[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

@@ -1,5 +1,5 @@
from functools import cached_property
from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional,
from typing import (Any, Dict, Iterable, List, Literal, Mapping, Optional, Set,
Tuple, TypedDict, Union)
import torch
@@ -1034,7 +1034,8 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal,
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"),
@@ -1044,6 +1045,7 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal,
(".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
@@ -1111,3 +1113,5 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal,
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
loaded_params.add(name)
return loaded_params