psaiops 0.0.2__py3-none-any.whl → 0.0.4__py3-none-any.whl

psaiops/score/attention/app.py

@@ -0,0 +1,149 @@
+import gradio
+
+import psaiops.score.attention.lib
+
+# META #########################################################################
+
+TITLE = '''Attention Scoring'''
+INTRO = '''Score each token according to the weights of the attention layers.\nThe model is fixed to "openai/gpt-oss-20b" for now.'''
+STYLE = ''''''
+
+MODEL = 'openai/gpt-oss-20b'
+
+# COLORS #######################################################################
+
+def create_color_map() -> dict:
+    return {
+        '-1': '#00ff00',
+        **{str(__i): '#{:02x}0000'.format(__i) for __i in range(101)}}
+
+# INTRO ########################################################################
+
+def create_intro_block(intro: str) -> dict:
+    __intro = gradio.Markdown(intro)
+    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')
+    return {
+        'model_block': __model_dd,
+        'layer_block': __layer_sl,
+        'head_block': __head_sl}
+
+# 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)
+    return {
+        'tokens_block': __tokens,
+        'topk_block': __topk,
+        'topp_block': __topp}
+
+# DISPLAY ######################################################################
+
+def create_display_block() -> dict:
+    __display = gradio.Radio(label='Display', value='Tokens', choices=['Tokens', 'Indexes'], 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)
+    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())
+    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'
+    return {
+        'process_block': __process,
+        'position_block': __position}
+
+# STATE ########################################################################
+
+def create_state() -> dict:
+    return {'attention_state': gradio.State(None), 'token_state': gradio.State(None)}
+
+# LAYOUT #######################################################################
+
+def create_layout(intro: str=INTRO) -> dict:
+    __fields = {}
+    __fields.update(create_intro_block(intro=intro))
+    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():
+                __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():
+                    __fields.update(create_model_block())
+                with gradio.Row():
+                    __fields.update(create_sampling_block())
+                with gradio.Row():
+                    __fields.update(create_display_block())
+    return __fields
+
+# EVENTS #######################################################################
+
+def update_layer_range(value: int, 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
+
+# APP ##########################################################################
+
+def create_app(title: str=TITLE, intro: str=INTRO, style: str=STYLE, model: str=MODEL) -> gradio.Blocks:
+    __fields = {}
+    with gradio.Blocks(theme=gradio.themes.Soft(), title=title, css=style) as __app:
+        # init
+        __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)
+        # create the UI
+        __fields.update(create_layout(intro=intro))
+        # init the state
+        __fields.update(create_state())
+        # fetch the relevant fields
+        __button = __fields['process_block']
+        # wire the input fields
+        __button.click(
+            fn=psaiops.score.attention.lib.score_tokens,
+            inputs=[__model, __tokenizer] + [__fields[__k] for __k in ['input_block', 'tokens_block', 'topk_block', 'topp_block', 'position_block', 'layer_block', 'head_block']] + [__device],
+            outputs=__fields['output_block'],
+            queue=False,
+            show_progress='full')
+        # gradio application
+        return __app
+
+# MAIN #########################################################################
+
+if __name__ == '__main__':
+    __app = create_app()
+    __app.launch(share=True, debug=True)

