Remove unnecessary load_weights methods (#44589)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-06-29 01:52:23 -07:00
committed by GitHub
parent db28ae2d07
commit 5051698e41
54 changed files with 821 additions and 1975 deletions
@@ -576,7 +576,8 @@ class BitsAndBytesModelLoader(BaseModelLoader):
# For some models like Molmo, we need to use hf_to_vllm_mapper
# to ensure correct loading of weights.
if hf_to_vllm_mapper := getattr(model, "hf_to_vllm_mapper", None):
self.weight_mapper = lambda name: hf_to_vllm_mapper._map_name(name)
unstacked_mapper = hf_to_vllm_mapper.get_unstacked_mapper()
self.weight_mapper = lambda name, m=unstacked_mapper: m._map_name(name)
self._get_bnb_target_modules(model)
self._classify_module_sharding(model)
@@ -131,8 +131,11 @@ def initialize_online_processing(layer: torch.nn.Module):
# Track loading progress to determine when to process/copy
info.load_numel = 0
info.load_numel_total = get_layer_size(layer)
_wrap_parameters_weight_loader(layer)
# Wrap each parameter's weight loader
def _wrap_parameters_weight_loader(layer: torch.nn.Module) -> None:
"""Wrap each parameter's weight loader."""
# Note that nested wrapping will occur for shared tensors
for name, tensor in get_layer_tensors(layer).items():
if name in SKIP_TENSORS:
@@ -168,6 +171,12 @@ def make_online_process_loader(layer: torch.nn.Module, param_name: str) -> Calla
logger.debug("%s: Excessive loading", layer.__class__.__name__)
return
# Re-run on each load: layers may register parameters later (e.g., `bias`).
# Wrap late parameters and refresh `load_numel_total` so processing waits
# until all parameters are loaded.
info.load_numel_total = get_layer_size(layer)
_wrap_parameters_weight_loader(layer)
# Bind and normalize arguments
bound_args = loader_signature.bind(*args, **kwargs)
bound_args.apply_defaults()
+1 -1
View File
@@ -290,6 +290,6 @@ def configure_quant_config(
# pass mappings by reference to quant_config
if hf_to_vllm_mapper is not None:
quant_config.apply_vllm_mapper(hf_to_vllm_mapper)
quant_config.apply_vllm_mapper(hf_to_vllm_mapper.get_unstacked_mapper())
if packed_mapping is not None:
quant_config.packed_modules_mapping = packed_mapping