Publish Cellpose-SAM WQuant W2A16-G64

README.md

@@ -0,0 +1,43 @@
+---
+library_name: cellpose
+pipeline_tag: image-segmentation
+base_model: mouseland/cellpose-sam
+base_model_relation: quantized
+license: bsd-3-clause
+tags:
+- cellpose
+- cellpose-sam
+- microscopy
+- quantization
+- image-segmentation
+---
+
+# Cellpose-SAM WQuant — W2A16-G64
+
+Ready-to-load `W2A16-G64` quantized variant of
+[Cellpose-SAM cpsam_v2](https://huggingface.co/mouseland/cellpose-sam).
+
+## Load
+
+```python
+import runpy
+from huggingface_hub import hf_hub_download
+
+loader = runpy.run_path(
+    hf_hub_download("capicu-ai/cellpose-sam-wquant-w2a16-g64", "runtime_loader.py")
+)
+model = loader["load_model"]()
+```
+
+Pass `device="cuda"` to `load_model` for CUDA.
+
+## Result
+
+- Weight size: 96.3 MiB
+- Compression versus FP32: 12.08x
+- Evaluated paired samples: 20
+- Complete bounded-panel criterion met: No
+- Reload verification normalized RMSE: 0.0
+
+`results.json` and `model_registry.json` contain the compact machine-readable
+record.

config.json

@@ -0,0 +1,1921 @@
+{
+  "schema_version": 1,
+  "base_model": "cpsam_v2",
+  "scheme": "W2A16-G64",
+  "replacements": [
+    {
+      "activation_bits": null,
+      "kind": "conv2d",
+      "name": "encoder.patch_embed.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 196608,
+        "group_size": 64,
+        "inner": 192,
+        "n_groups": 3,
+        "original_shape": [
+          1024,
+          3,
+          8,
+          8
+        ],
+        "padded_inner": 192,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.0.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.0.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.0.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.0.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.1.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.1.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.1.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.1.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.2.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.2.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.2.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.2.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.3.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.3.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.3.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.3.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.4.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.4.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.4.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.4.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.5.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.5.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.5.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.5.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.6.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.6.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.6.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.6.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.7.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.7.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.7.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.7.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.8.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.8.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.8.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.8.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.9.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.9.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.9.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.9.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.10.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.10.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.10.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.10.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.11.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.11.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.11.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.11.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.12.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.12.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.12.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.12.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.13.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.13.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.13.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.13.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.14.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.14.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.14.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.14.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.15.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.15.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.15.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.15.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.16.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.16.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.16.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.16.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.17.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.17.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.17.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.17.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.18.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.18.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.18.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.18.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.19.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.19.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.19.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.19.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.20.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.20.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.20.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.20.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.21.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.21.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.21.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.21.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.22.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.22.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.22.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.22.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.23.attn.qkv",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 3145728,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          3072,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 3072
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.23.attn.proj",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 1048576,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          1024,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.23.mlp.lin1",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          4096,
+          1024
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 4096
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "linear",
+      "name": "encoder.blocks.23.mlp.lin2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 4194304,
+        "group_size": 64,
+        "inner": 4096,
+        "n_groups": 64,
+        "original_shape": [
+          1024,
+          4096
+        ],
+        "padded_inner": 4096,
+        "qmax": 1,
+        "rows": 1024
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "conv2d",
+      "name": "encoder.neck.0",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 262144,
+        "group_size": 64,
+        "inner": 1024,
+        "n_groups": 16,
+        "original_shape": [
+          256,
+          1024,
+          1,
+          1
+        ],
+        "padded_inner": 1024,
+        "qmax": 1,
+        "rows": 256
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "conv2d",
+      "name": "encoder.neck.2",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 589824,
+        "group_size": 64,
+        "inner": 2304,
+        "n_groups": 36,
+        "original_shape": [
+          256,
+          256,
+          3,
+          3
+        ],
+        "padded_inner": 2304,
+        "qmax": 1,
+        "rows": 256
+      }
+    },
+    {
+      "activation_bits": null,
+      "kind": "conv2d",
+      "name": "out",
+      "quant_state": {
+        "bits": 2,
+        "code_count": 49152,
+        "group_size": 64,
+        "inner": 256,
+        "n_groups": 4,
+        "original_shape": [
+          192,
+          256,
+          1,
+          1
+        ],
+        "padded_inner": 256,
+        "qmax": 1,
+        "rows": 192
+      }
+    }
+  ],
+  "weights": {
+    "bytes": 100926616,
+    "file": "model.safetensors",
+    "sha256": "edae03a6cacce60ad6e3078a50febdd306ffb3f717041d8c059c373cf236d6d2",
+    "tensor_count": 450
+  }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edae03a6cacce60ad6e3078a50febdd306ffb3f717041d8c059c373cf236d6d2
+size 100926616

