psaiops 0.0.13__py3-none-any.whl → 0.4.0__py3-none-any.whl

psaiops/compose/contrast/lib.py

@@ -0,0 +1,143 @@
+import functools
+
+import torch
+import torch.nn.modules
+
+import mlable.shapes
+import psaiops.common.tokenizer
+
+# HOOK #########################################################################
+
+def capture_hidden_activation(
+    module: torch.nn.modules.Module,
+    inputs: torch.Tensor,
+    outputs: torch.Tensor,
+    index: int,
+    captured: dict,
+) -> None:
+    captured[index] = outputs # (B, S, E)
+
+# MASKS ########################################################################
+
+def compute_sequence_mask(
+    tokens: torch.Tensor, # (B, S)
+    # masks: torch.Tensor, # (B, S)
+) -> torch.Tensor:
+    __shape = mlable.shapes.divide(tokens.shape, axis=0, factor=2, insert=True)
+    # group the samples two by two
+    __data = tokens.reshape(__shape)
+    # compare each sample with its neighbor
+    __masks = __data[:, :1]  != __data[:, 1:]
+    # apply the same mask to both samples
+    return __masks.expand(__shape).reshape(tokens.shape)
+
+# REDUCTION ####################################################################
+
+def compute_delta_activation(
+    data: torch.Tensor, # (B, S, E)
+    masks: torch.Tensor, # (B, S,)
+    signs: torch.Tensor, # (B,)
+    keepdim: bool=True,
+) -> torch.Tensor:
+    __dtype = data.dtype
+    __device = data.device
+    __dim0, __dim1, __dim2 = tuple(data.shape)
+    # sign each sample along the batch axis
+    __shape = tuple(mlable.shapes.filter(data.shape, axes=[0]))
+    __signs = signs.to(dtype=__dtype, device=__device).view(__shape)
+    # combine along the batch axis to keep the shortest mask on the sequence axis
+    __shape = tuple(mlable.shapes.filter(data.shape, axes=[0, 1]))
+    __masks = masks.to(dtype=__dtype, device=__device).view(__shape)
+    # mean factor: half the signs size along the batch axis and the number of positions kept along the sequence axis
+    __factor = (0.5 * float(__dim0) * __masks.sum(dim=1, keepdim=True)).clamp(min=1.0)
+    # take the difference along the batch axis and the average along the sequence axis
+    return (data * __signs * __masks / __factor).sum(dim=[0, 1], keepdim=keepdim)
+
+# DELTA ########################################################################
+
+def add_delta_activation(
+    module: torch.nn.modules.Module,
+    inputs: torch.Tensor,
+    outputs: torch.Tensor,
+    delta: torch.Tensor,
+    alpha: torch.Tensor,
+    beta: torch.Tensor,
+) -> torch.Tensor:
+    # expand the single feature axis of the delta
+    __shape = mlable.shapes.filter(outputs.shape, axes=[-1])
+    # rescale the delta
+    return alpha * outputs + beta * delta.view(__shape)
+
+# MAIN #########################################################################
+
+def steer_model_output(
+    positive_str: str,
+    negative_str: str,
+    prompt_str: str,
+    positive_rate: float,
+    negative_rate: float,
+    prompt_rate: float,
+    token_num: int,
+    topk_num: int,
+    topp_num: float,
+    layer_idx: int,
+    device_str: str,
+    model_obj: object,
+    tokenizer_obj: object,
+) -> str:
+    # parse & sanitize
+    __prompt0 = positive_str.strip()
+    __prompt1 = negative_str.strip()
+    __prompt2 = prompt_str.strip()
+    __alpha0 = max(0.0, float(positive_rate))
+    __alpha1 = max(0.0, float(negative_rate))
+    __alpha2 = max(0.0, float(prompt_rate))
+    __count = max(1, int(token_num))
+    __topk = max(1, int(topk_num))
+    __topp = max(0.0, float(topp_num))
+    __index = max(0, int(layer_idx))
+    # store hidden states
+    __captured = {}
+    # stop if inputs are missing
+    if not (__prompt0 and __prompt1 and __prompt2):
+        return ''
+    # tokenize the 2 prompts and pad to same length
+    __inputs = psaiops.common.tokenizer.preprocess_token_ids(tokenizer=tokenizer_obj, prompts=(__prompt0, __prompt1), device=device_str)
+    # forward hook to capture output hidden state
+    __hook = functools.partial(capture_hidden_activation, index=__index, captured=__captured)
+    # attach to the model
+    __handle = model_obj.model.layers[__index].register_forward_hook(__hook)
+    with torch.no_grad():
+        # inference mode
+        model_obj.eval().to(device_str)
+        # prefill with a single forward
+        __outputs = model_obj(**__inputs, use_cache=True, output_attentions=False, output_hidden_states=False, return_dict=True)
+    # stop capturing activations
+    __handle.remove()
+    # select only the positions where the tokens differ
+    __masks = compute_sequence_mask(tokens=__inputs['input_ids'])
+    # activation delta at layer L
+    __delta = compute_delta_activation(data=__captured[__index], masks=__masks, signs=torch.Tensor([1, -1]), keepdim=False)
+    # add the delta on every forward pass
+    __hook = functools.partial(add_delta_activation, alpha=__alpha2, beta=0.5 * (__alpha0 + __alpha1), delta=__delta)
