We don't publish
your competitive advantage.

IFLoading a model with transformers raises ValueError: 'Due to a serious vulnerability... we now require users to upgrade torch to at least v2.6' on systems where torch 2.6 is unavailable (e.g., x86 Mac).

THENPin transformers to version 4.41.0 and numpy to 1.26.4 as a temporary workaround. Alternatively, convert the model checkpoint to safetensors format using a machine that supports torch 2.6 (e.g., Colab), then load the safetensors file with transformers without torch version restrictions.

IFCreating a ModernBert model with flash attention enabled results in TypeError: RotaryEmbedding.__init__() got an unexpected keyword argument 'pos_idx_in_fp32'

THENDowngrade flash-attn to version 2.7.4.post1, which still supports the pos_idx_in_fp32 parameter. Alternatively, remove the `pos_idx_in_fp32=True` argument from the `ModernBertUnpaddedRotaryEmbedding.__init__` super() call in the transformers source code, as the parameter was deprecated and removed in newer flash-attn versions.

IFWhen using langchain v0.0.323 with openai v1.0.1, ChatOpenAI raises AttributeError: 'module' object has no attribute 'ChatCompletion'.

THENUpgrade langchain to version >=0.0.331 (which uses the new openai client) or downgrade openai to <1.0.0. If upgrading langchain is not possible, switch to the new openai interface directly by using ChatCompletion.create instead of the deprecated attribute.

IFRunning vLLM >=0.10.0 with Gemma-2 models on H100 NVL GPUs causes RuntimeError: 'This flash attention build does not support tanh softcapping.'

THENDowngrade vLLM to version 0.9.2, which supports the tanh softcapping needed for Gemma-2 models on H100 GPUs. Alternatively, use A100 GPUs which work with newer vLLM versions.

IFWhen using ChatOpenAI with a custom httpx.Client or AsyncClient and openai>=1.14.2, a pydantic validation error occurs because langchain-openai<0.1.0 type-checks http_client incorrectly against both client types simultaneously.

THENUpgrade langchain-openai to >=0.1.0 (released around April 2024). The fix in PR #19164 resolves the sync client issue; for async clients, ensure the keyword argument is explicitly named 'http_async_client' rather than 'http_client'.

IFImporting transformers fails with AttributeError: module 'torch.utils._pytree' has no attribute 'register_pytree_node' on PyTorch versions below 2.0 (e.g., 1.11).

THENUpgrade PyTorch to a supported version (>=2.0 recommended) to resolve the compatibility issue. If upgrading is not possible, check for a transformers version that includes a proper version guard or patch (see PR #29364).

IFImportError related to missing classes when using langchain with third-party libraries like llama-index

THENEnsure version compatibility either by downgrading langchain to 0.0.340 or 0.0.148, or by upgrading llama-index to its latest version. Conflicting dependencies between langchain and llama-index can cause import errors for classes like BaseCache or Document.

IFLiteLLM v1.81.3-stable.opus-4-6 deployed on Amazon Bedrock fails to start with ImportError: cannot import name 'is_claude_4_5_on_bedrock' from litellm.llms.bedrock.common_utils.

THENUpgrade LiteLLM to v1.81.13 or later, which includes updated support for the new Amazon Bedrock model naming pattern for Claude Opus 4.6 (e.g., 'global.anthropic.claude-opus-4-6-v1'). The older version referenced a removed import.

IFWhen using transformers version 4.50.0, a RuntimeError occurs: 'register_pytree_node() got an unexpected keyword argument flatten_with_keys_fn'.

THENDowngrade transformers to version 4.49.0 to avoid this compatibility issue. Pin the version in your environment or requirements file (e.g., using `pip install transformers==4.49.0`). This resolves the error on Linux and Windows.

IFDeploying LiteLLM v1.81.3-stable.opus-4.6 fails with ImportError because the import of 'is_claude_4_5_on_bedrock' does not exist in the bedrock common_utils module, due to a changed naming pattern for Claude Opus 4.6 in Amazon Bedrock (e.g., global.anthropic.claude-opus-4-6-v1).

THENUpgrade to LiteLLM v1.81.13 or later, which includes the updated model mapping for Claude Opus 4.6. Avoid using pre-release or version-tagged builds (e.g., opus-4.6) that may reference outdated internal functions. Always verify the release notes or confirm with the latest stable version before deploying new model support.

IFWhen using transformers version 4.52.2 with PyTorch version less than 2.5, importing any model (e.g., RobertaModel) or even running `transformers env` raises NameError: name 'Replicate' is not defined.

THENUpgrade PyTorch to version 2.5 or later to satisfy the conditional import, or downgrade transformers to version 4.52.1 which did not exhibit this bug. Alternatively, apply the hotfix from PR #38262 that guards the use of `Replicate` behind the same version check used for its import.

IFWhen using langgraph-api==0.2.128 and langgraph-runtime-inmem==0.6.11, running 'langgraph dev' and chatting via API causes error: Configuration.__init__() got an unexpected keyword argument 'host'.

THENUpgrade langgraph-api to v0.2.129 or later and langgraph-runtime-inmem to v0.6.12 or later to fix the bug where request headers are incorrectly passed to the Configuration class.