From 16dba2ba11043143e4f6c9301b563987145d0116 Mon Sep 17 00:00:00 2001 From: ds5t5 Date: Mon, 25 Sep 2023 16:03:14 -0700 Subject: [PATCH] resolve comments --- convert-refact-hf-to-gguf.py | 82 ++++++++++++++++++++++++------------ llama.cpp | 2 +- 2 files changed, 57 insertions(+), 27 deletions(-) diff --git a/convert-refact-hf-to-gguf.py b/convert-refact-hf-to-gguf.py index 892b470c47c419..5a876c248cb3a9 100755 --- a/convert-refact-hf-to-gguf.py +++ b/convert-refact-hf-to-gguf.py @@ -1,5 +1,5 @@ #!/usr/bin/env python3 -# HF falcon--> gguf conversion +# HF refact--> gguf conversion from __future__ import annotations @@ -15,8 +15,8 @@ import torch from transformers import AutoTokenizer # type: ignore[import] -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) +if "NO_LOCAL_GGUF" not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf")) import gguf @@ -31,13 +31,17 @@ def bytes_to_unicode(): 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)) + 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) + cs.append(2**8 + n) n += 1 return dict(zip(bs, (chr(n) for n in cs))) @@ -54,32 +58,41 @@ def count_model_parts(dir_model: Path) -> int: def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a Refact model to a GGML compatible file") + parser = argparse.ArgumentParser( + description="Convert a Refact model to a GGML compatible file" + ) parser.add_argument( - "--vocab-only", action="store_true", + "--vocab-only", + action="store_true", help="extract only the vocab", ) parser.add_argument( - "--outfile", type=Path, + "--outfile", + type=Path, help="path to write to; default: based on input", ) parser.add_argument( - "model", type=Path, + "model", + type=Path, help="directory containing model file, or model file itself (*.bin)", ) parser.add_argument( - "ftype", type=int, choices=[0, 1], default=1, nargs='?', + "ftype", + type=int, + choices=[0, 1], + default=1, + nargs="?", help="output format - use 0 for float32, 1 for float16", ) return parser.parse_args() + args = parse_args() dir_model = args.model ftype = args.ftype if not dir_model.is_dir(): - - print(f'Error: {args.model} is not a directory', file = sys.stderr) + print(f"Error: {args.model} is not a directory", file=sys.stderr) sys.exit(1) # possible tensor data types @@ -93,9 +106,9 @@ def parse_args() -> argparse.Namespace: fname_out = args.outfile else: # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' + fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf" -print("gguf: loading model "+dir_model.name) +print("gguf: loading model " + dir_model.name) with open(dir_model / "config.json", "r", encoding="utf-8") as f: hparams = json.load(f) @@ -108,7 +121,7 @@ def parse_args() -> argparse.Namespace: # get number of model parts num_parts = count_model_parts(dir_model) -ARCH=gguf.MODEL_ARCH.REFACT +ARCH = gguf.MODEL_ARCH.REFACT gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) print("gguf: get model metadata") @@ -142,9 +155,9 @@ def parse_args() -> argparse.Namespace: scores: list[float] = [] toktypes: list[int] = [] -tokenizer_json_file = dir_model / 'tokenizer.json' +tokenizer_json_file = dir_model / "tokenizer.json" if not tokenizer_json_file.is_file(): - print(f'Error: Missing {tokenizer_json_file}', file = sys.stderr) + print(f"Error: Missing {tokenizer_json_file}", file=sys.stderr) sys.exit(1) # gpt2 tokenizer @@ -157,7 +170,11 @@ def parse_args() -> argparse.Namespace: # The number of tokens in tokenizer.json can differ from the expected vocab size. # This causes downstream issues with mismatched tensor sizes when running the inference -vocab_size = hparams["vocab_size"] if "vocab_size" in hparams else len(tokenizer_json["model"]["vocab"]) +vocab_size = ( + hparams["vocab_size"] + if "vocab_size" in hparams + else len(tokenizer_json["model"]["vocab"]) +) tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) @@ -176,29 +193,29 @@ def parse_args() -> argparse.Namespace: if ord(c) < 256: # single byte character text.append(byte_decoder[ord(c)]) else: # multibyte special token character - text.extend(c.encode('utf-8')) + text.extend(c.encode("utf-8")) else: print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.") pad_token = f"[PAD{i}]".encode("utf8") text = bytearray(pad_token) tokens.append(text) - scores.append(0.0) # dymmy + scores.append(0.0) # dymmy toktypes.append(gguf.TokenType.NORMAL) # dummy gguf_writer.add_token_list(tokens) gguf_writer.add_token_scores(scores) gguf_writer.add_token_types(toktypes) -special_vocab = gguf.SpecialVocab(dir_model, load_merges = True) +special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) special_vocab.add_to_gguf(gguf_writer) # TENSORS -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) +tensor_map = gguf.get_tensor_name_map(ARCH, block_count) # params for qkv transform -n_head = hparams["n_head"] +n_head = hparams["n_head"] n_head_kv = 1 head_dim = hparams["n_embd"] // n_head @@ -230,7 +247,7 @@ def parse_args() -> argparse.Namespace: data = data.squeeze().numpy() # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", )) + new_name = tensor_map.get_name(name, try_suffixes=(".weight",)) if new_name is None: print("Can not map tensor '" + name + "'") sys.exit() @@ -247,10 +264,23 @@ def parse_args() -> argparse.Namespace: data = data.astype(np.float32) # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + if ( + ftype == 1 + and data_dtype == np.float32 + and name.endswith(".weight") + and n_dims == 2 + ): data = data.astype(np.float16) - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + print( + new_name + + ", n_dims = " + + str(n_dims) + + ", " + + str(old_dtype) + + " --> " + + str(data.dtype) + ) gguf_writer.add_tensor(new_name, data) diff --git a/llama.cpp b/llama.cpp index ffd6e9ad239849..045a42baca3f2f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4129,7 +4129,7 @@ static bool llama_eval_internal( // If all tensors can be run on the GPU then using more than 1 thread is detrimental. const bool full_offload_supported = model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_BAICHUAN || - model.arch == LLM_ARCH_FALCON || + model.arch == LLM_ARCH_FALCON || model.arch == LLM_ARCH_REFACT; const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3; if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) {