v0.5.1.post1

Patch to fix A16W8 issues

v0.5.1

Various fixes related MXFP:

• Fix FP8 quantization range by @mobicham in #38
• MXFP4/NVFP4 update by @mobicham in #40
  Full Changelog: 0.5.0...0.5.1

v0.5.0

• Rocm integration by @mobicham in #35
• Improve memory usage by @mobicham in #36
• Add MXFP support to gemlite by @mobicham in #37
• Faster mxpf8 activation quantization
• Improve helper
• Bug fixes

v0.4.8

• Improve and clean-up helper.py to make it compatible with torchao/vllm.
• Restore ptrs in autotune for gemvs that was causing the first forward pass to be incorrect.
• Disable fp8 rounding.
• Update caches.

v0.4.7

This release is mainly focusing on improving performance:

• Faster GEMVs via fp16 acc and output caching.
• Better GEMM/Split-K performance with improved autotuning.
• Faster autotuning mode to avoid long startup time.

What's Changed

• Improved Perf by @mobicham in #34
  Full Changelog: 0.4.6...0.4.7

v0.4.6

This release is mainly focusing on vllm V1 (torch.compile) support.

What's Changed

• Add support for vllm compile by @mobicham in #32
  Full Changelog: 0.4.5...0.4.6

v0.4.5

• Update caches for 48GB gpus (Qwen2 VL/Llama3 8B)
• Add cpu-side packing
• Relax min size to 32
• fp16 acc fix
• add persistent SPLIT_K version
• fix tl.contiguous hint
• make m,n block sizes safe
• add BitNet support in helper
• add custom load_state_dict to allow weight serialization
• Update swizzle

v0.4.4

• Add bfloat16 support to gemlite kernels by @mobicham in #24

v0.4.3

• Add faster packing / unpacking utils
• Set MIN_SIZE = 64 for Gemma 3
• Update caches

v0.4.2.post1

• Avoid recompilation when the batch-size M changes: dcc2455
• Expose autotune M logic via set_autotune_setting(): 37dab27
• Fix bug related to config caching that was ignoring the pre-loaded cache: 3c4ab53