Upload folder using huggingface_hub

model-00001-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe3791a5f2f0773e38adf0c8d1da47e2f73a7f391b87427af2c13840660c38fa
+size 4840397588

model-00002-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:486db20e6ae740b78145eaacdd773a9f412c8cf2d218c22b53fd1b73a4ad685d
+size 4838824240

model-00003-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfdf23318e2bb71f62222dcd9e4916b96e802b869a62f80c691acca9e5a1b23a
+size 4838824296

model-00004-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3fa234ad7f37850b16954b2add5b5a766e74a3e6ade87da851dbb9459a71065
+size 4838824296

model-00005-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a59d7c1f22faa38f0c00ca0df5fa00bbdb3a45fa2439a9b7dcbcea04b7cc7fe
+size 4838824296

model-00006-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36469f8d727a9017afce8172c25852ae6e996ac9c8f4c7b68b9b712826c13652
+size 1692142452

model.safetensors.index.json

@@ -0,0 +1,338 @@
+{
+  "metadata": {
+    "total_size": 25887801592
+  },
+  "weight_map": {
+    "lm_head.bias": "model-00006-of-00006.safetensors",
+    "lm_head.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.0.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.0.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.0.attn.out_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.0.attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.0.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.0.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.0.mlp.fc_in.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.0.mlp.fc_in.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.0.mlp.fc_out.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.0.mlp.fc_out.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.1.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.1.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.1.attn.out_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.1.attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.1.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.1.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.1.mlp.fc_in.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.1.mlp.fc_in.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.1.mlp.fc_out.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.1.mlp.fc_out.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.10.attn.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.10.attn.masked_bias": "model-00002-of-00006.safetensors",
+    "transformer.h.10.attn.out_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.10.attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.10.ln_1.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.10.ln_1.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.10.mlp.fc_in.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.10.mlp.fc_in.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.10.mlp.fc_out.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.10.mlp.fc_out.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.11.attn.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.11.attn.masked_bias": "model-00002-of-00006.safetensors",
+    "transformer.h.11.attn.out_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.11.attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.11.ln_1.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.11.ln_1.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.11.mlp.fc_in.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.11.mlp.fc_in.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.11.mlp.fc_out.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.11.mlp.fc_out.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.12.attn.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.12.attn.masked_bias": "model-00002-of-00006.safetensors",
+    "transformer.h.12.attn.out_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.12.attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.12.ln_1.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.12.ln_1.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.12.mlp.fc_in.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.12.mlp.fc_in.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.12.mlp.fc_out.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.12.mlp.fc_out.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.13.attn.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.13.attn.masked_bias": "model-00003-of-00006.safetensors",
+    "transformer.h.13.attn.out_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.13.attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.13.ln_1.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.13.ln_1.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.13.mlp.fc_in.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.13.mlp.fc_in.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.13.mlp.fc_out.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.13.mlp.fc_out.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.14.attn.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.14.attn.masked_bias": "model-00003-of-00006.safetensors",
+    "transformer.h.14.attn.out_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.14.attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.14.ln_1.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.14.ln_1.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.14.mlp.fc_in.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.14.mlp.fc_in.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.14.mlp.fc_out.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.14.mlp.fc_out.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.15.attn.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.15.attn.masked_bias": "model-00003-of-00006.safetensors",
+    "transformer.h.15.attn.out_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.15.attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.15.ln_1.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.15.ln_1.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.15.mlp.fc_in.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.15.mlp.fc_in.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.15.mlp.fc_out.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.15.mlp.fc_out.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.16.attn.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.16.attn.masked_bias": "model-00003-of-00006.safetensors",
+    "transformer.h.16.attn.out_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.16.attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.16.ln_1.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.16.ln_1.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.16.mlp.fc_in.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.16.mlp.fc_in.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.16.mlp.fc_out.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.16.mlp.fc_out.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.17.attn.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.17.attn.masked_bias": "model-00003-of-00006.safetensors",
+    "transformer.h.17.attn.out_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.17.attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.17.ln_1.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.17.ln_1.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.17.mlp.fc_in.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.17.mlp.fc_in.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.17.mlp.fc_out.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.17.mlp.fc_out.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.18.attn.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.18.attn.masked_bias": "model-00003-of-00006.safetensors",
+    "transformer.h.18.attn.out_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.18.attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.18.ln_1.bias": "model-00003-of-00006.safetensors",
+    "transformer.h.18.ln_1.weight": "model-00003-of-00006.safetensors",
+    "transformer.h.18.mlp.fc_in.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.18.mlp.fc_in.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.18.mlp.fc_out.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.18.mlp.fc_out.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.19.attn.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.19.attn.masked_bias": "model-00004-of-00006.safetensors",
+    "transformer.h.19.attn.out_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.19.attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.19.ln_1.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.19.ln_1.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.19.mlp.fc_in.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.19.mlp.fc_in.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.19.mlp.fc_out.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.19.mlp.fc_out.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.2.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.2.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.2.attn.out_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.2.attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.2.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.2.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.2.mlp.fc_in.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.2.mlp.fc_in.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.2.mlp.fc_out.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.2.mlp.fc_out.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.20.attn.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.20.attn.masked_bias": "model-00004-of-00006.safetensors",
+    "transformer.h.20.attn.out_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.20.attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.20.ln_1.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.20.ln_1.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.20.mlp.fc_in.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.20.mlp.fc_in.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.20.mlp.fc_out.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.20.mlp.fc_out.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.21.attn.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.21.attn.masked_bias": "model-00004-of-00006.safetensors",
+    "transformer.h.21.attn.out_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.21.attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.21.ln_1.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.21.ln_1.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.21.mlp.fc_in.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.21.mlp.fc_in.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.21.mlp.fc_out.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.21.mlp.fc_out.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.22.attn.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.22.attn.masked_bias": "model-00004-of-00006.safetensors",
+    "transformer.h.22.attn.out_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.22.attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.22.ln_1.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.22.ln_1.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.22.mlp.fc_in.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.22.mlp.fc_in.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.22.mlp.fc_out.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.22.mlp.fc_out.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.23.attn.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.23.attn.masked_bias": "model-00004-of-00006.safetensors",
