add run_sh

test_fp32_with_bf16/run.sh

@@ -0,0 +1,7 @@
+!/bin/bash
+for dir in test_flash_attention/ test_fused_linear/ test_layernorm/ test_matmul/ test_silu/ test_vocab_parallel_embedding/
+do
+   cd $dir
+   ./run.sh &
+   cd ..
+done

test_fp32_with_bf16/test_flash_attention/case.txt

@@ -0,0 +1,8 @@
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-0
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-1
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-2
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-3
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-4
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-5
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-6
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fa-fp32-fp32-bf16-bf16-fp32-eager_tmp_111-eager_tmp_112-eager_tmp_113-eager_tmp_114-eager_tmp_115-pp-11-mp-7

test_fp32_with_bf16/test_flash_attention/run.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=1
+export NVIDIA_TF32_OVERRIDE=0
+export LD_LIBRARY_PATH=/usr/local/cuda/compat:$LD_LIBRARY_PATH:/usr/lib64/:/usr/local/lib/
+rm new_log_test_flash_attention_fp32vs_bfp16
+# 读取字符串
+while IFS= read -r line
+do
+    # 调用 train.py 并传递字符串作为参数
+    python test_flash_attention_fp32vs_bfp16.py "$line" 2>&1|tee >> new_log_test_flash_attention_fp32vs_bfp16
+
+done < case.txt

test_fp32_with_bf16/test_flash_attention/test_flash_attention_fp32vs_bfp16.py

@@ -28,6 +28,8 @@
     convert_dtype_to_torch_type,
     np_assert_accuracy
 )
+
+niuliling_path = None  # 全局变量
 def get_triangle_upper_mask(shape):
     mask = paddle.full(shape=shape, fill_value=-np.inf)
     mask.stop_gradient = True
@@ -56,26 +58,47 @@ def setUp(self):
 
     def init_np_inputs_and_dout(self):
         # init np array 
-        self.np_x = np.random.random(size=[1,8192,14,128]).astype("float32") - 0.5
-        self.np_dout = np.random.random(size=[1,8192,14,128]).astype("float32") - 0.5
+        data_xwb = np.load(niuliling_path+".npz")
+        data_dout = np.load(niuliling_path+".npy")
+
+        self.np_q = data_xwb["query"].astype("float32")
+        self.np_k = data_xwb["key"].astype("float32")
+        self.np_v = data_xwb["value"].astype("float32")
+
+        print("shape q ", self.np_q.shape) # = data_xwb["query"].astype("float32")
+        print("shape k", self.np_k.shape) # = data_xwb["key"].astype("float32")
+        print("shape v", self.np_v.shape) # = data_xwb["value"].astype("float32")
+        self.np_dout = data_dout.astype("float32")
 
     def gen_eager_inputs_and_dout(self):
         x_eager = paddle.to_tensor(
-            self.np_x,
+            self.np_q,
             dtype="float32",
             place="gpu",
         )
         x_eager.stop_gradient = False
+        k_eager = paddle.to_tensor(
+            self.np_k,
+            dtype="float32",
+            place="gpu",
+        )
+        k_eager.stop_gradient = False
+        v_eager = paddle.to_tensor(
+            self.np_v,
+            dtype="float32",
+            place="gpu",
+        )
+        v_eager.stop_gradient = False
         dout_eager = paddle.to_tensor(
             self.np_dout,
             dtype="float32",
             place="gpu",
         )
         dout_eager.stop_gradient = False
-        return x_eager, dout_eager
-    def cal_bfp16_res(self, x, dout):
-        out, _ = flash_attention(x, x, x, 0.0, True, False)
-        out_grads = paddle.grad([out], [x], grad_outputs=[dout])
+        return x_eager, k_eager, v_eager, dout_eager
+    def cal_bfp16_res(self, x, k, v, dout):
+        out, _ = flash_attention(x, k, v, 0.0, True, False)
+        out_grads = paddle.grad([out], [x,k,v], grad_outputs=[dout])
         return out, out_grads
 
