MCPcopy Create free account

hub / github.com/RiddleHe/llm-interp / functions

Functions204 in github.com/RiddleHe/llm-interp

⨍ Functions 204◇ Types & classes 0
↓ 12 callersFunction_print_list
(lst)
attention_sink/analyze_sink.py:4254
↓ 12 callersFunction_topk_energy_dims
( Z, k=3, min_median_mass=0.55, )
attention_sink/analyze_sink.py:365
↓ 12 callersFunctionprefill
(model, input_ids, position_ids, return_hidden_states: bool = False)
attention_sink/analyze_sink.py:1220
↓ 10 callersFunction_plot_progression_multi
(series, outdir, key, ylabel, title, fname, mode=None)
attention_sink/analyze_sink.py:3925
↓ 9 callersFunction_cloud_radius
(Z)
attention_sink/analyze_sink.py:549
↓ 9 callersFunction_topk_vals_mu_sd_str
(Z, k=3, mode="abs")
attention_sink/analyze_sink.py:577
↓ 7 callersFunction_mean_vec_norm
(X)
attention_sink/analyze_sink.py:535
↓ 6 callersFunction_fracmass_on_set
(Z, idx_set)
attention_sink/analyze_sink.py:340
↓ 6 callersFunction_get
(key, comp)
attention_sink/analyze_sink.py:4284
↓ 6 callersFunction_k_for_energy_frac_str
(X, fracs=(0.25, 0.75, 0.99))
attention_sink/analyze_sink.py:443
↓ 6 callersFunction_mean_activation_entropy
(Z)
attention_sink/analyze_sink.py:568
↓ 6 callersFunction_vec_norm_mu_sd_str
(X)
attention_sink/analyze_sink.py:540
↓ 5 callersFunction_collect_mlp_internals_for_layer
(model, layer_idx, tok_idxs, tokenized)
attention_sink/analyze_sink.py:905
↓ 5 callersFunction_filter_ranked_act_outliers_by_top1
(A, ranked_dims, min_rel_to_top1)
attention_sink/analyze_sink.py:1440
↓ 5 callersFunction_fmt_mu_sd
(mu, sd)
attention_sink/analyze_sink.py:476
↓ 5 callersFunction_print_metrics_table
(metrics, col_names, title=None, col_w=20)
attention_sink/analyze_sink.py:4261
↓ 5 callersFunction_rank_act_outlier_dims
(A, mult_global=50.0, mult_dim=5.0, positive_only=False)
attention_sink/analyze_sink.py:1426
↓ 5 callersFunction_reduce_headwise
(xs)
attention_sink/analyze_sink.py:491
↓ 5 callersFunction_scan_sink_attention_only
(model, tokenized, scan_layers)
attention_sink/analyze_sink.py:1292
↓ 4 callersFunction_capture_qkv_multi
(model, layer_idxs, cache, capture_attn_out=False, capture_mlp_out=False)
attention_sink/analyze_sink.py:1233
↓ 4 callersFunctionapply_rotary_pos_emb
(q, k, cos, sin, position_ids=None, unsqueeze_dim=1)
thinking_budget/rope.py:14
↓ 3 callersFunction_bias_3d_projection
(X, bias_vec, topk=2)
attention_sink/analyze_sink.py:294
↓ 3 callersFunction_collect_mlp_grid_mean_tensors
( model, tokenized, layer_idx, max_tok=9, )
attention_sink/analyze_sink.py:990
↓ 3 callersFunction_draw_bar
(ax, vals, ylabel, panel_title, stds=None)
attention_sink/analyze_sink.py:2243
↓ 3 callersFunction_draw_stack
(ax, stack_key, xlabel)
attention_sink/analyze_sink.py:3805
↓ 3 callersFunction_get_hidden_from_hook_args
(args, kwargs)
attention_sink/analyze_sink.py:86
↓ 3 callersFunction_k_at_cdf_threshold
(mass_sorted, thr)
attention_sink/analyze_sink.py:659
↓ 3 callersFunction_mean_dir_decomposition
(XQ, XK, VhQ, label, k)
attention_sink/analyze_sink.py:333
↓ 3 callersFunction_mean_or_zero
(xs)
attention_sink/analyze_sink.py:547
↓ 3 callersFunction_mean_std
(acc, sq, odim_int)
attention_sink/analyze_sink.py:2591
↓ 3 callersFunction_mean_topk_column_norm
(Wd, idx_set)
attention_sink/analyze_sink.py:597
↓ 3 callersFunction_orth_basis
(Q)
attention_sink/analyze_sink.py:4813
↓ 3 callersFunction_plot_panel
(ax, data_key, std_key, panel_title, ylabel, baseline_vals, baseline_label)
attention_sink/analyze_sink.py:2654
↓ 3 callersFunction_pos_mu_sd_str
(Z)
attention_sink/analyze_sink.py:531
↓ 3 callersFunction_rms
(x, dim=-1, keepdim=False, eps=EPS)
attention_sink/analyze_sink.py:520
↓ 3 callersFunction_run_tok0_subspace_ablation
(Q_basis=None)
attention_sink/analyze_sink.py:4875
↓ 3 callersFunction_topk_activated_set
(Z, k)
attention_sink/analyze_sink.py:398
↓ 3 callersFunction_yerr_from_var
(stats_by_layer, L, key)
attention_sink/analyze_sink.py:4020
↓ 3 callersFunctiondiff_stats
(a, b)
attention_nondeterministic/analyze_attention_nondeterministic.py:121
↓ 2 callersFunction_collect_gate_with_ablation
