kierandidi commited on
Commit
b49852c
·
verified ·
1 Parent(s): 13e6983

Document validated A100 and L40S AOT profile coverage

Browse files
Files changed (1) hide show
  1. README.md +12 -10
README.md CHANGED
@@ -73,14 +73,15 @@ are included for inspection.
73
 
74
  ## Compiled profiles
75
 
76
- `aoti/a100/` contains eight batch-1 profiles (128–1,024 tokens) built in the
77
- dedicated Space's GLIBC 2.35 runtime. `aoti/l40s/` contains the validated
78
- length and batch profiles built on Pasco L40S nodes (GLIBC 2.43); its manifest
79
- lists the available shapes. These are compiled
80
- trunk packages, not another copy of the weights. Their manifests pin the model,
81
- Torch, CUDA, GPU capability and minimum GLIBC; use a bundle only on a matching
82
- runtime. The dedicated Space uses the A100 bundle, while the ZeroGPU Space uses
83
- its separate Blackwell bundle.
 
84
 
85
  ## Files and provenance
86
 
@@ -103,8 +104,9 @@ collection with component-specific terms. See its card before using or redistrib
103
  any component.
104
 
105
  Generated structures require downstream scientific validation. The Spaces omit
106
- external refolding metrics. The three examples above were run on the dedicated
107
- A100 Space; larger batches and sustained load need separate qualification. No
 
108
  benchmark scores are claimed by this card.
109
 
110
  The RFD4 model and weights carry the repository's BSD-3-Clause license. Optional
 
73
 
74
  ## Compiled profiles
75
 
76
+ `aoti/a100/` contains 60 validated profiles: batches 1, 2, 4 and 8 across
77
+ 15 token buckets from 128 to 1,024. `aoti/l40s/` contains 58 profiles; batch 8
78
+ at 896 and 1,024 tokens exceeds the L40S's 48 GB memory. The loader splits
79
+ those requests into smaller compiled batches without falling back to eager
80
+ execution. Both bundles were built in an Ubuntu 22 runtime with GLIBC 2.35.
81
+ Their manifests pin the model, Torch, CUDA and GPU capability and list every
82
+ available shape. These are compiled trunk packages; model weights remain in
83
+ `model.safetensors`. The dedicated Space bundles the A100 profiles, while the
84
+ ZeroGPU Space uses a separate Blackwell bundle.
85
 
86
  ## Files and provenance
87
 
 
104
  any component.
105
 
106
  Generated structures require downstream scientific validation. The Spaces omit
107
+ external refolding metrics. The three examples above were run on the dedicated A100 Space. Compiler-disabled
108
+ native generation also produced two CIFs in one A100 compiled batch and four in
109
+ one L40S compiled batch. Sustained load needs separate qualification. No
110
  benchmark scores are claimed by this card.
111
 
112
  The RFD4 model and weights carry the repository's BSD-3-Clause license. Optional