File size: 8,519 Bytes
6ec9472 | 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 | import json, os, sys
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
sys.path.insert(0, os.path.dirname(__file__))
RES = os.path.join(os.path.dirname(__file__), "..", "results")
FIG = os.path.join(os.path.dirname(__file__), "..", "paper", "figs")
plt.rcParams.update({
"font.size": 8, "axes.labelsize": 8, "axes.titlesize": 8.5,
"legend.fontsize": 7, "xtick.labelsize": 7, "ytick.labelsize": 7,
"figure.dpi": 200, "savefig.dpi": 200, "axes.grid": True,
"grid.alpha": 0.25, "grid.linewidth": 0.5, "lines.linewidth": 1.3,
"axes.spines.top": False, "axes.spines.right": False,
"font.family": "serif", "mathtext.fontset": "cm",
})
C = ["#1b3a6b", "#c1440e", "#2e7d32", "#6a1b9a", "#c98a00", "#00695c"]
def load(n):
p = os.path.join(RES, n)
return json.load(open(p)) if os.path.exists(p) else None
def fig_amplification():
pr = load("projection.json")
if not pr:
return
a = pr["amplification"]
bits = [x["bits"] for x in a]
frac = [x["frac"] * 100 for x in a]
size = [x["model_gb"] for x in a]
fig, ax = plt.subplots(1, 2, figsize=(6.9, 2.35))
ax[0].plot(bits, size, "o-", color=C[0])
ax[0].axhline(294, ls="--", c=C[1], lw=1)
ax[0].text(9, 320, "free NVMe (294 GB)", color=C[1], fontsize=6.5)
ax[0].axhline(24, ls=":", c=C[2], lw=1)
ax[0].text(9, 27, "usable DRAM (24 GB)", color=C[2], fontsize=6.5)
ax[0].set_yscale("log"); ax[0].set_xlabel("weight rate (bits/parameter)")
ax[0].set_ylabel("model footprint (GB)")
ax[0].set_title("(a) 1.05T-parameter footprint")
ax[1].plot(bits, frac, "o-", color=C[0])
ax[1].set_xlabel("weight rate (bits/parameter)")
ax[1].set_ylabel("expert slots resident in 24 GB (%)")
ax[1].set_title("(b) DRAM cache capacity")
for b, f in zip(bits, frac):
if b in (16, 1.5):
ax[1].annotate(f"{f:.1f}%", (b, f), textcoords="offset points",
xytext=(4, 4), fontsize=6.5)
fig.tight_layout(); fig.savefig(os.path.join(FIG, "amplification.pdf"))
plt.close(fig)
def fig_io():
io = load("io_bench.json")
if not io:
return
fig, ax = plt.subplots(figsize=(3.4, 2.35))
for i, t in enumerate([1, 2, 4, 8]):
pts = sorted([(r["block_kb"], r["mb_s"] / 1000) for r in io["random"]
if r["threads"] == t])
ax.plot([p[0] for p in pts], [p[1] for p in pts], "o-", color=C[i],
label=f"{t} thread" + ("s" if t > 1 else ""), ms=3)
ax.set_xscale("log", base=2)
ax.set_xlabel("read block size (KiB)")
ax.set_ylabel("random-read bandwidth (GB/s)")
ax.axhline(io["host"]["seq_read_mb_s"] / 1000, ls="--", c="k", lw=0.9)
ax.text(80, io["host"]["seq_read_mb_s"] / 1000 + 0.15, "sequential",
fontsize=6.5)
ax.legend(loc="lower right")
fig.tight_layout(); fig.savefig(os.path.join(FIG, "io.pdf")); plt.close(fig)
def fig_cache():
cp = load("cache_policy.json")
cv = load("cache_validation.json")
if not cp or not cv:
return
fr = load("routing_freq.json")
fig, ax = plt.subplots(1, 3, figsize=(6.9, 2.25))
F = np.array([fr[str(l)] for l in range(cp["layers"])])
for l in range(0, cp["layers"], 3):
ax[0].plot(np.arange(1, F.shape[1] + 1), np.sort(F[l])[::-1],
color=C[0], alpha=0.35, lw=0.8)
s = cp["zipf_s"]
r = np.arange(1, F.shape[1] + 1)
z = r ** (-s); z = z / z.sum()
ax[0].plot(r, z, "--", color=C[1], lw=1.4, label=f"Zipf $s$={s:.2f}")
ax[0].set_xscale("log"); ax[0].set_yscale("log")
ax[0].set_xlabel("expert rank"); ax[0].set_ylabel("activation probability")
ax[0].set_title("(a) expert popularity"); ax[0].legend()
h = cp["policies"]
x = [r["frac"] * 100 for r in h]
ax[1].plot(x, [r["lru"] * 100 for r in h], "o-", color=C[0],
label="LRU", ms=3)
ax[1].plot(x, [r["static"] * 100 for r in h], "^-", color=C[2],
label="popularity-pinned", ms=3)
ax[1].plot(x, [r["hybrid"] * 100 for r in h], "d-", color=C[4],
label="hybrid (75% pinned)", ms=3)