Collect gate_out, optionally with self-attend-only ablation. If attn_ablate_idxs is None or empty, collects under normal causal attention. R
attention_sink/analyze_sink.py:2468
↓ 2 callersFunction_collect_residual_decomp_to_layer
(model, layer_idx, tok_idxs, tokenized)
attention_sink/analyze_sink.py:697
↓ 2 callersFunction_collect_residual_stream_components
Collect mean of each residual-stream component up to a target point. target: str like "mlp.1" or "attn.2". "mlp.L" -> components are embed
attention_sink/analyze_sink.py:3035
↓ 2 callersFunction_compute_dots
(uo, di, uo_sq, di_sq, w_row)
attention_sink/analyze_sink.py:2330
↓ 2 callersFunction_coords_uncentered
(X)
attention_sink/analyze_sink.py:5095
↓ 2 callersFunction_dims_title_block
(ranked_dims, label="top dims", max_show=12, per_line=4)
attention_sink/analyze_sink.py:1531
↓ 2 callersFunction_dot_contrib_energy_per_dim
(X, w_row, eps=EPS)
attention_sink/analyze_sink.py:645
↓ 2 callersFunction_draw
(ax, xlim=None, show_labels=True)
attention_sink/analyze_sink.py:3638
↓ 2 callersFunction_draw_contrib_bar
( ax, bar_vals_0, top_dims_0, bar_vals_8, top_dims_8, rank_l
attention_sink/analyze_sink.py:2074
↓ 2 callersFunction_draw_panel
(ax, ung, gat, ung_std, gat_std, panel_title)
attention_sink/analyze_sink.py:2361
↓ 2 callersFunction_headwise_cloud_radius_mu_var
(vecs_h)
attention_sink/analyze_sink.py:506
↓ 2 callersFunction_install_self_attend_only_hooks
Install hooks that force specified attention layers to self-attend only. Modifies the attention mask so each token can only attend to itself.
attention_sink/analyze_sink.py:2426
↓ 2 callersFunction_pick_all_heads_with_caches
(model, attns, layer_idx)
attention_sink/analyze_sink.py:1356
↓ 2 callersFunction_pick_mid_in_bin
(bin_idx, used)
attention_sink/analyze_sink.py:1629
↓ 2 callersFunction_pick_proj_background_input_dims
(M, focus_dims, total_bins=8, total_target=8)
attention_sink/analyze_sink.py:1611
↓ 2 callersFunction_plot_projection_histogram_stacked
( p0_list, pb_list, labels, out_path, title, bins=60 )
attention_sink/analyze_sink.py:3397
↓ 2 callersFunction_plot_z_topk_energy_stacks
( Z_list, tok_labels, topk, out_path, title, )
attention_sink/analyze_sink.py:3596
↓ 2 callersFunction_rank_proj_dims_from_focus_vectors
( M, focus_input_dims, focus_scalars, min_rel_to_top1, )
attention_sink/analyze_sink.py:1459
↓ 2 callersFunction_register_block_raw_captures
(model, L, cache)
attention_sink/analyze_sink.py:238
↓ 2 callersFunction_run_tok0_preln_ablation
(do_ablate)
attention_sink/analyze_sink.py:5022
↓ 2 callersFunction_set_panel_title
(ax, name, subtitle, fontsize=10)
attention_sink/analyze_sink.py:1553
↓ 2 callersFunction_shares_for_token
(t)
attention_sink/analyze_sink.py:5166
↓ 2 callersFunction_squash_rank_axis
(ranks, D, thr, tail_frac=0.10)
attention_sink/analyze_sink.py:677
↓ 2 callersFunction_tok0_special_dims_from_act
(A, mult_global=50.0, mult_dim=5.0, positive_only=False)
attention_sink/analyze_sink.py:1364
↓ 2 callersFunction_top_contrib_payload
(target_dim, contrib_top_k)
attention_sink/analyze_sink.py:2050
↓ 2 callersFunction_topk_list_from_scores
(scores, k)
attention_sink/analyze_sink.py:353
↓ 2 callersFunctionadd_common_determinism_flags
()
attention_nondeterministic/analyze_attention_nondeterministic.py:22
↓ 2 callersFunctionconpute_scores_from_scratch
(last_n_query_vectors, keys, total_steps=None, model=None, normalize=False, rerope_current_key=0)
thinking_budget/analyze_seed_thinking_budget.py:98
↓ 2 callersFunctionget_logit_rank
(logits, token_id)
thinking_budget/analyze_seed_thinking_budget.py:58
↓ 2 callersFunctionload_model
(model_id, attn_implementation, dtype)
attention_nondeterministic/analyze_attention_nondeterministic.py:94
↓ 2 callersFunctionmake_tokens
(tokenizer, target_len)
attention_nondeterministic/analyze_attention_nondeterministic.py:111
↓ 2 callersFunctionpick_head
(attentions, layer_idx, head_idx)
attention_sink/analyze_sink.py:1347
↓ 2 callersFunctionrerope_single_key_inplace
(keys, model, key_index, p_old=None, p_new=None)
thinking_budget/analyze_seed_thinking_budget.py:142
↓ 2 callersFunctionrotate_half
Rotates half the hidden dims of the input.