+    # attach to the model
+    __handle = model_obj.model.layers[__index].register_forward_hook(__hook)
+    # now process the user input
+    __inputs = psaiops.common.tokenizer.preprocess_token_ids(tokenizer=tokenizer_obj, prompts=(prompt_str,), device=device_str)
+    # generate the new with tampered activations
+    with torch.no_grad():
+        __outputs = model_obj.generate(
+            **__inputs,
+            max_new_tokens=__count,
+            do_sample=(0.0 < __topp < 1.0) or (__topk > 0),
+            top_k=__topk if (__topk > 0) else None,
+            top_p=__topp if (0.0 < __topp <= 1.0) else None,
+            return_dict_in_generate=True,
+            output_hidden_states=False,
+            output_attentions=False,
+            output_scores=False,
+            use_cache=True)
+    # stop altering the activations
+    __handle.remove()
+    # single string
+    return tokenizer_obj.decode(__outputs.sequences[0], skip_special_tokens=True)

psaiops/compose/maths/app.py

@@ -0,0 +1,323 @@
+import functools
+import itertools
+
+import gradio
+import pandas
+import torch
+import torch.cuda
+
+import psaiops.common.model
+import psaiops.common.tokenizer
+
+# META #########################################################################
+
+MODEL = 'openai/gpt-oss-20b'
+
+STYLE = '''.giga-text input { font-size: 32px; }'''
+TITLE = '''Activation Maths'''
+INTRO = '''Compose prompts in the latent space.\nUnder construction, only "openai/gpt-oss-20b" is available for now.'''
+
+COUNT = 8
+
+# COLORS #######################################################################
+
+def create_color_map() -> dict:
+    return {
+        '-1': '#004444',
+        **{str(__i): '#{:02x}0000'.format(int(2.55 * __i)) for __i in range(101)}}
+
+# INTRO ########################################################################
+
+def create_intro_block(intro: str) -> dict:
+    __intro = gradio.Markdown(intro, line_breaks=True)
+    return {'intro_block': __intro}
+
+# MODEL ########################################################################
+
+def create_model_block() -> dict:
+    __model = gradio.Dropdown(label='Model ID', 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 = gradio.Slider(label='Layer Depth', value=12, minimum=0, maximum=23, step=1, scale=1, interactive=True)
+    return {
+        'model_block': __model,
+        'layer_block': __layer,}
+
+# SAMPLING #####################################################################
+
+def create_sampling_block() -> dict:
+    __tokens = gradio.Slider(label='Tokens', value=32, 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,}
+
+# REDUCTION ####################################################################
+
+def create_reduction_block() -> dict:
+    __from = gradio.Slider(label='Average From', value=0, minimum=0, maximum=256, step=1, scale=1, interactive=True)
+    __to = gradio.Slider(label='Average To', value=256, minimum=0, maximum=256, step=1, scale=1, interactive=True)
+    return {
+        'from_block': __from,
+        'to_block': __to,}
+
+# INPUTS #######################################################################
+
+def create_inputs_row(operation: str='', index: int=0) -> dict:
+    with gradio.Row(equal_height=True, visible=(index == 0)) as __row:
+        __operation = gradio.Dropdown(
+            label=f'Operation',
+            value='' if (index == 0) else operation,
+            choices=(index == 0) * [''] + ['+', '-', 'x', '.', 'µ', '='],
+            elem_classes='giga-text',
+            scale=1,
+            show_label=(index == 0),
+            allow_custom_value=False,
+            multiselect=False,
+            interactive=(index != 0),
+            visible=(index == 0))
+        __alpha = gradio.Slider(
+            label='Factor',
+            value=1.0,
+            minimum=0.0,
+            maximum=8.0,
+            step=0.1,
+            scale=1,
+            show_label=(index == 0),
+            interactive=True,
+            visible=(index == 0))
+        __input = gradio.Textbox(
+            label=f'Prompt',
+            value='',
+            placeholder='Some text.',
+            lines=2,
+            max_lines=2,
+            scale=8,
+            show_label=(index == 0),
+            show_copy_button=True,
+            interactive=True,
+            visible=(index == 0))
+        __delete = gradio.Button(
+            value='✖',
+            variant='secondary',
+            size='lg',
+            scale=1,
+            interactive=(index != 0),
+            visible=(index == 0))
+    return {
+        f'row_{index}_block': __row,
+        f'operation_{index}_block': __operation,
+        f'factor_{index}_block': __alpha,
+        f'prompt_{index}_block': __input,
+        f'button_{index}_block': __delete,}
+
+# OUTPUTS ######################################################################
+
