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

22
tests/test_llm.py
View file

@ -86,6 +86,28 @@ def test_max_tokens_is_sent_to_openrouter():
assert completions["openrouter"].kwargs["max_tokens"] == llm.LLM_MAX_TOKENS >= 16000
def test_solver_step_uses_higher_max_tokens():
# beobachtet am Portfoliomanagement: 4× je ~26k Reasoning-Tokens am 32k-Limit abgeschnitten.
# DeepSeek V4.1 Flash unterstützt bis 393k Output-Tokens (OpenRouter) – die Solver-Stufe bekommt
# ein eigenes, höheres Limit, die übrigen Stufen bleiben beim globalen.
s_solver, c_solver = fake_session()
s_solver.create(str, [{"role": "user", "content": "x"}], step="Solver-Code")
assert c_solver["openrouter"].kwargs["max_tokens"] == llm.LLM_MAX_TOKENS_SOLVER > llm.LLM_MAX_TOKENS
s_repair, c_repair = fake_session()
s_repair.create(str, [{"role": "user", "content": "y"}], step="Solver-Reparatur")
assert c_repair["openrouter"].kwargs["max_tokens"] == llm.LLM_MAX_TOKENS_SOLVER
s_extra, c_extra = fake_session()
s_extra.create(str, [{"role": "user", "content": "z"}], step="Extraktion")
assert c_extra["openrouter"].kwargs["max_tokens"] == llm.LLM_MAX_TOKENS
def test_solver_max_tokens_zero_falls_back_to_global(monkeypatch):
monkeypatch.setattr(llm, "LLM_MAX_TOKENS_SOLVER", 0)
session, completions = fake_session()
session.create(str, [{"role": "user", "content": "x"}], step="Solver-Code")
assert completions["openrouter"].kwargs["max_tokens"] == llm.LLM_MAX_TOKENS
# --- Modell je Pipeline-Stufe (LLM_MODEL_EXTRACTION, LLM_MODEL_SOLVER, ...) ---
def test_step_model_falls_back_to_session_model(monkeypatch):