Hooks

ablate_pos_keep_self_attn_hook(module, args, kwargs, output, ablation_queries)

Hook function to ablate the tokens in the attention mask but keeping the self attn weigths. It will set to 0 the row of tokens to ablate except for the las position

Source code in easyroutine/interpretability/hooks.py

def ablate_pos_keep_self_attn_hook(
    module,
    args,
    kwargs,
    output,
    ablation_queries: pd.DataFrame,
):
    r"""
    Hook function to ablate the tokens in the attention
    mask but keeping the self attn weigths.
    It will set to 0 the row of tokens to ablate except for
    the las position
    """
    b = process_args_kwargs_output(args, kwargs, output)
    Warning("This function is deprecated. Use ablate_attn_mat_hook instead")
    attn_matrix = b.data
    # initial_shape = attn_matrix.shape

    for head, pos in zip(
        ablation_queries["head"], ablation_queries["pos_token_to_ablate"]
    ):
        attn_matrix[0, head, pos, :-1] = 0

    b.copy_(attn_matrix)

    return b

attention_pattern_head(module, args, kwargs, output, token_indexes, layer, cache, head='all', act_on_input=False, attn_pattern_avg='none', attn_pattern_row_partition=None)

Hook function to extract the attention pattern of the heads. It will extract the attention pattern. As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)

Parameters:

Name	Type	Description	Default
-	args	the input args of the hook function	required
-	kwargs	the input kwargs of the hook function	required
-	output	the output of the hook function	required
-	token_indexes(List[Tuple])	the indexes of the tokens to extract the attention pattern	required
-	layer (int	the layer of the model	required
-	cache (ActivationCache	the cache where to save the activations	required
-	head (Union[str, int]	the head of the model. If "all" is passed, it will extract all the heads of the layer	required
-	attn_pattern_avg (Literal["mean", "sum", "baseline_ratio", "none"]	the method to average the attention pattern	required
-	attn_pattern_row_partition (List[int]	the indexes of the tokens to partition the attention pattern	required

Avg strategies

If the attn_pattern_avg is not "none", the attention pattern is divided in blocks and the average value of each block is computed, using the method specified in attn_pattern_avg. The idea is to partition the attention pattern into groups of tokens, and then compute a single average value for each group. The pattern is divided into len(attn_pattern_row_partition) x len(token_indexes) blocks, where each block is a subset of the attention pattern. Each block B_ij is defined to have the indeces of the rows in attn_pattern_row_partition[i] and the columns in token_indexes[j]. If attn_pattern_row_partition is None, then the rows are the same as token_indexes.

0| a_00 0 0 0 0 0 0 token_indexes = [(1,3), (4)] 1| a_10 a_11 0 0 0 0 0 attn_pattern_row_partition = [(0,1)] 2| a_20 a_21 a_22 0 0 0 0 3| a_30 a_31 a_32 a_33 0 0 0 4| a_40 a_41 a_42 a_43 a_44 0 0 5| a_50 a_51 a_52 a_53 a_54 a_55 0 6| a_60 a_61 a_62 a_63 a_64 a_65 a_66

---

0    1     2   3     4    5    6

• Block B_00:
  • Rows: 0,1
  • Columns: 1,2,3
  • Block: [a_01, a_02, a_03, a_11, a_12, a_13]
• Block B_01:
  • Rows: 0,1
  • Columns: 4
  • Block: [a_04, a_14]

If attn_pattern_avg is "mean", the average value for each block is computed as the mean of the block, so the output will be: batch n_row_blocks n_col_blocks So in this case, the output will be: batch 1 2 where the first value is the average of the first block and the second value is the average of the second block.

The method to compute a single value for each block is specified by the attn_pattern_avg parameter, and can be one of the following: - "mean": Compute the mean of the block. - "sum": Compute the sum of the block. - "baseline_ratio": Compute the ratio of the observed average attention to the expected average attention. The expected average attention is computed by assuming that attention is uniform across the block. So, for each row in attn_pattern_row_partition, we compute the fraction of allowed keys that belong to token_indexes. The expected average attention is the sum of these fractions divided by the number of rows. The final ratio is the observed average attention divided by the expected average attention.

Source code in easyroutine/interpretability/hooks.py

def attention_pattern_head(
    module,
    args,
    kwargs,
    output,
    token_indexes,
    layer,
    cache,
    head: Union[str, int] = "all",
    act_on_input=False,
    attn_pattern_avg: Literal["mean", "sum", "baseline_ratio", "none"] = "none",
    attn_pattern_row_partition=None,
):
    """
    Hook function to extract the attention pattern of the heads. It will extract the attention pattern.
    As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)

    Arguments:
        - args: the input args of the hook function
        - kwargs: the input kwargs of the hook function
        - output: the output of the hook function
        - token_indexes (List[Tuple]) : the indexes of the tokens to extract the attention pattern
        - layer (int): the layer of the model
        - cache (ActivationCache): the cache where to save the activations
        - head (Union[str, int]): the head of the model. If "all" is passed, it will extract all the heads of the layer
        - attn_pattern_avg (Literal["mean", "sum", "baseline_ratio", "none"]): the method to average the attention pattern
        - attn_pattern_row_partition (List[int]): the indexes of the tokens to partition the attention pattern

    Avg strategies:
        If the attn_pattern_avg is not "none", the attention pattern is divided in blocks and the average value of each block is computed, using the method specified in attn_pattern_avg.
        The idea is to partition the attention pattern into groups of tokens, and then compute a single average value for each group. The pattern is divided into len(attn_pattern_row_partition) x len(token_indexes) blocks, where each block is a subset of the attention pattern. Each block B_ij is defined to have the indeces of the rows in attn_pattern_row_partition[i] and the columns in token_indexes[j]. If attn_pattern_row_partition is None, then the rows are the same as token_indexes.

