psaiops 0.0.17__tar.gz → 0.1.1__tar.gz

{psaiops-0.0.17 → psaiops-0.1.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: psaiops
-Version: 0.0.17
+Version: 0.1.1
 Summary: Web apps to inspect & engineer NN activations.
 License: .github/LICENSE.md
 Author: apehex

{psaiops-0.0.17 → psaiops-0.1.1}/psaiops/score/attention/app.py

@@ -18,7 +18,7 @@ MODEL = 'openai/gpt-oss-20b'
 
 def create_color_map() -> dict:
     return {
-        '-1': '#00ff00',
+        '-1': '#004444',
         **{str(__i): '#{:02x}0000'.format(int(2.55 * __i)) for __i in range(101)}}
 
 # INTRO ########################################################################
@@ -30,14 +30,8 @@ def create_intro_block(intro: str) -> dict:
 # MODEL ########################################################################
 
 def create_model_block() -> dict:
-    __model_dd = gradio.Dropdown(label='Model', value='openai/gpt-oss-20b', choices=['openai/gpt-oss-20b'], scale=1, allow_custom_value=False, multiselect=False, interactive=True) # 'openai/gpt-oss-120b'
-    __layer_sl = gradio.Slider(label='Layer Depth', value=12, minimum=-1, maximum=23, step=1, scale=1, interactive=True) # info='-1 to average on all layers'
-    __head_sl = gradio.Slider(label='Attention Head', value=-1, minimum=-1, maximum=63, step=1, scale=1, interactive=True) # info='-1 to average on all heads'
-    __model_dd.change(fn=update_layer_range, inputs=[__layer_sl, __model_dd], outputs=__layer_sl, queue=False, show_progress='hidden')
-    return {
-        'model_block': __model_dd,
-        'layer_block': __layer_sl,
-        'head_block': __head_sl}
+    __model = gradio.Dropdown(label='Model', value='openai/gpt-oss-20b', choices=['openai/gpt-oss-20b'], scale=1, allow_custom_value=False, multiselect=False, interactive=True) # 'openai/gpt-oss-120b'
+    return {'model_block': __model,}
 
 # SAMPLING #####################################################################
 
@@ -53,14 +47,14 @@ def create_sampling_block() -> dict:
 # TARGET #######################################################################
 
 def create_target_block() -> dict:
-    __target = gradio.Radio(label='Score', value='Inputs', choices=['Inputs', 'Outputs', 'Both'], scale=1, interactive=True)
+    __target = gradio.Radio(label='Score', value='Inputs', choices=['Inputs', 'Everything'], scale=1, interactive=True)
     return {'target_block': __target}
 
 # DISPLAY ######################################################################
 
-def create_display_block() -> dict:
-    __display = gradio.Radio(label='Display', value='Tokens', choices=['Tokens', 'Indexes'], scale=1, interactive=True)
-    return {'display_block': __display}
+# def create_display_block() -> dict:
+#     __display = gradio.Radio(label='Display', value='Tokens', choices=['Tokens', 'Indexes'], scale=1, interactive=True)
+#     return {'display_block': __display}
 
 # INPUTS #######################################################################
 
@@ -71,17 +65,25 @@ def create_inputs_block() -> dict:
 # OUTPUTS ######################################################################
 
 def create_outputs_block() -> dict:
-    __output = gradio.HighlightedText(label='Scores', value='', scale=1, interactive=False, show_legend=False, show_inline_category=False, combine_adjacent=True, color_map=create_color_map(), elem_classes='white-text')
+    __output = gradio.HighlightedText(label='Scores', value='', scale=1, interactive=False, show_legend=False, show_inline_category=False, combine_adjacent=False, color_map=create_color_map(), elem_classes='white-text')
     return {'output_block': __output}
 
+# SELECT #######################################################################
+
+def create_selection_block() -> dict:
+    __position = gradio.Slider(label='Token Position', value=-1, minimum=-1, maximum=15, step=1, scale=1, interactive=True) # info='-1 to average on all tokens'
+    __layer = gradio.Slider(label='Layer Depth', value=12, minimum=-1, maximum=23, step=1, scale=1, interactive=True) # info='-1 to average on all layers'
+    __head = gradio.Slider(label='Attention Head', value=-1, minimum=-1, maximum=63, step=1, scale=1, interactive=True) # info='-1 to average on all heads'
+    return {
+        'position_block': __position,
+        'layer_block': __layer,
+        'head_block': __head,}
+
 # ACTIONS ######################################################################
 
 def create_actions_block() -> dict:
     __process = gradio.Button('Process', variant='primary', size='lg', scale=1, interactive=True)
-    __position = gradio.Slider(label='Position', value=-1, minimum=-1, maximum=15, step=1, scale=1, interactive=True) # info='-1 to average on all tokens'
-    return {
-        'process_block': __process,
-        'position_block': __position}
+    return {'process_block': __process,}
 
