Harsha001 commited on
Commit
9890076
·
verified ·
1 Parent(s): 18fb385

Card: stock comparison rows

Browse files
Files changed (1) hide show
  1. README.md +6 -0
README.md CHANGED
@@ -36,6 +36,12 @@ KL divergence against the original weights (Thinking Machines' NVFP4 experts kep
36
  | Original weights (reference, not published) | 22.9 GB | 6.743 | | |
37
  | **Q8_0_ROCMFPX** | 26.9 GB | 6.741 | 0.0039 | 97.2% |
38
  | **Q4_0_ROCMFP4_STRIX_O8** | 15.5 GB | 6.784 | 0.0120 | 95.4% |
 
 
 
 
 
 
39
 
40
  The Q8 file is larger than the original checkpoint because the original experts are 4-bit NVFP4. 8 bits keep them
41
  almost exactly; the 4-bit ROCmFP4 file is the size to pick when memory is tight.
 
36
  | Original weights (reference, not published) | 22.9 GB | 6.743 | | |
37
  | **Q8_0_ROCMFPX** | 26.9 GB | 6.741 | 0.0039 | 97.2% |
38
  | **Q4_0_ROCMFP4_STRIX_O8** | 15.5 GB | 6.784 | 0.0120 | 95.4% |
39
+ | stock Q8_0 (comparison, not uploaded) | 27.7 GB | 6.739 | 0.0035 | 97.5% |
40
+ | stock Q4_K_M (comparison, not uploaded) | 17.4 GB | 6.819 | 0.0174 | 94.1% |
41
+
42
+ At 4 bits the ROCmFP4 file is smaller than stock Q4_K_M and closer to the original on every measure. At 8 bits stock
43
+ Q8_0 is a little closer to the original (0.0035 against 0.0039) and 0.8 GB larger; the ROCmFPX file is the one built for
44
+ the AMD kernels.
45
 
46
  The Q8 file is larger than the original checkpoint because the original experts are 4-bit NVFP4. 8 bits keep them
47
  almost exactly; the 4-bit ROCmFP4 file is the size to pick when memory is tight.