[Misc] Add uninitialized params tracking for AutoWeightsLoader (#10327)

Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2024-11-18 09:07:46 +08:00
committed by GitHub
parent d1557e66d3
commit c4e464333e
74 changed files with 454 additions and 185 deletions

View File

@@ -20,7 +20,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only OLMo model compatible with HuggingFace weights."""
from typing import Iterable, List, Optional, Tuple, Union
from typing import Iterable, List, Optional, Set, Tuple, Union
import torch
from torch import nn
@@ -356,7 +356,8 @@ class OlmoForCausalLM(nn.Module, SupportsPP):
next_tokens = self.sampler(logits, sampling_metadata)
return next_tokens
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
stacked_params_mapping = [
# (param_name, shard_name, shard_id)
("qkv_proj", "q_proj", "q"),
@@ -366,6 +367,7 @@ class OlmoForCausalLM(nn.Module, SupportsPP):
("gate_up_proj", "up_proj", 1),
]
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
@@ -402,3 +404,5 @@ class OlmoForCausalLM(nn.Module, SupportsPP):
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params