+ add gpt2 convert tool

.github/workflows/sync-code.yml

@@ -0,0 +1,23 @@
+name: 同步代码到七牛
+
+on:
+  push:
+    branches: [ main ]
+
+jobs:
+  sync_data:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+      - name: 执行同步脚本
+        if: ${{  steps.parse_body.outputs.sync_type == 'huggingface' }}
+        run: |
+          pip install -r requirements.txt
+          for d in tools; do
+            bash ./sync_hf.sh "$d" "playground/$d"
+          done
+        env:
+          QUNIU_ACCESS_TOKEN: ${{ secrets.QUNIU_ACCESS_TOKEN }}
+          QUNIU_SECRET_KEY: ${{ secrets.QUNIU_SECRET_KEY }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}

.gitpod.yml

@@ -0,0 +1,10 @@
+# This configuration file was automatically generated by Gitpod.
+# Please adjust to your needs (see https://www.gitpod.io/docs/introduction/learn-gitpod/gitpod-yaml)
+# and commit this file to your remote git repository to share the goodness with others.
+
+# Learn more from ready-to-use templates: https://www.gitpod.io/docs/introduction/getting-started/quickstart
+
+tasks:
+  - init: pip install -r requirements.txt
+
+

tools/megatron/convert_megatron_gpt2_to_pytorch.py

@@ -0,0 +1,314 @@
+import argparse
+import os
+import re
+import zipfile
+
+import torch
+
+from transformers import AutoTokenizer, GPT2Config
+
+
+def recursive_print(name, val, spaces=0):
+    # Format the message.
+    if name is None:
+        msg = None
+    else:
+        fmt = "." * max(0, spaces - 2) + "# {:" + str(50 - spaces) + "s}"
+        msg = fmt.format(name)
+
+    # Print and recurse (if needed).
+    if isinstance(val, dict):
+        if msg is not None:
+            print(msg)
+        for k in val.keys():
+            recursive_print(k, val[k], spaces + 2)
+    elif isinstance(val, torch.Tensor):
+        print(msg, ":", val.size())
+    else:
+        print(msg, ":", val)
+
+
+def fix_query_key_value_ordering(param, checkpoint_version, num_splits, num_heads, hidden_size):
+    # Permutes layout of param tensor to [num_splits * num_heads * hidden_size, :]
+    # for compatibility with later versions of NVIDIA Megatron-LM.
+    # The inverse operation is performed inside Megatron-LM to read checkpoints:
+    # https://github.com/NVIDIA/Megatron-LM/blob/v2.4/megatron/checkpointing.py#L209
+    # If param is the weight tensor of the self-attention block, the returned tensor
+    # will have to be transposed one more time to be read by HuggingFace GPT2.
+    input_shape = param.size()
+    if checkpoint_version == 1.0:
+        # version 1.0 stores [num_heads * hidden_size * num_splits, :]
+        saved_shape = (num_heads, hidden_size, num_splits) + input_shape[1:]
+        param = param.view(*saved_shape)
+        param = param.transpose(0, 2)
+        param = param.transpose(1, 2).contiguous()
+    elif checkpoint_version >= 2.0:
+        # other versions store [num_heads * num_splits * hidden_size, :]
+        saved_shape = (num_heads, num_splits, hidden_size) + input_shape[1:]
+        param = param.view(*saved_shape)
+        param = param.transpose(0, 1).contiguous()
+    param = param.view(*input_shape)
+    return param
+
+
+####################################################################################################
+
+
+def convert_megatron_checkpoint(args, input_state_dict, config):
+    # The converted output model.
+    output_state_dict = {}
+
+    # old versions did not store training args
+    ds_args = input_state_dict.get("args", None)
+    if ds_args is not None:
+        # do not make the user write a config file when the exact dimensions/sizes are already in the checkpoint
+        # from pprint import pprint
+        # pprint(vars(ds_args))
+
+        config.vocab_size = ds_args.padded_vocab_size
+        config.n_positions = ds_args.max_position_embeddings
+        config.n_embd = ds_args.hidden_size
+        config.n_layer = ds_args.num_layers
+        config.n_head = ds_args.num_attention_heads
+        config.n_inner = ds_args.ffn_hidden_size
+        # pprint(config)
+
+    # The number of heads.
+    heads = config.n_head
+    # The hidden_size per head.
+    hidden_size_per_head = config.n_embd // config.n_head
+    # Megatron-LM checkpoint version
+    if "checkpoint_version" in input_state_dict.keys():
+        checkpoint_version = input_state_dict["checkpoint_version"]
+    else:
+        checkpoint_version = 0.0