        0| a_00 0    0    0    0    0    0              token_indexes = [(1,3), (4)]
        1| a_10 a_11 0    0    0    0    0              attn_pattern_row_partition = [(0,1)]
        2| a_20 a_21 a_22 0    0    0    0
        3| a_30 a_31 a_32 a_33 0    0    0
        4| a_40 a_41 a_42 a_43 a_44 0    0
        5| a_50 a_51 a_52 a_53 a_54 a_55 0
        6| a_60 a_61 a_62 a_63 a_64 a_65 a_66
           ------------------------------
            0    1     2   3     4    5    6

        - Block B_00:
            - Rows: 0,1
            - Columns: 1,2,3
            - Block: [a_01, a_02, a_03, a_11, a_12, a_13]
        - Block B_01:
            - Rows: 0,1
            - Columns: 4
            - Block: [a_04, a_14]


        If attn_pattern_avg is "mean", the average value for each block is computed as the mean of the block, so the output will be: batch n_row_blocks n_col_blocks
        So in this case, the output will be: batch 1 2 where the first value is the average of the first block and the second value is the average of the second block.

        The method to compute a single value for each block is specified by the attn_pattern_avg parameter, and can be one of the following:
        - "mean": Compute the mean of the block.
        - "sum": Compute the sum of the block.
        - "baseline_ratio": Compute the ratio of the observed average attention to the expected average attention. The expected average attention is computed by assuming that attention is uniform across the block. So, for each row in attn_pattern_row_partition, we compute the fraction of allowed keys that belong to token_indexes. The expected average attention is the sum of these fractions divided by the number of rows. The final ratio is the observed average attention divided by the expected average attention.


    """
    # first get the attention pattern
    b = process_args_kwargs_output(args, kwargs, output)

    attn_pattern = b.data.detach().clone()  # (batch, num_heads,seq_len, seq_len)

    if head == "all":
        head_indices = range(attn_pattern.size(1))
    else:
        head_indices = [head]

    if attn_pattern_row_partition is not None:
        token_indexes_group1 = attn_pattern_row_partition
    else:
        token_indexes_group1 = token_indexes

    # For each token group (each tuple in token_indexes), compute a single average value.
    if attn_pattern_row_partition is not None:
        for h in head_indices:
            # For head h, pattern has shape [batch, seq_len, seq_len].
            group_avgs = []

            # Generate all combinations of groups
            for group1 in token_indexes_group1:
                for group2 in token_indexes:
                    # Extract the attention block for this combination.
                    attn_block = attn_pattern[:, h, list(group1), :][:, :, list(group2)]

                    # Depending on the selected averaging method, compute a metric.
                    if attn_pattern_avg == "mean":
                        # Simple mean over the block.
                        avg_val = torch.mean(attn_block, dim=(-2, -1))  # shape: [batch]

                    elif attn_pattern_avg == "sum":
                        # Simple sum over the block.
                        avg_val = torch.sum(attn_block, dim=(-2, -1))

                    elif attn_pattern_avg == "baseline_ratio":
                        # ---- Step 1. Compute the observed average attention in the block.
                        observed_val = torch.mean(
                            attn_block, dim=(-2, -1)
                        )  # shape: [batch]

                        # ---- Step 2. Compute the baseline expectation.
                        # For each row (i.e. token index) in group1, we calculate the fraction of allowed keys
                        # that belong to group2. Because the attention is lower-triangular,
                        # a row with index 'i' can only attend to tokens with indices <= i.
                        # Thus, for each row 'i' in group1, the expected fraction (if uniform) is:
                        #      (# of tokens in group2 with index <= i) / (i+1)
                        baseline_list = []
                        for i in group1:
                            # Count the number of tokens in group2 that are allowed for row i.
                            allowed_count = sum(1 for j in group2 if j <= i)
                            # Total keys available for row i (assuming indices start at 0).
                            total_allowed = i + 1
                            # Avoid division by zero (should not happen if i>=0).
                            baseline_ratio = (
                                allowed_count / total_allowed
                                if total_allowed > 0
                                else 0.0
                            )
                            baseline_list.append(baseline_ratio)

                        # Average the per-row baseline over all rows in group1.
                        # This represents the expected average attention to group2 if it were uniformly distributed.
                        baseline_val = sum(baseline_list) / len(baseline_list)

                        # ---- Step 3. Compute the final ratio.
                        # We compare the observed average attention to the baseline expectation.
                        # A value > 1 means that, on average, attention in this block is higher than expected.
                        # Expand baseline_val to match the batch shape for element-wise division.
                        baseline_tensor = torch.tensor(
                            baseline_val, device=observed_val.device
                        ).expand_as(observed_val)
                        avg_val = observed_val / baseline_tensor
                    else:
                        avg_val = attn_block

                    if attn_pattern_avg != "none":
                        # Append the computed metric for this block (keeping the batch dimension).
                        group_avgs.append(avg_val.unsqueeze(1))  # shape: [batch, 1]
                    else:
                        # If no averaging is requested, store the block directly.
                        group_avgs.append(attn_block)

            if attn_pattern_avg != "none":
                pattern_avg = einops.rearrange(
                    torch.cat(group_avgs, dim=1),
                    "batch (G1 G2) -> batch G1 G2",
                    G1=len(token_indexes_group1),
                    G2=len(token_indexes),
                )
            else:
                try:
                    pattern_avg = torch.cat(group_avgs, dim=1)
                except:
                    raise ValueError(
                        f"Error concatenating group_avgs with shapes {[x.shape for x in group_avgs]}"
                    )