model_registry.json

@@ -0,0 +1,9 @@
+{
+  "scheme": "W2A16-G64",
+  "path": "models/W2A16-G64",
+  "weights_bytes": 100926616,
+  "sha256": "edae03a6cacce60ad6e3078a50febdd306ffb3f717041d8c059c373cf236d6d2",
+  "verified": true,
+  "verification_normalized_rmse": 0.0,
+  "public_output_files": 20
+}

requirements.txt

@@ -0,0 +1,4 @@
+cellpose>=4.2.1.1
+huggingface_hub>=0.24
+safetensors>=0.4
+torch

results.json

@@ -0,0 +1,13 @@
+{
+  "all_datasets_covered": true,
+  "paired_samples": 20,
+  "f1_CIs_above_margin": false,
+  "all_strict_endpoint_CIs_above_margin": false,
+  "catastrophic_rate": 0.85,
+  "bounded_panel_claim_supported": false,
+  "replaced_operators": 100.0,
+  "model_storage_mib": 96.20679473876953,
+  "fp32_storage_mib": 1162.0671463012695,
+  "compression_ratio": 12.078846919873303,
+  "build_seconds": 0.35267456097062677
+}

runtime_loader.py

@@ -0,0 +1,169 @@
+"""Runtime-only loader for Capicu's quantized Cellpose-SAM exports."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+from huggingface_hub import snapshot_download
+from safetensors.torch import load_file
+from torch import nn
+
+REPO_ID = "capicu-ai/cellpose-sam-wquant-w2a16-g64"
+
+
+def _digest(path: Path) -> str:
+    value = hashlib.sha256()
+    with path.open("rb") as handle:
+        for block in iter(lambda: handle.read(8 * 1024 * 1024), b""):
+            value.update(block)
+    return value.hexdigest()
+
+
+def _restore(packed: torch.Tensor, scales: torch.Tensor, state: dict, device):
+    bits = int(state["bits"])
+    per_byte = 8 // bits
+    packed = packed.to(device=device, dtype=torch.uint8).flatten()
+    mask = (1 << bits) - 1
+    codes = torch.stack(
+        [(packed >> (index * bits)) & mask for index in range(per_byte)],
+        dim=1,
+    ).flatten()[: int(state["code_count"])]
+    signed = codes.to(torch.int16) - int(state["qmax"])
+    grouped = signed.reshape(
+        int(state["rows"]),
+        int(state["n_groups"]),
+        int(state["group_size"]),
+    )
+    restored = grouped.float() * scales.to(device).unsqueeze(-1)
+    restored = restored.reshape(int(state["rows"]), int(state["padded_inner"]))
+    return restored[..., : int(state["inner"])].reshape(state["original_shape"])
+
+
+def _activation(tensor: torch.Tensor, bits: int | None) -> torch.Tensor:
+    if bits is None:
+        return tensor
+    qmax = (1 << (bits - 1)) - 1
+    scale = tensor.detach().abs().amax().clamp_min(1e-12) / qmax
+    return torch.round(tensor / scale).clamp(-qmax, qmax) * scale
+
+
+class _Packed(nn.Module):
+    def _setup(self, packed, scales, bias, state, activation_bits):
+        self.register_buffer("packed_weight", packed.cpu())
+        self.register_buffer("weight_scales", scales.cpu())
+        self.register_buffer("bias", None if bias is None else bias.cpu())
+        self.state = state
+        self.activation_bits = activation_bits
+
+    def _weight(self, tensor):
+        return _restore(
+            self.packed_weight,
+            self.weight_scales,
+            self.state,
+            tensor.device,
+        ).to(dtype=tensor.dtype)
+
+
+class _Linear(_Packed):
+    def __init__(self, source, **payload):
+        super().__init__()
+        self.in_features = source.in_features
