mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-26 01:12:21 +00:00
2c718ea191
Anthropic's Messages API returns 400
(``temperature is deprecated for this model``) when ``temperature``,
``top_p``, or ``top_k`` is set to any non-default value on Claude
Opus 4.7 — per the Opus 4.7 migration guide. Previously
``AnthropicConfig.get_supported_openai_params`` unconditionally
advertised ``temperature`` and ``top_p`` for every Anthropic model,
so ``litellm.drop_params=True`` was a no-op on 4.7: litellm had no
reason to drop a param it believed was supported, the value was
forwarded to the API, and end users took a hard 400. The same shape
hits ``AmazonConverseConfig`` for Bedrock-routed Opus 4.7.
Fix follows the maintainers' existing capability-flag pattern
(``supports_*`` keys on model_prices_and_context_window.json read via
``_is_explicitly_disabled_factory``) rather than hardcoding model
strings in transformation logic:
* Mark every Anthropic-API-shape Opus 4.7 entry with
``supports_temperature: false`` / ``supports_top_p: false`` —
canonical Anthropic, all four Bedrock Converse regions
(anthropic/global/us/eu/au), Azure AI, and Vertex AI. Sonnet 4.6,
Haiku 4.5, Opus 4.6, and 3.x models are unchanged. OpenRouter and
Perplexity entries are deliberately left alone — they route through
different configs.
* New ``AnthropicConfig._param_explicitly_unsupported`` helper reads
``supports_{param}`` via the shared
``_is_explicitly_disabled_factory`` (same chain that drives
``_is_reasoning_effort_level_explicitly_disabled`` in
``OpenAIGPT5Config``). When the registry lookup misses — e.g. an
unreleased dated 4.7 snapshot — falls back to the existing
``_is_claude_4_7_model`` family check so future 4.7 variants are
covered before anyone updates the JSON.
* ``AnthropicConfig.get_supported_openai_params`` filters the helper
across every param (not just temperature) so future deprecations
only need a JSON entry. ``map_openai_params`` honours the same
helper inside the ``temperature`` / ``top_p`` branches so direct
callers don't leak the params either, even without ``drop_params``.
* ``AmazonConverseConfig.get_supported_openai_params`` mirrors the
filter so Bedrock-routed Opus 4.7 gets the same treatment.
The Databricks workspace endpoint surfaces the same Anthropic
deprecation (see the issue's follow-up comment); leaving that to a
separate PR keeps this change focused on the two configs the bug
report names.
The backup ``litellm/model_prices_and_context_window_backup.json``
absorbs upstream drift from earlier unrelated merges so the
``ci_cd/check_files_match.py`` byte-identity check passes — same
constraint that hit #26246.
Tests:
* New parametrized unit tests assert ``temperature`` / ``top_p`` are
filtered out of ``AnthropicConfig.get_supported_openai_params`` for
``claude-opus-4-7``, the dated variant
``claude-opus-4-7-20260416``, and the vendor-prefixed
``anthropic/claude-opus-4-7``.
* Family-fallback test covers an unreleased dated snapshot not in the
registry.
* Regression tests pin Sonnet 4.6 (both dated and alias), Haiku 4.5,
Sonnet 3.5, Sonnet 3.7, and Opus 4.6 to still advertise both
sampling params.
* ``map_openai_params`` test confirms explicit ``temperature=0.3,
top_p=0.9`` does not reach ``optional_params`` for Opus 4.7 but
does for Opus 4.6.
* End-to-end-shape test reproduces the issue: ``drop_params=True`` +
``temperature=0.1`` + ``transform_request`` no longer leaks
``temperature`` into the Anthropic request body.
* Bedrock Converse mirror suite covers all four region variants and
pins Sonnet 4.6 / Haiku 4.5 / 3.5-sonnet on Bedrock as regressions.
Existing Anthropic + Bedrock test suites stay green
(``pytest tests/test_litellm/llms/anthropic
tests/test_litellm/llms/bedrock/chat`` — 976 passed locally).
Fixes #26444
|
||
|---|---|---|
| .. | ||
| agent_tests | ||
| audio_tests | ||
| basic_proxy_startup_tests | ||
| batches_tests | ||
| benchmarks | ||
| code_coverage_tests | ||
| documentation_tests | ||
| enterprise | ||
| guardrails_tests | ||
| image_gen_tests | ||
| litellm | ||
| litellm-proxy-extras | ||
| litellm_core_utils | ||
| litellm_utils_tests | ||
| llm_responses_api_testing | ||
| llm_translation | ||
| load_tests | ||
| local_testing | ||
| logging_callback_tests | ||
| mcp_tests | ||
| multi_instance_e2e_tests | ||
| ocr_tests | ||
| old_proxy_tests/tests | ||
| openai_endpoints_tests | ||
| otel_tests | ||
| pass_through_tests | ||
| pass_through_unit_tests | ||
| proxy_admin_ui_tests | ||
| proxy_e2e_anthropic_messages_tests | ||
| proxy_security_tests | ||
| proxy_unit_tests | ||
| router_unit_tests | ||
| scim_tests | ||
| search_tests | ||
| spend_tracking_tests | ||
| store_model_in_db_tests | ||
| test_litellm | ||
| unified_google_tests | ||
| vector_store_tests | ||
| windows_tests | ||
| __init__.py | ||
| _flush_vcr_cache.py | ||
| _vcr_conftest_common.py | ||
| _vcr_redis_persister.py | ||
| eval_swe_bench.py | ||
| gettysburg.wav | ||
| large_text.py | ||
| openai_batch_completions.jsonl | ||
| README.MD | ||
| test_budget_management.py | ||
| test_callbacks_on_proxy.py | ||
| test_config.py | ||
| test_debug_warning.py | ||
| test_default_encoding_non_root.py | ||
| test_end_users.py | ||
| test_entrypoint.py | ||
| test_fallbacks.py | ||
| test_gpt5_azure_temperature_support.py | ||
| test_health.py | ||
| test_keys.py | ||
| test_litellm_proxy_responses_config.py | ||
| test_logging.conf | ||
| test_models.py | ||
| test_new_vector_store_endpoints.py | ||
| test_openai_endpoints.py | ||
| test_organizations.py | ||
| test_otel_thread_leak.py | ||
| test_passthrough_endpoints.py | ||
| test_presidio_latency.py | ||
| test_proxy_server_non_root.py | ||
| test_ratelimit.py | ||
| test_resource_cleanup.py | ||
| test_service_logger_otel.py | ||
| test_spend_logs.py | ||
| test_team.py | ||
| test_team_logging.py | ||
| test_team_members.py | ||
| test_users.py | ||
In total litellm runs 1000+ tests
[02/20/2025] Update:
To make it easier to contribute and map what behavior is tested,
we've started mapping the litellm directory in tests/test_litellm
This folder can only run mock tests.