[Dependency] Update Humming and remove NVFP4 scale workaround (#61556)

Signed-off-by: jinzhen.ljz <jinzhen.ljz@antgroup.com>
Co-authored-by: Codex <codex@openai.com>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Jinzhen Lin
2026-10-14 13:32:36 -07:00
committed by GitHub
co-authored by Codex Misha Goin Kevin H. Luu
parent 2e8a5930f6
commit fe98d90a9d
13 changed files with 54 additions and 47 deletions
@@ -584,7 +584,9 @@ class HummingLinearMethod(LinearMethodBase):
param_dtype=layer.param_dtype,
allow_fallback=self.quant_config.allow_input_schema_fallback,
)
self.layer_config = _hm.prepare_layer_config(
source_tensors = dict(layer.named_parameters())
self.layer_config, tensors = _hm.prepare_and_transform_humming_tensors(
tensors=source_tensors,
shape_n=layer.output_partition_sizes_sum,
shape_k=layer.input_size_per_partition,
weight_schema=self.weight_schema,
@@ -595,8 +597,6 @@ class HummingLinearMethod(LinearMethodBase):
torch_dtype=layer.param_dtype,
device=layer.weight.device,
)
source_tensors = dict(layer.named_parameters())
tensors = _hm.transform_humming_tensors(self.layer_config, source_tensors)
name = self.input_schema.static_tensor_scale_name
if name is not None:
tensors[name] = source_tensors[name]
@@ -610,7 +610,7 @@ class HummingLinearMethod(LinearMethodBase):
setattr(layer, name, param)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = _hm.create_humming_locks(layer.weight.device)
def apply(
self,