            # Add the pattern to the cache
            cache[f"pattern_L{layer}H{h}"] = pattern_avg
    else:
        # Without averaging, flatten token_indexes into one list.
        flatten_indexes = [item for tup in token_indexes for item in tup]
        for h in head_indices:
            # Slice both token dimensions using the flattened indexes.
            pattern_slice = attn_pattern[:, h, flatten_indexes][:, :, flatten_indexes]
            cache[f"pattern_L{layer}H{h}"] = pattern_slice

avg_attention_pattern_head(module, args, kwargs, output, token_indexes, layer, attn_pattern_current_avg, batch_idx, cache, avg=False, extract_avg_value=False, act_on_input=False)

Hook function to extract the average attention pattern of the heads. It will extract the attention pattern and then average it. As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)

Parameters:

Name	Type	Description	Default
-	b	the input of the hook function. It's the output of the attention pattern of the heads	required
-	s	the state of the hook function. It's the state of the model	required
-	layer	the layer of the model	required
-	head	the head of the model	required
-	attn_pattern_current_avg	the current average attention pattern	required

Source code in easyroutine/interpretability/hooks.py

def avg_attention_pattern_head(
    module,
    args,
    kwargs,
    output,
    token_indexes,
    layer,
    attn_pattern_current_avg,
    batch_idx,
    cache,
    avg: bool = False,
    extract_avg_value: bool = False,
    act_on_input=False,
):
    """
    Hook function to extract the average attention pattern of the heads. It will extract the attention pattern and then average it.
    As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)

    Args:
        - b: the input of the hook function. It's the output of the attention pattern of the heads
        - s: the state of the hook function. It's the state of the model
        - layer: the layer of the model
        - head: the head of the model
        - attn_pattern_current_avg: the current average attention pattern
    """
    # first get the attention pattern
    b = process_args_kwargs_output(args, kwargs, output)

    attn_pattern = b.data.detach().clone()  # (batch, num_heads,seq_len, seq_len)
    # attn_pattern = attn_pattern.to(torch.float32)
    num_heads = attn_pattern.size(1)

    token_indexes = [item for sublist in token_indexes for item in sublist]

    for head in range(num_heads):
        key = f"avg_pattern_L{layer}H{head}"
        if key not in attn_pattern_current_avg:
            attn_pattern_current_avg[key] = attn_pattern[:, head, token_indexes][
                :, :, token_indexes
            ]
        else:
            attn_pattern_current_avg[key] += (
                attn_pattern[:, head, token_indexes][:, :, token_indexes]
                - attn_pattern_current_avg[key]
            ) / (batch_idx + 1)
        attn_pattern_current_avg[key] = attn_pattern_current_avg[key]
        # var_key = f"M2_pattern_L{layer}H{head}"
        # if var_key not in attn_pattern_current_avg:
        #     attn_pattern_current_avg[var_key] = torch.zeros_like(attn_pattern[:, head])
        # attn_pattern_current_avg[var_key] = attn_pattern_current_avg[var_key] + (attn_pattern[:, head] - attn_pattern_current_avg[key]) * (attn_pattern[:, head] - attn_pattern_current_avg[var_key])

        if extract_avg_value:
            value_key = f"projected_value_L{layer}H{head}"
            try:
                values = cache[value_key]
            except KeyError:
                print(f"Values not found for {value_key}")
                return
            # get the attention pattern for the values
            value_norm = torch.norm(values, dim=-1)

            norm_matrix = (
                value_norm.unsqueeze(1).expand_as(attn_pattern[:, head]).transpose(1, 2)
            )

            norm_matrix = norm_matrix * attn_pattern[:, head]

            if value_key not in attn_pattern_current_avg:
                attn_pattern_current_avg[value_key] = norm_matrix[..., token_indexes, :]
            else:
                attn_pattern_current_avg[value_key] += (
                    norm_matrix[..., token_indexes, :]
                    - attn_pattern_current_avg[value_key]
                ) / (batch_idx + 1)

            # remove values from cache
            del cache[value_key]

avg_hook(module, args, kwargs, output, cache, cache_key, last_image_idx, end_image_idx)

It save the activations of the residual stream in the cache. It will save the activations in the cache (a global variable out the scope of the function)

Source code in easyroutine/interpretability/hooks.py

def avg_hook(
    module,
    args,
    kwargs,
    output,
    cache,
    cache_key,
    last_image_idx,
    end_image_idx,
):
    r"""
    It save the activations of the residual stream in the cache. It will save the activations in the cache (a global variable out the scope of the function)
    """
    b = process_args_kwargs_output(args, kwargs, output)

    img_avg = torch.mean(
        b.data.detach().clone()[:, 1 : last_image_idx + 1, :],
        dim=1,
    )
    text_avg = torch.mean(b.data.detach().clone()[:, end_image_idx:, :], dim=1)
    all_avg = torch.mean(b.data.detach().clone()[:, :, :], dim=1)

    cache[f"avg_{cache_key}"] = torch.cat(
        [img_avg.unsqueeze(1), text_avg.unsqueeze(1), all_avg.unsqueeze(1)], dim=1
    )

compute_statistics(tensor, dim=-1, keepdim=True, eps=1e-06)

Computes the mean, variance, and second moment of a given tensor along a specified dimension.