+
+    # The model.
+    model = input_state_dict["model"]
+    # The language model.
+    lm = model["language_model"]
+    # The embeddings.
+    embeddings = lm["embedding"]
+
+    # The word embeddings.
+    word_embeddings = embeddings["word_embeddings"]["weight"]
+    # Truncate the embedding table to vocab_size rows.
+    word_embeddings = word_embeddings[: config.vocab_size, :]
+    output_state_dict["transformer.wte.weight"] = word_embeddings
+
+    # The position embeddings.
+    pos_embeddings = embeddings["position_embeddings"]["weight"]
+    # Read the causal mask dimension (seqlen). [max_sequence_length, hidden_size]
+    n_positions = pos_embeddings.size(0)
+    if n_positions != config.n_positions:
+        raise ValueError(
+            f"pos_embeddings.max_sequence_length={n_positions} and config.n_positions={config.n_positions} don't match"
+        )
+    # Store the position embeddings.
+    output_state_dict["transformer.wpe.weight"] = pos_embeddings
+
+    # The transformer.
+    transformer = lm["transformer"] if "transformer" in lm.keys() else lm["encoder"]
+
+    # The regex to extract layer names.
+    layer_re = re.compile(r"layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+
+    # The simple map of names for "automated" rules.
+    megatron_to_transformers = {
+        "attention.dense": ".attn.c_proj.",
+        "self_attention.dense": ".attn.c_proj.",
+        "mlp.dense_h_to_4h": ".mlp.c_fc.",
+        "mlp.dense_4h_to_h": ".mlp.c_proj.",
+    }
+
+    # Extract the layers.
+    for key, val in transformer.items():
+        # Match the name.
+        m = layer_re.match(key)
+
+        # Stop if that's not a layer
+        if m is None:
+            break
+
+        # The index of the layer.
+        layer_idx = int(m.group(1))
+        # The name of the operation.
+        op_name = m.group(2)
+        # Is it a weight or a bias?
+        weight_or_bias = m.group(3)
+
+        # The name of the layer.
+        layer_name = f"transformer.h.{layer_idx}"
+
+        # For layernorm(s), simply store the layer norm.
+        if op_name.endswith("_norm"):
+            ln_name = "ln_1" if op_name.startswith("input") else "ln_2"
+            output_state_dict[layer_name + "." + ln_name + "." + weight_or_bias] = val
+
+        # Transpose the QKV matrix.
+        elif (
+            op_name == "attention.query_key_value" or op_name == "self_attention.query_key_value"
+        ) and weight_or_bias == "weight":
+            # Insert a tensor of 1x1xDxD bias.
+            causal_mask = torch.tril(torch.ones((n_positions, n_positions), dtype=torch.float16)).view(
+                1, 1, n_positions, n_positions
+            )
+            output_state_dict[layer_name + ".attn.bias"] = causal_mask
+
+            # Insert a "dummy" tensor for masked_bias.
+            masked_bias = torch.tensor(-1e4, dtype=torch.float16)
+            output_state_dict[layer_name + ".attn.masked_bias"] = masked_bias
+
+            out_val = fix_query_key_value_ordering(val, checkpoint_version, 3, heads, hidden_size_per_head)
+            # Megatron stores (3*D) x D but transformers-GPT2 expects D x 3*D.
+            out_val = out_val.transpose(0, 1).contiguous()
+            # Store.
+            output_state_dict[layer_name + ".attn.c_attn.weight"] = out_val
+
+        # Transpose the bias.
+        elif (
+            op_name == "attention.query_key_value" or op_name == "self_attention.query_key_value"
+        ) and weight_or_bias == "bias":
+            out_val = fix_query_key_value_ordering(val, checkpoint_version, 3, heads, hidden_size_per_head)
+            # Store. No change of shape.
+            output_state_dict[layer_name + ".attn.c_attn.bias"] = out_val
+
+        # Transpose the weights.
+        elif weight_or_bias == "weight":
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + "weight"] = val.transpose(0, 1)
+
+        # Copy the bias.
+        elif weight_or_bias == "bias":
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + "bias"] = val
+
+    # DEBUG.
+    assert config.n_layer == layer_idx + 1
+
+    # The final layernorm.
+    output_state_dict["transformer.ln_f.weight"] = transformer["final_norm.weight"]
+    output_state_dict["transformer.ln_f.bias"] = transformer["final_norm.bias"]
+
+    # For LM head, transformers' wants the matrix to weight embeddings.
