Upload folder using huggingface_hub

converted/config.json

@@ -0,0 +1,18 @@
+{
+    "dim": 4096,
+    "hidden_dim": 14336,
+    "head_dim": 128,
+    "n_layers": 32,
+    "n_heads": 32,
+    "n_kv_heads": 8,
+    "vocab_size": 32000,
+    "norm_eps": 1e-05,
+    "rope_theta": 1000000,
+    "max_batch_size": 32,
+    "max_seq_len": 4096,
+    "moe": {
+        "num_experts_per_tok": 2,
+        "num_experts": 8
+    },
+    "rope_scaling": null
+}

converted/consolidated.00-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22823d37d977c654db9c5a278a9f9b77deeb5f4ba4fce67ffd431a0047fb453b
+size 14886724427

converted/consolidated.01-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5bf9a705ffe440358f3dd60f0175e3ab4d1e4d2b9c8dddf2281de086a6023a1
+size 14886724427

converted/consolidated.02-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1298bcac5f348e9de6a78cba896f3913977152ba5f3f907c538dd5fea52d7386
+size 14886724427

converted/consolidated.03-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e87fc5c25e31ba27beb1de763ea6ec6398a7bd0a93b125ab9c1baa2fe0f8cf7
+size 14886724427

converted/consolidated.04-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ca45cf96dd0580a597b17c7189acb6913fe1e7b0deb791f9dcd65a80c819458
+size 14886724427

converted/consolidated.05-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4589bf33b4836634e7df1867cdd90c919def1c2ea144c3654c5d258e4024a1a8
+size 14886724427

converted/consolidated.06-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e72cb4cabbce6d07b5613f0e197033711a23753f828b6d936fc6866052ad80bf
+size 14886724427

converted/consolidated.07-of-08.model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfad68f6f5f6b4c9d0c6703e6dd24a07f32640006df51a1cd9d14953242b3cc5
+size 14886724427

