From 43009d3929bbafe2cf63d104ac8945db9a9bf267 Mon Sep 17 00:00:00 2001
From: zhuceHW <zhuce@huawei.com>
Date: Thu, 28 Aug 2025 11:27:42 +0800
Subject: [PATCH] 1. inductor no more need decomposition op white list, remove
 it 2.fix erfc decomposition logic 3.remove redundant lowering overload op

---
 torch_npu/_inductor/decomposition.py    | 26 ++-----------------------
 torch_npu/_inductor/lowering.py         | 10 ++++------
 torch_npu/_inductor/lowering_op_list.py |  8 +++-----
 3 files changed, 9 insertions(+), 35 deletions(-)

diff --git a/torch_npu/_inductor/decomposition.py b/torch_npu/_inductor/decomposition.py
index b9c725f3ff..8770f2b42c 100644
--- a/torch_npu/_inductor/decomposition.py
+++ b/torch_npu/_inductor/decomposition.py
@@ -1,32 +1,10 @@
 import torch._ops
-from torch._inductor.decomposition import decompositions, pw_cast_for_opmath
-from torch._inductor.decomposition import register_decomposition
-
-from .lowering import _init_set
+from torch._inductor.decomposition import pw_cast_for_opmath, register_decomposition
 
 aten = torch.ops.aten
 
-DECOMPOSITION_OVERLOAD_OP = [
-    aten._log_softmax,
-    aten.nll_loss_forward,
-    # aten.gelu_backward,
-    # aten.gelu,
-    aten.nll_loss_backward,
-    aten._log_softmax_backward_data,
-    aten.embedding_dense_backward,
-    aten.addmm,
-    aten.gelu
-]
-
 
 def _register_npu_inductor_decompositons():
-    overload_op_set = set()
-    _init_set(DECOMPOSITION_OVERLOAD_OP, overload_op_set)
-
-    for op in overload_op_set:
-        if (op in decompositions):
-            del decompositions[op]
-
     @register_decomposition([aten.scatter.src])
     @pw_cast_for_opmath
     def scatter_src(self, input_tensor, dim, index_tensor, source_tensor):
@@ -45,5 +23,5 @@ def _register_npu_inductor_decompositons():
 
     @register_decomposition([aten.erfc])
     def erfc(x):
-        tensor = torch.ones_like(x) - torch.exp(x)
+        tensor = torch.ones_like(x) - torch.erf(x)
         return tensor
diff --git a/torch_npu/_inductor/lowering.py b/torch_npu/_inductor/lowering.py
index 2b47e091af..4448a0b059 100644
--- a/torch_npu/_inductor/lowering.py
+++ b/torch_npu/_inductor/lowering.py
@@ -126,8 +126,10 @@ def _register_npu_inductor_fallbacks():
                 if flag:
                     continue
                 else:
-                    make_fallback(op)
                     FALLBACK_LIST.append(op)
+    for op in FALLBACK_LIST:
+        make_fallback(op)
+
     # 把需要overload的op在lowering里删除
     for op in overload_op_set:
         if op in lowerings:
@@ -262,8 +264,4 @@ def _register_npu_inductor_fallbacks():
 
     @register_lowering(aten.cat)
     def cat(inputs, dim=0):
-        return fallback_handler(aten.cat.default)(inputs, dim)
-
-    make_fallback(aten._log_softmax)
-    make_fallback(aten.gather)
-    make_fallback(aten.nll_loss_forward)
+        return fallback_handler(aten.cat.default)(inputs, dim)
\ No newline at end of file
diff --git a/torch_npu/_inductor/lowering_op_list.py b/torch_npu/_inductor/lowering_op_list.py
index db9c427e60..c101457f5e 100644
--- a/torch_npu/_inductor/lowering_op_list.py
+++ b/torch_npu/_inductor/lowering_op_list.py
@@ -81,7 +81,9 @@ GENERATE_LIST2 = [
     "foreach"
 ]
 
-FALLBACK_LIST = []
+FALLBACK_LIST = [
+    aten.gather,
+]
 
 # Delete these op in lowering list and then update lowering list with new lowering,
 # otherwise, it will not use npu overload lowering.
@@ -99,9 +101,5 @@ LOWERING_OVERLOAD_OP = [
     aten.var,
 
     aten.embedding,
-    aten.split,
-    aten.split_with_sizes,
-    aten.nll_loss_forward,
-    aten.gather,
     aten.cat,
 ]
-- 
Gitee