+    "transformer.h.23.attn.out_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.23.attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.23.ln_1.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.23.ln_1.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.23.mlp.fc_in.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.23.mlp.fc_in.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.23.mlp.fc_out.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.23.mlp.fc_out.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.24.attn.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.24.attn.masked_bias": "model-00004-of-00006.safetensors",
+    "transformer.h.24.attn.out_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.24.attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.24.ln_1.bias": "model-00004-of-00006.safetensors",
+    "transformer.h.24.ln_1.weight": "model-00004-of-00006.safetensors",
+    "transformer.h.24.mlp.fc_in.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.24.mlp.fc_in.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.24.mlp.fc_out.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.24.mlp.fc_out.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.25.attn.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.25.attn.masked_bias": "model-00005-of-00006.safetensors",
+    "transformer.h.25.attn.out_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.25.attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.25.ln_1.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.25.ln_1.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.25.mlp.fc_in.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.25.mlp.fc_in.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.25.mlp.fc_out.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.25.mlp.fc_out.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.26.attn.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.26.attn.masked_bias": "model-00005-of-00006.safetensors",
+    "transformer.h.26.attn.out_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.26.attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.26.ln_1.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.26.ln_1.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.26.mlp.fc_in.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.26.mlp.fc_in.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.26.mlp.fc_out.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.26.mlp.fc_out.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.27.attn.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.27.attn.masked_bias": "model-00005-of-00006.safetensors",
+    "transformer.h.27.attn.out_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.27.attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.27.ln_1.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.27.ln_1.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.27.mlp.fc_in.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.27.mlp.fc_in.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.27.mlp.fc_out.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.27.mlp.fc_out.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.28.attn.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.28.attn.masked_bias": "model-00005-of-00006.safetensors",
+    "transformer.h.28.attn.out_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.28.attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.28.ln_1.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.28.ln_1.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.28.mlp.fc_in.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.28.mlp.fc_in.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.28.mlp.fc_out.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.28.mlp.fc_out.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.29.attn.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.29.attn.masked_bias": "model-00005-of-00006.safetensors",
+    "transformer.h.29.attn.out_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.29.attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.29.ln_1.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.29.ln_1.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.29.mlp.fc_in.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.29.mlp.fc_in.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.29.mlp.fc_out.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.29.mlp.fc_out.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.3.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.3.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.3.attn.out_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.3.attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.3.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.3.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.3.mlp.fc_in.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.3.mlp.fc_in.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.3.mlp.fc_out.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.3.mlp.fc_out.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.30.attn.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.30.attn.masked_bias": "model-00005-of-00006.safetensors",
+    "transformer.h.30.attn.out_proj.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.30.attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.30.ln_1.bias": "model-00005-of-00006.safetensors",
+    "transformer.h.30.ln_1.weight": "model-00005-of-00006.safetensors",
+    "transformer.h.30.mlp.fc_in.bias": "model-00006-of-00006.safetensors",
+    "transformer.h.30.mlp.fc_in.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.30.mlp.fc_out.bias": "model-00006-of-00006.safetensors",
+    "transformer.h.30.mlp.fc_out.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.31.attn.bias": "model-00006-of-00006.safetensors",
+    "transformer.h.31.attn.masked_bias": "model-00006-of-00006.safetensors",
+    "transformer.h.31.attn.out_proj.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.31.attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.31.ln_1.bias": "model-00006-of-00006.safetensors",
+    "transformer.h.31.ln_1.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.31.mlp.fc_in.bias": "model-00006-of-00006.safetensors",
+    "transformer.h.31.mlp.fc_in.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.31.mlp.fc_out.bias": "model-00006-of-00006.safetensors",
+    "transformer.h.31.mlp.fc_out.weight": "model-00006-of-00006.safetensors",
+    "transformer.h.4.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.4.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.4.attn.out_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.4.attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.4.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.4.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.4.mlp.fc_in.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.4.mlp.fc_in.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.4.mlp.fc_out.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.4.mlp.fc_out.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.5.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.5.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.5.attn.out_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.5.attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.5.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.5.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.5.mlp.fc_in.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.5.mlp.fc_in.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.5.mlp.fc_out.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.5.mlp.fc_out.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.6.attn.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.6.attn.masked_bias": "model-00001-of-00006.safetensors",
+    "transformer.h.6.attn.out_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.6.attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.6.ln_1.bias": "model-00001-of-00006.safetensors",
+    "transformer.h.6.ln_1.weight": "model-00001-of-00006.safetensors",
+    "transformer.h.6.mlp.fc_in.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.6.mlp.fc_in.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.6.mlp.fc_out.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.6.mlp.fc_out.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.7.attn.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.7.attn.masked_bias": "model-00002-of-00006.safetensors",
+    "transformer.h.7.attn.out_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.7.attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.7.ln_1.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.7.ln_1.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.7.mlp.fc_in.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.7.mlp.fc_in.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.7.mlp.fc_out.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.7.mlp.fc_out.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.8.attn.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.8.attn.masked_bias": "model-00002-of-00006.safetensors",
+    "transformer.h.8.attn.out_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.8.attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.8.ln_1.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.8.ln_1.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.8.mlp.fc_in.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.8.mlp.fc_in.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.8.mlp.fc_out.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.8.mlp.fc_out.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.9.attn.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.9.attn.masked_bias": "model-00002-of-00006.safetensors",
+    "transformer.h.9.attn.out_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.9.attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.9.ln_1.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.9.ln_1.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.9.mlp.fc_in.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.9.mlp.fc_in.weight": "model-00002-of-00006.safetensors",
+    "transformer.h.9.mlp.fc_out.bias": "model-00002-of-00006.safetensors",
+    "transformer.h.9.mlp.fc_out.weight": "model-00002-of-00006.safetensors",
+    "transformer.ln_f.bias": "model-00006-of-00006.safetensors",
+    "transformer.ln_f.weight": "model-00006-of-00006.safetensors",
+    "transformer.structure.attn_pool.mlp.0.bias": "model-00006-of-00006.safetensors",
+    "transformer.structure.attn_pool.mlp.0.weight": "model-00006-of-00006.safetensors",
+    "transformer.structure.attn_pool.mlp.1.bias": "model-00006-of-00006.safetensors",
+    "transformer.structure.attn_pool.mlp.1.weight": "model-00006-of-00006.safetensors",
+    "transformer.structure.attn_pool.mlp.3.bias": "model-00006-of-00006.safetensors",
+    "transformer.structure.attn_pool.mlp.3.weight": "model-00006-of-00006.safetensors",
+    "transformer.wte.weight": "model-00001-of-00006.safetensors"
+  }
+}

