yitongl commited on
Commit
6dd6b86
·
verified ·
1 Parent(s): 1d0c0cc

Document standalone inference helper

Browse files
Files changed (1) hide show
  1. README.md +7 -0
README.md CHANGED
@@ -10,6 +10,7 @@ Contents:
10
  - `transformer/config.json`
11
  - `transformer/diffusion_pytorch_model.safetensors`
12
  - `backend_snapshot/`
 
13
 
14
  Training run:
15
 
@@ -32,6 +33,7 @@ It also includes the inference entrypoint snapshot and an example script:
32
 
33
  - `backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_700.sh`
34
  - `backend_snapshot/training_attention_settings.json`
 
35
 
36
  Attention setup for this checkpoint:
37
 
@@ -39,3 +41,8 @@ Attention setup for this checkpoint:
39
  tiles, group-local P quant, dropped-tile mean compensation
40
  - cross-attention: dense SDPA fallback, not FP4/sparse
41
  - force-dense paths: dense SDPA
 
 
 
 
 
 
10
  - `transformer/config.json`
11
  - `transformer/diffusion_pytorch_model.safetensors`
12
  - `backend_snapshot/`
13
+ - `standalone_inference/`
14
 
15
  Training run:
16
 
 
33
 
34
  - `backend_snapshot/scripts/inference/run_sfp4_ours_p_checkpoint_700.sh`
35
  - `backend_snapshot/training_attention_settings.json`
36
+ - `standalone_inference/`
37
 
38
  Attention setup for this checkpoint:
39
 
 
41
  tiles, group-local P quant, dropped-tile mean compensation
42
  - cross-attention: dense SDPA fallback, not FP4/sparse
43
  - force-dense paths: dense SDPA
44
+
45
+ `standalone_inference/` is a portable helper for normal inference. It contains
46
+ an overlay installer, a runner that downloads/loads the checkpoint-700
47
+ transformer weights, and the sparse FP4 backend files required by this
48
+ checkpoint.