For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo module

per_channel_grouped_4bit

Provides per-channel grouped 4-bit quantization and K-quant dequantization kernels.

Structs

• ​block_Q4_K: Represents a single Q4_K quantization block with grouped 4-bit values.
• ​block_Q6_K: Represents a single Q6_K quantization block with grouped 6-bit values.
• ​block_QK_K: Defines the element count per block shared by K-quant formats.
• ​Q4sym: Q4sym: compresses values of type float_dtype to 4bit unsigned integers which have been dynamically symmetrically quantized with the given scale factor.

Functions

• ​calculate_symmetric_vector: Symmetrically quantizes the given SIMD vector data with input type input_dtype and simd_width elements, assuming we want the results to fit in an unsigned integer of size output_bits.
• ​q4_k_dequantize_impl: Dequantizes a Q4_K encoded tensor into a float32 output tensor.
• ​q6_k_dequantize_impl: Dequantizes a Q6_K encoded tensor into a float32 output tensor.
• ​scale_min_k4: Extracts the 6-bit quantized scale and min for group g from a Q4_K block.

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit