Releases: dropbox/gemlite
Releases · dropbox/gemlite
v0.5.1.post1
Compare
Sorry, something went wrong.
No results found
Patch to fix A16W8 issues
v0.5.1
Compare
Sorry, something went wrong.
No results found
Various fixes related MXFP:
v0.5.0
Compare
Sorry, something went wrong.
No results found
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
Compare
Sorry, something went wrong.
No results found
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
Compare
Sorry, something went wrong.
No results found
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
v0.4.6
Compare
Sorry, something went wrong.
No results found
This release is mainly focusing on vllm V1 (torch.compile) support.
What's Changed
v0.4.5
Compare
Sorry, something went wrong.
No results found
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
Compare
Sorry, something went wrong.
No results found
v0.4.3
Compare
Sorry, something went wrong.
No results found
Add faster packing / unpacking utils
Set MIN_SIZE = 64 for Gemma 3
Update caches
v0.4.2.post1
Compare
Sorry, something went wrong.
No results found
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