+def create_outputs_block() -> dict:
+    __output = gradio.Textbox(label='= Total', value='', placeholder='Some text.', lines=2, max_lines=8, scale=1, show_label=True, show_copy_button=True, interactive=False)
+    return {'output_block': __output}
+
+# ACTIONS ######################################################################
+
+def create_actions_block() -> dict:
+    __add = gradio.Button(value='Add', variant='primary', size='lg', scale=1, interactive=True)
+    __process = gradio.Button(value='Process', variant='primary', size='lg', scale=1, interactive=True)
+    return {
+        'show_block': __add,
+        'process_block': __process,}
+
+# TABLE ########################################################################
+
+def create_table_block() -> dict:
+    __table = gradio.DataFrame(label='Summary', type='numpy', headers=None,  row_count=4, col_count=256, scale=1, interactive=False)
+    return {'table_block': __table,}
+
+# STATE ########################################################################
+
+def default_state(visible: bool=False) -> dict:
+    return {'visible': visible, 'operation': '+', 'factor': 1.0, 'prompt': ''}
+
+def create_state(limit: int=COUNT) -> dict:
+    return {
+        'cache_block': gradio.State(
+            [default_state(True)] + [default_state(False) for _ in range(limit - 1)])}
+
+# LAYOUT #######################################################################
+
+def create_layout(intro: str=INTRO, limit: int=COUNT) -> dict:
+    __fields = {}
+    __fields.update(create_intro_block(intro=intro))
+    with gradio.Tabs():
+        with gradio.Tab('Equation') as __main_tab:
+            __fields.update({'main_tab': __main_tab})
+            for __i in range(limit):
+                __fields.update(create_inputs_row(operation='+', index=__i))
+            with gradio.Row(equal_height=True):
+                __fields.update(create_outputs_block())
+            with gradio.Row(equal_height=True):
+                __fields.update(create_actions_block())
+        with gradio.Tab('Details') as __details_tab:
+            __fields.update({'details_tab': __details_tab})
+            with gradio.Row(equal_height=True):
+                __fields.update(create_table_block())
+        with gradio.Tab('Settings') as __settings_tab:
+            __fields.update({'settings_tab': __settings_tab})
+            with gradio.Column(scale=1):
+                with gradio.Row(equal_height=True):
+                    __fields.update(create_model_block())
+                with gradio.Row(equal_height=True):
+                    __fields.update(create_sampling_block())
+                with gradio.Row(equal_height=True):
+                    __fields.update(create_reduction_block())
+                    # __fields.update(create_display_block())
+    return __fields
+
+# DYNAMIC ######################################################################
+
+def get_input_rows(inputs: dict, limit: int=COUNT) -> list:
+    return list(itertools.chain.from_iterable([
+        [
+            inputs.get(f'row_{__i}_block', None),
+            inputs.get(f'operation_{__i}_block', None),
+            inputs.get(f'factor_{__i}_block', None),
+            inputs.get(f'prompt_{__i}_block', None),
+            inputs.get(f'button_{__i}_block', None),]
+        for __i in range(limit)]))
+
+def render_input_rows(rows: list) -> list:
+    return list(itertools.chain.from_iterable([
+        [
+            gradio.update(visible=__r.get('visible', False)),
+            gradio.update(visible=__r.get('visible', False), value=__r.get('operation', '')),
+            gradio.update(visible=__r.get('visible', False), value=__r.get('factor', 1.0)),
+            gradio.update(visible=__r.get('visible', False), value=__r.get('prompt', '')),
+            gradio.update(visible=__r.get('visible', False))]
+        for __r in rows]))
+
+def show_input_row(rows: list) -> tuple:
+    __count = 0
+    __rows = list(rows)
+    for __i in range(len(__rows)):
+        # count the number of hidden rows (before changing their state)
+        __count = __count + int(not __rows[__i]['visible'])
+        # all the visible rows stay the same and the first hidden row is toggled
+        __rows[__i]['visible'] = __rows[__i]['visible'] or (__count < 2)
+    # update state and components
+    return __rows, *render_input_rows(__rows)
+
+def hide_input_row(rows: list, index: int) -> tuple:
+    __rows = list(rows)
+    # always show the first row
+    if 0 < index < len(__rows):
+        # remove the target row
+        __rows.pop(index)
+        # keep the number of rows constant
+        __rows.append({'visible': False, 'operation': '+', 'factor': 1.0, 'prompt': ''})
+    # update state and components
+    return __rows, *render_input_rows(__rows)
+
+# EVENTS #######################################################################
+