modeling_InstructProGen.py

@@ -0,0 +1,847 @@
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.generation.configuration_utils import GenerationConfig
+from transformers.generation.logits_process import LogitsProcessorList
+from transformers.generation.stopping_criteria import StoppingCriteriaList
+from transformers.generation.streamers import BaseStreamer
+from transformers.generation.utils import GenerateOutput
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+from .configuration_progen import ProGenConfig
+
+logger = logging.get_logger(__name__)
+
+from .structure import StructureTransformer
+# from .structure_ligand import StructureTransformer
+
+
+import math
+
+# Inverse dim formula to find dim based on number of rotations
+def _yarn_find_correction_dim(num_rotations, dim, base=10000, max_position_embeddings=2048):
+    return (dim * math.log(max_position_embeddings/(num_rotations * 2 * math.pi)))/(2 * math.log(base))
+
+# Find dim range bounds based on rotations
+def _yarn_find_correction_range(low_rot, high_rot, dim, base=10000, max_position_embeddings=2048):
+    low = math.floor(_yarn_find_correction_dim(
+        low_rot, dim, base, max_position_embeddings))
+    high = math.ceil(_yarn_find_correction_dim(
+        high_rot, dim, base, max_position_embeddings))
+    return max(low, 0), min(high, dim-1)  # Clamp values just in case
+
+def _yarn_linear_ramp_mask(min, max, dim):
+    if min == max:
+        max += 0.001  # Prevent singularity
+
+    linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
+    ramp_func = torch.clamp(linear_func, 0, 1)
+    return ramp_func
+
+def _yarn_get_mscale(scale=1):
+    if scale <= 1:
+        return 1.0
+    return 0.1 * math.log(scale) + 1.0
+
+def yarn(x, seq_len=None, seq_dim=1, base=10000, scale=4, original_max_position_embeddings=1024, extrapolation_factor=1, attn_factor=1, beta_fast=4, beta_slow=0.5):
+    dim = x.shape[-1]
+    pos_freqs = base ** (torch.arange(0, dim, 2).float().to(x.device) / dim)
+    inv_freq_extrapolation = 1.0 / pos_freqs
+    inv_freq_interpolation = 1.0 / (scale * pos_freqs)
+
+    low, high = _yarn_find_correction_range(beta_fast, beta_slow, dim, base, original_max_position_embeddings)
+
+    inv_freq_mask = (1 - _yarn_linear_ramp_mask(low, high, dim // 2).float().to(x.device)) * extrapolation_factor # Get n-d rotational scaling corrected for extrapolation
+    inv_freq = inv_freq_interpolation * (1 - inv_freq_mask) + inv_freq_extrapolation * inv_freq_mask
+
+    mscale = float(_yarn_get_mscale(scale) * attn_factor) # Get n-d magnitude scaling corrected for interpolation
+
+    # Build here to make `torch.jit.trace` work.
+    if seq_len is None:
+        seq_len = x.shape[seq_dim]
+    t = torch.arange(seq_len, device=x.device, dtype=x.dtype)
+    freqs = torch.einsum("i,j->ij", t, inv_freq)
+    # torch.save(freqs, 'yarn_freq.pt')
+    return torch.sin(freqs)*mscale, torch.cos(freqs)*mscale
+
+def fixed_pos_embedding(x, seq_dim=1, seq_len=None):
+    dim = x.shape[-1]
+    if seq_len is None:
+        seq_len = x.shape[seq_dim]
+    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))
+    sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq).to(x.device).float()
+    # return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)
+    return torch.sin(sinusoid_inp).to(x.dtype), torch.cos(sinusoid_inp).to(x.dtype)
+
+def rotate_every_two(x):
+    x1 = x[:, :, :, ::2]
+    x2 = x[:, :, :, 1::2]
+    x = torch.stack((-x2, x1), axis=-1)
+    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')
+
+
+def apply_rotary_pos_emb(x, sincos, offset=0, position_ids=None):
+    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)
+    if position_ids is not None:
+        sin, cos = map(lambda t: t[None, :, None, :].repeat_interleave(2, 3), sincos)
+        sin = sin.take_along_dim(position_ids.unsqueeze(-1).unsqueeze(-1), dim=1)
+        cos = cos.take_along_dim(position_ids.unsqueeze(-1).unsqueeze(-1), dim=1)
+    else:
+        sin, cos = map(lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(2, 3), sincos)
+    return (x * cos) + (rotate_every_two(x) * sin)
+
+
+class ProGenAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        max_positions = config.max_position_embeddings
+        self.yarn = config.yarn
+        self.register_buffer(
+            "bias",
+            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
+                1, 1, max_positions, max_positions
+            ),
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e9))
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.embed_dim = config.hidden_size
+        self.num_attention_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_attention_heads
+        if self.head_dim * self.num_attention_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and `num_attention_heads`: {self.num_attention_heads})."
+            )
+        self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())
+        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)
+
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.rotary_dim = None
+        if config.rotary_dim is not None:
+            self.rotary_dim = config.rotary_dim
+
+    def _split_heads(self, x, n_head, dim_head, mp_num):
+        reshaped = x.reshape(x.shape[:-1] + (n_head//mp_num, dim_head))
+        reshaped = reshaped.reshape(x.shape[:-2] + (-1, ) + reshaped.shape[-1:])
+        return reshaped
+
+    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
+        """
+        Merges attn_head_size dim and num_attn_heads dim into n_ctx
+        """
+        if len(tensor.shape) == 5:
+            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
+        elif len(tensor.shape) == 4:
+            tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        else:
+            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
+        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def _attn(
+        self,
+        query,
+        key,
+        value,
+        attention_mask=None,
+        head_mask=None,
+    ):
+
+        # compute causal mask from causal mask buffer
+        query_length, key_length = query.size(-2), key.size(-2)
+        causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
+
+        # Keep the attention weights computation in fp32 to avoid overflow issues
+        query = query.to(torch.float32)
+        key = key.to(torch.float32)
+
+        attn_weights = torch.matmul(query, key.transpose(-1, -2))
+
+        attn_weights = attn_weights / self.scale_attn
+        attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))
+
+        if attention_mask is not None:
+            # Apply the attention mask
+            attn_weights = attn_weights + attention_mask
+
+        attn_weights = nn.Softmax(dim=-1)(attn_weights)
+        attn_weights = attn_weights.to(value.dtype)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = torch.matmul(attn_weights, value)
+
+        return attn_output, attn_weights
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        layer_past=None,
