Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-07 11:23:56 +01:00
Code Issues Packages Projects Releases Wiki Activity

gguf : support custom alignment value

Browse Source
This commit is contained in:
M. Yusuf Sarıgöz 2023-07-31 09:59:36 +03:00
parent 6b3a7b9f4f
commit 7aa0a0e7f7
3 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
constants.py
View File

@ -5,6 +5,7 @@ GGUF_DEFAULT_ALIGNMENT = 32
# general # general
KEY_GENERAL_ARCHITECTURE = "general.architecture" KEY_GENERAL_ARCHITECTURE = "general.architecture"
KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version" KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version"
KEY_GENERAL_ALIGNMENT = "general.alignment"
KEY_GENERAL_NAME = "general.name" KEY_GENERAL_NAME = "general.name"
KEY_GENERAL_AUTHOR = "general.author" KEY_GENERAL_AUTHOR = "general.author"
KEY_GENERAL_URL = "general.url" KEY_GENERAL_URL = "general.url"

5
ggml.c
View File

@ -18543,7 +18543,10 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
ctx->alignment = GGUF_DEFAULT_ALIGNMENT; ctx->alignment = GGUF_DEFAULT_ALIGNMENT;
// TODO: determine new alignment from kv if available int alignment_idx = gguf_find_key(ctx, "general.alignment");
if (alignment_idx != -1) {
ctx->alignment = gguf_get_u32(ctx, alignment_idx);
}
// we require the data section to be aligned, so take into account any padding // we require the data section to be aligned, so take into account any padding
{ {

6
gguf.py
View File

@ -220,6 +220,9 @@ class GGUFWriter:
self.write_uint32( self.write_uint32(
constants.KEY_GENERAL_QUANTIZATION_VERSION, quantization_version) constants.KEY_GENERAL_QUANTIZATION_VERSION, quantization_version)
def write_custom_alignment(self, alignment: int):
self.write_uint32(constants.KEY_GENERAL_ALIGNMENT, alignment)
def write_context_length(self, llm: str, length: int): def write_context_length(self, llm: str, length: int):
self.write_uint32( self.write_uint32(
constants.KEY_LLM_CONTEXT_LENGTH.format(llm=llm), length) constants.KEY_LLM_CONTEXT_LENGTH.format(llm=llm), length)
@ -292,11 +295,12 @@ class GGUFWriter:
if __name__ == "__main__": if __name__ == "__main__":
# Example usage with a file # Example usage with a file
gguf_writer = GGUFWriter.open("example.gguf") gguf_writer = GGUFWriter.open("example.gguf")
gguf_writer.write_header(2, 3) gguf_writer.write_header(2, 4)
gguf_writer.write_architecture("llama") gguf_writer.write_architecture("llama")
gguf_writer.write_uint32("answer", 42) # Write a 32-bit integer gguf_writer.write_uint32("answer", 42) # Write a 32-bit integer
gguf_writer.write_float32("answer_in_float", 42.0) # Write a 32-bit float gguf_writer.write_float32("answer_in_float", 42.0) # Write a 32-bit float
gguf_writer.write_custom_alignment(64)
tensor1 = np.ones((32,), dtype=np.float32) * 100.0 tensor1 = np.ones((32,), dtype=np.float32) * 100.0
tensor2 = np.ones((32,), dtype=np.float32) * 101.0 tensor2 = np.ones((32,), dtype=np.float32) * 101.0
gguf_writer.write_tensor_info("tensor0", tensor1) gguf_writer.write_tensor_info("tensor0", tensor1)