From cee7215b24344a71ed9619b5911dcfa209a0f59e Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sun, 13 Sep 2026 04:04:33 +0000 Subject: [PATCH] feat(model_info): opt-in field-level backfill from fallback generalization rules Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../fallback_generalizations.py | 31 ++++- ...odel_prices_and_context_window_backup.json | 4 + litellm/utils.py | 6 + model_prices_and_context_window.json | 4 + .../test_fallback_generalizations.py | 124 +++++++++++++++++- 5 files changed, 166 insertions(+), 3 deletions(-) diff --git a/litellm/litellm_core_utils/fallback_generalizations.py b/litellm/litellm_core_utils/fallback_generalizations.py index 7739fc82c77..a5360309a70 100644 --- a/litellm/litellm_core_utils/fallback_generalizations.py +++ b/litellm/litellm_core_utils/fallback_generalizations.py @@ -34,6 +34,11 @@ rules never mix the two and never use ``extends``. A rule whose Rules are only consulted after exact and case-insensitive lookups miss, so an exact cost-map entry always takes precedence over any rule. +Rules flagged with ``backfill_exact_entries: true`` also fill only keys missing +from an exact cost-map entry, while values already present on the entry win on +conflict. Only flagged capability rules participate in this backfill; routing +rules never do. + Patterns are matched case-insensitively with ``re.search`` and are not implicitly anchored: a rule must include ``^`` and ``$`` to bind to the whole model name, otherwise it matches as a substring. Keeping anchoring in the regex makes the rule @@ -57,6 +62,7 @@ PATTERN_FIELD: Final = "pattern" MODEL_INFO_FIELD: Final = "model_info" PROVIDER_KEY: Final = "litellm_provider" LEGACY_EXTENDS_FIELD: Final = "extends" +BACKFILL_FIELD: Final = "backfill_exact_entries" def _resolve_legacy_extends(rules: list) -> list: @@ -98,6 +104,7 @@ class _RoutingRule: class _CapabilityRule: pattern: re.Pattern model_info: dict + backfill_exact_entries: bool _CompiledRule = _RoutingRule | _CapabilityRule @@ -125,8 +132,9 @@ def _compile_rule(rule: object) -> tuple[_CompiledRule, ...]: e, ) return () + backfill: Final = rule.get(BACKFILL_FIELD) is True if PROVIDER_KEY not in model_info: - return (_CapabilityRule(pattern=compiled, model_info=model_info),) + return (_CapabilityRule(pattern=compiled, model_info=model_info, backfill_exact_entries=backfill),) provider: Final = model_info[PROVIDER_KEY] if not isinstance(provider, str): verbose_logger.warning( @@ -140,7 +148,7 @@ def _compile_rule(rule: object) -> tuple[_CompiledRule, ...]: return (_RoutingRule(pattern=compiled, provider=provider),) return ( _RoutingRule(pattern=compiled, provider=provider), - _CapabilityRule(pattern=compiled, model_info=model_info), + _CapabilityRule(pattern=compiled, model_info=model_info, backfill_exact_entries=backfill), ) @@ -151,6 +159,7 @@ class _FallbackGeneralizations: self.rules: list = [] self.routing_rules: tuple = () self.capability_rules: tuple = () + self.backfill_rules: tuple = () def set_rules(self, rules: list | None) -> None: installed: Final = rules if isinstance(rules, list) else [] @@ -158,6 +167,7 @@ class _FallbackGeneralizations: self.rules = installed self.routing_rules = tuple(rule for rule in compiled if isinstance(rule, _RoutingRule)) self.capability_rules = tuple(rule for rule in compiled if isinstance(rule, _CapabilityRule)) + self.backfill_rules = tuple(rule for rule in self.capability_rules if rule.backfill_exact_entries) def match_routing(self, model: str) -> str | None: if not model: @@ -175,6 +185,14 @@ class _FallbackGeneralizations: return None return {key: value for model_info in matched for key, value in model_info.items()} + def match_backfill(self, model: str) -> dict | None: + if not model: + return None + matched = tuple(rule.model_info for rule in self.backfill_rules if rule.pattern.search(model) is not None) + if not matched: + return None + return {key: value for model_info in matched for key, value in model_info.items()} + _registry: Final = _FallbackGeneralizations() @@ -210,3 +228,12 @@ def match_capability_generalizations(model: str) -> dict | None: capability rule matches. O(number of rules); only call once exact lookups have missed. """ return _registry.match_capabilities(model) + + +def match_backfill_generalizations(model: str) -> dict | None: + """Return the union of flagged capability rules matching ``model``. + + Later rules override earlier ones on key conflicts. Returns ``None`` when no + flagged rule