Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MiniMax-H3-fp8
|
| 2 |
+
|
| 3 |
+
FP8 (float8_e4m3fn) quantized version of MiniMax H3, for use in ComfyUI.
|
| 4 |
+
|
| 5 |
+
Details
|
| 6 |
+
Format: weight-only scaled FP8 (float8_e4m3fn), per-layer weight_scale metadata — same contract ComfyUI uses natively for other fp8_scaled checkpoints.
|
| 7 |
+
Quantized: attention (qkv_proj, out_proj), MLP (fc1, fc2), and AdaLN modulation (adaln_proj) linear layers across the transformer blocks.
|
| 8 |
+
Kept in original precision: first/last 2 transformer blocks, normalization layers, time embedder, patch-embed projections, and output heads.
|
| 9 |
+
Not official: this is an unofficial conversion. MiniMax/Comfy-Org's own released quantization for this model is int8_convrot, not FP8.
|
| 10 |
+
Usage
|
| 11 |
+
|
| 12 |
+
Drop the .safetensors file into ComfyUI/models/diffusion_models/ and load it like the standard MiniMax H3 checkpoint.
|
| 13 |
+
|
| 14 |
+
Base Model
|
| 15 |
+
|
| 16 |
+
Original weights: MiniMaxAI/MiniMax-H3
|