docs: checkpoint verification
Browse files- reports/quantization.md +10 -0
reports/quantization.md
CHANGED
|
@@ -1,5 +1,15 @@
|
|
| 1 |
# FlowHigh AX650 quantization report
|
| 2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
## Target
|
| 4 |
|
| 5 |
- Board: AX650N, NPU3, axengine 2.12.0s, compiler 8.0.
|
|
|
|
| 1 |
# FlowHigh AX650 quantization report
|
| 2 |
|
| 3 |
+
## Checkpoint
|
| 4 |
+
|
| 5 |
+
- Model: `FLowHigh_indep_adaptive_400k`
|
| 6 |
+
- Checkpoint size: 481,490,826 bytes (460 MiB), includes optimizer state.
|
| 7 |
+
- Weights: Flow 35.4M + BigVGAN 14.0M = 49.4M parameters (FP32 weights about 198 MB).
|
| 8 |
+
- Source: official Google Drive folder from the FLowHigh README; local copy from the HF mirror `kiriyamaX/FLowHigh_indep_adaptive_400k`.
|
| 9 |
+
- The 38 MB Flow axmodel is the S8-quantized 35.4M-parameter flow transformer; the 20 MB BigVGAN axmodel is separate.
|
| 10 |
+
|
| 11 |
+
Official demo `p364_055` 8 kHz: our PyTorch output vs the official Euler output has mel LSD 0.411 and waveform corr 0.998; the board output vs official has corr 0.997, mel LSD 0.515. This confirms the checkpoint/pipeline, not a wrong download.
|
| 12 |
+
|
| 13 |
## Target
|
| 14 |
|
| 15 |
- Board: AX650N, NPU3, axengine 2.12.0s, compiler 8.0.
|