Upload folder using huggingface_hub
Browse files
deepseek_v4/deepseek_v4_mask_layer0_sliding_attention.svg
ADDED
|
|
deepseek_v4/deepseek_v4_mask_layer1_compressed_sparse_attention.svg
ADDED
|
|
deepseek_v4/deepseek_v4_mask_layer2_heavily_compressed_attention.svg
ADDED
|
|
deepseek_v4/visualize_attention_masks.py
ADDED
|
@@ -0,0 +1,563 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Visualise the per-layer-type attention mask DeepSeek-V4 actually feeds to
|
| 2 |
+
`eager_attention_forward`, for the three attention layer types (sliding, CSA, HCA).
|
| 3 |
+
Generates the SVGs embedded in `docs/source/en/model_doc/deepseek_v4.md`.
|
| 4 |
+
|
| 5 |
+
We build a tiny V4 model (small enough that 16 source tokens Γ all compressed
|
| 6 |
+
slots fits on screen), forward a dummy batch through it, wrap each attention
|
| 7 |
+
layer's forward with a thin shim that replays the mask-extension logic
|
| 8 |
+
(`cat([sliding_mask, block_bias])`) and captures the exact post-cat mask, then
|
| 9 |
+
render each layer's mask in either:
|
| 10 |
+
|
| 11 |
+
* the `β ` / `β¬` ANSI-coloured style of `transformers.utils.attention_visualizer`
|
| 12 |
+
(default; suppressed when stdout isn't a TTY or `NO_COLOR=1` is set), or
|
| 13 |
+
* an SVG grid suitable for embedding in markdown (with `--svg <DIR>`).
|
| 14 |
+
|
| 15 |
+
For CSA layers the visualizer remaps the literal `[S, SΒ·k]` flat-slot `block_bias`
|
| 16 |
+
back to the more readable `[S, T_entries]` entry-visibility view (each compressor
|
| 17 |
+
column is then a compressed *entry* rather than a gather slot), and shades cells
|
| 18 |
+
red when an entry is causally available but the indexer's top-k did not pick it.
|
| 19 |
+
|
| 20 |
+
Run from the repository root::
|
| 21 |
+
|
| 22 |
+
python docs/source/en/imgs/deepseek_v4/visualize_attention_masks.py
|
| 23 |
+
python docs/source/en/imgs/deepseek_v4/visualize_attention_masks.py \\
|
| 24 |
+
--svg docs/source/en/imgs/deepseek_v4
|
| 25 |
+
"""
|
| 26 |
+
|
| 27 |
+
from __future__ import annotations
|
| 28 |
+
|
| 29 |
+
import argparse
|
| 30 |
+
import os
|
| 31 |
+
from pathlib import Path
|
| 32 |
+
|
| 33 |
+
import torch
|
| 34 |
+
import torch.nn.functional as F
|
| 35 |
+
|
| 36 |
+
from transformers import DeepseekV4Config, DeepseekV4ForCausalLM
|
| 37 |
+
from transformers.models.deepseek_v4.modeling_deepseek_v4 import (
|
| 38 |
+
DeepseekV4CSACompressor,
|
| 39 |
+
DeepseekV4HCACompressor,
|
| 40 |
+
DeepseekV4Indexer,
|
| 41 |
+
DeepseekV4PreTrainedModel,
|
| 42 |
+
)
|
| 43 |
+
from transformers.utils.output_capturing import OutputRecorder
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
# Tiny config so the visualised matrices fit in a terminal. Compress rates are dialled
|
| 47 |
+
# down (m_csa=4, m_hca=8 β vs the real 4/128) so HCA actually emits an entry at S=16.
|
| 48 |
+
CFG = DeepseekV4Config(
|
| 49 |
+
vocab_size=64,
|
| 50 |
+
hidden_size=64,
|
| 51 |
+
moe_intermediate_size=64,
|
| 52 |
+
num_hidden_layers=3,
|
| 53 |
+
num_attention_heads=4,
|
| 54 |
+
num_key_value_heads=1,
|
| 55 |
+
head_dim=32,
|
| 56 |
+
partial_rotary_factor=8 / 32,
|
| 57 |
+
q_lora_rank=32,
|
| 58 |
+
o_groups=2,
|
| 59 |
+
o_lora_rank=16,
|
| 60 |
+
num_experts_per_tok=2,
|
| 61 |
+
n_routed_experts=4,
|
| 62 |
+
n_shared_experts=1,
|
| 63 |
+
mlp_layer_types=["moe", "moe", "moe"],
|
| 64 |
+
layer_types=["sliding_attention", "compressed_sparse_attention", "heavily_compressed_attention"],
|
| 65 |
+
compress_rates={"compressed_sparse_attention": 4, "heavily_compressed_attention": 8},
|
| 66 |
+
sliding_window=8,
|
| 67 |
+
hc_mult=2,
|
| 68 |
+
hc_sinkhorn_iters=3,
|
| 69 |
+
hc_eps=1.0e-6,
|
| 70 |
+
index_n_heads=2,
|
| 71 |
+
index_head_dim=16,
|
| 72 |
+
index_topk=2,
|
| 73 |
+
num_nextn_predict_layers=0,
|
| 74 |
+
swiglu_limit=10.0,
|
| 75 |
+
rope_theta=10000.0,
|
| 76 |
+
compress_rope_theta=160000.0,
|
| 77 |
+
max_position_embeddings=64,
|
| 78 |
+
)
