feat: converter scripts from hf

gpt4all-quantizer/converter/convert_bert_hf_to_ggml.py

@@ -0,0 +1,101 @@
+import sys
+import struct
+import json
+import numpy as np
+
+from transformers import AutoModel, AutoTokenizer
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+with open(dir_model + "/vocab.txt", "r", encoding="utf-8") as f:
+    vocab = f.readlines()
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+model = AutoModel.from_pretrained(dir_model, low_cpu_mem_usage=True)
+print (model)
+
+print(tokenizer.encode('I believe the meaning of life is'))
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+fout = open(fname_out, "wb")
+
+print(hparams)
+
+fout.write(struct.pack("i", 0x62657274)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["max_position_embeddings"]))
+fout.write(struct.pack("i", hparams["hidden_size"]))
+fout.write(struct.pack("i", hparams["intermediate_size"]))
+fout.write(struct.pack("i", hparams["num_attention_heads"]))
+fout.write(struct.pack("i", hparams["num_hidden_layers"]))
+fout.write(struct.pack("i", ftype))
+
+for i in range(hparams["vocab_size"]):
+    text = vocab[i][:-1] # strips newline at the end
+    #print(f"{i}:{text}")
+    data = bytes(text, 'utf-8')
+    fout.write(struct.pack("i", len(data)))
+    fout.write(data)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    if name in ['embeddings.position_ids', 'pooler.dense.weight', 'pooler.dense.bias']:
+        continue
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    if ftype == 1 and name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            l_type = 1
+    else:
+        l_type = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), l_type))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

gpt4all-quantizer/converter/convert_falcon_hf_to_ggml.py

@@ -0,0 +1,143 @@
+# Based on: https://github.com/KerfuffleV2/ggml-falcon/blob/feat-improve-falcon-convert-hf/examples/falcon/convert-hf-to-ggml.py
+# Convert Hugging Face fine-tuned bloom-like models to ggml format
+#
+# Usage:
+#
+#   python3 convert_falcon_hf_to_ggml.py model_directory output_directory [use-f32]
+#
+# This script is similar to "convert-pt-to-ggml.py"
+#
+
+import io
+import os
+import sys
+import struct
+import json
+import code
+import torch
+import numpy as np
+import gc
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 3:
+    print("INFO: GGML V1 files produced are meant to be finalized through examples/falcon_quantize which will bring them to latest version and precision of choice");
+    print("Usage: python convert_falcon_hf_to_ggml.py model_directory output_directory [use-f32]")
+    print("  model_directory: name of the directory and model you convert (it should be a subdirectory)")
+    print("  output-directory: directory where the output file will be written")
+    print("  use-f32:    if present, use float32 instead of float16 (f32 is recommended)")
+    sys.exit(1)
+
+# num_parts = int(sys.argv[1])
+dir_model = sys.argv[1] # name and dir of model
+dir_out = sys.argv[2]   # output directory
+
+# make sure the output directory exists
+os.makedirs(dir_out, exist_ok=True)
+
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+ftype = 1
+if len(sys.argv) > 3:
+    ftype = 0
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+# print(tokenizer)
+config = AutoConfig.from_pretrained(dir_model, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(dir_model, trust_remote_code=True, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True)
+hparams = config.to_dict()
+
+n_head = hparams["n_head"]
+n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+head_dim = hparams["hidden_size"] // n_head
+print("* Loading model from: ", dir_model)
+
+fname_out = dir_out + f"/ggml-model-{dir_model.split('/')[-1]}-{ftype_str[ftype]}.bin"
+fout = open(fname_out, "wb")
+fout.write(struct.pack("i", 0x67676a74)) # magic: ggmf in hex (version 1) - possibly change to ggfc ?
+fout.write(struct.pack("i", 1)) # version
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["hidden_size"]))
+fout.write(struct.pack("i", n_head))
+fout.write(struct.pack("i", n_head_kv))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", 40 if "n_head_kv" in hparams else 7)) # obsolete field that breaks ggml compatibility - todo again remove one day
+fout.write(struct.pack("i", ftype))
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+for i in range(hparams["vocab_size"]):
+    text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+    fout.write(struct.pack("f", 0.0)) # falcon uses bpe on RefinedWeb - no probability scores used
+
+model = model.state_dict()
+for name in model.keys():
+    src = name
+    # The original query_key_value tensor contains n_head_kv "kv groups",
+    # each consisting of n_head/n_head_kv query weights followed by one key
+    # and one value weight (shared by all query heads in the kv group).
+    # This layout makes it a big pain to work with in GGML.
+    # So we rearrange them here,, so that we have n_head query weights
+    # followed by n_head_kv key weights followed by n_head_kv value weights,
+    # in contiguous fashion.
+
+    if "query_key_value" in src:
+        qkv = model[src].view(
+            n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+
+        q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
+        k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+        v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+
+        model[src] = torch.cat((q,k,v)).reshape_as(model[src])
+    data = model[src].squeeze()
+    n_dims = len(data.shape)
+    # default type is fp32
+    ftype_cur = 1 if ftype == 1 and n_dims > 1 else 0
+    data = data.to(dtype = torch.float16 if ftype_cur == 1 else torch.float32).numpy()
+    print(f'  |', name, data.shape, '->', data.dtype)
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str)
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

