File size: 6,632 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 | """Emit paper/numbers.tex so every quantity quoted in the paper is read
directly from the measurement artefacts rather than transcribed by hand."""
import glob, json, os, sys
RES = os.path.join(os.path.dirname(__file__), "..", "results")
OUT = os.path.join(os.path.dirname(__file__), "..", "paper", "numbers.tex")
def load(n, default=None):
p = os.path.join(RES, n)
return json.load(open(p)) if os.path.exists(p) else default
def fmt(x, d=2):
return f"{x:,.{d}f}"
def main():
M = {}
io = load("io_bench.json")
if io:
rnd = io["random"]
best = max(rnd, key=lambda r: r["mb_s"])
M["SeqRead"] = fmt(io["host"]["seq_read_mb_s"] / 1000, 2)
M["RandBest"] = fmt(best["mb_s"] / 1000, 2)
M["RandBestBlock"] = str(best["block_kb"] // 1024)
M["RandBestThreads"] = str(best["threads"])
one = [r for r in rnd if r["block_kb"] == 1024 and r["threads"] == 1][0]
M["RandOneThread"] = fmt(one["mb_s"] / 1000, 2)
M["RamCopy"] = fmt(io["host"]["ram_copy_GBs"], 1)
M["PCIe"] = fmt(io["host"]["h2d_16MB_GBs"], 2)
r64 = [r for r in rnd if r["block_kb"] == 64 and r["threads"] == 1][0]
M["RandSmall"] = fmt(r64["mb_s"] / 1000, 2)
db = load("decode_bench.json")
if db:
d = [x for x in db["decode"] if abs(x["bits"] - 1.5) < 1e-6][0]
M["DecodePacked"] = fmt(d["packed_GBs"], 2)
M["DecodeFpEquiv"] = fmt(d["fp16_equiv_GBs"], 1)
M["DecodeGw"] = fmt(d["weights_per_s"] / 1e9, 2)
M["PeakTflops"] = fmt(max(v["tflops"] for v in db["matmul"].values()), 2)
M["GPUName"] = db["gpu"]
pr = load("projection.json")
if pr:
P = pr["params"]
M["TotalParams"] = fmt(P["total"] / 1e9, 1)
M["ActiveParams"] = fmt(P["active"] / 1e9, 1)
M["ExpertSlots"] = f"{int(P['n_slots']):,}"
amp = {round(a["bits"], 2): a for a in pr["amplification"]}
M["FootprintFP"] = fmt(amp[16]["model_gb"], 0)
M["FootprintOnePFive"] = fmt(amp[1.5]["model_gb"], 0)
M["FootprintTwo"] = fmt(amp[2.0]["model_gb"], 0)
M["FootprintOne"] = fmt(amp[1.0]["model_gb"], 0)
M["ExpertMBfp"] = fmt(amp[16]["expert_mb"], 1)
M["ExpertMB"] = fmt(amp[1.5]["expert_mb"], 2)
M["CacheFracFP"] = fmt(amp[16]["frac"] * 100, 2)
M["CacheFrac"] = fmt(amp[1.5]["frac"] * 100, 2)
M["CacheExpertsFP"] = f"{amp[16]['dram_experts']:,}"
M["CacheExperts"] = f"{amp[1.5]['dram_experts']:,}"
M["Amplification"] = fmt(amp[1.5]["frac"] / amp[16]["frac"], 1)
M["ZipfS"] = fmt(pr["zipf_s"], 2)
rows = {(r["rate_bits"], r["batch"]): r for r in pr["projection"]}
b1 = rows[(1.5, 1)]
M["HitRate"] = fmt(b1["hit_rate"] * 100, 1)
M["TokSecOne"] = fmt(b1["tok_s"], 2)
M["MBperTok"] = fmt(b1["bytes_per_token_mb"], 0)
M["IOatExpert"] = fmt(b1["io_bw_gbs"], 2)
M["ResidentGB"] = fmt(b1["resident_gb"], 2)
M["TokSecEight"] = fmt(rows[(1.5, 8)]["tok_s"], 2)
M["TokSecBatch"] = fmt(rows[(1.5, 32)]["tok_s"], 2)
M["TokSecTwoBit"] = fmt(rows[(2.0, 1)]["tok_s"], 2)