     # def cal_fp32_res(self, x, dout):
@@ -87,16 +110,16 @@ def cal_bfp16_res(self, x, dout):
     #     out_grads = map_structure(lambda x: paddle.cast(x, dtype="float32"), out_grads)
     #     return out, out_grads
 
-    def cal_fp32_res(self, x, dout):
-        out = multi_head_attention(x, x, x)
-        out_grads = paddle.grad([out], [x], grad_outputs=[dout])
+    def cal_fp32_res(self, x, k, v, dout):
+        out = multi_head_attention(x, k, v)
+        out_grads = paddle.grad([out], [x,k,v], grad_outputs=[dout])
         return out, out_grads
 
     def test_flash_atten_fp32vsbfp16_mode1(self):
-        x_bfp16, dout_bfp16 = map_structure(lambda x: paddle.cast(x, dtype="bfloat16"), self.gen_eager_inputs_and_dout())
-        x_fp32, dout_fp32 = paddle.cast(x_bfp16,"float32"), paddle.cast(dout_bfp16,"float32")
-        out_fp32, out_grads_fp32 = self.cal_fp32_res(x_fp32, dout_fp32)
-        out_bfp16, out_grads_bfp16 = self.cal_bfp16_res(x_bfp16, dout_bfp16)
+        x_bfp16, k_bfp16, v_bfp16, dout_bfp16 = map_structure(lambda x: paddle.cast(x, dtype="bfloat16"), self.gen_eager_inputs_and_dout())
+        x_fp32, k_fp32, v_fp32, dout_fp32 = paddle.cast(x_bfp16,"float32"), paddle.cast(k_bfp16,"float32"),paddle.cast(v_bfp16,"float32"),paddle.cast(dout_bfp16,"float32")
+        out_fp32, out_grads_fp32 = self.cal_fp32_res(x_fp32,k_fp32, v_fp32, dout_fp32)
+        out_bfp16, out_grads_bfp16 = self.cal_bfp16_res(x_bfp16, k_bfp16, v_bfp16, dout_bfp16)
         pt_out_bfp16 = paddle.cast(out_bfp16, "float32")
         pt_out_grads_bfp16 = map_structure(lambda x: paddle.cast(x, dtype="float32"), out_grads_bfp16)
         try:
@@ -132,11 +155,11 @@ def test_flash_atten_fp32vsbfp16_mode1(self):
             print(e)
 
     def test_flash_atten_fp32vsbfp16_mode2(self):
-        x_bfp16, dout_bfp16 = map_structure(lambda x: paddle.cast(x, dtype="bfloat16"), self.gen_eager_inputs_and_dout())
-        x_fp32, dout_fp32 = paddle.cast(x_bfp16,"float32"), paddle.cast(dout_bfp16,"float32")
-        out_fp32, out_grads_fp32 = self.cal_fp32_res(x_fp32, dout_fp32)
+        x_bfp16, k_bfp16, v_bfp16, dout_bfp16 = map_structure(lambda x: paddle.cast(x, dtype="bfloat16"), self.gen_eager_inputs_and_dout())
+        x_fp32, k_fp32, v_fp32, dout_fp32 = paddle.cast(x_bfp16,"float32"), paddle.cast(k_bfp16,"float32"),paddle.cast(v_bfp16,"float32"),paddle.cast(dout_bfp16,"float32")
+        out_fp32, out_grads_fp32 = self.cal_fp32_res(x_fp32, k_fp32, v_fp32,dout_fp32)
         out_grads_fp32 =  map_structure(lambda x: paddle.cast(paddle.cast(x,"bfloat16"),"float32"), out_grads_fp32)
-        out_bfp16, out_grads_bfp16 = self.cal_bfp16_res(x_bfp16, dout_bfp16)
+        out_bfp16, out_grads_bfp16 = self.cal_bfp16_res(x_bfp16, k_bfp16, v_bfp16, dout_bfp16)
         pt_out_bfp16 = paddle.cast(out_bfp16, "float32")
         pt_out_grads_bfp16 = map_structure(lambda x: paddle.cast(x, dtype="float32"), out_grads_bfp16)
         try:
@@ -172,5 +195,14 @@ def test_flash_atten_fp32vsbfp16_mode2(self):
             print(e)
 