 # STATE ########################################################################
 
@@ -102,6 +104,8 @@ def create_layout(intro: str=INTRO) -> dict:
             with gradio.Row(equal_height=True):
                 __fields.update(create_inputs_block())
                 __fields.update(create_outputs_block())
+            with gradio.Row(equal_height=True):
+                __fields.update(create_selection_block())
             with gradio.Row(equal_height=True):
                 __fields.update(create_actions_block())
         with gradio.Tab('Settings') as __settings_tab:
@@ -113,7 +117,7 @@ def create_layout(intro: str=INTRO) -> dict:
                     __fields.update(create_sampling_block())
                 with gradio.Row(equal_height=True):
                     __fields.update(create_target_block())
-                    __fields.update(create_display_block())
+                    # __fields.update(create_display_block())
     return __fields
 
 # EVENTS #######################################################################
@@ -250,19 +254,37 @@ def create_app(title: str=TITLE, intro: str=INTRO, style: str=STYLE, model: str=
         # init the state
         __fields.update(create_state())
         # wire the input fields
+        __fields['tokens_block'].change(
+            fn=update_position_range,
+            inputs=[__fields[__k] for __k in ['position_block', 'tokens_block']],
+            outputs=__fields['position_block'],
+            queue=False,
+            show_progress='hidden')
+        __fields['model_block'].change(
+            fn=update_layer_range,
+            inputs=[__fields[__k] for __k in ['layer_block', 'model_block']],
+            outputs=__fields['layer_block'],
+            queue=False,
+            show_progress='hidden')
         __fields['process_block'].click(
             fn=__compute,
             inputs=[__fields[__k] for __k in ['tokens_block', 'topk_block', 'topp_block', 'position_block', 'layer_block', 'head_block', 'input_block']],
             outputs=[__fields[__k] for __k in ['output_block', 'input_state', 'output_state', 'attention_state']],
             queue=False,
             show_progress='full')
-        __fields['tokens_block'].change(
-            fn=update_position_range,
-            inputs=[__fields[__k] for __k in ['position_block', 'tokens_block']],
-            outputs=__fields['position_block'],
+        __fields['position_block'].change(
+            fn=update_text_highlight,
+            inputs=[__fields[__k] for __k in ['position_block', 'layer_block', 'head_block', 'input_state', 'output_state', 'attention_state']],
+            outputs=__fields['output_block'],
             queue=False,
             show_progress='hidden')
-        __fields['position_block'].change(
+        __fields['layer_block'].change(
+            fn=update_text_highlight,
+            inputs=[__fields[__k] for __k in ['position_block', 'layer_block', 'head_block', 'input_state', 'output_state', 'attention_state']],
+            outputs=__fields['output_block'],
+            queue=False,
+            show_progress='hidden')
+        __fields['head_block'].change(
             fn=update_text_highlight,
             inputs=[__fields[__k] for __k in ['position_block', 'layer_block', 'head_block', 'input_state', 'output_state', 'attention_state']],
             outputs=__fields['output_block'],

{psaiops-0.0.17 → psaiops-0.1.1}/psaiops/score/attention/lib.py

@@ -98,7 +98,7 @@ def reduce_attention_weights(
     __layer_slice = slice(None) if (__layer_idx < 0) else slice(__layer_idx, __layer_idx + 1)
     __sample_slice = slice(None)
     __head_slice = slice(None) if (__head_idx < 0) else slice(__head_idx, __head_idx + 1)
-    __token_slice = slice(input_dim, __output_dim) if (__token_idx < 0) else slice(input_dim + __token_idx, input_dim + __token_idx + 1)
+    __token_slice = slice(input_dim - 1, __output_dim) if (__token_idx < 0) else slice(input_dim + __token_idx - 1, input_dim + __token_idx)
     # filter the data
     __data = attention_data[__layer_slice, __sample_slice, __head_slice, __token_slice, slice(None)]
     # reduce all the axes but the last
@@ -119,7 +119,7 @@ def postprocess_attention_scores(
     __output_range = list(range(__output_dim - input_dim)) if (__token_idx < 0) else [__token_idx]
     __output_mask = torch.BoolTensor([__i in __output_range for __i in range(__output_dim - input_dim)])
     # normalize the scores
-    __input_scores = attention_data[__input_slice] / (attention_data[__input_slice].mean() + 1e-5)
+    __input_scores = attention_data[__input_slice] / (attention_data[__input_slice].max() + 1e-5)
     # round to obtain integer labels from 0 to 100
     __input_scores = torch.round(100.0 * __input_scores, decimals=0).type(torch.int32)
     # the generated tokens are not scored

{psaiops-0.0.17 → psaiops-0.1.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "psaiops"
-version = "0.0.17"
+version = "0.1.1"
 description = "Web apps to inspect & engineer NN activations."
 license = ".github/LICENSE.md"
 readme = ".github/README.md"