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
# ANSI colours, matching `transformers.utils.attention_visualizer`.
|
| 82 |
+
GREEN = "\033[92m"
|
| 83 |
+
YELLOW = "\033[93m"
|
| 84 |
+
DIM = "\033[90m"
|
| 85 |
+
RESET = "\033[0m"
|
| 86 |
+
BLACK_SQUARE = "β "
|
| 87 |
+
WHITE_SQUARE = "β¬"
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
def _meta_subtitle(meta: dict | None) -> str:
|
| 91 |
+
"""One-liner subtitle describing the layer's compressor config."""
|
| 92 |
+
if meta is None:
|
| 93 |
+
return ""
|
| 94 |
+
parts: list[str] = []
|
| 95 |
+
if meta.get("compress_rate") is not None:
|
| 96 |
+
parts.append(f"m={meta['compress_rate']}")
|
| 97 |
+
if meta.get("index_topk") is not None:
|
| 98 |
+
parts.append(f"k={meta['index_topk']}")
|
| 99 |
+
if meta.get("T_entries"):
|
| 100 |
+
parts.append(f"{meta['T_entries']} entries")
|
| 101 |
+
return " Β· ".join(parts)
|
| 102 |
+
|
| 103 |
+
|
| 104 |
+
def _render(
|
| 105 |
+
mask_2d: torch.Tensor,
|
| 106 |
+
query_labels: list[str],
|
| 107 |
+
kv_labels: list[str],
|
| 108 |
+
title: str,
|
| 109 |
+
*,
|
| 110 |
+
color: bool = True,
|
| 111 |
+
meta: dict | None = None,
|
| 112 |
+
) -> str:
|
| 113 |
+
"""Render a 2D additive mask (0 = visible, -inf = masked) as a β / β¬ grid.
|
| 114 |
+
|
| 115 |
+
`mask_2d` is `[S_q, S_kv]`. Columns past `len(query_labels)` are the
|
| 116 |
+
compressor / indexer entries the caller cat'd into the mask via
|
| 117 |
+
`cat([sliding_causal_mask, block_bias], dim=-1)`. When `color=False`,
|
| 118 |
+
diagonal / compressor slots are marked with the distinguishing glyphs
|
| 119 |
+
`β` and `β²` instead of ANSI colours so the output renders in markdown.
|
| 120 |
+
"""
|
| 121 |
+
visible = (mask_2d > -1e30).bool()
|
| 122 |
+
n_q, n_kv = visible.shape
|
| 123 |
+
sliding_kv = len(query_labels) # by construction kv labels begin with the sliding tokens
|
| 124 |
+
|
| 125 |
+
if color:
|
| 126 |
+
g, y, d, r = GREEN, YELLOW, DIM, RESET
|
| 127 |
+
visible_glyph = f"{g}{BLACK_SQUARE}{r}"
|
| 128 |
+
compr_glyph = f"{y}{BLACK_SQUARE}{r}"
|
| 129 |
+
else:
|
| 130 |
+
g = y = d = r = ""
|
| 131 |
+
visible_glyph = "β£"
|
| 132 |
+
compr_glyph = "β²"
|
| 133 |
+
|
| 134 |
+
max_q_label = max(len(q) for q in query_labels)
|
| 135 |
+
out: list[str] = []
|
| 136 |
+
out.append(title)
|
| 137 |
+
out.append("=" * len(title))
|
| 138 |
+
out.append(
|
| 139 |
+
f" shape=[{n_q}, {n_kv}] "
|
| 140 |
+
f"{visible_glyph} visible (sliding KV) {compr_glyph} visible (compressor entry) {WHITE_SQUARE} masked"
|
| 141 |
+
)
|
| 142 |
+
subtitle = _meta_subtitle(meta)
|
| 143 |
+
if subtitle:
|
| 144 |
+
out.append(f" {subtitle}")
|
| 145 |
+
if meta and meta.get("topk_picks") is not None:
|
| 146 |
+
# Append per-query indexer picks so warm-up sentinels and entry choices are auditable.
|
| 147 |
+
picks_lines = [f" indexer topk picks (entry id, -1 = warm-up sentinel):"]
|
| 148 |
+
for i, picks in enumerate(meta["topk_picks"][0]): # B=1
|
| 149 |
+
picks_lines.append(f" q{i:>2}: {picks}")
|
| 150 |
+
out.append("\n".join(picks_lines))