+        head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+        position_ids= None
+    ):
+
+        qkv = self.qkv_proj(hidden_states)
+        # TODO(enijkamp): factor out number of logical TPU-v3/v4 cores or make forward pass agnostic
+        # mp_num = 4
+        mp_num = 8
+        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))
+
+        local_dim = self.head_dim * self.num_attention_heads // mp_num
+        query, value, key = torch.split(qkv_split, local_dim, dim=-1)
+        query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+        key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+
+        value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)
+        value = value.permute(0, 2, 1, 3)
+
+        seq_len = key.shape[1] * 2 # for position ids greater than current
+        offset = 0
+
+        if layer_past is not None:
+            offset = layer_past[0].shape[-2]
+            seq_len += offset
+
+        if self.rotary_dim is not None:
+            k_rot = key[:, :, :, : self.rotary_dim]
+            k_pass = key[:, :, :, self.rotary_dim :]
+
+            q_rot = query[:, :, :, : self.rotary_dim]
+            q_pass = query[:, :, :, self.rotary_dim :]
+            if self.yarn:
+                sincos = yarn(k_rot, seq_dim=1, seq_len=seq_len)
+            else:
+                sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)
+            k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset, position_ids=position_ids)
+            q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=offset, position_ids=position_ids)
+
+            key = torch.cat([k_rot, k_pass], dim=-1)
+            query = torch.cat([q_rot, q_pass], dim=-1)
+        else:
+            if self.yarn:
+                sincos = yarn(k_rot, seq_dim=1, seq_len=seq_len)
+            else:
+                sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)
+            key = apply_rotary_pos_emb(key, sincos, offset=offset, position_ids=position_ids)
+            query = apply_rotary_pos_emb(query, sincos, offset=offset, position_ids=position_ids)
+
+        key = key.permute(0, 2, 1, 3)
+        query = query.permute(0, 2, 1, 3)
+
+        if layer_past is not None:
+            past_key = layer_past[0]
+            past_value = layer_past[1]
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        # compute self-attention: V x Softmax(QK^T)
+        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
+
+        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
+
+        attn_output = self.out_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+class ProGenMLP(nn.Module):
+    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * embed_dim
+        super().__init__()
+        embed_dim = config.n_embd
+
+        self.fc_in = nn.Linear(embed_dim, intermediate_size)
+        self.fc_out = nn.Linear(intermediate_size, embed_dim)
+
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states):
+        hidden_states = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc_out(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class ProGenBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = ProGenAttention(config)
+        self.mlp = ProGenMLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+        position_ids=None,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_outputs = self.attn(
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            position_ids=position_ids
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        hidden_states = attn_output + feed_forward_hidden_states + residual
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions)
+
+
+class ProGenPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ProGenConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    is_parallelizable = True
+    _no_split_modules = ["ProGenBlock"]
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear,)):
+            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, ProGenModel):
+            module.gradient_checkpointing = value
+
+class ProGenModel(ProGenPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embed_dim = config.n_embd
+        self.vocab_size = config.vocab_size
+        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([ProGenBlock(config) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)
+
+        self.gradient_checkpointing = False
+        self.structure = StructureTransformer(**config.structure)
+        
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        self.device_map = (
+            get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
+        )
+        assert_device_map(self.device_map, len(self.h))
+        self.model_parallel = True
+        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        self.wte = self.wte.to(self.first_device)
+        # Load onto devices
+        for k, v in self.device_map.items():
+            for block in v:
+                cuda_device = "cuda:" + str(k)
+                self.h[block] = self.h[block].to(cuda_device)
+        # ln_f to last
+        self.ln_f = self.ln_f.to(self.last_device)
+
+
+    def deparallelize(self):
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        self.wte = self.wte.to("cpu")
+        for index in range(len(self.h)):
+            self.h[index] = self.h[index].to("cpu")
+        self.ln_f = self.ln_f.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.wte
+
+    def set_input_embeddings(self, new_embeddings):
+        self.wte = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        query_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        if past_key_values is None:
+            # structure encode will check if input_ids contains valid 
+            # structure_embs: Tensor with size of batchsize * maxlen * width 
+            # structure_mask: Tensor with size of batchsize * maxlen * 1
+            # input_ids: new ids without structure path (ascii code)
+            # clone for keeping original input_ids not change
+            (structure_embs, structure_mask), input_ids = self.structure.encode(input_ids.clone())
+            
+        else:
+            structure_embs = None
+        
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # if token_type_ids is not None:
+        #     token_type_ids = token_type_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+
+
+
+        # Attention mask.
+        if attention_mask is not None:
+            assert batch_size > 0, "batch_size has to be defined and > 0"
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x num_attention_heads x N x N
+        # head_mask has shape n_layer x batch x num_attention_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids.int())
+        
+        if query_embeds is not None:
+            inputs_embeds = torch.cat([query_embeds, inputs_embeds], dim=1)
+            input_shape = inputs_embeds.size()[:-1]
+        
+        if structure_embs is not None:
+            try:
+                inputs_embeds_mask = (input_ids == 1).unsqueeze(-1) # bos to as structure placeholder