+        self.out_features = source.out_features
+        self._setup(**payload)
+
+    def forward(self, tensor):
+        tensor = _activation(tensor, self.activation_bits)
+        bias = None if self.bias is None else self.bias.to(tensor.device, tensor.dtype)
+        return F.linear(tensor, self._weight(tensor), bias)
+
+
+class _Conv2d(_Packed):
+    def __init__(self, source, **payload):
+        super().__init__()
+        self.stride = source.stride
+        self.padding = source.padding
+        self.dilation = source.dilation
+        self.groups = source.groups
+        self._setup(**payload)
+
+    def forward(self, tensor):
+        tensor = _activation(tensor, self.activation_bits)
+        bias = None if self.bias is None else self.bias.to(tensor.device, tensor.dtype)
+        return F.conv2d(
+            tensor,
+            self._weight(tensor),
+            bias,
+            self.stride,
+            self.padding,
+            self.dilation,
+            self.groups,
+        )
+
+
+def _replace(root: nn.Module, name: str, module: nn.Module) -> None:
+    parent_name, _, child_name = name.rpartition(".")
+    parent = root.get_submodule(parent_name) if parent_name else root
+    if child_name.isdigit() and isinstance(parent, (nn.Sequential, nn.ModuleList)):
+        parent[int(child_name)] = module
+    else:
+        setattr(parent, child_name, module)
+
+
+def load_model(
+    device: str | torch.device = "cpu",
+    local_repo: str | Path | None = None,
+):
+    """Download and load the ready-to-run quantized Cellpose-SAM model."""
+    root = (
+        Path(local_repo)
+        if local_repo is not None
+        else Path(
+            snapshot_download(
+                REPO_ID,
+                allow_patterns=[
+                    "config.json",
+                    "model.safetensors",
+                ],
+            )
+        )
+    )
+    manifest = json.loads((root / "config.json").read_text())
+    weights_path = root / "model.safetensors"
+    if _digest(weights_path) != manifest["weights"]["sha256"]:
+        raise ValueError("model checksum mismatch")
+
+    from cellpose import models as cellpose_models
+
+    target = torch.device(device)
+    model = cellpose_models.CellposeModel(
+        gpu=target.type == "cuda",
+        pretrained_model=manifest["base_model"],
+        device=target,
+        use_bfloat16=False,
+    )
+    state = load_file(str(weights_path), device="cpu")
+    network = model.net.cpu().eval()
+    for item in manifest["replacements"]:
+        name = item["name"]
+        source = network.get_submodule(name)
+        prefix = f"{name}."
+        payload = {
+            "packed": state[f"{prefix}packed_weight"],
+            "scales": state[f"{prefix}weight_scales"],
+            "bias": state.get(f"{prefix}bias"),
+            "state": item["quant_state"],
+            "activation_bits": item.get("activation_bits"),
+        }
+        replacement = (
+            _Linear(source, **payload)
+            if item["kind"] == "linear"
+            else _Conv2d(source, **payload)
+        )
+        _replace(network, name, replacement)
+    network.load_state_dict(state, strict=True)
+    model.net = network.to(target).eval()
+    return model