M["TokSecOneBit"] = fmt(rows[(1.0, 1)]["tok_s"], 2)
M["TokSecOneBitBatch"] = fmt(rows[(1.0, 32)]["tok_s"], 2)
fp16 = rows[(16.0, 1)]
M["TokSecFP"] = fmt(fp16["tok_s"], 2)
M["HitRateFP"] = fmt(fp16["hit_rate"] * 100, 1)
M["SlotsFP"] = f"{fp16['cap_slots']:,}"
M["SpeedupOverFP"] = fmt(b1["tok_s"] / fp16["tok_s"], 1)
M["AmpFactor"] = fmt(b1["tok_s"] / fp16["tok_s"] / (16.0 / 1.5), 2)
M["ReqDecode"] = fmt(b1["io_bw_gbs"], 2)
if "DecodePacked" in M:
M["DecodeGap"] = fmt(b1["io_bw_gbs"] / float(M["DecodePacked"]), 0)
M["FusedBw"] = fmt(b1["io_bw_gbs"] / 1.5 * 16 * 2, 0)
M["FusedRatio"] = fmt(b1["io_bw_gbs"] / 1.5 * 16 * 2 / 88.2, 2)
M["ZipfBias"] = fmt(pr.get("zipf_bias_pp", 0), 2)
cv = load("cache_validation.json")
if cv:
M["TraceTokens"] = f"{cv['tokens']:,}"
M["MassTopQ"] = fmt(cv["mass_top25pct"] * 100, 1)
M["ReuseMean"] = fmt(sum(cv["reuse_prev_token"]) / len(cv["reuse_prev_token"]) * 100, 1)
d = {r["batch"]: r for r in cv["distinct_per_batch"]}
for b, w_ in [(1, "One"), (8, "Eight"), (32, "ThirtyTwo")]:
if b in d:
M["Distinct" + w_] = fmt(d[b]["measured"], 1)
ws = cv["working_set"]
for k in ("1", "16", "64", "1024"):
if k in ws:
M["WorkSet" + {"1": "One", "16": "Sixteen", "64": "SixtyFour",
"1024": "Kilo"}[k]] = fmt(ws[k], 1)
cp = load("cache_policy.json")
if cp:
M["ZipfSmeas"] = fmt(cp["zipf_s"], 2)
M["TokenWS"] = f"{cp['token_working_set']:,}"
M["TokenWSFrac"] = fmt(cp["ws_frac"] * 100, 1)
M["Slots"] = f"{cp['n_slots']:,}"
M["StaticMAE"] = fmt(cp["mae_analytic_static"] * 100, 2)
M["ZipfMAE"] = fmt(cp["mae_analytic_zipf"] * 100, 2)
h = {round(r["frac"], 3): r for r in cp["policies"]}
for k, name in [(0.02, "Two"), (0.05, "Five"), (0.10, "Ten"),
(0.125, "Twelve"), (0.15, "Fifteen"), (0.25, "TwentyFive"),
(0.40, "Forty")]:
if k in h:
M["Lru" + name] = fmt(h[k]["lru"] * 100, 1)
M["Static" + name] = fmt(h[k]["static"] * 100, 1)
M["Hybrid" + name] = fmt(h[k]["hybrid"] * 100, 1)
fp = load("fp16_ppl.json")
if fp:
M["PPLfp"] = fmt(fp["ppl"], 2)
runs = {}
for p in glob.glob(os.path.join(RES, "quant_*.json")):
r = json.load(open(p))
runs[r["config"]["tag"]] = r
for tag, name in [("main15", "Ours"), ("main20", "OursTwo"),
("main10", "OursOne"), ("main25", "OursTwoFive"),
("noldlq15", "NoLdlq"), ("northt15", "NoRht"),
("rtn2", "RtnTwo"), ("rtn3", "RtnThree"),
("freq15", "Freq")]:
if tag in runs:
M["PPL" + name] = fmt(runs[tag]["ppl"], 2)
M["Bits" + name] = fmt(runs[tag]["avg_bits"], 2)
M["NumRuns"] = str(len(runs))
os.makedirs(os.path.dirname(OUT), exist_ok=True)
with open(OUT, "w") as f:
f.write("% auto-generated by code/gen_numbers.py -- do not edit\n")
for k, v in sorted(M.items()):
f.write("\\newcommand{\\n%s}{%s}\n" % (k, v))
print(f"wrote {OUT} with {len(M)} macros")
for k, v in sorted(M.items()):
print(f" {k} = {v}")
if __name__ == "__main__":
main()
|