+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_input_cache(cache: list, value: any, index: int, field: str) -> list:
+    __cache = list(cache)
+    __cache[index][field] = value
+    return __cache
+
+def update_operation_cache(cache: list, index: int, value: any) -> list:
+    return update_input_cache(cache=cache, index=int(index), value=str(value), field='operation')
+
+def update_factor_cache(cache: list, index: int, value: any) -> list:
+    return update_input_cache(cache=cache, index=int(index), value=float(value), field='factor')
+
+def update_prompt_cache(cache: list, index: int, value: any) -> list:
+    return update_input_cache(cache=cache, index=int(index), value=str(value), field='prompt')
+
+def update_table_data(tokenizer: object) -> callable:
+    # called with unpacked arguments
+    def __update_table_data(*prompts: list) -> list:
+        # array of token IDs
+        __outputs = tokenizer(prompts, return_tensors='pt', padding=True)
+        # array of token strings
+        __tokens = [tokenizer.convert_ids_to_tokens(__s) for __s in __outputs['input_ids']]
+        # shift the special characters
+        return [[__t.replace(chr(0x0120), ' ').replace(chr(0x010a), '\\n') for __t in __s] for __s in __tokens]
+    # fixed to a given tokenizer
+    return __update_table_data
+
+# APP ##########################################################################
+
+def create_app(title: str=TITLE, intro: str=INTRO, style: str=STYLE, limit: int=COUNT, model: str=MODEL) -> gradio.Blocks:
+    __inputs = {}
+    with gradio.Blocks(theme=gradio.themes.Soft(), title=title, css=style) as __app:
+        # load the model
+        __device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        # __model = psaiops.common.model.get_model(name=model, device=__device)
+        __tokenizer = psaiops.common.tokenizer.get_tokenizer(name=model, device=__device)
+        # create the UI
+        __inputs.update(create_layout(intro=intro, limit=limit))
+        # init the state
+        __inputs.update(create_state(limit=limit))
+        # apply the configuration
+        __format = update_table_data(tokenizer=__tokenizer)
+        # change the depth of the model
+        __inputs['model_block'].change(
+            fn=update_layer_range,
+            inputs=[__inputs[__k] for __k in ['layer_block', 'model_block']],
+            outputs=__inputs['layer_block'],
+            queue=False,
+            show_progress='hidden')
+        # show hidden row
+        __inputs['show_block'].click(
+            fn=show_input_row,
+            inputs=[__inputs['cache_block']],
+            outputs=[__inputs['cache_block']] + get_input_rows(inputs=__inputs, limit=limit),
+            queue=False,
+            show_progress='hidden')
+        # update the table
+        __inputs['details_tab'].select(
+            fn=__format,
+            inputs=[__inputs[f'prompt_{__i}_block'] for __i in range(limit)] + [__inputs['output_block']],
+            outputs=__inputs['table_block'],
+            queue=False,
+            show_progress='hidden')
+        # link each row of inputs to the cache
+        for __i in range(limit):
+            # update the target operation in the cache
+            __inputs[f'operation_{__i}_block'].change(
+                fn=update_operation_cache,
+                inputs=[__inputs['cache_block'], gradio.State(__i), __inputs[f'operation_{__i}_block']],
+                outputs=__inputs['cache_block'],
+                queue=False,
+                show_progress='hidden')
+            # update the target factor in the cache
+            __inputs[f'factor_{__i}_block'].change(
+                fn=update_factor_cache,
+                inputs=[__inputs['cache_block'], gradio.State(__i), __inputs[f'factor_{__i}_block']],
+                outputs=__inputs['cache_block'],
+                queue=False,
+                show_progress='hidden')
+            # update the target prompt in the cache
+            __inputs[f'prompt_{__i}_block'].change(
+                fn=update_prompt_cache,
+                inputs=[__inputs['cache_block'], gradio.State(__i), __inputs[f'prompt_{__i}_block']],
+                outputs=__inputs['cache_block'],
+                queue=False,
+                show_progress='hidden')
+            # hide the target row
+            __inputs[f'button_{__i}_block'].click(
+                fn=hide_input_row,
+                inputs=[__inputs['cache_block'], gradio.State(__i)],
+                outputs=[__inputs['cache_block']] + get_input_rows(inputs=__inputs, limit=limit),
+                queue=False,
+                show_progress='hidden')
+        # gradio application
+        return __app
+
+# MAIN #########################################################################
+
+if __name__ == '__main__':
+    __app = create_app()
+    __app.launch(share=True, debug=True)

psaiops/compose/maths/lib.py

@@ -0,0 +1 @@
+import torch