+                inputs_embeds = inputs_embeds.to(structure_embs.dtype) # change dtype manually, autocast won't change it.
+                inputs_embeds = inputs_embeds.masked_scatter(inputs_embeds_mask, structure_embs.masked_select(structure_mask.unsqueeze(-1)))
+                input_shape = inputs_embeds.size()[:-1]
+            except:
+                torch.save(input_ids, f'input_ids_{inputs_embeds.device}.pt')
+                torch.save(inputs_embeds, f'inputs_embeds_{inputs_embeds.device}.pt')
+                torch.save(structure_embs, f'structure_embs_{inputs_embeds.device}.pt')
+                
+
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+        
+        if position_ids is None:
+            if any(attention_mask[:, ..., 0] < 0): # padding left
+                position_ids = ((attention_mask >= 0).cumsum(dim=-1).squeeze() - 1).clamp(min=0)
+                position_ids = position_ids[:, past_length:past_length + input_shape[-1]]
+            else:
+                position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
+                position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        hidden_states = inputs_embeds
+
+        # disable token_type_ids
+        # if token_type_ids is not None:
+        #     token_type_embeds = self.wte(token_type_ids)
+        #     hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure layer_past is on same device as hidden_states (might not be correct)
+                if layer_past is not None:
+                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if isinstance(head_mask, torch.Tensor):
+                    head_mask = head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    # logger.warning(
+                    #     "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
+                    #     "`use_cache=False`..."
+                    # )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    head_mask[i],
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    position_ids=position_ids
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ProGenForCausalLM(ProGenPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head\.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = ProGenModel(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
+        self.init_weights()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.model_parallel = True
+
+    def deparallelize(self):
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def _update_model_kwargs_for_generation(self, *args, **kwargs):
+        model_kwargs = super()._update_model_kwargs_for_generation(*args, **kwargs)
+
+        position_ids = model_kwargs.get("position_ids", None)
+        attention_mask = model_kwargs.get("attention_mask", None)
+        use_cache = model_kwargs.get("use_cache", False)
+        
+        if attention_mask is not None and position_ids is not None:
+            # create position_ids on the fly for batch generation
+            position_ids = position_ids
+
+            # if past_key_values:
+            last_position = position_ids[:, -1:] # all position in a batch should be the sanme
+            current_position = last_position + 1
+            if not use_cache:
+                position_ids = torch.concat([position_ids, current_position], dim=-1)
+            else:
+                position_ids = current_position
+
+        else:
+            position_ids = None
+        model_kwargs['position_ids'] = position_ids
+        return model_kwargs
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
+        use_cache = kwargs.get("use_cache", False)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past_key_values:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        # if attention_mask is not None and position_ids is not None:
+        #     # create position_ids on the fly for batch generation
+        #     position_ids = position_ids
+        #     if input_ids[0, -1] != 3:
+        #     # if past_key_values:
+        #         last_position = position_ids[:, -1:] # all position in a batch should be the sanme
+        #         current_position = last_position + 1
+        #         if not use_cache:
+        #             position_ids = torch.concat([position_ids, current_position], dim=-1)
+        #         else:
+        #             position_ids = current_position
+
+        # else:
+        #     position_ids = None
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past_key_values,
+            "use_cache": use_cache,
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    # def prepare_inputs_for_generation(
+    #     self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    # ):
+    #     if past_key_values:
+    #         input_ids = input_ids[:, -1:]
+
+    #     position_ids = kwargs.get("position_ids")
+    #     if position_ids is not None:
+    #         last_position = position_ids[:, -1:] # all position in a batch should be the sanme
+    #         current_position = last_position + 1
+    #         position_ids = torch.concat([position_ids, current_position], dim=-1)
+
+    #     # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+    #     if inputs_embeds is not None and past_key_values is None:
+    #         model_inputs = {"inputs_embeds": inputs_embeds}
+    #     else:
+    #         model_inputs = {"input_ids": input_ids}
+
+    #     model_inputs.update(
+    #         {
+    #             "past_key_values": past_key_values,
+    #             "use_cache": kwargs.get("use_cache"),
+    #             "attention_mask": attention_mask,
+    #             "position_ids": position_ids
+    #         }
+    #     )
+    #     return model_inputs
+
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        query_embeds = None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
+            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            query_embeds=query_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        # make sure sampling in fp16 works correctly and
+        # compute loss in fp32 to match with mesh-tf version
+        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
+        lm_logits = self.lm_head(hidden_states).to(torch.float32)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past
+        )
+
+    # def generate(self, inputs: Tensor | None = None, generation_config: GenerationConfig | None = None, logits_processor: LogitsProcessorList | None = None, stopping_criteria: StoppingCriteriaList | None = None, prefix_allowed_tokens_fn: Callable[[int, Tensor], List[int]] | None = None, synced_gpus: bool | None = None, assistant_model: PreTrainedModel | None = None, streamer: BaseStreamer | None = None, negative_prompt_ids: Tensor | None = None, negative_prompt_attention_mask: Tensor | None = None, **kwargs) -> GenerateOutput | LongTensor:
+    #     return super().generate(inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, assistant_model, streamer, negative_prompt_ids, negative_prompt_attention_mask, **kwargs)

