psaiops 0.0.10__py3-none-any.whl → 0.0.11__py3-none-any.whl

psaiops/score/attention/app.py

@@ -24,16 +24,16 @@ def create_color_map() -> dict:
 # INTRO ########################################################################
 
 def create_intro_block(intro: str) -> dict:
-    __intro = gradio.Markdown(intro)
+    __intro = gradio.Markdown(intro, scale=1)
     return {'intro_block': __intro}
 
 # MODEL ########################################################################
 
 def create_model_block() -> dict:
-    __model_dd = gradio.Dropdown(label='Model', value='openai/gpt-oss-20b', choices=['openai/gpt-oss-20b'], 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, interactive=True) # info='-1 to average on all layers'
-    __head_sl = gradio.Slider(label='Attention Head', value=-1, minimum=-1, maximum=63, step=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')
+    __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, scale=1, queue=False, show_progress='hidden')
     return {
         'model_block': __model_dd,
         'layer_block': __layer_sl,
@@ -42,37 +42,43 @@ def create_model_block() -> dict:
 # SAMPLING #####################################################################
 
 def create_sampling_block() -> dict:
-    __tokens = gradio.Slider(label='Tokens', value=32, minimum=0, maximum=128, step=1, interactive=True)
-    __topk = gradio.Slider(label='Top K', value=4, minimum=1, maximum=8, step=1, interactive=True)
-    __topp = gradio.Slider(label='Top P', value=0.8, minimum=0.0, maximum=1.0, step=0.1, interactive=True)
+    __tokens = gradio.Slider(label='Tokens', value=16, minimum=1, maximum=128, step=1, scale=1, interactive=True)
+    __topk = gradio.Slider(label='Top K', value=4, minimum=1, maximum=8, step=1, scale=1, interactive=True)
+    __topp = gradio.Slider(label='Top P', value=0.9, minimum=0.0, maximum=1.0, step=0.1, scale=1, interactive=True)
     return {
         'tokens_block': __tokens,
         'topk_block': __topk,
         'topp_block': __topp}
 
+# TARGET #######################################################################
+
+def create_target_block() -> dict:
+    __target = gradio.Radio(label='Score', value='Inputs', choices=['Inputs', 'Outputs', 'Both'], scale=1, interactive=True)
+    return {'target_block': __target}
+
 # DISPLAY ######################################################################
 
 def create_display_block() -> dict:
-    __display = gradio.Radio(label='Display', value='Tokens', choices=['Tokens', 'Indexes'], interactive=True)
+    __display = gradio.Radio(label='Display', value='Tokens', choices=['Tokens', 'Indexes'], scale=1, interactive=True)
     return {'display_block': __display}
 
 # INPUTS #######################################################################
 
 def create_inputs_block() -> dict:
-    __input = gradio.Textbox(label='Prompt', value='', placeholder='A string of tokens to score.', lines=4, show_copy_button=True, interactive=True)
+    __input = gradio.Textbox(label='Prompt', value='', placeholder='A string of tokens to score.', lines=4, scale=1, show_copy_button=True, interactive=True)
     return {'input_block': __input}
 
 # OUTPUTS ######################################################################
 
 def create_outputs_block() -> dict:
-    __output = gradio.HighlightedText(label='Scores', value='', 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=True, color_map=create_color_map(), elem_classes='white-text')
     return {'output_block': __output}
 
 # ACTIONS ######################################################################
 
 def create_actions_block() -> dict:
-    __process = gradio.Button('Process', variant='primary', size='lg', interactive=True)
-    __position = gradio.Slider(label='Position', value=-1, minimum=-1, maximum=128, step=1, interactive=True) # info='-1 to average on all tokens'
+    __process = gradio.Button('Process', variant='primary', size='lg', scale=1, interactive=True)
+    __position = gradio.Slider(label='Position', value=-1, minimum=-1, maximum=128, step=1, scale=1, interactive=True) # info='-1 to average on all tokens'
     return {
         'process_block': __process,
         'position_block': __position}
@@ -80,7 +86,10 @@ def create_actions_block() -> dict:
 # STATE ########################################################################
 
 def create_state() -> dict:
-    return {'attention_state': gradio.State(None), 'token_state': gradio.State(None)}
+    return {
+        'input_state': gradio.State(None),
+        'output_state': gradio.State(None),
+        'attention_state': gradio.State(None),}
 
 # LAYOUT #######################################################################
 
@@ -90,36 +99,121 @@ def create_layout(intro: str=INTRO) -> dict:
     with gradio.Tabs():
         with gradio.Tab('Score Tokens') as __main_tab:
             __fields.update({'main_tab': __main_tab})
-            with gradio.Row():
-                with gradio.Column(scale=1):
-                    __fields.update(create_inputs_block())
-                with gradio.Column(scale=1):
-                    __fields.update(create_outputs_block())
-            with gradio.Row():
+            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_actions_block())
         with gradio.Tab('Settings') as __settings_tab:
             __fields.update({'settings_tab': __settings_tab})
             with gradio.Column(scale=1):
-                with gradio.Row():
+                with gradio.Row(equal_height=True):
                     __fields.update(create_model_block())
-                with gradio.Row():
+                with gradio.Row(equal_height=True):
                     __fields.update(create_sampling_block())
-                with gradio.Row():
+                with gradio.Row(equal_height=True):
+                    __fields.update(create_target_block())
                     __fields.update(create_display_block())
     return __fields
 