 if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print("请提供 data_path 参数")
+        sys.exit(1)
+
+    tmp = sys.argv[1]  # 设置全局变量 data_path
+    niuliling_path = tmp
+    print(tmp) 
+
+    del sys.argv[1]
     np.random.seed(2023)
     unittest.main()

test_fp32_with_bf16/test_fused_linear/case.txt

@@ -0,0 +1,8 @@
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_0.w_0-server_nlg_mask_lm_out_fc_0.b_0-eager_in_tmp_11-pp-11-mp-0
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_1.w_0-server_nlg_mask_lm_out_fc_1.b_0-eager_in_tmp_11-pp-11-mp-1
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_2.w_0-server_nlg_mask_lm_out_fc_2.b_0-eager_in_tmp_11-pp-11-mp-2
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_3.w_0-server_nlg_mask_lm_out_fc_3.b_0-eager_in_tmp_11-pp-11-mp-3
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_4.w_0-server_nlg_mask_lm_out_fc_4.b_0-eager_in_tmp_11-pp-11-mp-4
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_5.w_0-server_nlg_mask_lm_out_fc_5.b_0-eager_in_tmp_11-pp-11-mp-5
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_6.w_0-server_nlg_mask_lm_out_fc_6.b_0-eager_in_tmp_11-pp-11-mp-6
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/fl-fp32-bf16-bf16-bf16-eager_in_tmp_9-server_nlg_mask_lm_out_fc_7.w_0-server_nlg_mask_lm_out_fc_7.b_0-eager_in_tmp_11-pp-11-mp-7

test_fp32_with_bf16/test_fused_linear/run.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=2
+export NVIDIA_TF32_OVERRIDE=0
+export LD_LIBRARY_PATH=/usr/local/cuda/compat:$LD_LIBRARY_PATH:/usr/lib64/:/usr/local/lib/
+rm new_log_test_fused_linear_develop
+# 读取字符串
+while IFS= read -r line
+do
+    # 调用 train.py 并传递字符串作为参数
+    python test_fused_linear_develop.py "$line" 2>&1|tee >> new_log_test_fused_linear_develop
+
+done < case.txt

test_fp32_with_bf16/test_fused_linear/test_fused_linear_develop.py

@@ -13,6 +13,7 @@
 
 global_out = []
 global_dout = []
+niuliling_path = None  # 全局变量
 
 class TestFCDevelop(unittest.TestCase):
     def __init__(self, shape, dtype, test_mode):
@@ -28,10 +29,15 @@ def init_threshold(self):
 
     def init_np_inputs_and_dout(self):
         np.random.seed(123)
-        self.np_x = np.random.random(size=self.shape["x"]).astype("float32") - 0.5
-        self.np_w = np.random.random(size=self.shape["w"]).astype("float32") - 0.5
-        self.np_b = np.random.random(size=self.shape["b"]).astype("float32") - 0.5
-        self.np_dout = np.random.random(size=self.shape["dout"]).astype("float32") - 0.5
+        data_xwb = np.load(niuliling_path+".npz")
+        data_dout = np.load(niuliling_path+".npy")
+        self.np_x = data_xwb["x"]
+        self.np_w = data_xwb["weight"]
+        self.np_b = data_xwb["bias"]
+        print("shape x ", self.np_x.shape)
+        print("shape w ", self.np_w.shape)
+        print("shape b ", self.np_b.shape)
+        self.np_dout = data_dout.astype("float32")
 
