philtheriver commited on
Commit
4d438af
·
verified ·
1 Parent(s): 0efea2d

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +2 -3
README.md CHANGED
@@ -105,9 +105,8 @@ build-strix-rocmfp4/bin/llama-server \
105
 
106
  | Metric | Value |
107
  |--------|-------|
108
- | PPL vs `Q6_K` (code corpus) | _TODO — `scripts/sweep-rocmfpx-perplexity.sh`_ |
109
- | KL-divergence vs `Q6_K` | _TODO_ |
110
- | MTP draft acceptance (`--spec-draft-n-max 2`) | _TODO — `scripts/check-rocmfpx-mtp-smoke.sh`_ |
111
 
112
  > Reference: source BF16 perplexity on the calibration corpus measured during imatrix = **3.2374 ± 0.018**.
113
  > All numbers are hardware/driver/prompt-sensitive — reproduce before relying.
 
105
 
106
  | Metric | Value |
107
  |--------|-------|
108
+ | MTP decode (`draft-mtp`, `--spec-draft-n-max 2`) | ✅ active — ~18.6 tok/s decode on gfx1151 |
109
+ | PPL / KL vs `Q6_K` (code corpus) | not yet measured — optional `scripts/sweep-rocmfpx-perplexity.sh` |
 
110
 
111
  > Reference: source BF16 perplexity on the calibration corpus measured during imatrix = **3.2374 ± 0.018**.
112
  > All numbers are hardware/driver/prompt-sensitive — reproduce before relying.