[Migration] Migrate GGUF quantization support to plugin (#39612)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-06-12 12:02:21 -07:00
committed by GitHub
parent d6fd7ce8da
commit 6635279d8a
57 changed files with 72 additions and 9048 deletions
-18
View File
@@ -517,10 +517,6 @@ class OpenPanguEmbeddedAttention(nn.Module):
quant_config: QuantizationConfig | None,
) -> None:
is_neox_style = True
is_gguf = quant_config and quant_config.get_name() == "gguf"
if is_gguf and config.model_type == "PanguEmbedded":
is_neox_style = False
rope_parameters = config.rope_parameters or {}
if rope_parameters is not None and rope_parameters.get(
"mrope_interleaved", False
@@ -716,20 +712,6 @@ class OpenPanguSinkAttention(nn.Module):
# no need to narrow
is_sharded_weight = is_sharded_weight or use_bitsandbytes_4bit
# Special case for GGUF
is_gguf_weight = getattr(param, "is_gguf_weight", False)
is_gguf_weight_type = getattr(param, "is_gguf_weight_type", False)
if is_gguf_weight_type:
param.weight_type = loaded_weight.item()
# Materialize GGUF UninitializedParameter
if is_gguf_weight and isinstance(param, nn.UninitializedParameter):
final_shape = list(loaded_weight.shape)
if output_dim is not None:
assert final_shape[output_dim] % self.tp_size == 0
final_shape[output_dim] = final_shape[output_dim] // self.tp_size
param.materialize(final_shape, dtype=loaded_weight.dtype)
param_data = param.data
if output_dim is not None and not is_sharded_weight:
shard_size = param_data.shape[output_dim]