[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
+6 -2
View File
@@ -22,7 +22,7 @@
# limitations under the License.
"""Inference-only DeciLM model compatible with HuggingFace weights."""
from typing import Iterable, Tuple
from typing import Iterable, Set, Tuple
import torch
@@ -57,7 +57,8 @@ class DeciLMForCausalLM(LlamaForCausalLM):
delattr(config, "num_key_value_heads_per_layer")
super().__init__(vllm_config=vllm_config)
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"),
@@ -67,6 +68,7 @@ class DeciLMForCausalLM(LlamaForCausalLM):
("gate_up_proj", "up_proj", 1),
]
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if "rotary_emb.inv_freq" in name:
continue
@@ -97,6 +99,8 @@ class DeciLMForCausalLM(LlamaForCausalLM):
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
def _degroup_weight(self, loaded_weight: torch.Tensor) -> torch.Tensor:
hidden_size = self.config.hidden_size