-
[OPIK-7438] [BE] fix: gate sampling params for Anthropic adaptive-thinking models on the playground/inference path (#7575)
frostbyte_neo 发布于
2026-07-28 14:57:00 +00:00 - [OPIK-7438] [BE] fix: gate sampling params for Anthropic adaptive-thinking models on the playground/inference path
LlmProviderAnthropicMapper.toCreateMessageRequest forwarded temperature/top_p unconditionally, so
adaptive-thinking models (claude-sonnet-5, claude-opus-4-7/4-8) 400 with "temperature is deprecated for
this model" on API-created playground/inference/proxy requests that bypass the FE sanitizer (OPIK-6244).Gate temperature and top_p by AnthropicModelName.supportsSamplingParams(model) and whenever extended
thinking is enabled per-request via custom_parameters, mirroring the judge-path logic in
AnthropicClientGenerator (OPIK-7526 / PR #7531). top_k is not sent on this path (ChatCompletionRequest has
no top_k), so nothing to gate there.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- test(anthropic): cover claude-opus-4-7 and temperature-over-top_p precedence in sampling gate
Addresses review on #7575: add claude-opus-4-7 to the adaptive-model gating parameterization
(all three adaptive models now asserted to drop temperature/top_p), and add a case setting both
temperature and top_p on a sampling-capable model to guard the precedence branch (temperature kept,
top_p null).Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- refactor(anthropic): consolidate sampling-gate tests and make thinking type match case-insensitively
Address review on #7575 (andrescrz):
- Collapse the 8 SamplingParamsGating test methods into one @ParameterizedTest driven by
(description, request, expectedTemperature, expectedTopP), asserting via getTemperature()/getTopP()
accessors instead of direct field access. - Match thinking type case-insensitively (equalsIgnoreCase("disabled")) in both the mapper and the
judge-path AnthropicClientGenerator, keeping the two gates in parity.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- test(anthropic): assert full mapped payload and cover explicit null thinking type
Address follow-up review on #7575 (andrescrz):
- gatesSamplingParams now asserts actual vs expected AnthropicCreateMessageRequest via
usingRecursiveComparison() (full payload), not just temperature/top_p, so a collateral change to
any other mapped field is caught. - Add an explicit null thinking.type case asserting sampling params are still forwarded (null falls
through thinkingEnabled() as not-enabled, same path as blank/absent).
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
- test(anthropic): build expected payload independently of the mapper under test
Address review on #7575 (baz): the full-payload assertion built
expectedby re-calling
toCreateMessageRequest, so a collateral mapper bug would change actual and expected identically and
hide. Build expected directly from the known fixture (model, single user message, empty system,
default max_tokens, stream=false) plus the gate-derived temperature/top_p, so a regression in any
mapped field now fails the recursive comparison.Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
下载附件