converted/meta.json

@@ -0,0 +1,3 @@
+{
+  "llama_type": "mistral"
+}

converted/split.py

@@ -0,0 +1,144 @@
+weight_parallel_dim = {"llma.tok_embeddings.weight": 1, "llma.layers.0.attention.wq.weight": 0,
+                       "llma.layers.0.attention.wq.bias": 0, "llma.layers.0.attention.wk.weight": 0,
+                       "llma.layers.0.attention.wk.bias": 0, "llma.layers.0.attention.wv.weight": 0,
+                       "llma.layers.0.attention.wv.bias": 0, "llma.layers.0.attention.wo.weight": 1,
+                       "llma.layers.1.attention.wq.weight": 0, "llma.layers.1.attention.wq.bias": 0,
+                       "llma.layers.1.attention.wk.weight": 0, "llma.layers.1.attention.wk.bias": 0,
+                       "llma.layers.1.attention.wv.weight": 0, "llma.layers.1.attention.wv.bias": 0,
+                       "llma.layers.1.attention.wo.weight": 1, "llma.layers.2.attention.wq.weight": 0,
+                       "llma.layers.2.attention.wq.bias": 0, "llma.layers.2.attention.wk.weight": 0,
+                       "llma.layers.2.attention.wk.bias": 0, "llma.layers.2.attention.wv.weight": 0,
+                       "llma.layers.2.attention.wv.bias": 0, "llma.layers.2.attention.wo.weight": 1,
+                       "llma.layers.3.attention.wq.weight": 0, "llma.layers.3.attention.wq.bias": 0,
+                       "llma.layers.3.attention.wk.weight": 0, "llma.layers.3.attention.wk.bias": 0,
+                       "llma.layers.3.attention.wv.weight": 0, "llma.layers.3.attention.wv.bias": 0,
+                       "llma.layers.3.attention.wo.weight": 1, "llma.layers.4.attention.wq.weight": 0,
+                       "llma.layers.4.attention.wq.bias": 0, "llma.layers.4.attention.wk.weight": 0,
+                       "llma.layers.4.attention.wk.bias": 0, "llma.layers.4.attention.wv.weight": 0,
+                       "llma.layers.4.attention.wv.bias": 0, "llma.layers.4.attention.wo.weight": 1,
+                       "llma.layers.5.attention.wq.weight": 0, "llma.layers.5.attention.wq.bias": 0,
+                       "llma.layers.5.attention.wk.weight": 0, "llma.layers.5.attention.wk.bias": 0,
+                       "llma.layers.5.attention.wv.weight": 0, "llma.layers.5.attention.wv.bias": 0,
+                       "llma.layers.5.attention.wo.weight": 1, "llma.layers.6.attention.wq.weight": 0,
+                       "llma.layers.6.attention.wq.bias": 0, "llma.layers.6.attention.wk.weight": 0,
+                       "llma.layers.6.attention.wk.bias": 0, "llma.layers.6.attention.wv.weight": 0,
+                       "llma.layers.6.attention.wv.bias": 0, "llma.layers.6.attention.wo.weight": 1,
+                       "llma.layers.7.attention.wq.weight": 0, "llma.layers.7.attention.wq.bias": 0,
+                       "llma.layers.7.attention.wk.weight": 0, "llma.layers.7.attention.wk.bias": 0,
+                       "llma.layers.7.attention.wv.weight": 0, "llma.layers.7.attention.wv.bias": 0,
+                       "llma.layers.7.attention.wo.weight": 1, "llma.layers.8.attention.wq.weight": 0,
+                       "llma.layers.8.attention.wq.bias": 0, "llma.layers.8.attention.wk.weight": 0,
+                       "llma.layers.8.attention.wk.bias": 0, "llma.layers.8.attention.wv.weight": 0,
+                       "llma.layers.8.attention.wv.bias": 0, "llma.layers.8.attention.wo.weight": 1,
+                       "llma.layers.9.attention.wq.weight": 0, "llma.layers.9.attention.wq.bias": 0,
+                       "llma.layers.9.attention.wk.weight": 0, "llma.layers.9.attention.wk.bias": 0,
+                       "llma.layers.9.attention.wv.weight": 0, "llma.layers.9.attention.wv.bias": 0,
+                       "llma.layers.9.attention.wo.weight": 1, "llma.layers.10.attention.wq.weight": 0,
+                       "llma.layers.10.attention.wq.bias": 0, "llma.layers.10.attention.wk.weight": 0,
+                       "llma.layers.10.attention.wk.bias": 0, "llma.layers.10.attention.wv.weight": 0,
+                       "llma.layers.10.attention.wv.bias": 0, "llma.layers.10.attention.wo.weight": 1,
+                       "llma.layers.11.attention.wq.weight": 0, "llma.layers.11.attention.wq.bias": 0,
+                       "llma.layers.11.attention.wk.weight": 0, "llma.layers.11.attention.wk.bias": 0,