+    output_state_dict["lm_head.weight"] = word_embeddings
+
+    # It should be done!
+    return output_state_dict
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--print-checkpoint-structure", action="store_true")
+    parser.add_argument(
+        "path_to_checkpoint",
+        type=str,
+        help="Path to the checkpoint file (.zip archive or direct .pt file)",
+    )
+    parser.add_argument(
+        "output",
+        type=str,
+        help="Path to the pytorch model save",
+    )
+    parser.add_argument(
+        "--config_file",
+        default="",
+        type=str,
+        help="An optional config json file describing the pre-trained model.",
+    )
+    args = parser.parse_args()
+    os.makedirs(args.output, exist_ok=True)
+
+    # Extract the basename.
+    basename = args.output
+
+    # Load the model.
+    # the .zip is very optional, let's keep it for backward compatibility
+    print(f"Extracting PyTorch state dictionary from {args.path_to_checkpoint}")
+    if args.path_to_checkpoint.endswith(".zip"):
+        with zipfile.ZipFile(args.path_to_checkpoint, "r") as checkpoint:
+            with checkpoint.open("release/mp_rank_00/model_optim_rng.pt") as pytorch_dict:
+                input_state_dict = torch.load(pytorch_dict, map_location="cpu")
+    else:
+        input_state_dict = torch.load(args.path_to_checkpoint, map_location="cpu")
+
+    ds_args = input_state_dict.get("args", None)
+
+    # Read the config, or default to the model released by NVIDIA.
+    if args.config_file == "":
+        if ds_args is not None:
+            if ds_args.bias_gelu_fusion:
+                activation_function = "gelu_fast"
+            elif ds_args.openai_gelu:
+                activation_function = "gelu_new"
+            else:
+                activation_function = "gelu"
+        else:
+            # in the very early days this used to be "gelu_new"
+            activation_function = "gelu_new"
+
+        # Spell out all parameters in case the defaults change.
+        config = GPT2Config(
+            vocab_size=50257,
+            n_positions=1024,
+            n_embd=1024,
+            n_layer=24,
+            n_head=16,
+            n_inner=4096,
+            activation_function=activation_function,
+            resid_pdrop=0.1,
+            embd_pdrop=0.1,
+            attn_pdrop=0.1,
+            layer_norm_epsilon=1e-5,
+            initializer_range=0.02,
+            summary_type="cls_index",
+            summary_use_proj=True,
+            summary_activation=None,
+            summary_proj_to_labels=True,
+            summary_first_dropout=0.1,
+            scale_attn_weights=True,
+            use_cache=True,
+            bos_token_id=50256,
+            eos_token_id=50256,
+        )
+    else:
+        config = GPT2Config.from_json_file(args.config_file)
+
+    config.architectures = ["GPT2LMHeadModel"]
+
+    # Convert.
+    print("Converting...")
+
+    output_state_dict = convert_megatron_checkpoint(args, input_state_dict, config)
+
+    # Print the structure of converted state dict.
+    if args.print_checkpoint_structure:
+        recursive_print(None, output_state_dict)
+
+    # Add tokenizer class info to config
+    # see https://github.com/huggingface/transformers/issues/13906)
+    if ds_args is not None:
+        tokenizer_type = ds_args.tokenizer_type
+        if tokenizer_type == "GPT2BPETokenizer":
+            tokenizer_model_name = "openai-community/gpt2"
+        elif tokenizer_type == "PretrainedFromHF":
+            tokenizer_model_name = ds_args.tokenizer_name_or_path
+        else:
+            raise ValueError(f"Unrecognized tokenizer_type {tokenizer_type}")
+    else:
+        tokenizer_model_name = "openai-community/gpt2"
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_model_name)
+    tokenizer_class = type(tokenizer).__name__
+    config.tokenizer_class = tokenizer_class
+
+    # Store the config to file.
+    print("Saving config")
+    config.save_pretrained(basename)
+
+    # Save tokenizer based on args
+    print(f"Adding {tokenizer_class} tokenizer files")
+    tokenizer.save_pretrained(basename)
+
+    # Store the state_dict to file.
+    output_checkpoint_file = os.path.join(basename, "pytorch_model.bin")
+    print(f'Saving checkpoint to "{output_checkpoint_file}"')
+    torch.save(output_state_dict, output_checkpoint_file)