 # EVENTS #######################################################################
 
-def update_layer_range(value: int, model: str) -> dict:
+def update_layer_range(value: float, model: str) -> dict:
     return gradio.update(maximum=35, value=min(35, int(value))) if '120b' in model else gradio.update(maximum=23, value=min(23, int(value)))
 
-def update_position_range(value: int, dimension: int) -> dict:
-    return gradio.update(maximum=dimension - 1, value=min(dimension - 1, value))
-
-def update_output_value(
-    attention_data: torch.Tensor=None,
-    token_data: torch.Tensor=None,) -> torch.Tensor:
-    return
+def update_position_range(value: float, tokens: list) -> dict:
+    return gradio.update(maximum=len(tokens) - 1, value=min(len(tokens) - 1, int(value)))
+
+def update_computation_state(
+    token_num: float,
+    topk_num: float,
+    topp_num: float,
+    prompt_str: str,
+    device_str: str,
+    model_obj: object,
+    tokenizer_obj: object,
+) -> tuple:
+    # sanitize the inputs
+    __limit = max(1, min(128, int(token_num)))
+    __topk = max(1, min(128, int(token_num)))
+    __topp = max(0.0, min(1.0, float(token_num)))
+    __prompt = prompt_str.strip()
+    __device = device_str if (device_str in ['cpu', 'cuda']) else 'cpu'
+    # handle all exceptions at once
+    try:
+        # dictionary {'input_ids': _, 'attention_mask': _}
+        __inputs = psaiops.score.attention.lib.preprocess_token_ids(
+            tokenizer_obj=tokenizer_obj,
+            prompt_str=__prompt,
+            device_str=__device)
+        # parse the inputs
+        __input_dim = int(__inputs['input_ids'].shape[-1])
+        # tensor (1, T)
+        __outputs = psaiops.score.attention.lib.generate_token_ids(
+            model_obj=model_obj,
+            input_args=__inputs,
+            token_num=__limit,
+            topk_num=__topk,
+            topp_num=__topp)
+        # tensor (L, S, H, T, T)
+        __attentions = psaiops.score.attention.lib.compute_attention_weights(
+            model_obj=model_obj,
+            token_obj=__outputs)
+        # detokenize the IDs
+        __tokens = psaiops.score.attention.lib.postprocess_token_ids(
+            tokenizer_obj=tokenizer_obj,
+            token_obj=__outputs)
+        # update each component => (input, output, attention) states
+        return (
+            gradio.update(value=__tokens[:__input_dim]),
+            gradio.update(value=__tokens[__input_dim:]),
+            gradio.update(value=__attentions),)
+    except:
+        raise Exception('Attention generation aborted with an error.')
+    finally:
+        return (gradio.update(), gradio.update(), gradio.update())
+
+def update_text_highlight(
+    token_idx: float,
+    layer_idx: float,
+    head_idx: float,
+    input_data: list,
+    output_data: list,
+    attention_data: torch.Tensor,
+) -> dict:
+    # sanitize the inputs
+    __input_data = input_data or []
+    __output_data = output_data or []
+    __attention_data = attention_data or torch.empty(0)
+    __input_dim = len(__input_data)
+    __token_idx = max(0, min(__input_dim, int(token_idx)))
+    __layer_idx = max(0, int(layer_idx))
+    __head_idx = max(0, int(head_idx))
+    # exit if the data has not yet been computed
+    if (not __input_data) or (not __output_data) or (attention_data is None) or (len(attention_data) == 0):
+        return gradio.update()
+    # handle all exceptions at once
+    try:
+        # concat input and output tokens
+        __tokens = __input_data + __output_data
+        # reduce the layer, sample, head and output token axes => tensor (T,)
+        __scores = psaiops.score.attention.lib.reduce_attention_weights(
+            attention_data=__attention_data,
+            token_idx=__token_idx,
+            layer_idx=__layer_idx,
+            head_idx=__head_idx,
+            input_dim=__input_dim)
+        # translate the scores into integer labels
+        __labels = psaiops.score.attention.lib.postprocess_attention_scores(
+            attention_data=__scores,
+            input_dim=__input_dim,
+            token_idx=__token_idx)
+        # update the component with [(token, label), ...]
+        return gradio.update(value=list(zip(__tokens, __labels)))
+    except:
+        raise Exception('Attention reduction aborted with an error.')
+    finally:
+        return gradio.update()
 
