Add Lens-Turbo-3.8B-8bit DiT weights + model card

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+sample.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -6,10 +6,31 @@ tags: [mlx, text-to-image, diffusion, lens, lens-turbo, apple-silicon]
 base_model: microsoft/Lens-Turbo
 ---
 
-# Lens-Turbo-3.8B-8bit (MLX) — coming soon
+# Lens-Turbo-3.8B-8bit (MLX)
 
-MLX conversion of [microsoft/Lens-Turbo](https://huggingface.co/microsoft/Lens-Turbo) — the
-distilled 4-step (cfg 1.0) sibling of Lens, identical 3.8B DiT architecture. ~4.39 GB. Weights imminent.
+Apple **MLX** conversion of [microsoft/Lens-Turbo](https://huggingface.co/microsoft/Lens-Turbo) —
+the **distilled 4-step** sibling of Lens (identical 3.8B DiT architecture; sample at **4 steps,
+guidance 1.0**). int8 (group_size 64), ~4.39 GB. DiT-only (MIT); the GPT-OSS-20B encoder (Apache-2.0) and FLUX.2 VAE
+load from source. Architecture is byte-identical to base Lens, which is parity-locked vs the PT
+reference (DiT cosine 0.999999); this variant inherits that port.
 
-Base (20-step) family: [Lens 3.8B (MLX) collection](https://huggingface.co/collections/mlx-community/lens-38b-mlx-6a1c6846ca63123d871450f1).
-Code: [xocialize-code/lens-mlx](https://github.com/xocialize-code/lens-mlx).
+![sample](sample.png)
+
+## Usage
+
+```python
+from lens_mlx.pipeline_mlx import LensPipeline   # github.com/xocialize-code/lens-mlx
+
+# `base` = a microsoft/Lens snapshot (tokenizer + GPT-OSS encoder + FLUX.2 VAE).
+pipe = LensPipeline.from_pretrained(base, dit_repo="mlx-community/Lens-Turbo-3.8B-8bit")
+img = pipe("A serene lake below snow-capped mountains, golden hour.",
+           height=1024, width=1024, num_inference_steps=4, guidance_scale=1.0, seed=42)
+img.save("out.png")
+```
+
+> Tip: page weights into memory before the first forward (`mx.eval` the params) when loading
+> from slow/external storage, to avoid a Metal command-buffer watchdog timeout at large sizes.
+
+## License
+DiT weights MIT (from microsoft/Lens-Turbo) · GPT-OSS-20B encoder Apache-2.0 (not re-hosted) ·
+FLUX.2 VAE under its own terms (not re-hosted). Upstream: [microsoft/Lens-Turbo](https://huggingface.co/microsoft/Lens-Turbo).

config.json

@@ -0,0 +1,38 @@
+{
+  "_class_name": "LensTransformer2DModel",
+  "_diffusers_version": "0.37.1",
+  "attention_head_dim": 64,
+  "axes_dims_rope": [
+    8,
+    28,
+    28
+  ],
+  "enc_hidden_dim": 2880,
+  "gate_mlp": true,
+  "in_channels": 128,
+  "inner_dim": 1536,
+  "multi_layer_encoder_feature": true,
+  "num_attention_heads": 24,
+  "num_layers": 48,
+  "out_channels": 32,
+  "patch_size": 2,
+  "rms_norm": true,
+  "selected_layer_index": [
+    5,
+    11,
+    17,
+    23
+  ],
+  "mlx_format": true,
+  "quantization": {
+    "group_size": 64,
+    "bits": 8,
+    "keep_hi_precision": [
+      "img_in",
+      "txt_in",
+      "proj_out",
+      "time_text_embed",
+      "norm_out"
+    ]
+  }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30e44edf95750cf241b3b47dbd086984d472c7c9900036d4545a1079aadb0ef2
+size 4386667927