Remove legacy FC2 sidecar
Browse filesThe full input-major transformer supersedes the short-lived auxiliary layout. Keep only the original output-major checkpoint for generic runtimes and the full input-major checkpoint for H3ddle.
README.md
CHANGED
|
@@ -48,9 +48,8 @@ ConvRot metadata, reference conditioning, and non-core tensors are preserved.
|
|
| 48 |
This layout is selected by a versioned marker and validated against every
|
| 49 |
projection shape before H3ddle runs it. It is **not** compatible with runtimes
|
| 50 |
that assume the original output-major Comfy layout. The original transformer
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
checkpoint.
|
| 54 |
|
| 55 |
On a 32 GiB M1 Pro, a matched 512x896 reference-conditioned, eight-pass
|
| 56 |
generation reduced complete transformer time from 541.2 to 515.3 seconds:
|
|
@@ -142,7 +141,6 @@ of the MiniMax H3 Community License apply unchanged.
|
|
| 142 |
|---|---|---|
|
| 143 |
| `minimax_h3_ref2va_pruned_turbo_int8_convrot.safetensors` | 20,970,379,854 | `e64cef63bc2785bcd72e6103c52aa78c6cd2c4f9870a7ce79675083fd65cf2e7` |
|
| 144 |
| `minimax_h3_ref2va_pruned_turbo_int8_convrot_input_major.safetensors` | 20,970,380,012 | `5ca6696fe1cd9a8f254594ac67ee541f151b2377735dea3557364bd868270463` |
|
| 145 |
-
| `minimax_h3_ref2va_pruned_turbo_int8_convrot_fc2_input_major.safetensors` | 3,853,522,260 | `0ad6a5673abdf842c39d4d8de7c34c971a420b64bd5f79eb6f4331c5bfb5cd97` |
|
| 146 |
|
| 147 |
## Reproducibility references
|
| 148 |
|
|
@@ -153,8 +151,6 @@ the pipeline against the official file.
|
|
| 153 |
|
| 154 |
The full optimized layout is reproducible with
|
| 155 |
[`Scripts/repack-h3-input-major.py`](https://github.com/AlexanderIstomin/h3ddle/blob/main/Scripts/repack-h3-input-major.py).
|
| 156 |
-
The legacy FC2 sidecar is reproducible with
|
| 157 |
-
[`Scripts/optimize-h3-fc2-sidecar.py`](https://github.com/AlexanderIstomin/h3ddle/blob/main/Scripts/optimize-h3-fc2-sidecar.py).
|
| 158 |
|
| 159 |
## Contact
|
| 160 |
|
|
|
|
| 48 |
This layout is selected by a versioned marker and validated against every
|
| 49 |
projection shape before H3ddle runs it. It is **not** compatible with runtimes
|
| 50 |
that assume the original output-major Comfy layout. The original transformer
|
| 51 |
+
remains in this repository for those runtimes; current H3ddle managed installs
|
| 52 |
+
download only the full input-major checkpoint.
|
|
|
|
| 53 |
|
| 54 |
On a 32 GiB M1 Pro, a matched 512x896 reference-conditioned, eight-pass
|
| 55 |
generation reduced complete transformer time from 541.2 to 515.3 seconds:
|
|
|
|
| 141 |
|---|---|---|
|
| 142 |
| `minimax_h3_ref2va_pruned_turbo_int8_convrot.safetensors` | 20,970,379,854 | `e64cef63bc2785bcd72e6103c52aa78c6cd2c4f9870a7ce79675083fd65cf2e7` |
|
| 143 |
| `minimax_h3_ref2va_pruned_turbo_int8_convrot_input_major.safetensors` | 20,970,380,012 | `5ca6696fe1cd9a8f254594ac67ee541f151b2377735dea3557364bd868270463` |
|
|
|
|
| 144 |
|
| 145 |
## Reproducibility references
|
| 146 |
|
|
|
|
| 151 |
|
| 152 |
The full optimized layout is reproducible with
|
| 153 |
[`Scripts/repack-h3-input-major.py`](https://github.com/AlexanderIstomin/h3ddle/blob/main/Scripts/repack-h3-input-major.py).
|
|
|
|
|
|
|
| 154 |
|
| 155 |
## Contact
|
| 156 |
|
minimax_h3_ref2va_pruned_turbo_int8_convrot_fc2_input_major.safetensors
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:0ad6a5673abdf842c39d4d8de7c34c971a420b64bd5f79eb6f4331c5bfb5cd97
|
| 3 |
-
size 3853522260
|
|
|
|
|
|
|
|
|
|
|
|