File size: 4,310 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 | """Generate the paper's LaTeX tables directly from the measurement artefacts."""
import glob, json, os
RES = os.path.join(os.path.dirname(__file__), "..", "results")
PAP = os.path.join(os.path.dirname(__file__), "..", "paper")
def load(n, d=None):
p = os.path.join(RES, n)
return json.load(open(p)) if os.path.exists(p) else d
def w(name, s):
open(os.path.join(PAP, name), "w").write(s)
print("wrote", name)
def tab_quality():
runs = {}
for p in glob.glob(os.path.join(RES, "quant_*.json")):
r = json.load(open(p))
runs[r["config"]["tag"]] = r
fp = load("fp16_ppl.json")
order = [
("bf16 (reference)", None, None),
("RTN uniform, group 128", "rtn3", "scalar baseline"),
("RTN uniform, group 128", "rtn2", "scalar baseline"),
("RVQ, data-free", "noldlq15", "no LDLQ"),
("RVQ + LDLQ, no rotation", "northt15", "no incoherence proc."),
("RVQ + RHT + LDLQ (ours)", "main10", ""),
("RVQ + RHT + LDLQ (ours)", "main15", ""),
("RVQ + RHT + LDLQ (ours)", "main20", ""),
("\\quad + frequency-cond. alloc.", "freq15", "rate-matched"),
]
L = ["\\begin{tabular}{llrr}", "\\toprule",
"Method & Note & Bits/weight & PPL $\\downarrow$ \\\\", "\\midrule"]
for name, tag, note in order:
if tag is None:
if fp:
L.append(f"{name} & --- & 16.00 & {fp['ppl']:.2f} \\\\")
L.append("\\midrule")
continue
if tag in runs:
r = runs[tag]
L.append(f"{name} & {note} & {r['avg_bits']:.2f} & {r['ppl']:.2f} \\\\")
L += ["\\bottomrule", "\\end{tabular}"]
w("tab_quality.tex", "\n".join(L))
def tab_amp():
pr = load("projection.json")
if not pr:
return
L = ["\\begin{tabular}{rrrrrl}", "\\toprule",
"Rate & Footprint & Expert & Resident & Cache & Fits \\\\",
"(bits) & (GB) & (MB) & slots & fraction & 294\\,GB? \\\\", "\\midrule"]
for a in pr["amplification"]:
fits = "yes" if a["model_gb"] < 294 else "\\textbf{no}"
L.append(f"{a['bits']:.1f} & {a['model_gb']:,.0f} & {a['expert_mb']:.2f} & "
f"{a['dram_experts']:,} & {a['frac']*100:.2f}\\% & {fits} \\\\")
L += ["\\bottomrule", "\\end{tabular}"]
w("tab_amp.tex", "\n".join(L))
def tab_proj():
pr = load("projection.json")
if not pr:
return
ws = pr.get("token_working_set", 512)
rows = [r for r in pr["projection"] if r["batch"] in (1, 32)]
L = ["\\begin{tabular}{rrrrrrr}", "\\toprule",
"Rate & Cache & Recency & Hit & Fetch & \\multicolumn{2}{c}{Tokens/s} \\\\",
"\\cmidrule(lr){6-7}",
"(bits) & slots & viable? & rate & MB/token & batch 1 & batch 32 \\\\",
"\\midrule"]
seen = set()
for r in sorted(rows, key=lambda x: -x["rate_bits"]):
if r["rate_bits"] in seen:
continue
seen.add(r["rate_bits"])
b32 = next(x for x in pr["projection"]
if x["rate_bits"] == r["rate_bits"] and x["batch"] == 32)
ok = "yes" if r["cap_slots"] >= ws else "\\textbf{no}"
L.append(f"{r['rate_bits']:.1f} & {r['cap_slots']:,} & {ok} & "
f"{r['hit_rate']*100:.1f}\\% & {r['bytes_per_token_mb']:,.0f} & "
f"{r['tok_s']:.2f} & {b32['tok_s']:.2f} \\\\")
L += ["\\bottomrule", "\\end{tabular}"]
w("tab_proj.tex", "\n".join(L))
def tab_policy():
cp = load("cache_policy.json")
if not cp:
return
ws = cp["token_working_set"]
L = ["\\begin{tabular}{rrrrrr}", "\\toprule",
"Capacity & Slots & LRU & Static-freq. & Hybrid & Analytic \\\\",
"\\midrule"]
for r in cp["policies"]:
mark = "$^\\dagger$" if r["cap"] < ws else ""
L.append(f"{r['frac']*100:.1f}\\%{mark} & {r['cap']:,} & "
f"{r['lru']*100:.1f}\\% & {r['static']*100:.1f}\\% & "
f"{r['hybrid']*100:.1f}\\% & {r['analytic_static']*100:.1f}\\% \\\\")
L += ["\\bottomrule", "\\end{tabular}"]
w("tab_policy.tex", "\n".join(L))
if __name__ == "__main__":
for f in [tab_quality, tab_amp, tab_proj, tab_policy]:
try:
f()
except Exception as e:
print("skip", f.__name__, type(e).__name__, e)
|