|
| 151 |
+
|
| 152 |
+
# Column headers (kv indices, stacked vertically across `width` lines).
|
| 153 |
+
width = max(2, len(str(n_kv - 1)))
|
| 154 |
+
for line in range(width):
|
| 155 |
+
header_chars = []
|
| 156 |
+
for j in range(n_kv):
|
| 157 |
+
label = str(j).rjust(width)
|
| 158 |
+
ch = label[line] if line < len(label) else " "
|
| 159 |
+
# Tint compressor-section header so it visually separates from the sliding section.
|
| 160 |
+
header_chars.append(f"{d}{ch}{r}" if j >= sliding_kv else ch)
|
| 161 |
+
prefix = " " * (max_q_label + 6)
|
| 162 |
+
out.append(prefix + " ".join(header_chars))
|
| 163 |
+
|
| 164 |
+
# Body rows.
|
| 165 |
+
for i in range(n_q):
|
| 166 |
+
row_cells = []
|
| 167 |
+
for j in range(n_kv):
|
| 168 |
+
if not visible[i, j]:
|
| 169 |
+
row_cells.append(WHITE_SQUARE)
|
| 170 |
+
elif j >= sliding_kv:
|
| 171 |
+
row_cells.append(compr_glyph)
|
| 172 |
+
else:
|
| 173 |
+
row_cells.append(visible_glyph)
|
| 174 |
+
out.append(f"{query_labels[i].rjust(max_q_label)} : {str(i).rjust(2)} " + " ".join(row_cells))
|
| 175 |
+
|
| 176 |
+
out.append(f" {compr_glyph} columns past index {sliding_kv - 1} are compressor / indexer slots (block_bias).")
|
| 177 |
+
return "\n".join(out)
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
# SVG palette + geometry. Bigger cells so we can fit token-string labels comfortably,
|
| 181 |
+
# and compressor columns are drawn `m` cells wide so the visual encodes the fact that
|
| 182 |
+
# one compressed entry summarizes m source tokens (rather than a 1:1 KV slot like the
|
| 183 |
+
# sliding section).
|
| 184 |
+
SVG_CELL = 24
|
| 185 |
+
SVG_GAP = 2
|
| 186 |
+
SVG_PAD_LEFT = 120 # roomy enough for the longest row-label token ("morning") at 11pt
|
| 187 |
+
SVG_PAD_TOP = 180 # title/shape/meta (60px) + C_w + pos + rotated tokens (~60px) + a touch.
|
| 188 |
+
SVG_PAD_RIGHT = 48
|
| 189 |
+
SVG_PAD_BOTTOM = 64
|
| 190 |
+
SVG_COLORS = {
|
| 191 |
+
# Dark-mode palette: black background, light text.
|
| 192 |
+
"background": "#0b1220", # near-black with a hint of blue
|
| 193 |
+
"visible": "#22c55e", # green-500, pops against black
|
| 194 |
+
"masked": "#1e293b", # slate-800 β dim cells that are visible against bg but recede
|
| 195 |
+
"indexer_masked": "#ef4444",# red-500, also pops
|
| 196 |
+
"compressor": "#22c55e", # same green as visible (single-color attended-to)
|
| 197 |
+
"separator": "#475569", # slate-600
|
| 198 |
+
"text": "#f8fafc", # slate-50 β main text on black
|
| 199 |
+
"muted": "#94a3b8", # slate-400 β secondary text
|
| 200 |
+
}
|
| 201 |
+
|
| 202 |
+
# Static word list used as token labels on the SVG axes. Real token semantics don't
|
| 203 |
+
# affect the mask (which depends only on positions), so these labels are decorative.
|
| 204 |
+
SVG_TOKENS = [
|
| 205 |
+
"The", "quick", "brown", "fox", "jumps", "over", "the", "lazy",
|
| 206 |
+
"dog", "ate", "a", "small", "red", "fish", "this", "morning",
|
| 207 |
+
]
|
| 208 |
+
|
| 209 |
+
|
| 210 |
+
def _render_svg(mask_2d: torch.Tensor, title: str, sliding_kv: int, meta: dict | None = None) -> str:
|
| 211 |
+
"""Render the mask as a standalone SVG document.
|
| 212 |
+
|
| 213 |
+
Rows = queries (top β bottom). The sliding K=V section is one cell per
|
| 214 |
+
source position (key index). The compressor section is `T_entries`
|
| 215 |
+
blocks each `mΒ·cell` wide β the wide rectangle is the visual hint that
|
| 216 |
+
one compressed entry summarizes `m` source tokens (rather than 1:1).
|
| 217 |
+
|
| 218 |
+
Cells are coloured: green = visible (the query attends to this slot),
|
| 219 |
+
light gray = masked. Compressor entries are tinted amber when visible
|
| 220 |
+
to a given query and pale amber when masked, so the compressor section
|
| 221 |
+
reads as a distinct "compression band" even at a glance.