matches. O(number of rules); only call once exact lookups have matched. + """ + return _registry.match_backfill(model) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 89d582db151..6ba3c240a94 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -57717,6 +57717,7 @@ { "name": "claude-adaptive-thinking", "pattern": "claude-[a-z]+-(?:4[-._](?:[6-9]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", + "backfill_exact_entries": true, "description": "Claude at version 4.6 or higher, in any id shape that contains claude--: minors 4.6 through 4.99, any later major-minor, and bare 5+ majors so a new family shaped like claude-fable-5 matches. Requiring the claude- prefix keeps non-Claude names such as team-sonnet-5-1 out. The minor is capped at two digits so an 8-digit date suffix such as claude-opus-4-20250514 is never read as a >= 4.6 minor. Turns on adaptive thinking for new versions and new families with no code change.", "model_info": { "supports_adaptive_thinking": true @@ -57725,6 +57726,7 @@ { "name": "claude-legacy-thinking", "pattern": "claude-[a-z]+-4[-._]6(?!\\d)", + "backfill_exact_entries": true, "description": "Claude at version 4.6 exactly, in any id shape that contains claude--4-6 (dotted and underscored minors included, dated releases such as claude-sonnet-4-6-20260219 too). The 4.6 family is adaptive-thinking yet still accepts legacy thinking.type=enabled with budget_tokens, so the caller's hard budget cap is forwarded verbatim instead of being rewritten to an uncapped output_config.effort. The lookahead keeps two-digit minors such as 4-60 from matching. 4.7+ and 5+ majors reject the legacy shape and stay on the adaptive translation.", "model_info": { "supports_legacy_thinking": true @@ -57741,6 +57743,7 @@ { "name": "claude-mid-conversation-system", "pattern": "claude-[a-z]+-(?:4[-._](?:[89]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", + "backfill_exact_entries": true, "description": "Claude at version 4.8 or higher, in any id shape that contains claude--: minors 4.8 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Anthropic introduced mid-conversation system messages with Opus 4.8 and every newer Claude keeps them; 4.7 and below reject the system role inside messages.", "model_info": { "supports_mid_conversation_system": true @@ -57757,6 +57760,7 @@ { "name": "openai-reasoning-family-baseline", "pattern": "^(?!.*search-api)(?:[a-z0-9_.-]+/)*(?:ft:)?(?:o[1-9]\\d*(?![a-z0-9])|gpt-[5-9](?:\\.\\d+)?(?![0-9.])|(?:gpt-\\d+(?:\\.\\d+)?(?:-[a-z0-9]+)*-)?(?:codex|deep-research|chat-latest)(?![a-z0-9]))", + "backfill_exact_entries": true, "description": "OpenAI reasoning families by id shape, under any provider namespace and with an optional ft: prefix: the o-series (o1, o3-pro, o4-mini), gpt-5 through gpt-9 majors including dotted minors and suffixed variants (gpt-5.5-cyber, gpt-6-astra), and the codex, deep-research and chat-latest lines when standalone or on a gpt base. gpt-5-search-api is excluded because it is a search-only surface. Every model here is a reasoning model, and the Responses API drops the caller's reasoning param for any mapped OpenAI model whose info lacks supports_reasoning, so an id the registry has not named yet keeps its reasoning settings instead of silently losing them. Rules lose to exact entries. Carries no mode and no pricing, so cost stays on the standard unpriced behavior.", "model_info": { "supports_reasoning": true diff --git a/litellm/utils.py b/litellm/utils.py index 04139a124b6..87f270fa757 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -83,6 +83,7 @@ from litellm.constants import ( ) from litellm.litellm_core_utils.core_helpers import normalize_drop_params from litellm.litellm_core_utils.fallback_generalizations import ( + match_backfill_generalizations, match_capability_generalizations, ) from litellm.litellm_core_utils.sensitive_data_masker import redact_credentials_in_payload @@ -5819,6 +5820,11 @@ def _get_model_info_helper( ): _model_info = None + if _model_info is not None and key is not None: + backfill: Final = match_backfill_generalizations(key) + if backfill is not None: + _model_info = {**{k: v for k, v in backfill.items() if k not in _model_info}, **_model_info} + if _model_info is None: generalization: Final = _get_model_info_from_generalization( model=model, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 89d582db151..6ba3c240a94 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -57717,6 +57717,7 @@ { "name": "claude-adaptive-thinking", "pattern": "claude-[a-z]+-(?:4[-._](?:[6-9]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", + "backfill_exact_entries": true, "description": "Claude at version 4.6 or higher, in any id shape that contains claude--: minors 4.6 through 4.99, any later major-minor, and bare 5+ majors so a new family shaped like claude-fable-5 matches. Requiring the claude- prefix keeps non-Claude names such as team-sonnet-5-1 out. The minor is capped at two digits so an 8-digit date suffix such as claude-opus-4-20250514 is never read as a >= 4.6 minor. Turns on adaptive thinking for new versions and new families with no code change.", "model_info": { "supports_adaptive_thinking": true @@ -57725,6 +57726,7 @@ { "name": "claude-legacy-thinking", "pattern": "claude-[a-z]+-4[-._]6(?!\\d)", + "backfill_exact_entries": true, "description": "Claude at version 4.6 exactly, in any id shape that contains claude--4-6 (dotted and underscored minors included, dated releases such as claude-sonnet-4-6-20260219 too). The 4.6 family is adaptive-thinking yet still accepts legacy thinking.type=enabled with budget_tokens, so the caller's hard budget cap is forwarded verbatim instead of being rewritten to an uncapped output_config.effort. The lookahead keeps two-digit minors such as 4-60 from matching. 4.7+ and 5+ majors reject the legacy shape and stay on the adaptive translation.", "model_info": { "supports_legacy_thinking": true @@ -57741,6 +57743,7 @@ { "name": "claude-mid-conversation-system", "pattern": "claude-[a-z]+-(?:4[-._](?:[89]|[1-9]\\d)(?!\\d)|(?:[5-9]|[1-9]\\d)(?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", + "backfill_exact_entries": true, "description": "Claude at version 4.8 or higher, in any id shape that contains claude--: minors 4.8 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Anthropic introduced mid-conversation system messages with Opus 4.8 and every newer Claude keeps them; 4.7 and below reject the system role inside messages.", "model_info": { "supports_mid_conversation_system": true @@ -57757,6 +57760,7 @@ { "name": "openai-reasoning-family-baseline", "pattern": "^(?!.*search-api)(?:[a-z0-9_.-]+/)*(?:ft:)?(?:o[1-9]\\d*(?![a-z0-9])|gpt-[5-9](?:\\.\\d+)?(?![0-9.])|(?:gpt-\\d+(?:\\.\\d+)?(?:-[a-z0-9]+)*-)?(?:codex|deep-research|chat-latest)(?![a-z0-9]))", + "backfill_exact_entries": true, "description": "OpenAI reasoning families by id shape, under any provider namespace and with an optional ft: prefix: the o-series (o1, o3-pro, o4-mini), gpt-5 through gpt-9 majors including dotted minors and suffixed variants (gpt-5.5-cyber, gpt-6-astra), and the codex, deep-research and chat-latest lines when standalone or on a gpt base. gpt-5-search-api is excluded because it is a search-only surface. Every model here is a reasoning model, and the Responses API drops the caller's reasoning param for any mapped OpenAI model whose info lacks supports_reasoning, so an id the registry has not named yet keeps its reasoning settings instead of silently losing them. Rules lose to exact entries. Carries no mode and no pricing, so cost stays on the standard unpriced behavior.", "model_info": { "supports_reasoning": true diff --git a/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py b/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py index b6e656f282a..0221b9bbffc 100644 --- a/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py +++ b/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py @@ -11,11 +11,11 @@ import logging import pytest - import litellm from litellm._logging import verbose_logger from litellm.litellm_core_utils.fallback_generalizations import ( get_fallback_generalization_rules, + match_backfill_generalizations, match_capability_generalizations, match_routing_generalization, set_fallback_generalizations, @@ -116,6 +116,42 @@ def test_capability_union_is_last_wins_in_file_order(restore_generalizations): } +def test_backfill_requires_per_rule_opt_in(restore_generalizations): + restore_generalizations( + [ + {"name": "base", "pattern": r"^acme-", "model_info": {"supports_reasoning": True}}, + { + "name": "opt-in", + "pattern": r"^acme-", + "backfill_exact_entries": True, + "model_info": {"supports_vision": True}, + }, + ] + ) + assert match_backfill_generalizations("acme-1") == {"supports_vision": True} + assert match_capability_generalizations("acme-1") == { + "supports_reasoning": True, + "supports_vision": True, + } + + restore_generalizations( + [{"name": "base", "pattern": r"^acme-", "model_info": {"supports_reasoning": True}}] + ) + assert match_backfill_generalizations("acme-1") is None + + restore_generalizations( + [ + { + "name": "route", + "pattern": r"^acme-", + "backfill_exact_entries": True, + "model_info": {"litellm_provider": "openai"}, + } + ] + ) + assert match_backfill_generalizations("acme-1") is None + + def test_routing_rules_are_excluded_from_capability_results(restore_generalizations): restore_generalizations( [ @@ -299,6 +335,58 @@ def test_exact_entry_takes_precedence_over_rule(restore_generalizations): assert info["input_cost_per_token"] != 999.0 +def test_exact_entries_backfill_only_missing_fields(restore_generalizations, monkeypatch): + monkeypatch.setattr( + litellm, + "model_cost", + { + **litellm.model_cost, + "acme-full": { + "input_cost_per_token": 1e-6, + "output_cost_per_token": 2e-6, + "litellm_provider": "openai", + "mode": "chat", + "max_tokens": 7, + "supports_reasoning": False, + }, + "acme-bare": { + "input_cost_per_token": 3e-6, + "output_cost_per_token": 4e-6, + "litellm_provider": "openai", + "mode": "chat", + }, + }, + ) + restore_generalizations( + [ + { + "name": "acme-backfill", + "pattern": r"^acme-", + "backfill_exact_entries": True, + "model_info": {"supports_reasoning": True, "max_tokens": 5}, + } + ] + ) + litellm.get_model_info.cache_clear() + + full = litellm.get_model_info("acme-full", custom_llm_provider="openai") + assert full["supports_reasoning"] is False + assert full["max_tokens"] == 7 + + bare = litellm.get_model_info("acme-bare", custom_llm_provider="openai") + assert bare["supports_reasoning"] is True + assert bare["max_tokens"] == 5 + assert bare["input_cost_per_token"] == 3e-6 + assert bare["key"] == "acme-bare" + + restore_generalizations( + [{"name": "acme-backfill", "pattern": r"^acme-", "model_info": {"supports_reasoning": True, "max_tokens": 5}}] + ) + litellm.get_model_info.cache_clear() + unflagged = litellm.get_model_info("acme-bare", custom_llm_provider="openai") + assert unflagged.get("supports_reasoning") is None + + # --------------------------------------------------------------------------- # # Shipped rules (bundled cost map) # --------------------------------------------------------------------------- # @@ -605,6 +693,10 @@ def test_shipped_wandb_rule_loses_to_mapped_non_reasoning_entries(shipped_cost_m assert litellm.supports_reasoning(model=model, custom_llm_provider="wandb") is False, model +def test_shipped_wandb_rule_does_not_backfill_mapped_entries(shipped_cost_map): + assert match_backfill_generalizations("wandb/meta-llama/Llama-3.1-8B-Instruct") is None + + def test_shipped_wandb_rule_is_anchored_to_the_wandb_namespace(shipped_cost_map): """``^wandb/`` is anchored, so it cannot leak onto another provider's ids.""" assert match_capability_generalizations("wandb/some-new-model") == {"supports_reasoning": True} @@ -722,3 +814,33 @@ def test_shipped_openai_reasoning_rule_skips_non_reasoning_gpt_ids(shipped_cost_ def test_shipped_openai_reasoning_rule_loses_to_mapped_entries(shipped_cost_map): assert "gpt-5-search-api" in litellm.model_cost assert litellm.supports_reasoning(model="gpt-5-search-api", custom_llm_provider="openai") is False + + +@pytest.mark.parametrize( + "model,provider", + [ + ("azure/us/o1-2024-12-17", "azure"), + ("github_copilot/gpt-5", "github_copilot"), + ], +) +def test_shipped_openai_reasoning_rule_backfills_mapped_entries(shipped_cost_map, model, provider): + assert model in litellm.model_cost + raw_entry = litellm.model_cost[model] + assert "supports_reasoning" not in raw_entry + model_without_provider = model.removeprefix(f"{provider}/") + assert litellm.supports_reasoning(model=model_without_provider, custom_llm_provider=provider) is True + info = litellm.get_model_info(model=model_without_provider, custom_llm_provider=provider) + assert info["input_cost_per_token"] == raw_entry.get("input_cost_per_token", 0) + + +def test_shipped_claude_thinking_rules_backfill_without_family_limits(shipped_cost_map): + model = "perplexity/anthropic/claude-sonnet-4-6" + assert model in litellm.model_cost + raw_entry = litellm.model_cost[model] + assert "supports_adaptive_thinking" not in raw_entry + assert "max_input_tokens" not in raw_entry + + info = litellm.get_model_info(model="anthropic/claude-sonnet-4-6", custom_llm_provider="perplexity") + assert info["supports_adaptive_thinking"] is True + assert info["supports_legacy_thinking"] is True + assert info.get("max_input_tokens") is None