mirror of
https://github.com/kuhyx/testsAndMisc.git
synced 2026-11-10 20:47:03 +01:00
Krzysztof kuhy Rudnicki
1ba202c567
Some checks are pending
dependency freshness / dependencies on newest stable (push) Waiting to run
md-naming / markdown naming convention (push) Waiting to run
Pre-commit checks / pre-commit (${{ matrix.group }}) (mypy, mypy) (push) Waiting to run
Pre-commit checks / pre-commit (${{ matrix.group }}) (pylint, pylint) (push) Waiting to run
Pre-commit checks / pre-commit (${{ matrix.group }}) (rest, ) (push) Waiting to run
Pre-commit checks / pre-commit (${{ matrix.group }}) (shellcheck, shellcheck) (push) Waiting to run
Python tests / test (push) Waiting to run
repo-contract / repo contract (push) Waiting to run
Shell tests / shell-tests (push) Waiting to run
Shell tests / lib-tests (push) Waiting to run
Shell tests / arch-tests (push) Waiting to run
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JThjboXTeZ3yJPBzxMx5wS
80 lines
3.1 KiB
Python
80 lines
3.1 KiB
Python
"""Carry cost: what a tool result costs by staying in context afterwards.
|
|
|
|
A result of N tokens is cache-read again on every later API call of the same
|
|
file until a compaction drops it (or the session ends), so its real price is
|
|
N x (cache-read price of each later call) -- usually far more than the call
|
|
that fetched it. A compaction shows up as the context shrinking below
|
|
``COMPACTION_DROP_RATIO`` of the previous call's, the same rule token_audit's
|
|
image-cost model uses.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import TYPE_CHECKING, Protocol
|
|
|
|
from python_pkg.token_audit.model import COMPACTION_DROP_RATIO
|
|
from python_pkg.token_audit.pricing import PER_MILLION, WRITE_5M
|
|
|
|
if TYPE_CHECKING:
|
|
from python_pkg.token_audit.pricing import PriceBook
|
|
|
|
|
|
class _Row(Protocol):
|
|
"""Anything with creation / carry accumulators."""
|
|
|
|
create_usd: float
|
|
carry_usd: float
|
|
carry_calls: int
|
|
|
|
|
|
class Tracker:
|
|
"""Collects calls and results in order, settles carry at the end."""
|
|
|
|
def __init__(self) -> None:
|
|
"""Start with no calls and no results."""
|
|
self.calls: list[tuple[str, int]] = []
|
|
self.writes: list[float] = []
|
|
self.results: list[tuple[int, int, _Row]] = []
|
|
|
|
def turn(self, model: str, context: int, write: float = WRITE_5M) -> None:
|
|
"""Note one API call, in order, with its cache-write price multiplier."""
|
|
self.calls.append((model, context))
|
|
self.writes.append(write)
|
|
|
|
def result(self, row: _Row, tokens: int) -> None:
|
|
"""Note a result arriving before call number ``len(self.calls)``."""
|
|
self.results.append((len(self.calls), tokens, row))
|
|
|
|
def spans(self) -> list[tuple[int, int, int, _Row]]:
|
|
"""``(first, end, tokens, row)``: the calls ``first..end-1`` carry it."""
|
|
# cut[i]: index of the first call at or after i that starts a new
|
|
# context (after a compaction), else len(calls).
|
|
size = len(self.calls)
|
|
cut = [size] * (size + 1)
|
|
for index in range(size - 1, -1, -1):
|
|
dropped = index > 0 and self.calls[index][1] < (
|
|
COMPACTION_DROP_RATIO * self.calls[index - 1][1]
|
|
)
|
|
cut[index] = index if dropped else cut[index + 1]
|
|
return [
|
|
(start, cut[start + 1], tokens, row)
|
|
for start, tokens, row in self.results
|
|
if start < size
|
|
]
|
|
|
|
def settle(self, book: PriceBook) -> None:
|
|
"""Add each result's creation and carry to its row.
|
|
|
|
Creation: the first call after the result writes it to cache, at that
|
|
call's write multiplier of input price. Carry: every call from there
|
|
to the next compaction re-reads it at the cache-read price.
|
|
"""
|
|
prefix = [0.0]
|
|
for model, _ in self.calls:
|
|
prefix.append(prefix[-1] + book.for_model(model).cache_read)
|
|
for start, end, tokens, row in self.spans():
|
|
price = book.for_model(self.calls[start][0]).input
|
|
row.create_usd += tokens * self.writes[start] * price / PER_MILLION
|
|
row.carry_usd += tokens * (prefix[end] - prefix[start]) / PER_MILLION
|
|
row.carry_calls += end - start
|