|
| 222 |
+
"""
|
| 223 |
+
visible = (mask_2d > -1e30).bool()
|
| 224 |
+
n_q, n_kv = visible.shape
|
| 225 |
+
n_compr = n_kv - sliding_kv
|
| 226 |
+
m = (meta or {}).get("compress_rate") or 1
|
| 227 |
+
stride = SVG_CELL + SVG_GAP
|
| 228 |
+
compr_block_w = m * SVG_CELL + (m - 1) * SVG_GAP
|
| 229 |
+
|
| 230 |
+
# Per-column x positions and widths. Sliding cells: 1Γcell. Compressor entries:
|
| 231 |
+
# `m`Γcell wide each, with a small gap between sections.
|
| 232 |
+
section_gap = SVG_GAP * 4
|
| 233 |
+
col_x: list[float] = []
|
| 234 |
+
col_w: list[float] = []
|
| 235 |
+
cursor = float(SVG_PAD_LEFT)
|
| 236 |
+
for j in range(sliding_kv):
|
| 237 |
+
col_x.append(cursor)
|
| 238 |
+
col_w.append(SVG_CELL)
|
| 239 |
+
cursor += stride
|
| 240 |
+
if n_compr > 0:
|
| 241 |
+
cursor += section_gap
|
| 242 |
+
for _w in range(n_compr):
|
| 243 |
+
col_x.append(cursor)
|
| 244 |
+
col_w.append(compr_block_w)
|
| 245 |
+
cursor += compr_block_w + stride
|
| 246 |
+
grid_w = cursor - SVG_PAD_LEFT - stride
|
| 247 |
+
grid_h = n_q * stride - SVG_GAP
|
| 248 |
+
w = SVG_PAD_LEFT + grid_w + SVG_PAD_RIGHT
|
| 249 |
+
h = SVG_PAD_TOP + grid_h + SVG_PAD_BOTTOM
|
| 250 |
+
|
| 251 |
+
subtitle = _meta_subtitle(meta)
|
| 252 |
+
# Explicit `width` / `height` (matching viewBox) so renderers don't fall back to the
|
| 253 |
+
# 150Γ150 SVG default when the file is embedded without sizing CSS. `preserveAspectRatio`
|
| 254 |
+
# keeps it sharp when scaled.
|
| 255 |
+
elems: list[str] = [
|
| 256 |
+
f'<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 {w} {h}" '
|
| 257 |
+
f'width="{w}" height="{h}" preserveAspectRatio="xMinYMin meet" '
|
| 258 |
+
f'font-family="ui-monospace, SFMono-Regular, Menlo, monospace" font-size="11">',
|
| 259 |
+
f' <rect x="0" y="0" width="{w}" height="{h}" fill="{SVG_COLORS["background"]}" />',
|
| 260 |
+
f' <text x="{SVG_PAD_LEFT}" y="20" font-size="15" font-weight="600" fill="{SVG_COLORS["text"]}">{title}</text>',
|
| 261 |
+
f' <text x="{SVG_PAD_LEFT}" y="40" fill="{SVG_COLORS["muted"]}">'
|
| 262 |
+
f'shape=[{n_q}, {n_kv}] keys β queries β</text>',
|
| 263 |
+
]
|
| 264 |
+
if subtitle:
|
| 265 |
+
elems.append(
|
| 266 |
+
f' <text x="{SVG_PAD_LEFT}" y="60" fill="{SVG_COLORS["compressor"]}">{subtitle}</text>'
|
| 267 |
+
)
|
| 268 |
+
|
| 269 |
+
# Column headers, top-down within the `SVG_PAD_TOP` gutter (above the grid line):
|
| 270 |
+
# y = 78 : `C_w` (centered over the compressor block, bold green)
|
| 271 |
+
# y = 92 : `pos aβb` source-position range for the compressor block
|
| 272 |
+
# y = SVG_PAD_TOP - 70 = 270 : numeric column index (sliding) β just above the
|
| 273 |
+
# tallest rotated token so it reads as a tick label.
|
| 274 |
+
# y = SVG_PAD_TOP - 60 .. SVG_PAD_TOP : rotated (-90Β°) per-position token labels
|
| 275 |
+
# (sliding + per-source-position inside blocks)
|
| 276 |
+
# y = SVG_PAD_TOP : grid begins.
|
| 277 |
+
idx_y = SVG_PAD_TOP - 70
|
| 278 |
+
token_baseline_y = SVG_PAD_TOP - 8
|
| 279 |
+
# Vertical extent of the rotated-token area we want the dashed separators to span β
|
| 280 |
+
# just the token region, not the full title/subtitle gutter.
|
| 281 |
+
token_top_y = SVG_PAD_TOP - 60
|
| 282 |
+
for j in range(sliding_kv):
|
| 283 |
+
cx = col_x[j] + col_w[j] / 2
|
| 284 |
+
elems.append(
|
| 285 |
+
f' <text x="{cx:.1f}" y="{idx_y}" text-anchor="middle" '
|
| 286 |
+
f'fill="{SVG_COLORS["muted"]}" font-size="10">{j}</text>'
|
| 287 |
+
)
|
| 288 |
+
tok = SVG_TOKENS[j] if j < len(SVG_TOKENS) else f"t{j}"
|
| 289 |
+
elems.append(
|
| 290 |
+
f' <text x="{cx:.1f}" y="{token_baseline_y}" text-anchor="start" '
|
| 291 |
+
f'transform="rotate(-90 {cx:.1f} {token_baseline_y})" '
|
| 292 |
+
f'fill="{SVG_COLORS["text"]}" font-size="12" font-style="italic">{tok}</text>'
|
| 293 |
+
)
|
| 294 |
+
for w_idx in range(n_compr):
|
| 295 |
+
j = sliding_kv + w_idx
|
| 296 |
+
cx = col_x[j] + col_w[j] / 2
|
| 297 |
+
block_x = col_x[j]