Parameters:

Name	Type	Description	Default
tensor	Tensor	Input tensor.	required
dim	int	Dimension along which to compute statistics (default: -1).	-1
keepdim	bool	Whether to keep the reduced dimension (default: True).	True
eps	float	Small constant for numerical stability.	1e-06

Returns:

Name	Type	Description
tuple		(mean, variance, second_moment)

Source code in easyroutine/interpretability/hooks.py

def compute_statistics(tensor, dim=-1, keepdim=True, eps=1e-6):
    """
    Computes the mean, variance, and second moment of a given tensor along a specified dimension.

    Args:
        tensor (torch.Tensor): Input tensor.
        dim (int): Dimension along which to compute statistics (default: -1).
        keepdim (bool): Whether to keep the reduced dimension (default: True).
        eps (float): Small constant for numerical stability.

    Returns:
        tuple: (mean, variance, second_moment)
    """
    mean = tensor.mean(dim=dim, keepdim=keepdim)  # Compute mean
    second_moment = tensor.pow(2).mean(
        dim=dim, keepdim=keepdim
    )  # Compute second moment
    variance = second_moment - mean.pow(2)  # Compute variance using E[X²] - (E[X])²

    return mean.squeeze(-1), variance.squeeze(-1), second_moment.squeeze(-1)

create_dynamic_hook(pyvene_hook, **kwargs)

DEPRECATED: pyvene is not used anymore. This function is used to create a dynamic hook. It is a wrapper around the pyvene_hook function.

Source code in easyroutine/interpretability/hooks.py

def create_dynamic_hook(pyvene_hook: Callable, **kwargs):
    r"""
    DEPRECATED: pyvene is not used anymore.
    This function is used to create a dynamic hook. It is a wrapper around the pyvene_hook function.
    """
    partial_hook = partial(pyvene_hook, **kwargs)

    def wrap(*args, **kwargs):
        return partial_hook(*args, **kwargs)

    return wrap

embed_hook(module, args, kwargs, output, token_indexes, cache, cache_key)

Hook function to extract the embeddings of the tokens. It will save the embeddings in the cache (a global variable out the scope of the function)

Source code in easyroutine/interpretability/hooks.py

def embed_hook(module, args, kwargs, output, token_indexes, cache, cache_key):
    r"""
    Hook function to extract the embeddings of the tokens. It will save the embeddings in the cache (a global variable out the scope of the function)
    """
    b = process_args_kwargs_output(args, kwargs, output)
    cache[cache_key] = []
    for token_index in token_indexes:
        cache[cache_key].append(b.data.detach().clone()[..., list(token_index)])
    cache[cache_key] = tuple(cache[cache_key])

input_embedding_hook(module, args, kwargs, output, cache, cache_key, token_indexes, keep_gradient=False, avg=False)

Hook to capture the output of the embedding layer, enable gradients, and store it in the cache.

Source code in easyroutine/interpretability/hooks.py

def input_embedding_hook(
    module, args, kwargs, output, cache, cache_key, token_indexes, keep_gradient: bool = False, avg: bool = False
):
    r"""
    Hook to capture the output of the embedding layer, enable gradients, and store it in the cache.
    """
    embeddings_tensor = process_args_kwargs_output(args, kwargs, output)

    if keep_gradient:
        # Enable gradient tracking for the embeddings tensor
        embeddings_tensor.requires_grad_(True).retain_grad()
        cache[cache_key] = embeddings_tensor # we slice in the end if keep gradient
        return restore_same_args_kwargs_output(embeddings_tensor, args, kwargs, output)  # Return the original (potentially modified in-place) output structure
    if avg:
        token_avgs = []
        for token_tuple in token_indexes:
            # Slice out the tokens specified by the tuple.
            token_slice = embeddings_tensor[:, list(token_tuple), :]
            # Average over the token dimension (dim=1) and keep that dimension.
            token_avg = torch.mean(token_slice, dim=1, keepdim=True)
            token_avgs.append(token_avg)
        cache[cache_key] = (
            torch.cat(token_avgs, dim=1)  # Store the tensor that's part of the graph
        )
    else:
        flatten_indexes = [item for tup in token_indexes for item in tup]
        cache[cache_key] = (
            embeddings_tensor[:,flatten_indexes,:]  # Store the tensor that's part of the graph
        )

    return (
        output  # Return the original (potentially modified in-place) output structure
    )

intervention_attn_mat_hook(module, args, kwargs, output, q_positions, k_positions, head, multiplication_value, patching_values=None, apply_softmax=False)

Hook function to ablate the tokens in the attention mask. It will set to 0 the value vector of the tokens to ablate