         # convert np array dtype
         if self.dtype == "float16":
@@ -122,10 +128,20 @@ def test_eager_accuracy(self):
         del out_grads_eager
 
 if __name__ == '__main__':
-    x = [[1, 8192, 14336], [1, 8192, 14336], [1, 8192, 1792], [1, 8192, 4816]]
-    w = [[14336, 5376], [14336, 9632], [1792, 14336], [4816, 14336]]
-    b = [[5376], [9632], [14336], [14336]]
-    dout = [[1, 8192, 5376], [1, 8192, 9632], [1, 8192, 14336], [1, 8192, 14336]]
+    if len(sys.argv) < 2:
+        print("请提供 data_path 参数")
+        sys.exit(1)
+
+    tmp = sys.argv[1]  # 设置全局变量 data_path
+    niuliling_path = tmp
+    print(tmp) 
+    del sys.argv[1]
+
+
+    x = [[1, 8192, 14336]] #, [1, 8192, 14336], [1, 8192, 1792], [1, 8192, 4816]]
+    w = [[14336, 5376]] #, [14336, 9632], [1792, 14336], [4816, 14336]]
+    b = [[5376], [9632]] #, [14336], [14336]]
+    dout = [[1, 8192, 5376]]#, [1, 8192, 9632], [1, 8192, 14336], [1, 8192, 14336]]
     shape_list = []
     for i in range(len(x)):
         shape = {}
@@ -134,13 +150,15 @@ def test_eager_accuracy(self):
         shape["b"] = b[i]
         shape["dout"] = dout[i]
         shape_list.append(shape)
+
     for test_mode in [1,2]:
         if test_mode == 1:
             atol = 1e-2
         elif test_mode == 2:
             atol = 1e-6
-        print("test_mode_{test_mode} start*************************************************************************" \
-            .format(test_mode=test_mode))
+        #print("test_mode_{test_mode} start*************************************************************************" \
+        #    .format(test_mode=test_mode))
+        print(tmp, "test_mode", test_mode) 
         for shape in shape_list:
             global_out.clear()
             global_dout.clear()

test_fp32_with_bf16/test_layernorm/case.txt

@@ -0,0 +1,8 @@
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-0
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-1
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-2
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-3
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-4
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-5
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-6
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/ln-bf16-fp32-fp32-fp32-eager_in_tmp_0-encoder_layer_119_pre_att__layer_norm_scale_0.w_0-encoder_layer_119_pre_att_layer_norm_bias_0.b_0-eager_tmp_205-pp-11-mp-7

test_fp32_with_bf16/test_layernorm/run.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=1
+export NVIDIA_TF32_OVERRIDE=0
+export LD_LIBRARY_PATH=/usr/local/cuda/compat:$LD_LIBRARY_PATH:/usr/lib64/:/usr/local/lib/
+rm new_log_test_layernorm_fp32vsbfp16
+# 读取字符串
+while IFS= read -r line
+do
+    # 调用 train.py 并传递字符串作为参数
+    python test_layernorm_fp32vsbfp16.py "$line" 2>&1|tee >> new_log_test_layernorm_fp32vsbfp16
+
+done < case.txt

test_fp32_with_bf16/test_layernorm/test_layernorm_fp32vsbfp16.py

@@ -8,16 +8,19 @@
 from utils import TOLERANCE, convert_dtype_to_torch_type, np_assert_accuracy
 
 
+niuliling_path = None  # 全局变量
 class TestLayerNormFP32vsBFP16(unittest.TestCase):
     def setUp(self):
         self.init_np_inputs_and_dout()
 
     def init_np_inputs_and_dout(self):
         # init np array 
-        self.np_x = np.random.random(size=[1, 8192, 14336]).astype("float32") - 0.5
-        self.np_w = np.random.random(size=[14336]).astype("float32") - 0.5
-        self.np_b= np.random.random(size=[14336]).astype("float32") - 0.5
-        self.np_dout = np.random.random(size=[1, 8192, 14336]).astype("float32") - 0.5
+        data_xwb = np.load(niuliling_path+".npz")
+        data_dout = np.load(niuliling_path+".npy")
+        self.np_x = data_xwb["x"].astype("float32")
+        self.np_w = data_xwb["weight"].astype("float32")
+        self.np_b = data_xwb["bias"].astype("float32")
+        self.np_dout = data_dout.astype("float32")
 