+                       "llma.layers.11.attention.wv.weight": 0, "llma.layers.11.attention.wv.bias": 0,
+                       "llma.layers.11.attention.wo.weight": 1, "llma.layers.12.attention.wq.weight": 0,
+                       "llma.layers.12.attention.wq.bias": 0, "llma.layers.12.attention.wk.weight": 0,
+                       "llma.layers.12.attention.wk.bias": 0, "llma.layers.12.attention.wv.weight": 0,
+                       "llma.layers.12.attention.wv.bias": 0, "llma.layers.12.attention.wo.weight": 1,
+                       "llma.layers.13.attention.wq.weight": 0, "llma.layers.13.attention.wq.bias": 0,
+                       "llma.layers.13.attention.wk.weight": 0, "llma.layers.13.attention.wk.bias": 0,
+                       "llma.layers.13.attention.wv.weight": 0, "llma.layers.13.attention.wv.bias": 0,
+                       "llma.layers.13.attention.wo.weight": 1, "llma.layers.14.attention.wq.weight": 0,
+                       "llma.layers.14.attention.wq.bias": 0, "llma.layers.14.attention.wk.weight": 0,
+                       "llma.layers.14.attention.wk.bias": 0, "llma.layers.14.attention.wv.weight": 0,
+                       "llma.layers.14.attention.wv.bias": 0, "llma.layers.14.attention.wo.weight": 1,
+                       "llma.layers.15.attention.wq.weight": 0, "llma.layers.15.attention.wq.bias": 0,
+                       "llma.layers.15.attention.wk.weight": 0, "llma.layers.15.attention.wk.bias": 0,
+                       "llma.layers.15.attention.wv.weight": 0, "llma.layers.15.attention.wv.bias": 0,
+                       "llma.layers.15.attention.wo.weight": 1, "llma.layers.16.attention.wq.weight": 0,
+                       "llma.layers.16.attention.wq.bias": 0, "llma.layers.16.attention.wk.weight": 0,
+                       "llma.layers.16.attention.wk.bias": 0, "llma.layers.16.attention.wv.weight": 0,
+                       "llma.layers.16.attention.wv.bias": 0, "llma.layers.16.attention.wo.weight": 1,
+                       "llma.layers.17.attention.wq.weight": 0, "llma.layers.17.attention.wq.bias": 0,
+                       "llma.layers.17.attention.wk.weight": 0, "llma.layers.17.attention.wk.bias": 0,
+                       "llma.layers.17.attention.wv.weight": 0, "llma.layers.17.attention.wv.bias": 0,
+                       "llma.layers.17.attention.wo.weight": 1, "llma.layers.18.attention.wq.weight": 0,
+                       "llma.layers.18.attention.wq.bias": 0, "llma.layers.18.attention.wk.weight": 0,
+                       "llma.layers.18.attention.wk.bias": 0, "llma.layers.18.attention.wv.weight": 0,
+                       "llma.layers.18.attention.wv.bias": 0, "llma.layers.18.attention.wo.weight": 1,
+                       "llma.layers.19.attention.wq.weight": 0, "llma.layers.19.attention.wq.bias": 0,
+                       "llma.layers.19.attention.wk.weight": 0, "llma.layers.19.attention.wk.bias": 0,
+                       "llma.layers.19.attention.wv.weight": 0, "llma.layers.19.attention.wv.bias": 0,
+                       "llma.layers.19.attention.wo.weight": 1, "llma.layers.20.attention.wq.weight": 0,
+                       "llma.layers.20.attention.wq.bias": 0, "llma.layers.20.attention.wk.weight": 0,
+                       "llma.layers.20.attention.wk.bias": 0, "llma.layers.20.attention.wv.weight": 0,
+                       "llma.layers.20.attention.wv.bias": 0, "llma.layers.20.attention.wo.weight": 1,
+                       "llma.layers.21.attention.wq.weight": 0, "llma.layers.21.attention.wq.bias": 0,
+                       "llma.layers.21.attention.wk.weight": 0, "llma.layers.21.attention.wk.bias": 0,
+                       "llma.layers.21.attention.wv.weight": 0, "llma.layers.21.attention.wv.bias": 0,
+                       "llma.layers.21.attention.wo.weight": 1, "llma.layers.22.attention.wq.weight": 0,
+                       "llma.layers.22.attention.wq.bias": 0, "llma.layers.22.attention.wk.weight": 0,
+                       "llma.layers.22.attention.wk.bias": 0, "llma.layers.22.attention.wv.weight": 0,
+                       "llma.layers.22.attention.wv.bias": 0, "llma.layers.22.attention.wo.weight": 1,
+                       "llma.layers.23.attention.wq.weight": 0, "llma.layers.23.attention.wq.bias": 0,