psaiops/score/attention/lib.py

@@ -0,0 +1,191 @@
+import functools
+
+import torch
+import transformers
+
+import deformers.models.openai.gptoss
+
+# LOAD #########################################################################
+
+@functools.lru_cache(maxsize=4)
+def get_tokenizer(name: str, device: str='cpu'):
+    return transformers.AutoTokenizer.from_pretrained(
+        name,
+        use_fast=True,
+        dtype='auto',
+        device_map=device)
+
+@functools.lru_cache(maxsize=2)
+def get_model(name: str, device: str='cpu'):
+    __model = deformers.models.openai.gptoss.GptOssForCausalInference.from_pretrained(
+        name,
+        dtype='auto',
+        device_map=device)
+    # toggle the inference mode (not training)
+    __model.eval()
+    # transformers model
+    return __model
+
+# PREPROCESS #####################################################################
+
+def preprocess_token_ids(
+    tokenizer_obj: object,
+    prompt_str: str,
+    device_str: str='cpu'
+) -> dict:
+    # tokenize
+    __inputs = tokenizer_obj(prompt_str, return_tensors='pt')
+    # move to the main device
+    return {__k: __v.to(device_str) for __k, __v in __inputs.items()}
+
+# GENERATE #######################################################################
+
+def generate_token_ids(
+    model_obj: object,
+    input_args: dict,
+    token_num: int,
+    topk_num: int = 4,
+    topp_num: float = 0.9,
+) -> torch.Tensor:
+    # generate completion
+    with torch.no_grad():
+        __outputs = model_obj.generate(
+            **input_args,
+            max_new_tokens=token_num,
+            do_sample=(0.0 < topp_num < 1.0) or (topk_num > 0),
+            top_k=topk_num if (topk_num > 0) else None,
+            top_p=topp_num if (0.0 < topp_num < 1.0) else None,
+            return_dict_in_generate=True,
+            output_hidden_states=False,
+            output_attentions=False,
+            output_scores=False,
+            early_stopping=True,
+            use_cache=True)
+    # full sequence
+    return __outputs.sequences # (1, T)
+
+# COMPUTE ########################################################################
+
+def compute_attention_weights(
+    model_obj: object,
+    token_obj: torch.Tensor,
+) -> torch.Tensor:
+    # process the full sequence
+    with torch.no_grad():
+        __outputs = model_obj(
+            input_ids=token_obj,
+            output_attentions=True,
+            return_dict=True)
+    # parse the outputs
+    return torch.stack(__outputs.attentions, dim=0)
+
+# REDUCE #######################################################################
+
+def reduce_attention_weights(
+    attention_data: torch.Tensor,
+    token_idx: int, # -1 => avg over all tokens
+    layer_idx: int, # -1 => avg over layers
+    head_idx: int, # -1 => avg over heads
+    input_dim: int,
+) -> torch.Tensor:
+    # parse
+    __layer_dim, __batch_dim, __head_dim, __output_dim, __output_dim = tuple(attention_data.shape) # L, B, H, T, T
+    __layer_idx = min(layer_idx, __layer_dim)
+    __head_idx = min(head_idx, __head_dim)
+    __token_idx = min(token_idx, __output_dim - input_dim - 1) # T = I + O
+    # select the relevant data along each axis
+    __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)
+    # filter the data
+    __data = attention_data[__layer_slice, __sample_slice, __head_slice, __token_slice, slice(None)]
+    # reduce all the axes but the last
+    return __data.mean(dim=tuple(range(len(__data.shape) - 1)))
+
+# FORMAT #########################################################################
+
+def postprocess_attention_scores(
+    attention_data: torch.Tensor, # (T,)
+    input_dim: int,
+    token_idx: int,
+) -> list:
+    __output_dim = int(attention_data.shape[-1])
+    # isolate the scores of the input prompt
+    __input_slice = slice(0, input_dim)
+    # mask the token that were used to compute the scores
+    __token_idx = min(token_idx, __output_dim - input_dim - 1) # T = I + O
+    __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].sum() + 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
+    __output_scores = torch.where(__output_mask, -1, 0).type(torch.int32)
+    # native list of serialized integers
+    return [str(__i) for __i in __input_scores.tolist() + __output_scores.tolist()] # (I,) + (O,) = (T,)
+
+# POSTPROCESS ####################################################################
+
+def postprocess_token_ids(
+    tokenizer_obj: object,
+    token_obj: torch.Tensor,
+) -> list:
+    # remove the batch axis
+    __indices = token_obj.squeeze().tolist()
+    # back to token strings
+    __tokens = tokenizer_obj.convert_ids_to_tokens(__indices)
+    # normalize the tokens
+    return [__t.replace('Ġ', ' ') for __t in __tokens]
+
+# COMPUTE ########################################################################
+
+def score_tokens(
+    model_obj: object,
+    tokenizer_obj: object,
+    prompt_str: str,
+    token_num: int=32,
+    topk_num: int = 4,
+    topp_num: float = 0.9,
+    token_idx: int, # -1 => avg over all tokens
+    layer_idx: int,   # -1 => avg over layers
+    head_idx: int,    # -1 => avg over heads
+    device_str: str='cuda',
+) -> list:
+    # dictionary {'input_ids': _, 'attention_mask': _}
+    __inputs = preprocess_token_ids(
+        tokenizer_obj=tokenizer_obj,
+        prompt_str=prompt_str,
+        device_str=device_str)
+    # parse the inputs
+    __input_dim = int(__inputs['input_ids'].shape[-1])
+    # tensor (1, T)
+    __outputs = generate_token_ids(
+        model_obj=model_obj,
+        input_args=__inputs,
+        token_num=token_num,
+        topk_num=topk_num,
+        topp_num=topp_num)
+    # tensor (L, S, H, T, T)
+    __attentions = compute_attention_weights(
+        model_obj=model_obj,
+        token_obj=__outputs)
+    # reduce the layer, sample, head and output token axes => tensor (T,)
+    __scores = reduce_attention_weights(
+        __attentions,
+        token_idx=token_idx,
+        layer_idx=layer_idx,
+        head_idx=head_idx,
+        input_dim=__input_dim)
+    # translate the scores into integer labels
+    __labels = postprocess_attention_scores(
+        __scores,
+        input_dim=__input_dim,
+        token_idx=token_idx)
+    # detokenize the IDs
+    __tokens = postprocess_token_ids(
+        tokenizer_obj=__tokenizer,
+        token_obj=__outputs)
+    # match tokens and labels for the HighlightedText field
+    return list(zip(__tokens, __labels))

