File size: 9,701 Bytes
60b21d3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 | # SPDX-FileCopyrightText: 2025 Stanford University, ETH Zurich, and the project authors (see CONTRIBUTORS.md)
# SPDX-FileCopyrightText: 2025 This source file is part of the OpenTSLM open-source project.
#
# SPDX-License-Identifier: MIT
"""
Paper-style plots: memory usage scaling with N for different lengths (L).
- Rows = config (SoftPrompt, Flamingo)
- Cols = sequence lengths (L) [excluding L=1]
- Hue = base model
- Y-axis sharing logic:
* Flamingo: all panels share y-axis
* SoftPrompt: all panels have independent y-axes
- OOM cases (peak_cuda_reserved_gb > OOM_THRESHOLD) are shown by extending the
line upward and marking with a red X + "OOM".
"""
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import matplotlib
import re
from matplotlib.lines import Line2D
OOM_THRESHOLD = 180 # GB
def parse_model_name(llm_id, model_type):
"""Return base_model, config (SoftPrompt or Flamingo)."""
if llm_id.startswith("meta-llama/"):
base_name = llm_id.replace("meta-llama/", "")
elif llm_id.startswith("google/"):
base_name = llm_id.replace("google/", "")
else:
base_name = llm_id
# Normalize base model names to match expected order
if "Llama-3.2-1B" in base_name:
base_name = "Llama-3.2-1B"
elif "Llama-3.2-3B" in base_name:
base_name = "Llama-3.2-3B"
elif "gemma-3-270m" in base_name:
base_name = "Gemma-3-270M"
elif "gemma-3-1b-pt" in base_name:
base_name = "Gemma-3-1B-pt"
if model_type == "EmbedHealthSP":
type_name = "SoftPrompt"
elif model_type == "EmbedHealthFlamingo":
type_name = "Flamingo"
else:
type_name = model_type
return base_name, type_name
def parse_simulation_dataset(name):
"""Parse Simulation dataset name like 'Simulation-L10-N5' → (L=10, N=5)."""
match = re.match(r"Simulation-L(\d+)-N(\d+)", name)
if match:
return int(match.group(1)), int(match.group(2))
return None, None
def plot_memory_usage_paper(csv_file="memory_simulation.csv"):
# Publication style
plt.style.use("seaborn-v0_8-white")
matplotlib.rcParams.update({
"font.family": "serif",
"font.serif": ["Palatino", "Times New Roman", "DejaVu Serif"],
"font.size": 18,
"axes.labelsize": 20,
"axes.titlesize": 20,
"legend.fontsize": 17,
"xtick.labelsize": 17,
"ytick.labelsize": 17,
"axes.linewidth": 0.6,
"axes.edgecolor": "0.15",
})
# Load & preprocess
df = pd.read_csv(csv_file)
df["peak_cuda_reserved_gb"] = df["peak_cuda_reserved_gb"].replace(-1, pd.NA)
df = df[df["dataset"].str.startswith("Simulation-")]
df[["base_model", "config"]] = df.apply(
lambda row: pd.Series(parse_model_name(row["llm_id"], row["model"])), axis=1
)
df[["L", "N"]] = df["dataset"].apply(
lambda s: pd.Series(parse_simulation_dataset(s))
)
df = df.dropna(subset=["L", "N"])
df["L"] = df["L"].astype(int)
df["N"] = df["N"].astype(int)
df = df[df["L"] != 1]
df = df.sort_values(by=["base_model", "config", "L", "N"])
# Palette + markers - use consistent colors with the other script
# Define consistent order and colors for each model
model_order = ["Gemma-3-270M", "Gemma-3-1B-pt", "Llama-3.2-1B", "Llama-3.2-3B"]
color_map = {
"Gemma-3-270M": "#4477AA",
"Gemma-3-1B-pt": "#66CCEE",
"Llama-3.2-1B": "#228833",
"Llama-3.2-3B": "#CC6677"
}
# Get base models in the specified order
base_models = [bm for bm in model_order if bm in df["base_model"].unique()]
custom_palette = [color_map.get(bm, "#888888") for bm in base_models]
markers_dict = dict(zip(
base_models,
["o", "s", "^", "D", "p", "X", "*"]
))
# Unique sequence lengths
unique_L = sorted(df["L"].unique())
# Create subplot grid manually: 2 rows (SoftPrompt, Flamingo)
fig, axes = plt.subplots(
2, len(unique_L),
figsize=(3.2 * len(unique_L), 6),
sharex="col",
)
# Row mapping
row_map = {"SoftPrompt": 0, "Flamingo": 1}
# Precompute Flamingo y-lims
flamingo_df = df[df["config"] == "Flamingo"]