+                       "llma.layers.23.attention.wk.weight": 0, "llma.layers.23.attention.wk.bias": 0,
+                       "llma.layers.23.attention.wv.weight": 0, "llma.layers.23.attention.wv.bias": 0,
+                       "llma.layers.23.attention.wo.weight": 1, "llma.layers.24.attention.wq.weight": 0,
+                       "llma.layers.24.attention.wq.bias": 0, "llma.layers.24.attention.wk.weight": 0,
+                       "llma.layers.24.attention.wk.bias": 0, "llma.layers.24.attention.wv.weight": 0,
+                       "llma.layers.24.attention.wv.bias": 0, "llma.layers.24.attention.wo.weight": 1,
+                       "llma.layers.25.attention.wq.weight": 0, "llma.layers.25.attention.wq.bias": 0,
+                       "llma.layers.25.attention.wk.weight": 0, "llma.layers.25.attention.wk.bias": 0,
+                       "llma.layers.25.attention.wv.weight": 0, "llma.layers.25.attention.wv.bias": 0,
+                       "llma.layers.25.attention.wo.weight": 1, "llma.layers.26.attention.wq.weight": 0,
+                       "llma.layers.26.attention.wq.bias": 0, "llma.layers.26.attention.wk.weight": 0,
+                       "llma.layers.26.attention.wk.bias": 0, "llma.layers.26.attention.wv.weight": 0,
+                       "llma.layers.26.attention.wv.bias": 0, "llma.layers.26.attention.wo.weight": 1,
+                       "llma.layers.27.attention.wq.weight": 0, "llma.layers.27.attention.wq.bias": 0,
+                       "llma.layers.27.attention.wk.weight": 0, "llma.layers.27.attention.wk.bias": 0,
+                       "llma.layers.27.attention.wv.weight": 0, "llma.layers.27.attention.wv.bias": 0,
+                       "llma.layers.27.attention.wo.weight": 1, "llma.layers.28.attention.wq.weight": 0,
+                       "llma.layers.28.attention.wq.bias": 0, "llma.layers.28.attention.wk.weight": 0,
+                       "llma.layers.28.attention.wk.bias": 0, "llma.layers.28.attention.wv.weight": 0,
+                       "llma.layers.28.attention.wv.bias": 0, "llma.layers.28.attention.wo.weight": 1,
+                       "llma.layers.29.attention.wq.weight": 0, "llma.layers.29.attention.wq.bias": 0,
+                       "llma.layers.29.attention.wk.weight": 0, "llma.layers.29.attention.wk.bias": 0,
+                       "llma.layers.29.attention.wv.weight": 0, "llma.layers.29.attention.wv.bias": 0,
+                       "llma.layers.29.attention.wo.weight": 1, "llma.layers.30.attention.wq.weight": 0,
+                       "llma.layers.30.attention.wq.bias": 0, "llma.layers.30.attention.wk.weight": 0,
+                       "llma.layers.30.attention.wk.bias": 0, "llma.layers.30.attention.wv.weight": 0,
+                       "llma.layers.30.attention.wv.bias": 0, "llma.layers.30.attention.wo.weight": 1,
+                       "llma.layers.31.attention.wq.weight": 0, "llma.layers.31.attention.wq.bias": 0,
+                       "llma.layers.31.attention.wk.weight": 0, "llma.layers.31.attention.wk.bias": 0,
+                       "llma.layers.31.attention.wv.weight": 0, "llma.layers.31.attention.wv.bias": 0,
+                       "llma.layers.31.attention.wo.weight": 1, "llma.output.weight": 0, "llma.output.bias": 0}
+
+import torch
+from pathlib import Path
+
+Path("./converted").mkdir(exist_ok=True)
+
+ori = torch.load("consolidated.00.pth", map_location="cpu")
+ori = {"llma." + key: val for key, val in ori.items()}
+
+
+def func(rank=0):
+    shard_split_to = 8
+    split_ckpt = {}
+    for key, ori_param in ori.items():
+        if key in weight_parallel_dim:
+            split_ckpt[key] = torch.chunk(ori_param, shard_split_to, weight_parallel_dim[key])[
+                rank % shard_split_to].clone()
+            if rank == 0:
+                print(f"chunk {key}")
+        else:
+            if "experts." in key and int(key.split("experts.")[1].split(".")[0]) != rank:
+                continue
+            else:
+                split_ckpt[key] = ori_param
+                if rank == 0:
+                    print(f"inherit {key}")
+    torch.save({"model": split_ckpt}, f"converted/consolidated.{rank:02d}-of-08.model.pth")
+
+
+for r in range(8):
+    func(r)

converted/tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dadfd56d766715c61d2ef780a525ab43b8e6da4de6865bda3d95fdef5e134055
+size 493443