Oct 5 20 days ago

1. 10:55

   Table to RAM with 64 MB sequential reads (page-faulting the mmap over the mount was far too slow)

   logic65
   Verified signature

   e846b6b
2. 10:30

   Speed: grouped_mm experts inside the GPU call, table to RAM after packing, load index without table pieces, timing logs

   logic65
   Verified signature

   8792537
3. 10:08

   Log cgroup memory (anon/file/shmem) during model load

   logic65
   Verified signature

   7aefae7
4. 10:01

   Memory-map the n-gram table (ZeroGPU host memory limit is 104 GB; table + body overflowed it)

   logic65
   Verified signature

   b8a7b4e
5. 09:50

   Small files from the Hub, shards from the mount (size-checked); drop flash-linear-attention (no GPU at import on ZeroGPU)

   logic65
   Verified signature

   d3c524a
6. 09:40

   Whittle 45B ZeroGPU chat demo: stock transformers + whittle_load.py, model mounted at /models/w45

   logic65
   Verified signature

   35b139d
7. 09:40

   initial commit

   logic65
   Verified signature

   5598ac2