Source code in easyroutine/interpretability/hooks.py

def intervention_attn_mat_hook(
    module,
    args,
    kwargs,
    output,
    q_positions,
    k_positions,
    head,
    multiplication_value,
    patching_values: Optional[Union[str, torch.Tensor]] = None,
    apply_softmax: bool = False,
    # ablation_queries: pd.DataFrame,
):
    r"""
    Hook function to ablate the tokens in the attention
    mask. It will set to 0 the value vector of the
    tokens to ablate
    """
    # Get the shape of the attention matrix
    b = process_args_kwargs_output(args, kwargs, output)
    batch_size, num_heads, seq_len_q, seq_len_k = b.shape

    # Used during generation
    if seq_len_q < len(q_positions):
        q_positions = 0

    # Create boolean masks for queries and keys
    q_mask = torch.zeros(seq_len_q, dtype=torch.bool, device=b.device)
    q_mask[q_positions] = True  # Set positions to True

    k_mask = torch.zeros(seq_len_k, dtype=torch.bool, device=b.device)
    k_mask[k_positions] = True  # Set positions to TrueW

    # Create a 2D mask using outer product
    head_mask = torch.outer(q_mask, k_mask)  # Shape: (seq_len_q, seq_len_k)

    # Expand mask to match the dimensions of the attention matrix
    # Shape after expand: (batch_size, num_heads, seq_len_q, seq_len_k)
    # head_mask = (
    #     head_mask.unsqueeze(0).unsqueeze(0).expand(batch_size, num_heads, -1, -1)
    # )

    # select the head
    # head_mask = head_mask[:, head, :, :]

    if patching_values is None or patching_values == "ablation":
        logger.debug("No patching values provided, ablation will be performed")
        # Apply the ablation function directly to the attention matrix
        b[:, head, head_mask] = multiply_pattern(
            b[:, head, head_mask], multiplication_value
        )

    else:
        # Apply the patching values to the attention matrix
        logger.debug("Patching values provided, applying patching values")
        logger.debug(
            "Patching values shape: %s. It is expected to have shape seq_len x seq_len",
            patching_values.shape,
        )

        b[:, head, head_mask] = patching_values[head_mask]
    if apply_softmax:
        b[:, head] = torch.nn.functional.softmax(b[:, head], dim=-1)
    return b

intervention_heads_hook(module, args, kwargs, output, ablation_queries)

Hook function to ablate the heads in the attention mask. It will set to 0 the output of the heads to ablate

Source code in easyroutine/interpretability/hooks.py

def intervention_heads_hook(
    module,
    args,
    kwargs,
    output,
    ablation_queries: pd.DataFrame,
):
    r"""
    Hook function to ablate the heads in the attention
    mask. It will set to 0 the output of the heads to
    ablate
    """
    b = process_args_kwargs_output(args, kwargs, output)
    attention_matrix = b.clone().data

    for head in ablation_queries["head"]:
        attention_matrix[0, head, :, :] = 0

    b.copy_(attention_matrix)
    return b

intervention_resid_hook(module, args, kwargs, output, token_indexes, patching_values=None)

Hook function to ablate the tokens in the residual stream. It will set to 0 the value vector of the tokens to ablate

Source code in easyroutine/interpretability/hooks.py

def intervention_resid_hook(
    module,
    args,
    kwargs,
    output,
    token_indexes,
    patching_values: Optional[Union[str, torch.Tensor]] = None,
):
    r"""
    Hook function to ablate the tokens in the residual stream. It will set to 0 the value vector of the
    tokens to ablate
    """
    b = process_args_kwargs_output(args, kwargs, output)
    # detach b to avoid modifying the original tensor
    b = b.data.detach().clone()
    if patching_values is None or patching_values == "ablation":
        logger.debug(
            "No patching values provided, ablation will be performed on the residual stream"
        )
        b[..., token_indexes, :] = 0
    else:
        logger.debug(
            "Patching values provided, applying patching values to the residual stream"
        )
        b[..., list(token_indexes), :] = patching_values
    return restore_same_args_kwargs_output(b, args, kwargs, output)

multiply_pattern(tensor, multiplication_value)

Set the attention values to zero

Source code in easyroutine/interpretability/hooks.py

def multiply_pattern(tensor, multiplication_value):
    r"""
    Set the attention values to zero
    """
    # return torch.zeros_like(tensor) + multiplication_value
    return tensor * multiplication_value

projected_key_vectors_head(module, args, kwargs, output, layer, cache, token_indexes, num_attention_heads, num_key_value_heads, hidden_size, d_head, out_proj_weight, out_proj_bias, head='all', act_on_input=False, expand_head=True, avg=False)

Hook function to extract the key vectors of the heads and project them through the attention output matrix. This shows the contribution that keys in each position could have to the residual stream through the attention mechanism.

Like other hooks, it saves the activations in the cache.

Parameters:

Name	Type	Description	Default
b		the input of the hook function (output of the key vectors)	required
layer		the layer of the model	required
head	Union[str, int]	the head of the model. If "all" is passed, it will extract all the heads of the layer	'all'
expand_head	bool	bool to expand the head dimension when extracting the keys vectors	True

Source code in easyroutine/interpretability/hooks.py

def projected_key_vectors_head(
    module,
    args,
    kwargs,
    output,
    layer,
    cache,
    token_indexes,
    num_attention_heads: int,
    num_key_value_heads: int,
    hidden_size: int,
    d_head: int,
    out_proj_weight,
    out_proj_bias,
    head: Union[str, int] = "all",
    act_on_input=False,
    expand_head: bool = True,
    avg=False,
):
    r"""
    Hook function to extract the key vectors of the heads and project them through the attention output matrix.
    This shows the contribution that keys in each position could have to the residual stream through the attention mechanism.

    Like other hooks, it saves the activations in the cache.

    Args:
        b: the input of the hook function (output of the key vectors)
        layer: the layer of the model
        head: the head of the model. If "all" is passed, it will extract all the heads of the layer
        expand_head: bool to expand the head dimension when extracting the keys vectors
    """
    # Get the key vectors
    b = process_args_kwargs_output(args, kwargs, output)

    keys = b.data.detach().clone()  # (batch, num_heads, seq_len, head_dim)

    # Reshape the key vectors to have a separate dimension for the different heads
    keys = rearrange(
        keys,
        "batch seq_len (num_key_value_heads d_heads) -> batch num_key_value_heads seq_len d_heads",
        num_key_value_heads=num_key_value_heads,
        d_heads=d_head,
    )

    # If needed, repeat KV heads to match attention heads (for grouped query attention)
    keys = repeat_kv(keys, num_attention_heads // num_key_value_heads)

    keys = rearrange(
        keys,
        "batch num_head seq_len d_model -> batch seq_len num_head d_model",
    )

    # Reshape out_proj_weight to get the blocks for each head
    out_proj_weight = out_proj_weight.t().view(
        num_attention_heads,
        d_head,
        hidden_size,
    )

    # Apply bias if present
    if out_proj_bias is not None:
        out_proj_bias = out_proj_bias.view(1, 1, 1, hidden_size)

    # Apply the projection for each head
    projected_keys = einsum(
        keys,
        out_proj_weight,
        "batch seq_len num_head d_head, num_head d_head d_model -> batch seq_len num_head d_model",
    )
    if out_proj_bias is not None:
        projected_keys = projected_keys + out_proj_bias

    # Rearrange the tensor to have dimensions that we prefer
    projected_keys = rearrange(
        projected_keys,
        "batch seq_len num_head d_model -> batch num_head seq_len d_model",
    )

    # Process token indices
    if avg:
        # For each tuple, slice out the tokens and average over them
        token_avgs = []
        for token_tuple in token_indexes:
            token_slice = projected_keys[..., list(token_tuple), :]
            token_avg = torch.mean(token_slice, dim=-2, keepdim=True)
            token_avgs.append(token_avg)
        projected_keys = torch.cat(token_avgs, dim=-2)
    else:
        flatten_indexes = [item for tup in token_indexes for item in tup]
        projected_keys = projected_keys[..., flatten_indexes, :]

    # Save to cache based on selected heads
    if head == "all":
        for head_idx in range(num_attention_heads):
            cache[f"projected_key_L{layer}H{head_idx}"] = projected_keys[:, head_idx]
    else:
        cache[f"projected_key_L{layer}H{head}"] = projected_keys[:, int(head)]

projected_query_vectors_head(module, args, kwargs, output, layer, cache, token_indexes, num_attention_heads, num_key_value_heads, hidden_size, d_head, out_proj_weight, out_proj_bias, head='all', act_on_input=False, expand_head=True, avg=False)

Hook function to extract the query vectors of the heads and project them through the attention output matrix. This shows the contribution that queries in each position could have to the residual stream through the attention mechanism.

Like other hooks, it saves the activations in the cache.

Parameters:

Name	Type	Description	Default
b		the input of the hook function (output of the query vectors)	required
layer		the layer of the model	required
head	Union[str, int]	the head of the model. If "all" is passed, it will extract all the heads of the layer	'all'
expand_head	bool	bool to expand the head dimension when extracting the query vectors	True

Source code in easyroutine/interpretability/hooks.py

def projected_query_vectors_head(
    module,
    args,
    kwargs,
    output,
    layer,
    cache,
    token_indexes,
    num_attention_heads: int,
    num_key_value_heads: int,
    hidden_size: int,
    d_head: int,
    out_proj_weight,
    out_proj_bias,
    head: Union[str, int] = "all",
    act_on_input=False,
    expand_head: bool = True,
    avg=False,
):
    r"""
    Hook function to extract the query vectors of the heads and project them through the attention output matrix.
    This shows the contribution that queries in each position could have to the residual stream through the attention mechanism.

    Like other hooks, it saves the activations in the cache.

    Args:
        b: the input of the hook function (output of the query vectors)
        layer: the layer of the model
        head: the head of the model. If "all" is passed, it will extract all the heads of the layer
        expand_head: bool to expand the head dimension when extracting the query vectors
    """
    # Get the query vectors
    b = process_args_kwargs_output(args, kwargs, output)

    queries = b.data.detach().clone()  # (batch, seq_len, num_heads*d_head)

    # Reshape the query vectors to have a separate dimension for the heads
    queries = rearrange(
        queries,
        "batch seq_len (num_attention_heads d_heads) -> batch num_attention_heads seq_len d_heads",
        num_attention_heads=num_attention_heads,
        d_heads=d_head,
    )

    queries = rearrange(
        queries,
        "batch num_head seq_len d_model -> batch seq_len num_head d_model",
    )

    # Reshape out_proj_weight to get the blocks for each head
    out_proj_weight = out_proj_weight.t().view(
        num_attention_heads,
        d_head,
        hidden_size,
    )

    # Apply bias if present
    if out_proj_bias is not None:
        out_proj_bias = out_proj_bias.view(1, 1, 1, hidden_size)