gpt4all-quantizer/converter/convert_mpt_hf_to_ggml.py

@@ -0,0 +1,142 @@
+# Convert Hugging Face fine-tuned bloom-like models to ggml format
+#
+# Usage:
+#
+#   python3 models/convert-h5-to-ggml.py 
+#
+# This script is similar to "convert-pt-to-ggml.py"
+#
+
+import os
+import sys
+import struct
+import torch
+import numpy as np
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 3:
+    print("Usage: python convert-hf-to-ggml.py model_name dir-output [use-f32]")
+    print("  model_name: name of the model to convert. Example: 'bigscience/bloomz-560m'")
+    print("  dir-output: directory where the output file will be written")
+    print("  use-f32:    if present, use float32 instead of float16")
+    sys.exit(1)
+
+model_name = sys.argv[1]
+dir_out = sys.argv[2]
+
+# make sure the output directory exists
+os.makedirs(dir_out, exist_ok=True)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+ftype = 1
+if len(sys.argv) > 3:
+    ftype = 0
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+hparams = config.to_dict()
+print("Loading model: ", model_name)
+model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, config=config, torch_dtype=torch.float16 if ftype == 1 else torch.float32, low_cpu_mem_usage=True)
+print("Model loaded: ", model_name)
+
+
+fname_out = dir_out + f"/ggml-model-{model_name.split('/')[-1]}-{ftype_str[ftype]}.bin"
+fout = open(fname_out, "wb")
+vocab = tokenizer.vocab
+
+hparams["multiple_of"] = 1
+fout.write(struct.pack("I", 0x67676d6d)) # magic: ggml in hex
+fout.write(struct.pack("I", model.config.vocab_size))
+fout.write(struct.pack("I", model.config.max_seq_len))
+fout.write(struct.pack("I", model.config.n_layers))
+fout.write(struct.pack("I", model.config.n_heads))
+fout.write(struct.pack("I", model.config.d_model))
+fout.write(struct.pack("f", model.config.attn_config['alibi_bias_max']))
+clip_qkv = model.config.attn_config['clip_qkv']
+fout.write(struct.pack("f",  clip_qkv if clip_qkv is not None else 0))
+fout.write(struct.pack("I", ftype))
+
+# # Is this correct??
+# dot_token = tokenizer.encode(".")[0]
+# write tokens to ggml file 
+dot_token = tokenizer.encode('.')[0]
+fout.write(struct.pack("I", model.config.vocab_size))
+
+for i in range(model.config.vocab_size):
+    text = tokenizer.decode([dot_token, i]).encode('utf-8')
+    # remove the first byte (it's always '.')
+    text = text[1:]
+    enclen = len(text)
+    if i in tokenizer.all_special_ids:
+        print(f"special token: {text}")
+        enclen = enclen | 1<<31
+    fout.write(struct.pack("I", enclen))
+    fout.write(text)
+    
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0;
+    if ftype != 0:
+        # Keep token embeddings in fp32
+        if name[-7:] == ".weight" and n_dims == 2 and ".wte" not in name:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

gpt4all-quantizer/converter/convert_replit_hf_to_ggml.py

@@ -0,0 +1,137 @@
+import os
+from pathlib import Path
+import sys
+import struct
+import json
+import numpy as np
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from huggingface_hub import hf_hub_download
+import sentencepiece.sentencepiece_model_pb2 as model
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py model_name out_dir [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+
+# output in the same directory as the model
+model_name = sys.argv[1]
+out_dir = sys.argv[2]
+if not os.path.exists(out_dir):
+    os.mkdir(out_dir)
+
+fname_out = sys.argv[2] + "/ggml-replit-code-v1-3b.bin"
+
+if not os.path.exists(model_name):
+    hf_hub_download(repo_id=model_name, filename="config.json", local_dir=out_dir)
+    hf_hub_download(repo_id=model_name, filename="spiece.model", local_dir=out_dir)
+else:
+    # copy file from model_name to out_dir
+    os.system("cp " + model_name + "/config.json " + out_dir)
+    os.system("cp " + model_name + "/spiece.model " + out_dir)
+
+
+with open(out_dir + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+sp_proto = model.ModelProto()
+sp_proto.ParseFromString(open(Path(out_dir) / "spiece.model", "rb").read())
+
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 3:
+    ftype = int(sys.argv[3])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-replit-code-v1-3b-" + ftype_str[ftype] + ".bin"
+
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_name, low_cpu_mem_usage=True, trust_remote_code=True
+)
+# print (model)
+
+# print(tokenizer.encode('I believe the meaning of life is'))
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+fout = open(fname_out, "wb")
+
+print(hparams)
+
+fout.write(struct.pack("i", 0x7265706c))  # magic: repl in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["max_seq_len"]))
+fout.write(struct.pack("i", hparams["d_model"]))
+fout.write(struct.pack("i", hparams["n_heads"]))
+fout.write(struct.pack("i", hparams["n_layers"]))
+fout.write(struct.pack("i", ftype))
+
+
+# TODO: temporary hack to not deal with implementing the tokenizer
+for piece in sp_proto.pieces:
+    encoded_piece = piece.piece.encode("utf-8")
+    fout.write(struct.pack("i", len(encoded_piece)))
+    fout.write(encoded_piece)
+    fout.write(struct.pack("f", piece.score))
+
+name_mapping = {
+    "norm_1": "ln_1",
+    "norm_2": "ln_2",
+    "ffn": "mlp",
+    "up_proj": "mlp_up",
+    "down_proj": "mlp_down",
+    "norm_f":  "ln_f"
+}
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if ftype != 0:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    for x in ["norm_1", "norm_2", "ffn", "up_proj", "down_proj", "norm_f"]:
+        if x in name:
+            name = name.replace(x, name_mapping[x])
+            print("  Renaming to: " + name)
+    # header
+    str = name.encode("utf-8")
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str)
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")