|
| 298 |
+
a, b = w_idx * m, (w_idx + 1) * m - 1
|
| 299 |
+
# `C_w` + `pos aβb` sit on the same row as the sliding-section indices, hugging
|
| 300 |
+
# the top of the rotated-token area β same vertical band as the column ticks.
|
| 301 |
+
elems.append(
|
| 302 |
+
f' <text x="{cx:.1f}" y="{idx_y - 14}" text-anchor="middle" '
|
| 303 |
+
f'fill="{SVG_COLORS["visible"]}" font-size="13" font-weight="700">C{w_idx}</text>'
|
| 304 |
+
)
|
| 305 |
+
elems.append(
|
| 306 |
+
f' <text x="{cx:.1f}" y="{idx_y}" text-anchor="middle" '
|
| 307 |
+
f'fill="{SVG_COLORS["muted"]}" font-size="10">pos {a}β{b}</text>'
|
| 308 |
+
)
|
| 309 |
+
# One rotated label per source position, dashed vertical separators between
|
| 310 |
+
# sub-cells inside the block (only as tall as the rotated-token area, not the
|
| 311 |
+
# whole gutter β they're a hint that the wide green block bundles m positions).
|
| 312 |
+
for sub_i in range(m):
|
| 313 |
+
pos = a + sub_i
|
| 314 |
+
sub_cx = block_x + sub_i * stride + SVG_CELL / 2
|
| 315 |
+
tok = SVG_TOKENS[pos] if pos < len(SVG_TOKENS) else f"t{pos}"
|
| 316 |
+
elems.append(
|
| 317 |
+
f' <text x="{sub_cx:.1f}" y="{token_baseline_y}" text-anchor="start" '
|
| 318 |
+
f'transform="rotate(-90 {sub_cx:.1f} {token_baseline_y})" '
|
| 319 |
+
f'fill="{SVG_COLORS["text"]}" font-size="12" font-style="italic">{tok}</text>'
|
| 320 |
+
)
|
| 321 |
+
if sub_i > 0:
|
| 322 |
+
sep_x = block_x + sub_i * stride - SVG_GAP / 2
|
| 323 |
+
elems.append(
|
| 324 |
+
f' <line x1="{sep_x:.1f}" y1="{token_top_y}" '
|
| 325 |
+
f'x2="{sep_x:.1f}" y2="{SVG_PAD_TOP}" '
|
| 326 |
+
f'stroke="{SVG_COLORS["separator"]}" stroke-width="0.7" stroke-dasharray="3 3" />'
|
| 327 |
+
)
|
| 328 |
+
|
| 329 |
+
# Row labels: query index (small, muted) + token (italic, dark) on the left.
|
| 330 |
+
# `SVG_PAD_LEFT` is sized to fit the longest token in the static word list.
|
| 331 |
+
for i in range(n_q):
|
| 332 |
+
cy = SVG_PAD_TOP + i * stride + SVG_CELL / 2 + 4
|
| 333 |
+
elems.append(
|
| 334 |
+
f' <text x="6" y="{cy:.1f}" fill="{SVG_COLORS["muted"]}" font-size="10">q{i:>2}</text>'
|
| 335 |
+
)
|
| 336 |
+
tok = SVG_TOKENS[i] if i < len(SVG_TOKENS) else f"t{i}"
|
| 337 |
+
elems.append(
|
| 338 |
+
f' <text x="{SVG_PAD_LEFT - 8}" y="{cy:.1f}" text-anchor="end" '
|
| 339 |
+
f'fill="{SVG_COLORS["text"]}" font-size="12" font-style="italic">{tok}</text>'
|
| 340 |
+
)
|
| 341 |
+
|
| 342 |
+
# Dashed vertical separator between sliding K=V and compressor sections.
|
| 343 |
+
if 0 < sliding_kv < n_kv:
|
| 344 |
+
sep_x = col_x[sliding_kv - 1] + col_w[sliding_kv - 1] + section_gap / 2
|
| 345 |
+
elems.append(
|
| 346 |
+
f' <line x1="{sep_x:.1f}" y1="{SVG_PAD_TOP - 28}" x2="{sep_x:.1f}" '
|
| 347 |
+
f'y2="{SVG_PAD_TOP + grid_h:.1f}" stroke="{SVG_COLORS["separator"]}" '
|
| 348 |
+
f'stroke-width="1" stroke-dasharray="4 3" />'
|
| 349 |
+
)