structure.py

@@ -0,0 +1,315 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from collections import OrderedDict
+import math
+import requests
+from io import BytesIO
+from functools import partial
+import pickle
+from typing import Callable, Optional, Sequence, Tuple, List
+import numpy as np
+import os
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.nn.init import trunc_normal_
+from torchvision import transforms
+from torchvision.transforms import InterpolationMode
+
+class GLU(nn.Module):
+    def __init__(self,hidden_size):
+        super().__init__()
+        self.linear_proj = nn.Linear(hidden_size,hidden_size,bias=False)
+        self.norm1 = nn.LayerNorm(hidden_size)
+        self.act1 = nn.GELU()
+        self.act2 = nn.functional.silu
+        self.dense_h_to_4h = nn.Linear(hidden_size,hidden_size*4,bias=False)
+        self.gate_proj = nn.Linear(hidden_size,hidden_size*4,bias=False)
+        self.dense_4h_to_h = nn.Linear(hidden_size*4,hidden_size,bias=False)
+
+    def forward(self,x):
+        x = self.linear_proj(x)
+        x = self.act1(self.norm1(x))
+        x = self.act2(self.gate_proj(x))*self.dense_h_to_4h(x)
+        x = self.dense_4h_to_h(x)
+        return x
+def swiglu(x):
+    x = torch.chunk(x, 2, dim=-1)
+    return nn.functional.silu(x[0]) * x[1]
+
+class GLU_new(nn.Module):
+    def __init__(self,hidden_size, dropout=0.1):
+        super().__init__()
+        intermediate_size = int((4 * hidden_size * 2 / 3) / 64) * 64
+        intermediate_size = 1280
+
+        self.act = swiglu
+        self.dense_h_to_4h = nn.Linear(hidden_size, intermediate_size * 2, bias=False)
+        self.dense_4h_to_h = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.dropout = nn.Dropout(p=dropout)
+
+    def forward(self,x):
+        x = self.dense_h_to_4h(x)
+        x = self.act(x)
+        x = self.dense_4h_to_h(x)
+        x = self.dropout(x)
+        return x
+
+
+n_queries = 32
+def get_abs_pos(abs_pos, tgt_size):
+    # abs_pos: L, C
+    # tgt_size: M
+    # return: M, C
+    src_size = int(math.sqrt(abs_pos.size(0)))
+    tgt_size = int(math.sqrt(tgt_size))
+    dtype = abs_pos.dtype
+
+    if src_size != tgt_size:
+        return F.interpolate(
+            abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
+            size=(tgt_size, tgt_size),
+            mode="bicubic",
+            align_corners=False,
+        ).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype)
+    else:
+        return abs_pos
+
+from einops import rearrange, repeat
+
+def get_1d_sincos_pos_embed(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out) # (M, D/2)
+    emb_cos = np.cos(out) # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        kv_dim,
+        embed_dim,
+        num_heads=8,
+        n_queries=64,
+        max_seqlen=1024,
+        perceiver_resampler_positional_emb=True,
+        use_GLU=False,
+        bos_init=False,
+        dropout=0.0
+    ):
+        super().__init__()
+        self.perceiver_resampler_positional_emb = perceiver_resampler_positional_emb
+
+        if self.perceiver_resampler_positional_emb:
+            assert n_queries <= max_seqlen
+            self.stride = max_seqlen // n_queries
+            # self.nan_emb = nn.Parameter(torch.randn(1, kv_dim))
+            # nn.init.trunc_normal_(self.nan_emb, std=.02)
+            pos = np.arange(max_seqlen, dtype=np.float32)
+            self.register_buffer(
+                "pos_embed",
+                torch.from_numpy(get_1d_sincos_pos_embed(embed_dim, pos)).float()
+            )
+        self.latents = nn.Parameter(torch.randn(n_queries, embed_dim))
+        if bos_init:
+            self.latents.load('')
+        else:
+            nn.init.trunc_normal_(self.latents, std=1e-3)
+
+        self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
+        self.attn = nn.MultiheadAttention(embed_dim, num_heads, batch_first=True, dropout=dropout)
+        self.ln_q = nn.LayerNorm(embed_dim)
+        self.ln_kv = nn.LayerNorm(embed_dim)
+        self.ln_post = nn.LayerNorm(embed_dim)
+        if use_GLU:
+            print('GLU *********************************')
+            self.proj = GLU_new(embed_dim, dropout=dropout)
+        else:
+            self.proj = nn.Linear(embed_dim, embed_dim, bias=False)
+
+        self.apply(self._init_weights)
+    
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.trunc_normal_(m.weight, std=1e-3)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(self, struc_x):
+        """
+        Args:
+            x (torch.Tensor): protein structure features
+                shape (B, L, C)
+        Returns:
+            shape (B, n, C) where n is self.num_latents
+        """
+        x = struc_x["encoder_out"]
+        mask = struc_x["encoder_padding_mask"]
+
+
+        nan_mask = torch.isnan(x)
+        if nan_mask.any():
+            x = x.masked_fill(nan_mask, 0.0)
+            # nan_mask = nan_mask.sum(dim=-1).bool()
