sgyaqing commited on
Commit
9d4dd73
·
verified ·
1 Parent(s): ab131fe

27b tier: state the measurement condition first, then the weights-vs-graph-pools memory reason

Browse files
Files changed (1) hide show
  1. README.md +1 -1
README.md CHANGED
@@ -122,7 +122,7 @@ Measured on one RTX 4090 (48 GB) for `0.8b`/`4b` and one NVIDIA A100-SXM4-80GB f
122
 
123
  ### 27b (A100 80 GB)
124
 
125
- Measured with CUDA graphs disabled (`VERDICT_TORCH_GRAPHS=0`); with graphs enabled the static graph pools take the card to 98% and leave too little memory for batching — the image and video cells at c=8 and c=64 drop 61–72%.
126
 
127
  <table>
128
  <thead>
 
122
 
123
  ### 27b (A100 80 GB)
124
 
125
+ Measured with CUDA graphs disabled (`VERDICT_TORCH_GRAPHS=0`): the bf16 weights alone take most of the 80 GB card, leaving no room for the graph pools.
126
 
127
  <table>
128
  <thead>