thinking_budget/rope.py:7
↓ 1 callersFunction_append_suffix
(path, suffix)
attention_sink/analyze_sink.py:1211
↓ 1 callersFunction_bin_of_dim
(dim_idx)
attention_sink/analyze_sink.py:1623
↓ 1 callersFunction_collect_component_mean_activations
( model, tokenized, layer_idx, kind, )
attention_sink/analyze_sink.py:853
↓ 1 callersFunction_collect_mlp_in_gradients_for_single_zdim
( model, layer_idx, tok_idx, tokenized, z_dim, )
attention_sink/analyze_sink.py:1114
↓ 1 callersFunction_collect_raw_components_for_layer
(model, layer_idx, tok_idxs, tokenized)
attention_sink/analyze_sink.py:780
↓ 1 callersFunction_component_direction_stats
(R, A, M)
attention_sink/analyze_sink.py:560
↓ 1 callersFunction_compute_Wd_subspace
(model, layer_idx, vec_idx)
attention_sink/analyze_sink.py:410
↓ 1 callersFunction_disable_packed_sequence_splitting
()
attention_sink/analyze_sink.py:16
↓ 1 callersFunction_eff_rank_from_eigs
(eigs)
attention_sink/analyze_sink.py:4920
↓ 1 callersFunction_enable_fa2
()
attention_nondeterministic/analyze_attention_nondeterministic.py:40
↓ 1 callersFunction_fmt
(v, kind)
attention_sink/analyze_sink.py:4275
↓ 1 callersFunction_fmt_mu_sd_float
(mu, sd, mu_digits=2, sd_digits=2)
attention_sink/analyze_sink.py:471
↓ 1 callersFunction_focus_input_dims_from_act
(A, max_keep)
attention_sink/analyze_sink.py:1578
↓ 1 callersFunction_frac_energy_on_direction
(X, u)
attention_sink/analyze_sink.py:323
↓ 1 callersFunction_gram_diagnostics
(vs, names)
attention_sink/analyze_sink.py:4928
↓ 1 callersFunction_install_local_input_subspace_ablation_hook
( model, layer_idx, Q, tok_ablate=0, )
attention_sink/analyze_sink.py:201
↓ 1 callersFunction_install_lower_attn_hooks
(model, layers, factor=0.5, sink_k_idx=0, only_return_sink_value=False)
attention_sink/analyze_sink.py:100
↓ 1 callersFunction_install_mlp_out_ablation_hook
( model, layer_idx, mode, vec_idx=None, tok_ablate=0, tok_ref=8, clamp_alpha_max=1
attention_sink/analyze_sink.py:151
↓ 1 callersFunction_k_for_energy_frac
(X, fracs=(0.25, 0.75, 0.99))
attention_sink/analyze_sink.py:417
↓ 1 callersFunction_load_prompts_from_file
(path)
attention_sink/analyze_sink.py:66
↓ 1 callersFunction_majority_sign_over_threshold
(X, dim_idx, thr=0.75)
attention_sink/analyze_sink.py:666
↓ 1 callersFunction_make_tok0_ablation_preln_prehook
( Ue_basis, tok_ablate=0, )
attention_sink/analyze_sink.py:4839
↓ 1 callersFunction_make_tok0_subspace_ablation_mlp_prehook
(Q_basis, tok_ablate=0)
attention_sink/analyze_sink.py:4822
↓ 1 callersFunction_mk_attn_hook
(ii)
attention_sink/analyze_sink.py:726
↓ 1 callersFunction_mk_fwd_hook
(L_)
attention_sink/analyze_sink.py:1272
↓ 1 callersFunction_mk_hook
(L, attn_layer)
attention_sink/analyze_sink.py:1242
↓ 1 callersFunction_mk_mlp_fwd_hook
(L_)
attention_sink/analyze_sink.py:1282
next →1–100 of 204, ranked by callers