]> git.djapps.eu Git - pkg/ggml/sources/ggml/commitdiff
gpt-2 : convert h5 to ggml (#35)
authorCordeiro <redacted>
Wed, 29 Mar 2023 20:39:27 +0000 (15:39 -0500)
committerGitHub <redacted>
Wed, 29 Mar 2023 20:39:27 +0000 (23:39 +0300)
* Script to convert h5 to ggml adapted from gpt-j example

* Fix map tensors

* optimize

* rename headers to keep compatibility

* revert gpt-2/main.cpp

---------

Co-authored-by: Alan <redacted>
Co-authored-by: Alan <redacted>
Co-authored-by: ocordeiro <redacted>
examples/gpt-2/convert-h5-to-ggml.py [new file with mode: 0644]

diff --git a/examples/gpt-2/convert-h5-to-ggml.py b/examples/gpt-2/convert-h5-to-ggml.py
new file mode 100644 (file)
index 0000000..7ee3395
--- /dev/null
@@ -0,0 +1,195 @@
+# Convert GPT-2 h5 transformer model to ggml format
+#
+# Load the model using GPT2Model.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import struct
+import json
+import numpy as np
+import re
+
+from transformers import GPT2Model
+
+# 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 signficant 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) < 2:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    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 + "/vocab.json", "r") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/added_tokens.json", "r") as f:
+    encoder_added = json.load(f)
+
+with open(dir_model + "/config.json", "r") as f:
+    hparams = json.load(f)
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 2:
+    use_f16 = False
+    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+
+model = GPT2Model.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+#fout.write(struct.pack("i", hparams["rotary_dim"]))
+fout.write(struct.pack("i", use_f16))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for key in encoder_added:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 0;
+    if use_f16:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+
+    # for efficiency - transpose these matrices:
+    #  "transformer.h.*.mlp.c_proj.weight
+    if name.endswith(".mlp.c_proj.weight"):
+        print("  Transposing")
+        data = data.transpose()
+
+    # rename headers to keep compatibility
+    if name == "ln_f.weight":
+        name = "model/ln_f/g"
+    elif name == "ln_f.bias":
+        name = "model/ln_f/b"
+    elif name == "wte.weight":
+        name = "model/wte"
+    elif name == "wpe.weight":
+        name = "model/wpe"
+    elif re.match(r'h\.\d+\.ln_1\.weight', name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/g"
+    elif re.match(r"h\.\d+\.ln_1\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/b"
+    elif re.match(r"h\.\d+\.attn\.c_attn\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/w"
+    elif re.match(r"h\.\d+\.attn\.c_attn\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/b"
+    elif re.match(r"h\.\d+\.attn\.c_proj\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/w"
+    elif re.match(r"h.\d+.attn.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/b"
+    elif re.match(r"h.\d+.ln_2.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/g"
+    elif re.match(r"h.\d+.ln_2.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/b"
+    elif re.match(r"h.\d+.mlp.c_fc.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/w"
+    elif re.match(r"h.\d+.mlp.c_fc.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/b"
+    elif re.match(r"h.\d+.mlp.c_proj.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/w"
+    elif re.match(r"h.\d+.mlp.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/b"
+    else:
+        print("Unrecognized variable name. %s", name)
+
+    str = name.encode('utf-8')
+
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    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("")
\ No newline at end of file