Upload ModelOpt quantized FLUX.2 Klein 4B transformer variants

README.md

@@ -0,0 +1,41 @@
+---
+tags:
+- flux
+- quantization
+- modelopt
+- fp8
+- nvfp4
+- int8
+license: apache-2.0
+---
+
+# Quantized FLUX.2 Klein 4B Transformer (ModelOpt)
+
+This repo stores NVIDIA Model Optimizer checkpoints for FLUX.2 Klein 4B transformer quantization variants.
+
+## Contents
+
+- `fp8/transformer_modelopt.pt`
+- `fp8/transformer_modelopt_meta.json`
+- `w8a8/transformer_modelopt.pt`
+- `w8a8/transformer_modelopt_meta.json`
+- `nvfp4/transformer_modelopt.pt`
+- `nvfp4/transformer_modelopt_meta.json`
+
+## Restore into pipeline
+
+```python
+import modelopt.torch.opt as mto
+from klein_pipeline import Flux2KleinPipeline
+import torch
+
+pipe = Flux2KleinPipeline.from_pretrained(
+    "black-forest-labs/FLUX.2-klein-4B", torch_dtype=torch.bfloat16
+).to("cuda")
+
+ckpt = "fp8/transformer_modelopt.pt"  # or w8a8 / nvfp4
+mto.restore(pipe.transformer, ckpt)
+pipe.transformer.eval()
+```
+
+Uploaded from Modal volume `klein4B-assets`.

fp8/transformer_modelopt.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d001d1c3d47729a4f954488be91ed1c2e6d1900ccffb68689b5d599ad761c19d
+size 7751397212

fp8/transformer_modelopt_meta.json

@@ -0,0 +1,18 @@
+{
+  "model_id": "/models/FLUX.2-klein-4B",
+  "component": "transformer",
+  "backend": "nvidia-modelopt",
+  "config": "FP8_DEFAULT_CFG",
+  "variant": "fp8",
+  "dtype": "torch.bfloat16",
+  "calibration": {
+    "iters": 8,
+    "batch_size": 1,
+    "image_path": "/models/calib/blue_car_resize.jpeg",
+    "steps": 4,
+    "height": 576,
+    "width": 384,
+    "guidance_scale": 4.0
+  },
+  "version": 1
+}

nvfp4/transformer_modelopt.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d88a854fb402456cf1337c66b1b3c8714651c9bc8e53952fa063b67908f81b37
+size 7751397788

nvfp4/transformer_modelopt_meta.json

@@ -0,0 +1,18 @@
+{
+  "model_id": "/models/FLUX.2-klein-4B",
+  "component": "transformer",
+  "backend": "nvidia-modelopt",
+  "config": "NVFP4_DEFAULT_CFG",
+  "variant": "nvfp4",
+  "dtype": "torch.bfloat16",
+  "calibration": {
+    "iters": 8,
+    "batch_size": 1,
+    "image_path": "/models/calib/blue_car_resize.jpeg",
+    "steps": 4,
+    "height": 576,
+    "width": 384,
+    "guidance_scale": 4.0
+  },
+  "version": 1
+}

w8a8/transformer_modelopt.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddd72bcef95822f5101e64348c96d1bf1bb64f5028322ec1e40922519e95e4e5
+size 7756973568

w8a8/transformer_modelopt_meta.json

@@ -0,0 +1,18 @@
+{
+  "model_id": "/models/FLUX.2-klein-4B",
+  "component": "transformer",
+  "backend": "nvidia-modelopt",
+  "config": "INT8_SMOOTHQUANT_CFG",
+  "variant": "w8a8",
+  "dtype": "torch.bfloat16",
+  "calibration": {
+    "iters": 8,
+    "batch_size": 1,
+    "image_path": "/models/calib/blue_car_resize.jpeg",
+    "steps": 4,
+    "height": 576,
+    "width": 384,
+    "guidance_scale": 4.0
+  },
+  "version": 1
+}