+            # x[nan_mask] += self.nan_emb
+
+        x = self.kv_proj(x)
+        x = self.ln_kv(x)
+
+        b, seqlen = x.shape[:2]
+
+        latents = self.ln_q(self.latents)
+        if self.perceiver_resampler_positional_emb:
+            # TODO: interpolate
+            latents = latents + self.pos_embed[::self.stride].contiguous().to(latents.device)
+            pos_emb = self.pos_embed[:seqlen].unsqueeze(0).to(latents.device)
+            x = x + pos_emb.contiguous()
+        
+        # blocks
+        latents = repeat(latents, "n d -> b n d", b=b)
+        out = self.attn(latents, x, x, key_padding_mask=~mask)[0]
+
+        out = self.ln_post(out)
+        out = self.proj(out)
+
+        return out
+
+class mlp(nn.Module):
+    def __init__(self, width, output_dim, **kwargs):
+        super().__init__(**kwargs)
+
+        self.mlp = nn.Sequential(
+            nn.Linear(width, output_dim),
+            nn.LayerNorm(output_dim),
+            nn.GELU(),
+            nn.Linear(output_dim, output_dim)
+        )
+
+    def forward(self, struc_x):
+        x = struc_x["encoder_out"]
+        mask = struc_x["encoder_padding_mask"]
+        return self.mlp(x), mask
+
+class StructureTransformer(nn.Module):
+
+    def __init__(
+            self,
+            width: int = 640,
+            n_queries: int = 32,
+            output_dim: int = 4096,
+            embedding_keys=set(["mpnn_emb"]),
+            max_seqlen: int=1024,
+            num_heads: int=8,
+            structure_emb_path_prefix='structure_emb',
+            projector='mlp',
+            **kwargs
+    ):
+        super().__init__()
+
+        self.structure_emb_path_prefix = structure_emb_path_prefix
+        # self.transformer = None # replace None with a pretrained strucure encoder
+        self.embedding_keys = embedding_keys
+        self.max_seqlen = max_seqlen
+        self.width = width
+        self.n_queries = n_queries
+        if projector == 'mlp':
+            self.attn_pool = mlp(
+                width=width,
+                output_dim=output_dim,
+                **kwargs
+                )
+        else:
+            self.attn_pool = Resampler(
+                embed_dim=output_dim,
+                kv_dim=width,
+                n_queries=n_queries,
+                max_seqlen=max_seqlen,
+                num_heads=num_heads,
+                **kwargs
+            )
+
+    def prepare_structure(self, sample):
+        emb_pad = torch.zeros((self.max_seqlen, self.width))
+        emb_mask = torch.zeros((self.max_seqlen), dtype=bool)
+        
+        
+        ### domians ###
+        emb = []
+        for ek in self.embedding_keys:
+            if ek in sample:
+                if isinstance( sample[ek], List):
+                    emb.append(torch.cat(sample[ek]))
+                else:
+                    emb.append(sample[ek].squeeze())
+        # emb = [sample[ek] for ek in self.embedding_keys if ek in sample]
+        emb = torch.cat(emb, dim=-1)
+        
+        emb_pad[:len(emb)] = emb
+        emb_mask[:len(emb)] = 1
+        return emb_pad, emb_mask
+
+    def forward(self, x):
+
+        # x = self.transformer(x)
+        x = self.attn_pool(x)
+
+        return x
+
+    def encode(self, input_ids: List[torch.Tensor]):
+        structure_embs = []
+        structure_mask = []
+
+        for structure_path in input_ids:
+            
+            if structure_path[0] == 1: # bos token for bypassing DPO trainer
+                structure_path[0] = 0
+            if structure_path[0] == 0: # left padding
+                structure_path = structure_path[structure_path > 0]
+                
+            path_length = (structure_path>32).sum() # structure path should greater than 32 in ascii
+            structure_path = [chr(s) for s in structure_path[:path_length].int().tolist() if s > 0]
+
+            structure_path = os.path.join(self.structure_emb_path_prefix, ''.join(structure_path))
+            
+            if not os.path.exists(structure_path):
+                print(structure_path)
+                print('no structure found')
+                return None
+                
+            with open(structure_path, 'rb') as f:
+                structure, struc_mask = self.prepare_structure(pickle.load(f))
+                
+
+            structure_embs.append(structure)
+            structure_mask.append(struc_mask)
+
+        input_ids[input_ids > 32] = 1 # change ascii code back to <|bos|>
+        structure_embs = torch.stack(structure_embs, dim=0).to(
+            device=next(self.attn_pool.parameters()).device, 
+            dtype=next(self.attn_pool.parameters()).dtype)
+        structure_mask = torch.stack(structure_mask, dim=0).to(
+            device=next(self.attn_pool.parameters()).device)
+
+        return self({
+                'encoder_out': structure_embs,
+                'encoder_padding_mask': structure_mask
+            }), input_ids