ax[1].plot(x, [r["analytic_static"] * 100 for r in h], "s:", color=C[1],
label="analytic model", ms=3)
ws = cp["ws_frac"] * 100
ax[1].axvline(ws, ls="--", c=C[3], lw=1)
ax[1].text(ws + 2, 72, "per-token\nworking set", color=C[3], fontsize=6)
ax[1].set_xlabel("cache capacity (% of expert slots)")
ax[1].set_ylabel("hit rate (%)")
ax[1].set_title("(b) replacement policy"); ax[1].legend(loc="lower right")
d = cv["distinct_per_batch"]
ax[2].plot([r["batch"] for r in d], [r["measured"] for r in d], "o-",
color=C[0], label="measured", ms=3)
ax[2].plot([r["batch"] for r in d], [r["irm_measured_pop"] for r in d],
"s--", color=C[1], label="IRM model", ms=3)
ax[2].set_xscale("log", base=2)
ax[2].set_xlabel("tokens per batch")
ax[2].set_ylabel("distinct experts / layer")
ax[2].set_title("(c) batch amortisation"); ax[2].legend(loc="lower right")
fig.tight_layout(); fig.savefig(os.path.join(FIG, "cache.pdf")); plt.close(fig)
def fig_quality():
import glob
runs = []
for p in glob.glob(os.path.join(RES, "quant_*.json")):
runs.append(json.load(open(p)))
if not runs:
return
fig, ax = plt.subplots(figsize=(5.0, 3.05))
base = [r for r in runs if r["config"].get("tag", "").startswith("rtn")]
ours = sorted([r for r in runs if r["config"]["tag"].startswith("main")],
key=lambda r: r["avg_bits"])
abl = sorted([r for r in runs if r["config"]["tag"].startswith("northt")],
key=lambda r: r["avg_bits"])
noldl = sorted([r for r in runs if r["config"]["tag"].startswith("noldlq")],
key=lambda r: r["avg_bits"])
freq = sorted([r for r in runs if r["config"]["tag"].startswith("freq")],
key=lambda r: r["avg_bits"])
fp = load("fp16_ppl.json")
for grp, lab, st, c in [(ours, "RVQ + RHT + LDLQ (ours)", "o-", C[0]),
(freq, "+ frequency-conditioned alloc.", "D-", C[4]),
(noldl, "no LDLQ (data-free)", "s--", C[1]),
(abl, "no incoherence processing", "^--", C[2]),
(base, "RTN uniform", "v:", C[3])]:
if grp:
ax.plot([r["avg_bits"] for r in grp], [r["ppl"] for r in grp], st,
label=lab, color=c, ms=3)
if fp:
ax.axhline(fp["ppl"], ls="--", c="k", lw=0.9)
ax.text(3.32, fp["ppl"] * 1.12, f"bf16 = {fp['ppl']:.2f}", fontsize=6.5,
ha="right", va="bottom")
ax.set_yscale("log")
ax.set_xlim(0.85, 3.45)
ax.set_xlabel("average weight rate (bits/parameter)")
ax.set_ylabel("WikiText-2 perplexity")
ax.legend(loc="upper center", bbox_to_anchor=(0.5, -0.24), ncol=2,
frameon=False, fontsize=7)
fig.tight_layout()
fig.savefig(os.path.join(FIG, "quality.pdf"), bbox_inches="tight")
plt.close(fig)
def fig_throughput():
pr = load("projection.json")
if not pr:
return
fig, ax = plt.subplots(1, 2, figsize=(6.9, 2.35))
s = pr["sensitivity_1p5bit"]
ax[0].plot([x["hit_rate"] * 100 for x in s], [x["tok_s"] for x in s], "-",
color=C[0])
hr = pr["projection"][0]["hit_rate"] * 100
ax[0].set_xlabel("expert-cache hit rate (%)")
ax[0].set_ylabel("decode throughput (tokens/s)")
ax[0].set_yscale("log")
ax[0].set_title("(a) sensitivity at 1.5 bit, batch 1")
rows = pr["projection"]
bits = sorted(set(r["rate_bits"] for r in rows))
for i, B in enumerate([1, 8, 32]):
y = [next(r["tok_s"] for r in rows if r["rate_bits"] == b and r["batch"] == B)
for b in bits]
ax[1].plot(bits, y, "o-", color=C[i], label=f"batch {B}", ms=3)
ax[1].set_xlabel("weight rate (bits/parameter)")
ax[1].set_ylabel("decode throughput (tokens/s)")
ax[1].set_yscale("log"); ax[1].legend()
ax[1].set_title("(b) projected 1.05T throughput")
fig.tight_layout(); fig.savefig(os.path.join(FIG, "throughput.pdf"))
plt.close(fig)
if __name__ == "__main__":
os.makedirs(FIG, exist_ok=True)
for f in [fig_amplification, fig_io, fig_cache, fig_quality, fig_throughput]:
try:
f()
print("ok", f.__name__)
except Exception as e:
print("skip", f.__name__, type(e).__name__, e)
|