From 24f48833ab5a1fcacb15c1908e13a80b35fa0f31 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?M=2E=20Yusuf=20Sar=C4=B1g=C3=B6z?= Date: Sun, 13 Aug 2023 16:55:42 +0300 Subject: [PATCH] fix conflicts --- convert-llama-h5-to-gguf.py | 82 +++++++++++++++++-------------------- 1 file changed, 38 insertions(+), 44 deletions(-) diff --git a/convert-llama-h5-to-gguf.py b/convert-llama-h5-to-gguf.py index 574788ee0..cf9f6f802 100644 --- a/convert-llama-h5-to-gguf.py +++ b/convert-llama-h5-to-gguf.py @@ -23,7 +23,6 @@ def permute(weights: NDArray, n_head: int) -> NDArray: .swapaxes(1, 2) .reshape(weights.shape)) - def count_model_parts(dir_model: str) -> int: num_parts = 0 for filename in os.listdir(dir_model): @@ -34,7 +33,6 @@ def count_model_parts(dir_model: str) -> int: print("gguf: found " + str(num_parts) + " model parts") return num_parts - if len(sys.argv) < 3: print("Usage: convert-h5-to-ggml.py dir-model ftype\n") print(" ftype == 0 -> float32") @@ -188,7 +186,7 @@ else: ) for part_name in part_names: - print("gguf: loading model part '" + part_name + "'") + print("gguf: loading model part '"+ part_name + "'") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") for name in model_part.keys(): @@ -206,7 +204,7 @@ for part_name in part_names: # permute these if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"): - data = permute(data, head_count) + data = permute(data,head_count) # map tensor names if name.endswith(".weight") and name[:-7] in tensor_map: @@ -214,11 +212,11 @@ for part_name in part_names: elif name.endswith(".bias") and name[:-5] in tensor_map: name = tensor_map[name[:-5]] + ".bias" else: - print("Can not map tensor '" + name + "'") + print( "Can not map tensor '" + name + "'" ) sys.exit() n_dims = len(data.shape) - data_dtype = data.dtype + data_dtype = data.dtype # if f32 desired, convert any float16 to float32 if ftype == 0 and data.dtype == np.float16: @@ -255,60 +253,56 @@ else: ) for part_name in part_names: - print("gguf: loading model part '" + part_name + "'") + print("gguf: loading model part '"+ part_name + "'") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") for name in model_part.keys(): data = model_part[name] -<< << << < HEAD - n_dims = len(data.shape) - data_dtype = data.dtype -== == == = - old_dtype = data.dtype + + old_dtype = data.dtype - # we don't need these - if name.endswith(".rotary_emb.inv_freq"): - continue ->>>>>> > 17800cd80fec468411481dc34a51d42a936442f1 + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) + # convert any unsupported data types to float32 + if data.dtype != torch.float16 and data.dtype != torch.float32: + data = data.to(torch.float32) - data = data.squeeze().numpy() + data = data.squeeze().numpy() - # permute these - if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"): - data = permute(data, head_count) + # permute these + if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"): + data = permute(data, head_count) - # map tensor names - if name.endswith(".weight") and name[:-7] in tensor_map: - name = tensor_map[name[:-7]] + ".weight" - elif name.endswith(".bias") and name[:-5] in tensor_map: - name = tensor_map[name[:-5]] + ".bias" - else: - print("Can not map tensor '" + name + "'" ) - sys.exit() + # map tensor names + if name.endswith(".weight") and name[:-7] in tensor_map: + name = tensor_map[name[:-7]] + ".weight" + elif name.endswith(".bias") and name[:-5] in tensor_map: + name = tensor_map[name[:-5]] + ".bias" + else: + print( "Can not map tensor '" + name + "'" ) + sys.exit() - n_dims = len(data.shape) - data_dtype = data.dtype + n_dims = len(data.shape) + data_dtype = data.dtype - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data.dtype == np.float16: - data = data.astype(np.float32) + # if f32 desired, convert any float16 to float32 + if ftype == 0 and data.dtype == np.float16: + data = data.astype(np.float32) - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if ftype == 1 and data_dtype == np.float16 and n_dims == 1: + 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: - data = data.astype(np.float16) + # 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: + data = data.astype(np.float16) - print(name + ", shape " + str(len(data.shape)) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + print( name + ", shape " + str(len(data.shape)) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.write_tensor_to_file(data) + gguf_writer.write_tensor_to_file(data) gguf_writer.close()