    # Apply the projection for each head
    projected_queries = einsum(
        queries,
        out_proj_weight,
        "batch seq_len num_head d_head, num_head d_head d_model -> batch seq_len num_head d_model",
    )
    if out_proj_bias is not None:
        projected_queries = projected_queries + out_proj_bias

    # Rearrange the tensor to have dimensions that we prefer
    projected_queries = rearrange(
        projected_queries,
        "batch seq_len num_head d_model -> batch num_head seq_len d_model",
    )

    # Process token indices
    if avg:
        # For each tuple, slice out the tokens and average over them
        token_avgs = []
        for token_tuple in token_indexes:
            token_slice = projected_queries[..., list(token_tuple), :]
            token_avg = torch.mean(token_slice, dim=-2, keepdim=True)
            token_avgs.append(token_avg)
        projected_queries = torch.cat(token_avgs, dim=-2)
    else:
        flatten_indexes = [item for tup in token_indexes for item in tup]
        projected_queries = projected_queries[..., flatten_indexes, :]

    # Save to cache based on selected heads
    if head == "all":
        for head_idx in range(num_attention_heads):
            cache[f"projected_query_L{layer}H{head_idx}"] = projected_queries[
                :, head_idx
            ]
    else:
        cache[f"projected_query_L{layer}H{head}"] = projected_queries[:, int(head)]

projected_value_vectors_head(module, args, kwargs, output, layer, cache, token_indexes, num_attention_heads, num_key_value_heads, hidden_size, d_head, out_proj_weight, out_proj_bias, head='all', act_on_input=False, expand_head=True, avg=False)

Hook function to extract the values vectors of the heads. It will extract the values vectors and then project them with the final W_O projection As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)

Parameters:

Name	Type	Description	Default
b		the input of the hook function. It's the output of the values vectors of the heads	required
s		the state of the hook function. It's the state of the model	required
layer		the layer of the model	required
head	Union[str, int]	the head of the model. If "all" is passed, it will extract all the heads of the layer	'all'
expand_head	bool	bool to expand the head dimension when extracting the values vectors and the attention pattern. If true, in the cache we will have a key for each head, like "value_L0H0", "value_L0H1", ... while if False, we will have only one key for each layer, like "value_L0" and the dimension of the head will be taken into account in the tensor.	True

Source code in easyroutine/interpretability/hooks.py

def projected_value_vectors_head(
    module,
    args,
    kwargs,
    output,
    layer,
    cache,
    token_indexes,
    num_attention_heads: int,
    num_key_value_heads: int,
    hidden_size: int,
    d_head: int,
    out_proj_weight,
    out_proj_bias,
    head: Union[str, int] = "all",
    act_on_input=False,
    expand_head: bool = True,
    avg=False,
):
    r"""
    Hook function to extract the values vectors of the heads. It will extract the values vectors and then project them with the final W_O projection
    As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)

    Args:
        b: the input of the hook function. It's the output of the values vectors of the heads
        s: the state of the hook function. It's the state of the model
        layer: the layer of the model
        head: the head of the model. If "all" is passed, it will extract all the heads of the layer
        expand_head: bool to expand the head dimension when extracting the values vectors and the attention pattern. If true, in the cache we will have a key for each head, like "value_L0H0", "value_L0H1", ...
                        while if False, we will have only one key for each layer, like "value_L0" and the dimension of the head will be taken into account in the tensor.

    """
    # first get the values vectors
    b = process_args_kwargs_output(args, kwargs, output)

    values = b.data.detach().clone()  # (batch, num_heads,seq_len, head_dim)

    # reshape the values vectors to have a separate dimension for the different heads
    values = rearrange(
        values,
        "batch seq_len (num_key_value_heads d_heads) -> batch num_key_value_heads seq_len d_heads",
        num_key_value_heads=num_key_value_heads,
        d_heads=d_head,
    )

    #        "batch seq_len (num_key_value_heads d_heads) -> batch seq_len num_key_value_heads d_heads",

    values = repeat_kv(values, num_attention_heads // num_key_value_heads)

    values = rearrange(
        values,
        "batch num_head seq_len d_model -> batch seq_len num_head d_model",
    )

    # reshape in order to get the blocks for each head
    out_proj_weight = out_proj_weight.t().view(
        num_attention_heads,
        d_head,
        hidden_size,
    )

    # apply bias if present (No in Chameleon)
    if out_proj_bias is not None:
        out_proj_bias = out_proj_bias.view(1, 1, 1, hidden_size)

    # apply the projection for each head
    projected_values = einsum(
        values,
        out_proj_weight,
        "batch seq_len num_head d_head, num_head d_head d_model -> batch seq_len num_head d_model",
    )
    if out_proj_bias is not None:
        projected_values = projected_values + out_proj_bias

    # rearrange the tensor to have dimension that we like more
    projected_values = rearrange(
        projected_values,
        "batch seq_len num_head d_model -> batch num_head seq_len d_model",
    )

    # slice for token index
    # Assume projected_values has shape [batch, num_heads, tokens, d_model]
    if avg:
        # For each tuple, slice the tokens along dimension -2 and average over that token slice.
        token_avgs = []
        for token_tuple in token_indexes:
            # Slice out the tokens for this tuple.
            # Using ellipsis ensures we index the last two dimensions correctly.
            token_slice = projected_values[..., list(token_tuple), :]
            # Average over the token dimension (which is -2) while keeping that dimension.
            token_avg = torch.mean(token_slice, dim=-2, keepdim=True)
            token_avgs.append(token_avg)
        # Concatenate the averaged slices along the token dimension (-2).
        projected_values = torch.cat(token_avgs, dim=-2)
    else:
        # Flatten the list of token tuples into a single list of token indices.
        flatten_indexes = [item for tup in token_indexes for item in tup]
        projected_values = projected_values[..., flatten_indexes, :]