|
| 350 |
+
|
| 351 |
+
# Indexer-rejected mask (CSA only): causally available compressor entry that the
|
| 352 |
+
# indexer's top-k chose not to pick. Rendered red to distinguish from the plain
|
| 353 |
+
# "not yet causally ready" masking (light gray).
|
| 354 |
+
rejected = (meta or {}).get("indexer_rejected")
|
| 355 |
+
if rejected is not None:
|
| 356 |
+
if rejected.ndim == 4:
|
| 357 |
+
rejected = rejected[0, 0]
|
| 358 |
+
elif rejected.ndim == 3:
|
| 359 |
+
rejected = rejected[0]
|
| 360 |
+
|
| 361 |
+
# Cells.
|
| 362 |
+
for i in range(n_q):
|
| 363 |
+
for j in range(n_kv):
|
| 364 |
+
x = col_x[j]
|
| 365 |
+
y = SVG_PAD_TOP + i * stride
|
| 366 |
+
in_compressor = j >= sliding_kv
|
| 367 |
+
is_rejected = rejected is not None and bool(rejected[i, j])
|
| 368 |
+
if visible[i, j]:
|
| 369 |
+
fill = SVG_COLORS["compressor"] if in_compressor else SVG_COLORS["visible"]
|
| 370 |
+
elif is_rejected:
|
| 371 |
+
fill = SVG_COLORS["indexer_masked"]
|
| 372 |
+
else:
|
| 373 |
+
fill = SVG_COLORS["masked"]
|
| 374 |
+
elems.append(
|
| 375 |
+
f' <rect x="{x:.1f}" y="{y}" width="{col_w[j]:.1f}" height="{SVG_CELL}" '
|
| 376 |
+
f'rx="3" ry="3" fill="{fill}" />'
|
| 377 |
+
)
|
| 378 |
+
|
| 379 |
+
# Legend at the bottom. The red legend item is only meaningful for CSA (the only
|
| 380 |
+
# layer type that has an indexer + top-k pick step), so we include it only when
|
| 381 |
+
# the captured mask actually has rejected cells.
|
| 382 |
+
legend_y = SVG_PAD_TOP + grid_h + 30
|
| 383 |
+
legend_items = [
|
| 384 |
+
("attended-to", SVG_COLORS["visible"]),
|
| 385 |
+
("masked (not causally ready)", SVG_COLORS["masked"]),
|
| 386 |
+
]
|
| 387 |
+
if rejected is not None and bool(rejected.any()):
|
| 388 |
+
legend_items.append(("indexer-masked (top-k did not pick)", SVG_COLORS["indexer_masked"]))
|
| 389 |
+
x_cursor = SVG_PAD_LEFT
|
| 390 |
+
for label, color in legend_items:
|
| 391 |
+
elems.append(
|
| 392 |
+
f' <rect x="{x_cursor}" y="{legend_y - 12}" width="14" height="14" rx="3" ry="3" fill="{color}" />'
|
| 393 |
+
)
|
| 394 |
+
elems.append(
|
| 395 |
+
f' <text x="{x_cursor + 20}" y="{legend_y}" fill="{SVG_COLORS["text"]}">{label}</text>'
|
| 396 |
+
)
|
| 397 |
+
x_cursor += 20 + len(label) * 6.5 + 24
|
| 398 |
+
|
| 399 |
+
elems.append("</svg>")
|
| 400 |
+
return "\n".join(elems)
|
| 401 |
+
|
| 402 |
+
|
| 403 |
+
def _sliding_causal_mask(seq_len: int, sliding_window: int) -> torch.Tensor:
|
| 404 |
+
"""Standard `[1, 1, S, S]` sliding-window-causal additive mask (0 visible, -inf masked).
|
| 405 |
+
Identical to what `create_sliding_window_causal_mask` builds for the sliding section
|
| 406 |
+
of every V4 attention layer β we synthesise it here so the visualiser stays decoupled
|
| 407 |
+
from the model's internal mask-building."""
|
| 408 |
+
q = torch.arange(seq_len)
|
| 409 |
+
k = torch.arange(seq_len)
|
| 410 |
+
diff = q.view(-1, 1) - k.view(1, -1)
|
| 411 |
+
visible = (diff >= 0) & (diff < sliding_window)
|
| 412 |
+
return torch.where(visible, torch.zeros((), dtype=torch.float32), torch.full((), float("-inf"))).view(1, 1, seq_len, seq_len)
|
| 413 |
+
|
| 414 |
+
|
| 415 |
+
def main() -> None:
|
| 416 |
+
torch.manual_seed(0)
