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++
#!/usr/bin/env python3
"""A stub engine for exercising `strata tune run` end to end, on every OS.
Run it as ``[sys.executable, "stub_engine.py", "--serve", ...]`` (``strata_control.tuning.engine_command``
does that for any ``.py`` engine). It speaks the real serve protocol (READY / PP / T / DONE) and
makes its decode speed depend on one flag, so a sweep has something real to find. It is NOT a model
and its numbers mean nothing about any hardware -- it exists so the *sweep machinery* can be tested
without weights.
--kv int8 decodes at 200 tok/s
anything else (including --kv fp16) decodes at 100 tok/s
The prompt's length comes from the GEN line's token list, so the workload is genuinely the one the
caller asked for.
"""
import sys
def main(argv):
kv = "unknown"
for index, arg in enumerate(argv):
if arg == "--kv" and index + 1 < len(argv):
kv = argv[index + 1]
if "--help" in argv:
sys.stderr.write("usage:\n --kv KV\n --ple-io MODE\n")
return 0
out = sys.stdout
out.write("READY 32768 stop\n")
out.flush()
for raw in sys.stdin:
line = raw.strip()
if line == "QUIT" or not line:
return 0
if not line.startswith("GEN"):
continue
fields = line.split()
count = len(fields[-1].split(","))
try:
max_new = int(fields[1])
except (IndexError, ValueError):
max_new = 8
decode_ms = 50.0 if kv == "int8" else 100.0
out.write(f"PP {count} {count} 200.0 2560.0\n")
for i in range(max_new): # one T per token: the harness cross-checks the stream against DONE
out.write(f"T {i}\n")
out.write(f"DONE {max_new} {count} 200.0 {decode_ms} stop 0 0\n")
out.flush()
return 0
if __name__ == "__main__":
sys.exit(main(sys.argv[1:]))