AMajor/grok-1
1
0
Fork 0
mirror of https://github.com/xai-org/grok-1.git synced 2024-11-22 19:49:53 +03:00
Code Issues Packages Projects Releases Wiki Activity

path_tuple_to_string() -> path_tuple_to_string_cached(): Optimizing Regex Operations with Caching

Browse Source
This commit is contained in:
Madhav 2024-03-23 15:37:18 +05:30
parent 104affb51a
commit 10fd2278a8
1 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
checkpoint.py
View File

@ -125,6 +125,14 @@ def load_tensors(shaped_arrays, directory, mesh_config, tensor_indices=None):
return results
"""
For path_tuple_to_string(),
introducing a simple caching mechanism to avoid recomputing regex matches for paths that have already been processed.
"""
@lru_cache(maxsize=32) # Set maxsize to 32MB
def path_tuple_to_string_cached(path: tuple) -> str:
return path_tuple_to_string(path=path)
def path_tuple_to_string(path: tuple) -> str:
pieces = []
for elem in path:
@ -184,13 +192,13 @@ def replace_with_load_state(
) -> Any:
flatten_load, _ = jax.tree_util.tree_flatten_with_path(load_state)
flatten_init, structure_init = jax.tree_util.tree_flatten_with_path(init_state)
load_map = {path_tuple_to_string(path): tensor for path, tensor in flatten_load}
load_map = {path_tuple_to_string_cached(path): tensor for path, tensor in flatten_load}
replaced = []
num_replicas = 1
data_model_shards = math.prod(mesh_config)
for i, (init_path, tensor) in enumerate(flatten_init):
init_path_str = path_tuple_to_string(init_path)
init_path_str = path_tuple_to_string_cached(init_path)
load_path_str = get_load_path_str_cached(init_path_str, load_rename_rules, load_exclude_rules)
if load_path_str is None:
rank_logger.info(f"Excluded from restore: {init_path_str}.")