biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
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
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm/model_executor/models/qwen2.py
View File

@@ -21,7 +21,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen2 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
@@ -332,7 +332,8 @@ class Qwen2Model(nn.Module):
hidden_states, _ = self.norm(hidden_states, residual)
return hidden_states
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"),
@@ -342,6 +343,7 @@ class Qwen2Model(nn.Module):
("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
@@ -372,6 +374,8 @@ class Qwen2Model(nn.Module):
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params
class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
@@ -494,13 +498,14 @@ class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
) -> Optional[PoolerOutput]:
return self._pooler(hidden_states, pooling_metadata)
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
loader = AutoWeightsLoader(
self,
skip_prefixes=(["lm_head."]
if self.config.tie_word_embeddings else None),
)
loader.load_weights(weights)
return loader.load_weights(weights)
class Qwen2EmbeddingModel(nn.Module, SupportsLoRA, SupportsPP):
@@ -564,7 +569,8 @@ class Qwen2EmbeddingModel(nn.Module, SupportsLoRA, SupportsPP):
) -> Optional[PoolerOutput]:
return self._pooler(hidden_states, pooling_metadata)
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
def load_weights(self, weights: Iterable[Tuple[str,
torch.Tensor]]) -> Set[str]:
loader = AutoWeightsLoader(self,
ignore_unexpected_prefixes=["lm_head."])
loader.load_weights(weights)
return loader.load_weights(weights)