 # APP ##########################################################################
 
@@ -130,21 +224,40 @@ def create_app(title: str=TITLE, intro: str=INTRO, style: str=STYLE, model: str=
         __device = 'cuda' if torch.cuda.is_available() else 'cpu'
         __model = psaiops.score.attention.lib.get_model(name=model, device=__device)
         __tokenizer = psaiops.score.attention.lib.get_tokenizer(name=model, device=__device)
-        # adapt the scoring function
-        __score = functools.partial(psaiops.score.attention.lib.score_tokens, model_obj=__model, tokenizer_obj=__tokenizer, device_str=__device)
+        # adapt the computing function
+        __compute = functools.partial(update_computation_state, model_obj=__model, tokenizer_obj=__tokenizer, device_str=__device)
         # create the UI
         __fields.update(create_layout(intro=intro))
         # init the state
         __fields.update(create_state())
         # fetch the relevant fields
-        __button = __fields['process_block']
+        __button_block, __position_block, __output_block = (__fields['process_block'], __fields['position_block'], __fields['output_block'])
+        __output_state, __attention_state = (__fields['output_state'], __fields['attention_state'])
         # wire the input fields
-        __button.click(
-            fn=__score,
-            inputs=[__fields[__k] for __k in ['input_block', 'tokens_block', 'topk_block', 'topp_block', 'position_block', 'layer_block', 'head_block']],
-            outputs=__fields['output_block'],
+        __button_block.click(
+            fn=__compute,
+            inputs=[__fields[__k] for __k in ['tokens_block', 'topk_block', 'topp_block', 'input_block']],
+            outputs=[__fields[__k] for __k in ['input_state', 'output_state', 'attention_state']],
             queue=False,
             show_progress='full')
+        __output_state.change(
+            fn=update_position_range,
+            inputs=[__position_block, __output],
+            outputs=__position_block,
+            queue=False,
+            show_progress='hidden')
+        __attention_state.change(
+            fn=update_text_highlight,
+            inputs=[__fields[__k] for __k in ['position_block', 'layer_block', 'head_block', 'input_state', 'output_state', 'attention_state']],
+            outputs=__output_block,
+            queue=False,
+            show_progress='hidden')
+        __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=__output_block,
+            queue=False,
+            show_progress='hidden')
         # gradio application
         return __app
 

psaiops/score/attention/lib.py

@@ -28,6 +28,7 @@ def get_model(name: str, device: str='cpu'):
 
 # PREPROCESS #####################################################################
 
+@functools.lru_cache(maxsize=4)
 def preprocess_token_ids(
     tokenizer_obj: object,
     prompt_str: str,
@@ -137,7 +138,7 @@ def postprocess_token_ids(
     # back to token strings
     __tokens = tokenizer_obj.convert_ids_to_tokens(__indices)
     # normalize the tokens
-    return [__t.replace('Ġ', ' ') for __t in __tokens]
+    return [__t.replace(chr(0x0120), ' ').replace(chr(0x010a), '\n') for __t in __tokens]
 
 # COMPUTE ########################################################################
 
@@ -149,7 +150,6 @@ def score_tokens(
     token_idx: int,
     layer_idx: int,
     head_idx: int,
-    *,
     device_str: str,
     model_obj: object,
     tokenizer_obj: object,

{psaiops-0.0.10.dist-info → psaiops-0.0.11.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: psaiops
-Version: 0.0.10
+Version: 0.0.11
 Summary: Web apps to inspect & engineer NN activations.
 License: .github/LICENSE.md
 Author: apehex

{psaiops-0.0.10.dist-info → psaiops-0.0.11.dist-info}/RECORD

@@ -7,9 +7,9 @@ psaiops/elements/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,
 psaiops/elements/data.py,sha256=vGYeMN11uP9gs8rV6aSDffE_TeIX5PmdzWGwUpdGE2Y,906
 psaiops/score/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 psaiops/score/attention/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops/score/attention/app.py,sha256=Kp4Y4Qv9OkrezsqYdFLLWKmyAhC9BPv5J93-vQLRToM,7046
-psaiops/score/attention/lib.py,sha256=vWm8e_xUkM4ahgYQfaNK_kpU9eGkKAGHfgA-Sm1htmc,6869
+psaiops/score/attention/app.py,sha256=SS-cQ45HAv03rxGjzRNCV5BIfyP_Qry58bJ0b3OIxZU,11979
+psaiops/score/attention/lib.py,sha256=kPLdmWCYbpMnd7fi4ZeCvxiGzJuvTKYSXgRAm_w02KA,6928
 psaiops/steer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops-0.0.10.dist-info/METADATA,sha256=DP3q2K3C6jPQQI46mMY-7Vm-yQNJ04hG9ETqjyHLnLo,1222
-psaiops-0.0.10.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
-psaiops-0.0.10.dist-info/RECORD,,
+psaiops-0.0.11.dist-info/METADATA,sha256=s3jjhNDNG0Zm-pBoxrkuznmkuxJr-04Dxj6MbFsCNIA,1222
+psaiops-0.0.11.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
+psaiops-0.0.11.dist-info/RECORD,,