dschlueter/text_to_solver_program
Watch
1
0
Fork
You've already forked text_to_solver_program
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

feat: stufen-spezifisches Token-Limit für die Solver-Stufe (LLM_MAX_TOKENS_SOLVER)

Browse source 
Die Solver-Stufe erzeugt langen Python-Code plus Reasoning. Bei komplexen
Aufgaben (Portfoliomanagement: 4× je ~26k Reasoning-Tokens am 32k-Limit
abgeschnitten, $0,28 für nichts) reichte das globale Limit nicht. DeepSeek
V4.1 Flash unterstützt bis 393k Output-Tokens (OpenRouter).

Neu: LLM_MAX_TOKENS_SOLVER (Default 65536, 0 = globales Limit gilt auch
für den Solver). Wirkt nur für die Stufen 'Solver-Code' und 'Solver-Reparatur'.
CLI-Option --max-tokens-solver. 438 Tests grün.
This commit is contained in:
Dieter Schlüter 2026-09-18 21:48:18 +02:00
commit e7a2b8de81
1 changed files with 51 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

16
tests/test_cli.py
View file

@ -14,7 +14,8 @@ from test_resolve import gespeichertes_ergebnis
def einstellungen_zuruecksetzen(monkeypatch):
# main() setzt Modulkonstanten; monkeypatch stellt sie nach jedem Test wieder her
for modul, name in ((TextToModel, "UNDERSTAND_RETRIES"), (TextToModel, "INFEASIBLE_RECHECK"),
(llm, "REASONING_EFFORT"), (llm, "LLM_MAX_TOKENS"), (llm, "LLM_TIMEOUT"),
(llm, "REASONING_EFFORT"), (llm, "LLM_MAX_TOKENS"), (llm, "LLM_MAX_TOKENS_SOLVER"),
(llm, "LLM_TIMEOUT"),
(solver_stage, "SOLVER_TIMEOUT"), (solver_stage, "SANDBOX")):
monkeypatch.setattr(modul, name, getattr(modul, name))
@ -101,7 +102,8 @@ def test_options_reach_pipeline_and_modules(tmp_path, monkeypatch):
monkeypatch.setattr(TextToModel, "run_pipeline", fake_pipeline)
code = TextToModel.main([str(aufgabe), "-m", "x/y", "--max-revisions", "0", "--solver-retries", "4",
"--max-restarts", "0", "--understand-retries", "-1", "--infeasible-recheck", "0",
"--reasoning", "low", "--max-tokens", "1000", "--llm-timeout", "30",
"--reasoning", "low", "--max-tokens", "1000", "--max-tokens-solver", "2000",
"--llm-timeout", "30",
"--solver-timeout", "90", "--sandbox", "bwrap", "-o", str(tmp_path / "e.json")])
assert code == TextToModel.EXIT_REJECTED
assert gesehen["task"].startswith("Maximiere") and gesehen["task_dir"] == tmp_path
@ -109,16 +111,18 @@ def test_options_reach_pipeline_and_modules(tmp_path, monkeypatch):
== ("x/y", 0, 4, 0)
assert gesehen["solve_model"] is True
assert (TextToModel.UNDERSTAND_RETRIES, TextToModel.INFEASIBLE_RECHECK) == (-1, 0)
assert (llm.REASONING_EFFORT, llm.LLM_MAX_TOKENS, llm.LLM_TIMEOUT) == ("low", 1000, 30)
assert (llm.REASONING_EFFORT, llm.LLM_MAX_TOKENS, llm.LLM_MAX_TOKENS_SOLVER, llm.LLM_TIMEOUT) \
== ("low", 1000, 2000, 30)
assert (solver_stage.SOLVER_TIMEOUT, solver_stage.SANDBOX) == (90, "bwrap")
assert "90" in solver_stage.solver_prompt() # Zeitbudget im Prompt folgt der Option
def test_without_options_the_defaults_stay(tmp_path, monkeypatch):
vorher = (llm.LLM_MAX_TOKENS, solver_stage.SOLVER_TIMEOUT, solver_stage.SANDBOX, TextToModel.UNDERSTAND_RETRIES)
vorher = (llm.LLM_MAX_TOKENS, llm.LLM_MAX_TOKENS_SOLVER, solver_stage.SOLVER_TIMEOUT,
solver_stage.SANDBOX, TextToModel.UNDERSTAND_RETRIES)
monkeypatch.setattr(TextToModel, "run_pipeline", lambda task, **kw: PipelineResult(task=task, reliable=True))
aufgabe = tmp_path / "a.txt"
aufgabe.write_text("Minimiere x.", encoding="utf-8")
TextToModel.main([str(aufgabe), "-o", str(tmp_path / "e.json")])
assert (llm.LLM_MAX_TOKENS, solver_stage.SOLVER_TIMEOUT, solver_stage.SANDBOX,
TextToModel.UNDERSTAND_RETRIES) == vorher
assert (llm.LLM_MAX_TOKENS, llm.LLM_MAX_TOKENS_SOLVER, solver_stage.SOLVER_TIMEOUT,
solver_stage.SANDBOX, TextToModel.UNDERSTAND_RETRIES) == vorher