streaming conversion without pytorch

convert-pth-to-ggml.py

@@ -17,11 +17,16 @@
 # and vocabulary.
 #
 
+from collections import defaultdict
 import sys
 import json
 import struct
 import numpy as np
-import torch
+from tqdm import tqdm
+import zipfile
+import pickle
+import concurrent.futures
+
 from sentencepiece import SentencePieceProcessor
 
 if len(sys.argv) < 3:
@@ -73,19 +78,22 @@ hparams.update({"vocab_size": tokenizer.vocab_size()})
 
 n_parts = get_n_parts(hparams["dim"])
 
-print(hparams)
+print(f'Model params.json: {hparams}')
-print('n_parts = ', n_parts)
+print(f'Parts to process: {n_parts}')
 
-for p in range(n_parts):
-    print('Processing part ', p)
 
-    #fname_model = sys.argv[1] + "/consolidated.00.pth"
+def get_fname(p):
-    fname_model = sys.argv[1] + "/consolidated.0" + str(p) + ".pth"
+    fname = "/consolidated.0" + str(p) + ".pth"
+    return fname
+
+def process_part(p):
+    fname = get_fname(p)
+    fname_model = sys.argv[1] + fname
     fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
     if (p > 0):
         fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin" + "." + str(p)
 
-    model = torch.load(fname_model, map_location="cpu")
+    print(f"Processing part {fname}")
 
     fout = open(fname_out, "wb")
 
@@ -123,7 +131,58 @@ for p in range(n_parts):
             fout.write(struct.pack("i", len(text)))
             fout.write(text)
 
-    for k, v in model.items():
+
+    def load_model(fname):
+        class Tensor():
+            def __init__(self, shape, dtype, loadinfo):
+                self.shape = shape
+                self.dtype = dtype
+                self.loadinfo = loadinfo
+                # print(shape, dtype)
+
+            def numpy(self):
+                fname_model, base_name, storage_offset, k, shape, dtype = self.loadinfo
+                with zipfile.ZipFile(fname_model, 'r') as myzip:
+                    with myzip.open(f'{base_name}/data/{k}') as myfile:
+                        bytes_size = np.dtype(self.dtype).itemsize
+                        myfile.seek(storage_offset * bytes_size, 1)
+                        ret = np.empty(shape, dtype=dtype)
+                        myfile.readinto(ret.data)
+                        return ret
+
+        def my_unpickle(datapkl, fname_model, base_name):
+            def my_rebuild_tensor(storage, storage_offset, size, stride, requires_grad, backward_hooks, metadata=None):
+                storage_type = storage[1]
+                obj_key = storage[2]
+                return Tensor(shape=size, dtype=storage_type, loadinfo=(
+                    fname_model, base_name, storage_offset,
+                    obj_key, size, storage_type
+                ))
+
+            class MyUnpickler(pickle.Unpickler):
+                def find_class(self, *p):
+                    if p == ('torch', 'HalfStorage'): return np.float16
+                    if p == ('torch', 'FloatStorage'): return np.float32
+                    if p == ('torch._utils', '_rebuild_tensor_v2'): return my_rebuild_tensor
+                    if p == ('collections', 'OrderedDict'): return dict
+                    raise ValueError(f'Unrecognized pickle {p}')
+
+                def persistent_load(self, pid):
+                    return pid
+
+            return MyUnpickler(datapkl).load()
+
+        with zipfile.ZipFile(fname, 'r') as myzip:
+            base_name = myzip.namelist()[0].split('/', 1)[0]
+            # print(myzip.namelist())
+            with myzip.open(f'{base_name}/data.pkl') as myfile:
+                model = my_unpickle(myfile, fname, base_name)
+        return model
+
+    model = load_model(fname_model)
+
+    for k, v in (t := tqdm(model.items())):
+        t.set_description(f"Processing {k} with shape {tuple(v.shape)} and type {np.dtype(v.dtype)}")
         name = k
         shape = v.shape
 
@@ -131,11 +190,11 @@ for p in range(n_parts):
         if name[-5:] == "freqs":
             continue
 
-        print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
+        # print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
 
         #data = tf.train.load_variable(dir_model, name).squeeze()
         data = v.numpy().squeeze()
-        n_dims = len(data.shape);
+        n_dims = len(data.shape)
 
         # for efficiency - transpose some matrices
         # "model/h.*/attn/c_attn/w"
@@ -154,7 +213,7 @@ for p in range(n_parts):
         # default type is fp16
         ftype_cur = 1
         if ftype == 0 or n_dims == 1:
-            print("  Converting to float32")
+            # print("  Converting to float32")
             data = data.astype(np.float32)
             ftype_cur = 0
 
@@ -163,7 +222,7 @@ for p in range(n_parts):
         fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur))
         for i in range(n_dims):
             fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
-        fout.write(sname);
+        fout.write(sname)
 
         # data
         data.tofile(fout)
@@ -175,3 +234,10 @@ for p in range(n_parts):
 
     print("Done. Output file: " + fname_out + ", (part ", p, ")")
     print("")
+
+with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
+    futures = {executor.submit(process_part, p) for p in range(n_parts)}
+    for f in (concurrent.futures.as_completed(futures)):
+        if f.exception() is not None: raise f.exception()
+
+print("All done.")