Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
import json
import numpy as np
TV = {1: 39.96, 2: 43.21, 3: 48.39, 4: 53.68, 5: 59.40, 6: 66.00, 7: 74.66, 8: 80.34} # ms per verify window, gfx1151 IQ3_S (STRATA_SPEC_T_CYCLE run)
HOST = 0.07
PROD = {"chat_explain": (2.39, 43.33, 2.38, 43.97), "code_py": (2.27, 43.32, 2.22, 42.72), "code_cpp": (2.11, 40.60, 2.05, 39.93),
"math": (2.99, 52.38, 2.95, 52.29), "summarize": (2.81, 48.25, 2.75, 48.06), "json_extract": (3.25, 53.91, 3.22, 54.22),
"creative": (2.30, 42.77, 2.29, 42.65), "chat_advice": (1.94, 38.59, 1.92, 38.47)} # default tau, default tok/s, MTP-only tau, MTP-only tok/s
names = list(PROD)
A8 = json.load(open(r"C:\dflash2-work\anchors1_b8.json"))
A4 = json.load(open(r"C:\dflash2-work\anchors1_b4.json"))
SEQ = {n: json.load(open(r"C:\dflash2-work\p\%s.seq.json" % n)) for n in names}
def walk(rows, T_of, D, n):
by = {r["s"]: r for r in rows}
s = len(SEQ[n]["prompt"])
tok = 0
ms = 0.0
rounds = 0
hist = np.zeros(10)
while s in by:
r = by[s]
T = T_of(r)
a = 0
while a < T - 1 and a < len(r["ok"]) and r["ok"][a]:
a += 1
tok += a + 1
ms += TV[T] + D + HOST
rounds += 1
hist[T] += 1
s += a + 1
return tok / rounds, ms / rounds, 1000.0 * tok / ms, hist
def conf_T(theta, cap=8):
def f(r):
T = 1
for p in r["p"][:cap - 1]:
if p >= theta:
T += 1
else:
break
return T
return f
print("check, fixed b=8 walk tau vs the earlier walk: ", {n: round(walk(A8[n], lambda r: 8, 8, n)[0], 3) for n in names})
print("check, fixed b=4 drafter walk tau: ", {n: round(walk(A4[n], lambda r: 4, 8, n)[0], 3) for n in names})
out = {}
for D in (5.0, 8.0, 12.0):
print("\n=== drafter cost D = %.0f ms per round (verify cost + D + 0.07) ===" % D)
cfgs = [("fixed4 block-4 drafter", A4, lambda r: 4), ("fixed3 block-8 drafter", A8, lambda r: 3),
("fixed4 block-8 drafter", A8, lambda r: 4), ("fixed5", A8, lambda r: 5), ("fixed6", A8, lambda r: 6),
("fixed7", A8, lambda r: 7), ("fixed8", A8, lambda r: 8)]
for th in (0.3, 0.4, 0.5, 0.6, 0.7, 0.8):
for cap in (4, 5, 6, 8):
cfgs.append(("conf>=%.1f cap %d" % (th, cap), A8, conf_T(th, cap)))
print("%-26s" % "policy" + "".join("%15s" % n[:13] for n in names) + " mean-of-ratios")
best = None
for label, A, f in cfgs:
ratios, row = [], "%-26s" % label
for n in names:
tau, ms, tps, hist = walk(A[n], f, D, n)
ratios.append(tps / PROD[n][1])
row += "%15s" % ("%.2f/%.1f(%+.0f%%)" % (tau, tps, 100 * (tps / PROD[n][1] - 1)))
m = 100 * (np.mean(ratios) - 1)
if D == 8.0 or label.startswith("fixed") or True:
print(row + " %+6.1f%%" % m)
if best is None or m > best[0]:
best = (m, label)
out[(D, label)] = m
print("best:", best)