| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
Add the Chat input to Anthropic | 1 年前 | |
Add the Chat input to Anthropic | 1 年前 | |
Update Dottxt tests for new SDK interface - Update client import to DotTxt/AsyncDotTxt - Replace list_models() with models.list() and model.id - Replace model_name + model_revision fixtures with single model_name - Add async model fixtures and async test coverage - Add missing model validation tests - Fix test for wrong inference parameters to use model with identifier set | 4 个月前 | |
Create the inputs module and modify the treatment of multimodal inputs (#1677) Closes #1649 | 1 年前 | |
Add the Chat input to Gemini | 1 年前 | |
Use the JsonSchema conversion methods in models | 10 个月前 | |
fix: correct typo | 8 个月前 | |
fix(tokenizers): inline spiece underline constant Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> | 5 个月前 | |
test: simplify some tests | 8 个月前 | |
Add coverage for AsyncLMStudio's context lifecycle and from_lmstudio Three code paths in lmstudio.py had no test exercising them: - AsyncLMStudio.close(), which was never called by any existing test - the branch in generate_stream that skips re-entering the client context on a second call, since every existing test only streamed once per model instance - the success branches of from_lmstudio() for a plain Client and a plain AsyncClient, since the existing init tests only reach from_lmstudio() through the invalid-client path when no real server is configured lmstudio.Client and lmstudio.AsyncClient don't connect eagerly on construction, so the from_lmstudio() dispatch can be checked with a throwaway client pointed at a bogus host, without needing a live server or a mock. | 1 个月前 | |
Run LMStudio image input tests without a live server The image-handling branches of LMStudioTypeAdapter.format_input only need a real LM Studio server because they call lms.prepare_image, which uploads bytes to a running instance. Everything else in these tests is just checking that the adapter builds the right lmstudio.Chat structure, so stub prepare_image instead of skipping the tests whenever LMSTUDIO_SERVER_URL isn't set. This also adds a test for the invalid image asset branch inside chat-formatted input, mirroring the one that already exists for list-formatted input. | 1 个月前 | |
Improve exceptions handling | 3 个月前 | |
Make cosmethic changes to the Mistral model | 10 个月前 | |
Handle both HF tokenizer and mlx tokenizer wrapper for the MLXLM model | 4 个月前 | |
test: Add some initial tests | 8 个月前 | |
Pull Ollama vision model to fix failing multimodal tests | 2 个月前 | |
Fix typo in vision argument of Ollama | 7 个月前 | |
Improve exceptions handling | 3 个月前 | |
Add the Chat input to OpenAI | 1 年前 | |
Improve exceptions handling | 3 个月前 | |
fix(sglang): copy caller extra_body before merging structured outputs _pop("extra_body", {}) returns the caller's dict when present, so .update() leaked structured_outputs/regex into subsequent unconstrained calls. Mirror the VLLM fix (#1933) with dict(...). Fixes #1931 | 1 个月前 | |
Add the Chat input to SGLang | 1 年前 | |
Add tests for the batch method of models | 1 年前 | |
Add tests for the batch method of models | 1 年前 | |
fix: add tests for _check_hf_chat_template function in tokenizer module | 8 个月前 | |
Fix a bug in the Transformers tokenizer | 6 个月前 | |
Add the device_dtype init parameter to the transformers model | 10 个月前 | |
Add tests for mutiple assets usage in multimodal chat template format | 11 个月前 | |
Remove unused Llama tokenizer detection TransformerTokenizer.is_llama is write-only: convert_token_to_string() now always applies the <0x20> / SPIECE space workaround and no longer branches on the flag. This removes the unused Llama-tokenizer detection helper, the unused attribute assignment, and the tests that only exercised that dead path. get_llama_tokenizer_types() is not exported in __all__, but it is currently included in the generated API reference; happy to keep it as a deprecated compatibility shim if maintainers prefer preserving that import for one release. Closes #1874. | 2 个月前 | |
test: simplify some tests | 8 个月前 | |
fix: only set additionalProperties on object schemas set_additional_properties_false_json_schema keyed off any value equal to the string 'object' (match.value == 'object') and then mutated that node's parent. This wrongly injected 'additionalProperties': False into non-object schemas whenever a scalar value happened to equal 'object' -- for example a string field emitted by pydantic as {'type': 'string', 'const': 'object'} for a Literal['object'] discriminator, or a field with a 'default'/'title'/'description' of 'object'. The OpenAI structured-output API rejects 'additionalProperties' on non-object schemas, so a legitimate model would fail with a 400 error. Rewrite the helper to recurse over the schema and set 'additionalProperties' only on dicts whose 'type' is 'object', matching the JSON Schema keyword rather than an incidental string value. Nested object schemas are still handled. Add regression tests. | 1 个月前 | |
fix(vllm): copy caller's extra_body before merging structured_outputs `VLLM._build_client_args` and `AsyncVLLM._build_client_args` previously popped the caller-provided `extra_body` dict and called `.update(...)` on it directly. `dict.pop(key, {})` only returns a fresh dict when the key is *absent* — when the caller passes `extra_body=my_dict`, the same object comes back, so the subsequent `.update(output_type_args)` mutates the caller's dict in place. Reusing one `extra_body` dict across calls then leaks the previous call's `structured_outputs` constraint into later unconstrained calls, silently constraining generation to a stale regex/JSON schema. Wrap the pop in `dict(...)` so we own the dict before merging into it. Shallow copy is sufficient because `output_type_args` only adds top-level keys (`structured_outputs`, `response_format`). Adds a parametrized regression test over `sync_model`/`async_model` covering both the in-place mutation and the cross-call leak. Fixes #1931. | 1 个月前 | |
Refactor test_check_chat_template functions to streamline tests and remove redundancy | 4 个月前 | |
test: simplify some tests | 8 个月前 | |
fix(vllm): send structured_outputs instead of removed guided_* keys vLLM removed the per-request guided_json/guided_regex/guided_grammar extra_body keys in v0.12 and silently ignores unknown request fields, so from_vllm against a vLLM 0.23 server dropped the constraint and returned unconstrained output. Emit extra_body={"structured_outputs": {...}} with the json/regex/ grammar keys (whitespace_pattern nested alongside) instead, mirroring the offline path migrated in #1779. Fixes #1887 | 2 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 1 年前 | ||
| 1 年前 | ||
| 4 个月前 | ||
| 1 年前 | ||
| 1 年前 | ||
| 10 个月前 | ||
| 8 个月前 | ||
| 5 个月前 | ||
| 8 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 3 个月前 | ||
| 10 个月前 | ||
| 4 个月前 | ||
| 8 个月前 | ||
| 2 个月前 | ||
| 7 个月前 | ||
| 3 个月前 | ||
| 1 年前 | ||
| 3 个月前 | ||
| 1 个月前 | ||
| 1 年前 | ||
| 1 年前 | ||
| 1 年前 | ||
| 8 个月前 | ||
| 6 个月前 | ||
| 10 个月前 | ||
| 11 个月前 | ||
| 2 个月前 | ||
| 8 个月前 | ||
| 1 个月前 | ||
| 1 个月前 | ||
| 4 个月前 | ||
| 8 个月前 | ||
| 2 个月前 |