scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-11 22:51:28 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 124

fix(cost): apply fast pricing alias to thresholds

Browse source
This commit is contained in:
LH-kevin 2026-08-22 21:05:28 +08:00
parent 194c13384f
commit 11937bf88d
2 changed files with 150 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
litellm/litellm_core_utils/llm_cost_calc/utils.py
View file

@ -334,7 +334,10 @@ def _get_token_base_cost(
continue
threshold_key_tier = match.group("tier")
if threshold_key_tier is not None and threshold_key_tier != service_tier:
active_pricing_tier = (
None if service_tier is None else _SERVICE_TIER_TO_COST_KEY_SUFFIX.get(service_tier.lower())
)
if threshold_key_tier is not None and threshold_key_tier != active_pricing_tier:
continue
try:

146
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_threshold_cost_selection.py
View file

@ -197,6 +197,152 @@ def test_unequal_threshold_highest_standard_still_wins():
assert rates[1] == pytest.approx(1e-6)
@pytest.mark.parametrize(
("tier_field", "tier_value", "rate_index"),
[
("output_cost_per_token_above_200k_tokens_priority", 1.5e-6, 1),
("cache_read_input_token_cost_above_200k_tokens_priority", 4e-7, 4),
],
)
def test_fast_service_tier_uses_priority_qualified_threshold(tier_field, tier_value, rate_index):
"""fast aliases to priority pricing, so a priority-qualified threshold with no standard
sibling must apply to a fast request too."""
model_info = {
"input_cost_per_token": 1e-6,
"output_cost_per_token": 1e-6,
"cache_read_input_token_cost": 2e-7,
tier_field: tier_value,
}
usage = Usage(
prompt_tokens=250_000,
completion_tokens=1_000,
total_tokens=251_000,
)
rates = _get_token_base_cost(
model_info=model_info,
usage=usage,
service_tier="fast",
)
assert rates[rate_index] == pytest.approx(tier_value)
@pytest.mark.parametrize(
("service_tier", "expected"),
[
("priority", 1.5e-6),
("flex", 1e-6),
(None, 1e-6),
],
ids=["priority-guard", "flex-guard", "default-guard"],
)
def test_fast_alias_guard_tiers_for_priority_only_output_threshold(service_tier, expected):
"""priority must keep working under the alias; flex and default must not pick up the
priority rate."""
model_info = {
"input_cost_per_token": 1e-6,
"output_cost_per_token": 1e-6,
"output_cost_per_token_above_200k_tokens_priority": 1.5e-6,
}
usage = Usage(
prompt_tokens=250_000,
completion_tokens=1_000,
total_tokens=251_000,
)
rates = _get_token_base_cost(
model_info=model_info,
usage=usage,
service_tier=service_tier,
)
assert rates[1] == pytest.approx(expected)
@pytest.mark.parametrize("service_tier", ["priority", "fast"], ids=["priority", "fast"])
def test_fast_builtin_standard_plus_priority_threshold_resolves_priority(service_tier):
"""The builtin shape (standard sibling present) keeps resolving the priority variant via
the existing standard-key alias lookup."""
model_info = {
"input_cost_per_token": 1e-6,
"output_cost_per_token": 1e-6,
"output_cost_per_token_above_200k_tokens": 1e-6,
"output_cost_per_token_above_200k_tokens_priority": 1.5e-6,
}
usage = Usage(
prompt_tokens=250_000,
completion_tokens=1_000,
total_tokens=251_000,
)
rates = _get_token_base_cost(
model_info=model_info,
usage=usage,
service_tier=service_tier,
)
assert rates[1] == pytest.approx(1.5e-6)
@pytest.mark.parametrize(
"model_info",
[
{
"input_cost_per_token": 1e-6,
"output_cost_per_token": 1e-6,
"output_cost_per_token_above_200k_tokens": 1e-6,
"output_cost_per_token_above_200k_tokens_priority": 1.5e-6,
},
{
"input_cost_per_token": 1e-6,
"output_cost_per_token": 1e-6,
"output_cost_per_token_above_200k_tokens_priority": 1.5e-6,
"output_cost_per_token_above_200k_tokens": 1e-6,
},
],
ids=["standard-first", "priority-first"],
)
def test_fast_equal_threshold_matching_tier_wins_both_orders(model_info):
"""The Phase D equal-threshold tie-break must hold under the fast alias, independent of
insertion order."""
usage = Usage(
prompt_tokens=250_000,
completion_tokens=1_000,
total_tokens=251_000,
)
rates = _get_token_base_cost(
model_info=model_info,
usage=usage,
service_tier="fast",
)
assert rates[1] == pytest.approx(1.5e-6)
def test_fast_service_tier_uppercase_uses_priority_qualified_threshold():
"""The scanner must be as case-insensitive as the lookup path."""
model_info = {
"input_cost_per_token": 1e-6,
"output_cost_per_token": 1e-6,
"output_cost_per_token_above_200k_tokens_priority": 1.5e-6,
}
usage = Usage(
prompt_tokens=250_000,
completion_tokens=1_000,
total_tokens=251_000,
)
rates = _get_token_base_cost(
model_info=model_info,
usage=usage,
service_tier="FAST",
)
assert rates[1] == pytest.approx(1.5e-6)
@pytest.mark.parametrize(
("tier_field", "tier_value", "rate_index"),
[