v5.14.0

🐛 Bug Fixes

• Fixed GPTBigCode and GPTNeoX for the Transformers modelling backend for vLLM.
• Pinned the `kernels` dependency to a compatible version in the benchmark workflow.
• Made the DeepGEMM Triton fallback more robust when `CUDA_HOME` is unset or misconfigured.
• Updated SDPA prefill to leverage the FlashAttention kernel with `StaticCache` for performance gains.
• Fixed a misleading double-negative warning message for `synced_gpus` in continuous batching mode.
• Fixed crashes in greedy assisted generation with different tokenizers.
• Fixed a Flash Attention performance regression affecting models like Qwen3-VL.
• Resolved a MoE decode optimization bug where the grouped-to-batched matrix multiplication switch was not applied to experts residing in submodels.
• Simplified cache dispatch logic by introducing explicit layer-type mappings for sliding and static layers.
• Fixed read-only cache failures in CPU CI environments.
• Fixed MPS graph cache growth during variable-length batch training on Apple Silicon.
• Covered xet in CI for runtime errors.
• Added TokenizersBackend fallback.
• Resolved continuous batching XPU availability checks at runtime.
• Added `kernels_fallback_ok` kwarg to `is_flash_attn_N_available`.
• Added expectations for gemma4 tests on H100.
• DeepGEMM shouldn't pad on SM90.
• Fixed half-precision torch.compile crash in DETR-family sine position embeddings.
• Fixed hardcoded paths in siglip checkpoint/vocab loading.
• Updated AMD CI runner groups to amd-mi300.
• Pointed to Gemma 4 model in Gemma4ForCausalLM docstring example.
• Fixed Qwen Omni batched text postprocessing.
• Fixed AqlmConfig error messages to say "int" instead of "float".
• Fixed check for interactive stdout in `_style` function.
• Fixed `get_json_schema` crash on non-string docstring choices.
• Made `MODEL_IDS_TO_TOKENIZERS_BACKEND` capture all DeepSeek R1 distills.
• Updated doc preprocessing regex to prevent ReDoS.
• Shard on read Dtensor aware.
• Switched AMD daily CI to mi300 runners.
• Reduced processor test memory usage by using tiny Hub checkpoints.
• Torch compile backend defaults to "neuron".
• Fixed flash-attn Docker build broken by setuptools 83 removing pkg_resources.
• Fixed integration test values.
• Fixed DeepSpeed SP loss aggregation and LocalLayerRepository kwargs.
• Ran tests only for the top 10 download models.
• Fixed InputTokensDetails missing cache_write_tokens for openai>=2.34.0.
• Reverted "Trigger a scheduled run".
• Removed Executorch from CI until latest version is supported and fully tested on CI env.
• Became more defensive with `remap_legacy_layer_types` for custom models.
• Fixed DistributedConfig docstring for unimplemented sp_plan.
• Switched mlinter to 0.1.2.