tokenization_iPLM.py

@@ -0,0 +1,110 @@
+from typing import List, Optional, Union
+from transformers import PreTrainedTokenizerFast
+from tokenizers.processors import TemplateProcessing
+from tokenizers import Tokenizer
+from transformers.tokenization_utils_base import BatchEncoding, EncodedInput, PreTokenizedInput, TextInput, TruncationStrategy
+from transformers.utils import PaddingStrategy, TensorType
+import torch
+import numpy as np
+
+def create_tokenizer_custom(file):
+    with open(file, 'r') as f:
+        return Tokenizer.from_str(f.read())
+    
+
+class iPLMTokenizer(PreTrainedTokenizerFast):
+    def __init__(self, parallel=False, **kwargs):
+        super().__init__(tokenizer_object=create_tokenizer_custom(kwargs.get('tokenizer_file')), **kwargs)
+        self.add_special_tokens({'pad_token': '<|pad|>'})
+        self.parallel = parallel
+    def __call__(
+            self,
+            text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+            n_queries = -1, # -1 for vary-length prompt, int with larger than 0 for fix-length, 0 for no prompt
+            text_pair: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = None,
+            text_target: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+            text_pair_target: Optional[
+                Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]
+            ] = None,
+            add_special_tokens: bool = True,
+            padding: Union[bool, str, PaddingStrategy] = False,
+            truncation: Union[bool, str, TruncationStrategy] = None,
+            max_length: Optional[int] = None,
+            stride: int = 0,
+            is_split_into_words: bool = False,
+            pad_to_multiple_of: Optional[int] = None,
+            return_tensors: Optional[Union[str, TensorType]] = None,
+            return_token_type_ids: Optional[bool] = None,
+            return_attention_mask: Optional[bool] = None,
+            return_overflowing_tokens: bool = False,
+            return_special_tokens_mask: bool = False,
+            return_offsets_mapping: bool = False,
+            return_length: bool = False,
+            verbose: bool = True,
+            **kwargs,
+        ) -> BatchEncoding:
+        
+        if not isinstance(text, list):
+            text = [text]
+            batching = False
+        else:
+            batching = True
+        
+        # add prompt
+        text_with_prompt = []
+        for t in text:
+            prompt_length = 0
+            assert '|' in t, 'prompt not found'
+            
+            raw_text = t.split('|')[-1]
+
+            if n_queries > 0: # use fix length prompt
+                prompt_length = n_queries
+            elif n_queries < 0:
+                prompt_length = len(raw_text.replace('1', '').replace('2', ''))
+            
+            text_with_prompt.append('<|bos|>' * prompt_length + raw_text)
+        
+        batch = super().__call__(
+            text=text_with_prompt,
+            text_pair=text_pair,
+            text_target=text_target,
+            text_pair_target=text_pair_target,
+            add_special_tokens=add_special_tokens,
+            padding=padding,
+            truncation= truncation,
+            max_length=max_length,
+            stride=stride,
+            is_split_into_words=is_split_into_words,
+            pad_to_multiple_of=pad_to_multiple_of,
+            padding_side=None,
+            return_tensors=return_tensors,
+            return_token_type_ids=return_token_type_ids,
+            return_attention_mask=return_attention_mask,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_offsets_mapping=return_offsets_mapping,
+            return_length=return_length,
+            verbose=verbose,
+            **kwargs
+            )
+
+        # add structure ids
+        for i in range(len(text)):            
+            if '|' not in text[i]:
+                continue
+
+            structure_ids = text[i].split('|')[0]
+            if return_tensors is None:
+                for j in range(len(structure_ids)):
+                    batch['input_ids'][i][j] = ord(structure_ids[j])
+            else:
+                batch['input_ids'][i, :len(structure_ids)] = torch.tensor([ord(c) for c in structure_ids])
+
+        if "token_type_ids" in batch:
+            del batch["token_type_ids"]
+
+        if batching:
+            return batch
+        else:
+            return {k:v[0] for k, v in batch.items()}

tokenizer.json

@@ -0,0 +1,95 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|pad|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|bos|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|eos|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "vocab": {
+      "<|pad|>": 0,
+      "<|bos|>": 1,
+      "<|eos|>": 2,
+      "1": 3,
+      "2": 4,
+      "A": 5,
+      "B": 6,
+      "C": 7,
+      "D": 8,
+      "E": 9,
+      "F": 10,
+      "G": 11,
+      "H": 12,
+      "I": 13,
+      "K": 14,
+      "L": 15,
+      "M": 16,
+      "N": 17,
+      "O": 18,
+      "P": 19,
+      "Q": 20,
+      "R": 21,
+      "S": 22,
+      "T": 23,
+      "U": 24,
+      "V": 25,
+      "W": 26,
+      "X": 27,
+      "Y": 28,
+      "Z": 29
+    },
+    "merges": []
+  }
+}

tokenizer_config.json

@@ -0,0 +1,11 @@
+{
+  "n_queries": 256,
+  "use_structure": true,
+  "tokenizer_class": "iPLMTokenizer",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_iPLM.iPLMTokenizer",
+      null
+      ]
+  }
+}