Upload folder using huggingface_hub

config.json

@@ -0,0 +1,24 @@
+{
+    "apply_residual_connection_post_layernorm": false,
+    "architectures": [
+      "BloomModel"
+    ],
+    "attention_dropout": 0.0,
+    "attention_softmax_in_fp32": true,
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "hidden_dropout": 0.0,
+    "hidden_size": 4096,
+    "initializer_range": 0.02,
+    "layer_norm_epsilon": 1e-05,
+    "masked_softmax_fusion": true,
+    "model_type": "bloom",
+    "n_head": 32,
+    "n_layer": 32,
+    "pad_token_id": 0,
+    "pretraining_tp": 1,
+    "slow_but_exact": false,
+    "transformers_version": "4.26.0.dev0",
+    "use_cache": true,
+    "vocab_size": 32768
+  }

convert_weight.py

@@ -0,0 +1,72 @@
+import torch
+
+
+input_dir_path = "/scratch/project_462000086/norwegian_gpt/Megatron-DeepSpeed-fixed/checkpoints/global_step120000"
+output_dir_path = "/scratch/project_462000086/norwegian_gpt/Megatron-DeepSpeed-fixed/hf_pilot_checkpoint_120k"
+
+n_hidden = 4096
+n_heads = 32
+n_layers = 32
+n_tp = 4
+
+
+weights = {}
+
+# embedding
+embedding_weights = []
+for i in range(n_tp):
+    path = f"{input_dir_path}/layer_01-model_0{i}-model_states.pt"
+    checkpoint = torch.load(path)
+
+    embedding_weights.append(checkpoint["word_embeddings.weight"].bfloat16())
+
+    weights[f"transformer.word_embeddings_layernorm.weight"] = checkpoint["word_embeddings.norm.weight"].bfloat16()
+    weights[f"transformer.word_embeddings_layernorm.bias"] = checkpoint["word_embeddings.norm.bias"].bfloat16()
+
+weights[f"transformer.word_embeddings.weight"] = torch.cat(embedding_weights, dim=0)
+weights[f"lm_head.weight"] = torch.cat(embedding_weights, dim=0)
+del embedding_weights
+
+
+# transformer layers
+for layer in range(n_layers):
+    qkv_weights = []
+    qkv_biases = []
+    o_weights = []
+    up_weights = []
+    up_biases = []
+    down_weights = []
+
+    for i in range(n_tp):
+        path = f"{input_dir_path}/layer_{layer+3:02d}-model_0{i}-model_states.pt"
+        checkpoint = torch.load(path)
+
+        weights[f"transformer.h.{layer}.input_layernorm.weight"] = checkpoint["input_layernorm.weight"].bfloat16()
+        weights[f"transformer.h.{layer}.input_layernorm.bias"] = checkpoint["input_layernorm.bias"].bfloat16()
+        weights[f"transformer.h.{layer}.self_attention.dense.bias"] = checkpoint["self_attention.dense.bias"].bfloat16()
+        weights[f"transformer.h.{layer}.post_attention_layernorm.weight"] = checkpoint["post_attention_layernorm.weight"].bfloat16()
+        weights[f"transformer.h.{layer}.post_attention_layernorm.bias"] = checkpoint["post_attention_layernorm.bias"].bfloat16()
+        weights[f"transformer.h.{layer}.mlp.dense_4h_to_h.bias"] = checkpoint["mlp.dense_4h_to_h.bias"].bfloat16()
+
+        qkv_weights.append(checkpoint["self_attention.query_key_value.weight"].bfloat16())
+        qkv_biases.append(checkpoint["self_attention.query_key_value.bias"].bfloat16())
+        o_weights.append(checkpoint["self_attention.dense.weight"].bfloat16())
+        up_weights.append(checkpoint["mlp.dense_h_to_4h.weight"].bfloat16())
+        up_biases.append(checkpoint["mlp.dense_h_to_4h.bias"].bfloat16())
+        down_weights.append(checkpoint["mlp.dense_4h_to_h.weight"].bfloat16())
+
+    weights[f"transformer.h.{layer}.self_attention.query_key_value.weight"] = torch.cat(qkv_weights, dim=0)
+    weights[f"transformer.h.{layer}.self_attention.query_key_value.bias"] = torch.cat(qkv_biases, dim=0)
+    weights[f"transformer.h.{layer}.self_attention.dense.weight"] = torch.cat(o_weights, dim=1)
+    weights[f"transformer.h.{layer}.mlp.dense_h_to_4h.weight"] = torch.cat(up_weights, dim=0)
+    weights[f"transformer.h.{layer}.mlp.dense_h_to_4h.bias"] = torch.cat(up_biases, dim=0)
+    weights[f"transformer.h.{layer}.mlp.dense_4h_to_h.weight"] = torch.cat(down_weights, dim=1)
+
+# output layer norm
+path = f"{input_dir_path}/layer_36-model_00-model_states.pt"
+checkpoint = torch.load(path)
+
+weights[f"transformer.ln_f.bias"] = checkpoint["bias"].bfloat16()
+weights[f"transformer.ln_f.weight"] = checkpoint["weight"].bfloat16()
+
+torch.save(weights, f"{output_dir_path}/pytorch_model.bin")

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a7ec0e38e7550b6a3660412cba5f4b69a62c06ae54325265ed94123c0a58e7f
+size 13425343241

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "cls_token": "</s>"}

tokenizer_config.json

@@ -0,0 +1,2 @@
+{"unk_token": "<unk>", "eos_token": "</s>", "bos_token": "<s>", "tokenizer_class": "PreTrainedTokenizerFast"}
+