    # Post-process the value vectors by selecting heads.
    if head == "all":
        for head_idx in range(num_attention_heads):
            cache[f"projected_value_L{layer}H{head_idx}"] = projected_values[
                :, head_idx
            ]
    else:
        cache[f"projected_value_L{layer}H{head}"] = projected_values[:, int(head)]

query_key_value_hook(module, args, kwargs, output, cache, cache_key, token_indexes, layer, head_dim, num_key_value_groups, head='all', avg=False)

Same as save_resid_hook but for the query, key and value vectors, it just have a reshape to have the head dimension.

Source code in easyroutine/interpretability/hooks.py

def query_key_value_hook(
    module,
    args,
    kwargs,
    output,
    cache: ActivationCache,
    cache_key,
    token_indexes,
    layer,
    head_dim,
    num_key_value_groups: int,
    head: Union[str, int] = "all",
    avg: bool = False,
):
    r"""
    Same as save_resid_hook but for the query, key and value vectors, it just have a reshape to have the head dimension.
    """
    b = process_args_kwargs_output(args, kwargs, output)
    input_shape = b.shape[:-1]
    hidden_shape = (*input_shape, -1, head_dim)
    b = b.view(hidden_shape).transpose(1, 2)
    # cache[cache_key] = b.data.detach().clone()[..., token_index, :]

    info_string = "Shape: batch seq_len d_head"

    heads = [idx for idx in range(b.size(1))] if head == "all" else [head]
    for head_idx in heads:
        # Compute the group index for keys/values if needed.
        group_idx = head_idx // (b.size(1) // num_key_value_groups)
        # Decide whether to use group_idx or head_idx based on cache_key.
        if "values_" in cache_key or "keys_" in cache_key:
            # Select the slice corresponding to the group index.
            tensor_slice = b.data.detach().clone()[:, group_idx, ...]
        else:
            # Use the head index directly.
            tensor_slice = b.data.detach().clone()[:, head_idx, ...]

        # Process the token indexes.
        if avg:
            # For each token tuple, average over the tokens.
            # Note: After slicing, the token dimension is the first dimension of tensor_slice,
            # i.e. tensor_slice has shape (batch, tokens, d_head) so we average along dim=1.
            tokens_avgs = []
            for token_tuple in token_indexes:
                # Slice tokens using the token_tuple.
                token_subslice = tensor_slice[:, list(token_tuple), :]
                # Average over the token dimension (dim=1) and keep that dimension.
                token_avg = torch.mean(token_subslice, dim=1, keepdim=True)
                tokens_avgs.append(token_avg)
            # Concatenate the averages along the token dimension (dim=1).
            processed_tokens = torch.cat(tokens_avgs, dim=1)
        else:
            # Flatten the token indexes from the list of tuples.
            flatten_indexes = [item for tup in token_indexes for item in tup]
            processed_tokens = tensor_slice[:, flatten_indexes, :]

        # Build a unique key for the cache by including layer and head information.
        key = f"{cache_key}L{layer}H{head_idx}"
        cache.add_with_info(key, processed_tokens, info_string)

restore_same_args_kwargs_output(b, args, kwargs, output)

Inverse of process_args_kwargs_output

Source code in easyroutine/interpretability/hooks.py

def restore_same_args_kwargs_output(b, args, kwargs, output):
    """
    Inverse of process_args_kwargs_output
    """

    if output is not None:
        if isinstance(output, tuple):
            b = (b,) + output[1:]
    if output is None:
        if len(args) > 0:
            args = (b,) + args[1:]
        else:
            candidate_keys = ["hidden_states"]
            for key in candidate_keys:
                if key in kwargs:
                    kwargs[key] = b
                    break
        return args, kwargs
    return b  # type:ignore

save_resid_hook(module, args, kwargs, output, cache, cache_key, token_indexes, avg=False)

It save the activations of the residual stream in the cache. It will save the activations in the cache (a global variable out the scope of the function)

Source code in easyroutine/interpretability/hooks.py

def save_resid_hook(
    module,
    args,
    kwargs,
    output,
    cache: ActivationCache,
    cache_key,
    token_indexes,
    avg: bool = False,
):
    r"""
    It save the activations of the residual stream in the cache. It will save the activations in the cache (a global variable out the scope of the function)
    """
    b = process_args_kwargs_output(args, kwargs, output)

    # slice the tensor to get the activations of the token we want to extract
    if avg:
        token_avgs = []
        for token_index in token_indexes:
            slice_ = b.data.detach().clone()[..., list(token_index), :]
            token_avgs.append(torch.mean(slice_, dim=-2, keepdim=True))

        # cache[cache_key] = torch.cat(token_avgs, dim=-2)
        cache.add_with_info(
            cache_key,
            torch.cat(token_avgs, dim=-2),
            "Shape: batch avg_over_target_token_position, d_model",
        )

    else:
        flatten_indexes = [item for sublist in token_indexes for item in sublist]
        cache[cache_key] = b.data.detach().clone()[..., flatten_indexes, :]