|
| 417 |
+
|
| 418 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 419 |
+
# Extend `DeepseekV4PreTrainedModel._can_record_outputs` with the compressor /
|
| 420 |
+
# indexer outputs we need to reconstruct the visualised mask. The base class
|
| 421 |
+
# already registers `router_logits` / `hidden_states` / `attentions`; we add three
|
| 422 |
+
# more keys, each routed by `OutputRecorder(target_class=β¦, index=β¦)`:
|
| 423 |
+
#
|
| 424 |
+
# * `csa_block_bias` β `DeepseekV4CSACompressor.forward()` returns `(gathered, block_bias)`,
|
| 425 |
+
# we grab index 1 (the per-query bias `[B, 1, S, S*k]`).
|
| 426 |
+
# * `hca_block_bias` β `DeepseekV4HCACompressor.forward()` likewise (already entry-resolved).
|
| 427 |
+
# * `indexer_topk` β `DeepseekV4Indexer.forward()` returns a single tensor `[B, S, k]`.
|
| 428 |
+
#
|
| 429 |
+
# Setting this *before* instantiating the model is what `PreTrainedModel.__init__`
|
| 430 |
+
# picks up to populate `_CAN_RECORD_REGISTRY`. The `@capture_outputs` decorator on
|
| 431 |
+
# `DeepseekV4Model.forward` then installs the corresponding `register_forward_hook`
|
| 432 |
+
# calls lazily on first request and stashes the captures on the returned
|
| 433 |
+
# `ModelOutput`. We just flip `config.output_<key>=True` to switch them on.
|
| 434 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 435 |
+
DeepseekV4PreTrainedModel._can_record_outputs = {
|
| 436 |
+
**(DeepseekV4PreTrainedModel._can_record_outputs or {}),
|
| 437 |
+
"csa_block_bias": OutputRecorder(DeepseekV4CSACompressor, index=1),
|
| 438 |
+
"hca_block_bias": OutputRecorder(DeepseekV4HCACompressor, index=1),
|
| 439 |
+
"indexer_topk": OutputRecorder(DeepseekV4Indexer, index=0),
|
| 440 |
+
}
|
| 441 |
+
|
| 442 |
+
cfg = CFG
|
| 443 |
+
for key in ("csa_block_bias", "hca_block_bias", "indexer_topk"):
|
| 444 |
+
setattr(cfg, f"output_{key}", True)
|
| 445 |
+
full_model = DeepseekV4ForCausalLM(cfg).eval()
|
| 446 |
+
# `@capture_outputs` decorates `DeepseekV4Model.forward` (the inner model). The outer
|
| 447 |
+
# `DeepseekV4ForCausalLM.forward` builds a `CausalLMOutput` that only forwards logits
|
| 448 |
+
# and last_hidden_state, so our extra capture keys disappear if we go via the outer.
|
| 449 |
+
# Call the inner directly β we don't need logits to visualise masks.
|
| 450 |
+
inner_model = full_model.model
|
| 451 |
+
|
| 452 |
+
seq_len = 16
|
| 453 |
+
input_ids = torch.arange(seq_len).unsqueeze(0) % cfg.vocab_size # [1, S]
|
| 454 |
+
with torch.no_grad():
|
| 455 |
+
out = inner_model(input_ids=input_ids)
|
| 456 |
+
|
| 457 |
+
# Tuples (one entry per layer of the matching module class). With our 3-layer
|
| 458 |
+
# config: 1 CSA layer + 1 HCA layer + 1 sliding-only layer.
|
| 459 |
+
csa_block_biases = getattr(out, "csa_block_bias", ()) or ()
|
| 460 |
+
hca_block_biases = getattr(out, "hca_block_bias", ()) or ()
|
| 461 |
+
indexer_topks = getattr(out, "indexer_topk", ()) or ()
|
| 462 |
+
|
| 463 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 464 |
+
# Build the display mask + metadata per layer from the captured outputs.
|
| 465 |
+
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 466 |
+
sliding_mask = _sliding_causal_mask(seq_len, cfg.sliding_window) # [1, 1, S, S]
|
| 467 |
+
captured: dict[int, tuple[str, torch.Tensor, dict]] = {}
|
| 468 |
+
csa_iter = iter(csa_block_biases)
|
| 469 |
+
hca_iter = iter(hca_block_biases)
|
| 470 |
+
indexer_iter = iter(indexer_topks)
|
| 471 |
+
for layer_idx, layer in enumerate(inner_model.layers):
|
| 472 |
+
attn = layer.self_attn
|
| 473 |
+
layer_type = attn.layer_type
|
| 474 |
+
meta = {
|
| 475 |
+
"compress_rate": (attn.compressor.compress_rate if attn.compressor is not None else None),
|
| 476 |
+
"index_topk": (
|
| 477 |
+
attn.compressor.indexer.index_topk
|
| 478 |
+
if attn.compressor is not None and getattr(attn.compressor, "indexer", None) is not None
|
| 479 |
+
else None
|
| 480 |
+
),
|
| 481 |
+
"T_entries": seq_len // attn.compressor.compress_rate if attn.compressor is not None else 0,
|
| 482 |
+
"topk_picks": None,
|
| 483 |
+
"indexer_rejected": None,
|
| 484 |
+
}
|
| 485 |
+
if attn.compressor is None:
|
| 486 |
+
captured[layer_idx] = (layer_type, sliding_mask, meta)
|
| 487 |
+
continue
|
| 488 |
+
|
| 489 |
+
if layer_type == "heavily_compressed_attention":
|
| 490 |
+
block_bias = next(hca_iter).detach().cpu()
|
| 491 |
+
display_mask = torch.cat([sliding_mask, block_bias.to(sliding_mask.dtype)], dim=-1)
|
| 492 |
+
elif layer_type == "compressed_sparse_attention":
|
| 493 |
+
_ = next(csa_iter) # consume the (literal) `[B, 1, S, S*k]` flat-slot bias
|
| 494 |
+
topk = next(indexer_iter).detach().cpu() # [B, S, k]
|
| 495 |
+
meta["topk_picks"] = topk.tolist()
