functions.items():
         if isinstance(func, tvm.relax.Function):
@@ -152,12 +162,31 @@ def prim_func_usage_gen(mod: tvm.ir.IRModule):
                         functor = raw_args[0]
                         if isinstance(functor, tvm.ir.GlobalVar):
                             if isinstance(mod.functions[functor], tvm.tir.PrimFunc):
-                                args = [arg.struct_info for arg in raw_args[1:]] + [
-                                    binding.value.struct_info
-                                ]
+                                args = extract_shape(raw_args[1:]) + extract_shape(binding.value)
                                 yield gv, functor, args
 
 
+def extract_dynamic_var(func_dict):
+    dynamic_var_dict = {}
+    for gv in func_dict:
+        print(gv)
+        dynamic_var_dict[gv] = set()
+        for functor in func_dict[gv]:
+            for arg_list, _ in func_dict[gv][functor]:
+                for arg in arg_list:
+                    if isinstance(arg, tvm.relax.TensorStructInfo):
+                        for v in arg.shape.values:
+                            if isinstance(v, tvm.tir.Var):
+                                dynamic_var_dict[gv].add((v.name, v.dtype))
+                    elif isinstance(arg, tvm.relax.ShapeStructInfo):
+                        for v in arg.values:
+                            if isinstance(v, tvm.tir.Var):
+                                dynamic_var_dict[gv].add((v.name, v.dtype))
+                    else:
+                        raise NotImplementedError
+    return dynamic_var_dict
+
+
 def extract(mod: tvm.ir.IRModule):
     def update_records(records, new_args):
         for i, (args, count) in enumerate(records):
@@ -178,8 +207,10 @@ def update_records(records, new_args):
                 prim_func_dict[functor] = []
             update_records(prim_func_dict[functor], args)
             update_records(relax_func_dict[gv][functor], args)
+    dynamic_var_dict = extract_dynamic_var(relax_func_dict)
     print(prim_func_dict)
     print(relax_func_dict)
+    print(dynamic_var_dict)
 
 
 if __name__ == "__main__":