candle-core

• Add a scattered kv cache.

• Update some comments.

Bump the crate version to 0.9.0. (#2924)

• Retrieve the current positions for rotating KV caches.

• Add the function to the kv cache too.

• More testing.

Bump the crate version. (#2895)

Added

Modified

Added

• Added the Mistral 7b v0.1 model 983.
• Quantized version of the Mistral model 1009.
• Add the gelu-erf op and activation function 969.
• Add the mixformer/phi-v1.5 model 930.
• Add the sclice-scatter op 927.
• Add the Wuerstchen diffusion model 911.

Modified

• Support for simd128 intrinsics in some quantized vecdots 982.
• Optimize the index-select cuda kernel 976.
• Self-contained safetensor wrappers 946.

Added

• Support for top_p sampling 819.
• T5 model including decoding 864.
• 1-d upsampling 839.

Modified

• Bugfix for conv2d 820.
• Support tensor based indexing using .i 842.

Added

• Add some RNNs (GRU and LSTM) in candle-nn 674, 688.
• gguf v2 support 725.
• Quantized llama example in Python using the pyo3 api 716.
• candle-nn layer for conv2d-transposed 760.
• Add the Segment-Anything Model (SAM) as an example 773.
• TinyViT backbone for the segment anything example 787.
• Shape with holes support 770.

Modified

• Dilations are now supported in conv-transpose2d. 671.
• Interactive mode for the quantized model 690.
• Faster softmax operation 747.
• Faster convolution operations on CPU and CUDA via im2col 802.
• Moving some models to a more central location 796.

Added

• Add the powf op 664.
• Stable Diffusion XL support 647.
• Add the conv-transpose2d op 635.
• Refactor the VarBuilder api 627.
• Add some quantization command 625.
• Support more quantized types, e.g. Q2K, Q4K, Q5K... 586.
• Add pose estimation to the yolo example 589.
• Api to write GGUF files 585.
• Support more quantization types 580.
• Add EfficientNet as an example Computer Vision model 572.
• Add a group parameter to convolutions 566.
• New dtype: int64 563.
• Handling of the GGUF file format. 559.