     def gen_eager_inputs_and_dout(self):
         x = paddle.to_tensor(
@@ -133,5 +136,14 @@ def test_layernorm_fp32vsbfp16_mode2(self):
         except Exception as e:
             print(e)
 if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print("请提供 data_path 参数")
+        sys.exit(1)
+
+    tmp = sys.argv[1]  # 设置全局变量 data_path
+    niuliling_path = tmp
+    print(tmp) 
+    del sys.argv[1]
+
     np.random.seed(2023)
-    unittest.main()
+    unittest.main()

test_fp32_with_bf16/test_matmul/case.txt

@@ -0,0 +1,8 @@
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_0.w_0-eager_tmp_101-pp-11-mp-0
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_1.w_0-eager_tmp_101-pp-11-mp-1
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_2.w_0-eager_tmp_101-pp-11-mp-2
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_3.w_0-eager_tmp_101-pp-11-mp-3
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_4.w_0-eager_tmp_101-pp-11-mp-4
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_5.w_0-eager_tmp_101-pp-11-mp-5
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_6.w_0-eager_tmp_101-pp-11-mp-6
+/home/niuliling/more/PaddleApiTest/test_fp32_with_bf16/numpy_data/matmul-bf16-bf16-bf16-eager_tmp_100-encoder_layer_130_multi_head_att_output_fc_7.w_0-eager_tmp_101-pp-11-mp-7

test_fp32_with_bf16/test_matmul/run.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+export NVIDIA_TF32_OVERRIDE=0
+export LD_LIBRARY_PATH=/usr/local/cuda/compat:$LD_LIBRARY_PATH:/usr/lib64/:/usr/local/lib/
+# 读取字符串
+rm new_log_matmul_fp32vsbf16
+while IFS= read -r line
+do
+    # 调用 train.py 并传递字符串作为参数
+    python test_matmul_fp32vsbfp16.py "$line" 2>&1|tee >> new_log_matmul_fp32vsbf16
+done < case.txt

test_fp32_with_bf16/test_matmul/test_matmul_fp32vsbfp16.py

@@ -25,16 +25,22 @@
 from utils import (
     np_assert_accuracy,
 )
+niuliling_path = None  # 全局变量
 
 class TestMatmulFP32vsBFP16(unittest.TestCase):
     def setUp(self):
         self.init_np_inputs_and_dout()
 
     def init_np_inputs_and_dout(self):
         # init np array 
-        self.np_x = np.random.random(size=[1, 8192, 14336]).astype("float32") - 0.5
-        self.np_y = np.random.random(size=[14336, 12528]).astype("float32") - 0.5
-        self.np_dout = np.random.random(size=[1, 8192, 12528]).astype("float32") - 0.5
+        data_xy = np.load(niuliling_path+".npz")
+        data_dout = np.load(niuliling_path+".npy")
+
+        self.np_x = data_xy["x"].astype("float32")
+        self.np_y = data_xy["y"].astype("float32")
+        print("shape x", self.np_x.shape)
+        print("shape y", self.np_y.shape)
+        self.np_dout = data_dout.astype("float32")
 
     def gen_eager_inputs_and_dout(self):
         x = paddle.to_tensor(
@@ -56,6 +62,7 @@ def gen_eager_inputs_and_dout(self):
         )
         dout.stop_gradient = False
         return x, y, dout
+
     def cal_res(self, x, y, dout):
         out = paddle.matmul(x, y)
         out_grads = paddle.grad([out], [x, y], grad_outputs=[dout])
@@ -143,5 +150,13 @@ def test_matmul_fp32vsbfp16_mode2(self):
             print(e)
 
 if __name__ == '__main__':
+    if len(sys.argv) < 2:
+        print("请提供 data_path 参数")
+        sys.exit(1)
+
+    tmp = sys.argv[1]  # 设置全局变量 data_path
+    niuliling_path = tmp
+    print(tmp) 
+    del sys.argv[1]
     np.random.seed(2023)
     unittest.main()