|
| 496 |
+
# Remap the flat-slot view to the more readable entry-visibility view via
|
| 497 |
+
# the indexer's per-query top-k picks (one-hot + OR-reduce sidesteps the
|
| 498 |
+
# duplicate-index clobber that `scatter_` would hit on warm-up sentinels).
|
| 499 |
+
T_entries = meta["T_entries"]
|
| 500 |
+
valid = topk >= 0
|
| 501 |
+
safe = topk.clamp(min=0)
|
| 502 |
+
oh = F.one_hot(safe, num_classes=T_entries).bool() & valid.unsqueeze(-1)
|
| 503 |
+
em = oh.any(dim=-2) # [B, S, T_entries]
|
| 504 |
+
entry_bias = torch.where(em, torch.zeros((), dtype=torch.float32), torch.full((), float("-inf")))
|
| 505 |
+
display_mask = torch.cat([sliding_mask, entry_bias.unsqueeze(1)], dim=-1)
|
| 506 |
+
# Red-highlight cells: causally available entry that the indexer rejected.
|
| 507 |
+
causal_threshold = torch.arange(seq_len).add(1).floor_divide(attn.compressor.compress_rate)
|
| 508 |
+
entry_idx = torch.arange(T_entries)
|
| 509 |
+
causally_available = entry_idx.view(1, -1) < causal_threshold.view(-1, 1)
|
| 510 |
+
rejected = causally_available & ~em[0]
|
| 511 |
+
meta["indexer_rejected"] = torch.cat(
|
| 512 |
+
[torch.zeros((seq_len, seq_len), dtype=torch.bool), rejected], dim=-1
|
| 513 |
+
)
|
| 514 |
+
else:
|
| 515 |
+
display_mask = sliding_mask
|
| 516 |
+
|
| 517 |
+
captured[layer_idx] = (layer_type, display_mask, meta)
|
| 518 |
+
|
| 519 |
+
color = os.environ.get("NO_COLOR") is None and os.isatty(1)
|
| 520 |
+
|
| 521 |
+
svg_dir = main.svg_dir if hasattr(main, "svg_dir") else None
|
| 522 |
+
if svg_dir is not None:
|
| 523 |
+
svg_dir.mkdir(parents=True, exist_ok=True)
|
| 524 |
+
|
| 525 |
+
print()
|
| 526 |
+
for layer_idx in sorted(captured):
|
| 527 |
+
layer_type, mask, meta = captured[layer_idx]
|
| 528 |
+
if mask is None:
|
| 529 |
+
continue
|
| 530 |
+
if mask.ndim == 4:
|
| 531 |
+
mask_2d = mask[0, 0]
|
| 532 |
+
elif mask.ndim == 3:
|
| 533 |
+
mask_2d = mask[0]
|
| 534 |
+
else:
|
| 535 |
+
mask_2d = mask
|
| 536 |
+
|
| 537 |
+
q_labels = [f"q{i}" for i in range(mask_2d.shape[0])]
|
| 538 |
+
kv_labels_sliding = [f"k{i}" for i in range(seq_len)]
|
| 539 |
+
n_compressor = mask_2d.shape[1] - seq_len
|
| 540 |
+
kv_labels_compressor = [f"c{i}" for i in range(n_compressor)]
|
| 541 |
+
kv_labels = kv_labels_sliding + kv_labels_compressor
|
| 542 |
+
|
| 543 |
+
title = f"Layer {layer_idx}: {layer_type}"
|
| 544 |
+
print(_render(mask_2d, q_labels, kv_labels, title, color=color, meta=meta))
|
| 545 |
+
print()
|
| 546 |
+
if svg_dir is not None:
|
| 547 |
+
svg_path = svg_dir / f"deepseek_v4_mask_layer{layer_idx}_{layer_type}.svg"
|
| 548 |
+
svg_path.write_text(_render_svg(mask_2d, title, sliding_kv=seq_len, meta=meta))
|
| 549 |
+
print(f" wrote {svg_path}")
|
| 550 |
+
print()
|
| 551 |
+
|
| 552 |
+
|
| 553 |
+
if __name__ == "__main__":
|
| 554 |
+
parser = argparse.ArgumentParser(description=__doc__)
|
| 555 |
+
parser.add_argument(
|
| 556 |
+
"--svg",
|
| 557 |
+
type=Path,
|
| 558 |
+
default=None,
|
| 559 |
+
help="Directory to write one SVG per layer (in addition to the ANSI / plain-text terminal output).",
|
| 560 |
+
)
|
| 561 |
+
args = parser.parse_args()
|
| 562 |
+
main.svg_dir = args.svg # type: ignore[attr-defined]
|
| 563 |
+
main()
|