Update.

Readme.md

@@ -3,6 +3,27 @@
 
 ## data flow
 
+                        query
+                          |
+                      tokenizer  -> input_ids
+                          |
+ rotary_pos_emb       embedding  ->  [1, 6, 4096]   
+               \     /
+               GLMBlock x 28  ->  [6, 1, 4096]    <━━━┓
+                                                      ┃
+                RMSNorm  ->  [6, 1, 4096]             ┃
+                                                      ┃
+                 [-1:]  ->  [1, 1, 4096]              ┃
+                                                      ┃
+                Linear  ->  [1, 1, 65024]             ┃
+                                                      ┃
+                softmax  ->  [1, 65024]               ┃
+                                                      ┃
+               multinomial  ->  [1]                   ┃
+                                                      ┃
+          cat([input_ids, next_tokens])            ━━━┛
+
+
 input_ids = tokenizer.build_chat_input(query, history=history, role=role)
 
 for:

tools/show.py

@@ -55,3 +55,23 @@ def DumpListToFile(list, name="list"):
     for d in list:
         f.writelines("%s" % d + os.linesep)
     f.close()
+
+
+prob_true = 0
+prob_all = 0
+
+
+def ProbGE0(tensor: torch.tensor):
+    global prob_true
+    global prob_all
+    m = tensor.ge(0)
+    prob_true = prob_true + m.sum().item()
+    prob_all = prob_all + math.prod(tensor.shape)
+
+
+def DumpProb():
+    global prob_true
+    global prob_all
+    print("prob_true : " + str(prob_true))
+    print("prob_all : " + str(prob_all))
+    print("prob : " + str((prob_true * 100) / prob_all) + "%")

tools/test.py

@@ -11,3 +11,8 @@ show.DumpTensorToImage(radata, "test.png")
 
 radata = torch.randn(127, 127)
 show.DumpTensorToLog(radata, "test.log")
+
+
+radata = torch.randn(127, 127) - 0.5
+show.ProbGE0(radata)
+show.DumpProb()