{psaiops-0.0.2.dist-info → psaiops-0.0.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: psaiops
-Version: 0.0.2
+Version: 0.0.4
 Summary: Web apps to inspect & engineer NN activations.
 License: .github/LICENSE.md
 Author: apehex

psaiops-0.0.4.dist-info/RECORD

@@ -0,0 +1,15 @@
+psaiops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+psaiops/combine/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+psaiops/compose/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+psaiops/compose/contrast/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+psaiops/edit/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+psaiops/elements/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+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=3kltkxpbSNXWAGuzhO1cXNnGUOWNFBgUHA17wJMrICU,6803
+psaiops/score/attention/lib.py,sha256=MCpXql8INTdXnFkEBoyJp-unazC9nGbD7y81G2z3l9U,6958
+psaiops/steer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+psaiops-0.0.4.dist-info/METADATA,sha256=xOUeGbX3yfu06s2HLDPtYkHC8StVNUJB7DQ-kH_J0E0,1221
+psaiops-0.0.4.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
+psaiops-0.0.4.dist-info/RECORD,,

psaiops/score/app.py

@@ -1,27 +0,0 @@
-import gradio
-
-# META #########################################################################
-
-TITLE = '''Token Scoring'''
-INTRO = '''Score each input / output token according to a given metric.'''
-STYLE = ''''''
-
-# MODEL ########################################################################
-
-def create_model_tab() -> None:
-    pass
-
-# ROOT #########################################################################
-
-def create_root_block(title: str=TITLE, intro: str=INTRO, style: str=STYLE) -> gradio.Block:
-    with gradio.Blocks(theme=gradio.themes.Soft(), title=title, css=style) as __app:
-        with gradio.Row():
-            with gradio.Column(scale=1):
-                gradio.Markdown(intro)
-        return __app
-
-# MAIN #########################################################################
-
-if __name__ == '__main__':
-    __app = create_root_block()
-    __app.launch(share=True, debug=True)

psaiops-0.0.2.dist-info/RECORD

@@ -1,10 +0,0 @@
-psaiops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops/compose/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops/elements/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops/elements/data.py,sha256=vGYeMN11uP9gs8rV6aSDffE_TeIX5PmdzWGwUpdGE2Y,906
-psaiops/score/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops/score/app.py,sha256=7HfVEFdzJ1IarJ7A4FaWR8GeeGw2tttG993Qn-gzBmY,910
-psaiops/steer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-psaiops-0.0.2.dist-info/METADATA,sha256=Xu81-r81hohISmebG13Cftrr5tyETxei_Jar8ZxWCZM,1221
-psaiops-0.0.2.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
-psaiops-0.0.2.dist-info/RECORD,,