flamingo_ymin, flamingo_ymax = None, None
if not flamingo_df.empty:
flamingo_ymin = flamingo_df["peak_cuda_reserved_gb"].min(skipna=True)
flamingo_ymax = flamingo_df["peak_cuda_reserved_gb"].max(skipna=True)
flamingo_ymin = 0
flamingo_ymax = max(flamingo_ymax if flamingo_ymax else 0, OOM_THRESHOLD * 1.1)
# Iterate configs
for cfg in ["SoftPrompt", "Flamingo"]:
cfg_df = df[df["config"] == cfg]
for j, L in enumerate(unique_L):
ax = axes[row_map[cfg], j]
subdf = cfg_df[cfg_df["L"] == L]
for bm, sdf in subdf.groupby("base_model"):
sdf = sdf.sort_values("N")
# Split successful vs OOM runs
ok_df = sdf[sdf["peak_cuda_reserved_gb"] <= OOM_THRESHOLD]
oom_df = sdf[sdf["peak_cuda_reserved_gb"] > OOM_THRESHOLD]
# Normal line
if not ok_df.empty:
ax.plot(
ok_df["N"], ok_df["peak_cuda_reserved_gb"],
label=bm,
color=custom_palette[base_models.index(bm)],
marker=markers_dict[bm],
linewidth=2.2,
markersize=5,
alpha=0.9,
)
# OOM handling
if not oom_df.empty:
first_oom = oom_df.iloc[0]
last_ok_y = ok_df["peak_cuda_reserved_gb"].iloc[-1] if not ok_df.empty else OOM_THRESHOLD * 0.9
# extend line upward
ax.plot(
[ok_df["N"].iloc[-1] if not ok_df.empty else first_oom["N"], first_oom["N"]],
[last_ok_y, OOM_THRESHOLD * 1.05],
color=custom_palette[base_models.index(bm)],
linestyle="--",
linewidth=1.5,
alpha=0.8,
)
# red X marker at OOM
ax.scatter(
first_oom["N"], OOM_THRESHOLD * 1.05,
color="red",
marker="x",
s=70,
linewidth=2,
zorder=5,
)
ax.text(
first_oom["N"], OOM_THRESHOLD * 1.05,
"OOM", color="red", fontsize=9,
fontweight="bold", ha="center", va="bottom"
)
# Titles
if row_map[cfg] == 0:
ax.set_title(f"L = {L}", fontsize=19, fontweight="bold")
# Y labels only leftmost col
if j == 0:
ax.set_ylabel(
f"{cfg}",
fontsize=18,
fontweight="bold"
)
else:
ax.set_ylabel("")
# Styling
ax.set_facecolor("#F8F9FA")
ax.grid(True, which="major", linestyle="-", linewidth=0.4, alpha=0.5)
ax.grid(True, which="minor", linestyle=":", linewidth=0.3, alpha=0.3)
ax.minorticks_on()
# Remove individual x-axis labels - will add global one later
ax.set_xlabel("")
ax.set_xticks([1, 2, 3, 4, 5])
ax.tick_params(axis="x", which="both", labelbottom=True)
# Y-axis rules
if cfg == "Flamingo":
ax.set_ylim(0, 65)
elif j <= 2:
ax.set_ylim(0, 25)
# Legend (global, right side) - create in specified order
# Get handles and labels from first subplot
handles_dict = {}
labels_dict = {}
for handle, label in zip(*axes[0, 0].get_legend_handles_labels()):
handles_dict[label] = handle
labels_dict[label] = label
# Create ordered handles and labels
ordered_handles = []
ordered_labels = []
for bm in base_models:
if bm in handles_dict:
ordered_handles.append(handles_dict[bm])
ordered_labels.append(labels_dict[bm])
# Add OOM handle
oom_handle = Line2D([0], [0], color="red", marker="x", linestyle="--",
markersize=8, label="Out of Memory (OOM)")
ordered_handles.append(oom_handle)
ordered_labels.append("Out of Memory (OOM)")
# Legend in top left plot only
top_left_ax = axes[0, 0]
top_left_ax.legend(
ordered_handles, ordered_labels,
title=None,
loc="upper right",
frameon=True,
framealpha=0.95,
edgecolor="0.3",
fontsize=12,
)
# Add main vertical title
fig.text(0.02, 0.5, "Peak Memory (GB)", rotation=90, fontsize=20, fontweight="bold", ha="center", va="center")
# Add global x-axis label spanning the bottom row
fig.text(0.5, 0.01, "Number of Time Series (N)", fontsize=16, fontweight="bold", ha="center", va="center")
# Layout - no longer need space for bottom legend
plt.tight_layout(pad=0.5)
plt.subplots_adjust(left=0.08)
# Save
for fmt in ["png", "pdf"]:
plt.savefig(
f"memory_usage_paper.{fmt}",
dpi=300 if fmt == "png" else None,
bbox_inches="tight",
pad_inches=0,
facecolor="white",
format=fmt,
)
plt.show()
if __name__ == "__main__":
plot_memory_usage_paper()
|