some users report that this repo is now being flagged as malicious?

no idea why, but I am removing all prebuilt binaries except libopenblas. windows users can still obtain it from /releases and osx and linux users can rebuild from source code.

koboldcpp.py

@@ -302,6 +302,12 @@ def main(args):
     elif not args.noblas:
         print("Attempting to use OpenBLAS library for faster prompt ingestion. A compatible libopenblas.dll will be required.")
         use_blas = True
+
+    if args.psutil_set_threads:
+        import psutil
+        args.threads = psutil.cpu_count(logical=False)
+        print("Overriding thread count, using " + str(args.threads) + " threads instead.")
+
     init_library() # Note: if blas does not exist and is enabled, program will crash.
     ggml_selected_file = args.model_file
     embedded_kailite = None 
@@ -363,6 +369,7 @@ if __name__ == '__main__':
         physical_core_limit = int(os.cpu_count()/2)
     default_threads = (physical_core_limit if physical_core_limit<=3 else max(3,physical_core_limit-1))
     parser.add_argument("--threads", help="Use a custom number of threads if specified. Otherwise, uses an amount based on CPU cores", type=int, default=default_threads)
+    parser.add_argument("--psutil_set_threads", help="Experimental flag. If set, uses psutils to determine thread count based on physical cores.", action='store_true')
     parser.add_argument("--stream", help="Uses pseudo streaming", action='store_true')
     parser.add_argument("--noblas", help="Do not use OpenBLAS for accelerated prompt ingestion", action='store_true')
     args = parser.parse_args()

otherarch/convert_hf_gptj.py

@@ -0,0 +1,173 @@
+# Convert GPT-J-6B h5 transformer model to ggml format
+#
+# Load the model using GPTJForCausalLM.
+# 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 torch
+import numpy as np
+
+from transformers import GPTJForCausalLM
+
+# 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) < 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 + "/vocab.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
+    encoder_added = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# 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"
+
+
+model = GPTJForCausalLM.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", ftype))
+
+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_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 efficiency - transpose these matrices:
+    # (note - with latest ggml this is no longer more efficient, so disabling it)
+    #  "transformer.h.*.mlp.fc_in.weight"
+    #  "transformer.h.*.attn.out_proj.weight"
+    #  "transformer.h.*.attn.q_proj.weight"
+    #  "transformer.h.*.attn.k_proj.weight"
+    #  "transformer.h.*.attn.v_proj.weight"
+    #if name.endswith(".mlp.fc_in.weight")     or \
+    #   name.endswith(".attn.out_proj.weight") or \
+    #   name.endswith(".attn.q_proj.weight")   or \
+    #   name.endswith(".attn.k_proj.weight")   or \
+    #   name.endswith(".attn.v_proj.weight"):
+    #    print("  Transposing")
+    #    data = data.transpose()
+
+    # 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("")

otherarch/gptj_v2.cpp

@@ -143,7 +143,6 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
 
         ctx_size += (5 + 10*n_layer)*256; // object overhead
 
-        ctx_size = ctx_size * 3 / 2;
         printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
     }