[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
@@ -9,6 +9,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
kInt8DynamicTokenSym,
)
from vllm.platforms import current_platform
from vllm.utils.humming import create_humming_locks
from vllm.utils.import_utils import has_humming
from .MPLinearKernel import MPLinearKernel, MPLinearLayerConfig
@@ -71,7 +72,7 @@ class HummingLinearKernel(MPLinearKernel):
input_schema=input_schema,
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,
@@ -11,6 +11,7 @@ from vllm.model_executor.layers.quantization.utils.humming import (
quant_key_to_input_schema,
)
from vllm.platforms import current_platform
from vllm.utils.humming import create_humming_locks
from vllm.utils.import_utils import has_humming
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
@@ -60,7 +61,7 @@ class HummingMxFp4LinearKernel(MxFp4LinearKernel):
layer, quant_config, input_schema=input_schema
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,
@@ -12,6 +12,7 @@ from vllm.model_executor.layers.quantization.utils.humming import (
)
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp6E2M3Static
from vllm.platforms import current_platform
from vllm.utils.humming import create_humming_locks
from vllm.utils.import_utils import has_humming
from .base import MxFp6LinearKernel, MxFp6LinearLayerConfig
@@ -66,7 +67,7 @@ class HummingMxFp6LinearKernel(MxFp6LinearKernel):
layer, quant_config, input_schema=input_schema
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,
@@ -12,6 +12,7 @@ from vllm.model_executor.layers.quantization.utils.humming import (
)
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Dynamic
from vllm.platforms import current_platform
from vllm.utils.humming import create_humming_locks
from vllm.utils.import_utils import has_humming
from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig
@@ -57,7 +58,7 @@ class HummingMxfp8LinearKernel(Mxfp8LinearKernel):
layer, quant_config, input_schema=input_schema
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,
@@ -12,6 +12,7 @@ from vllm.model_executor.layers.quantization.utils.humming import (
)
from vllm.model_executor.layers.quantization.utils.quant_utils import kNvfp4Dynamic
from vllm.platforms import current_platform
from vllm.utils.humming import create_humming_locks
from vllm.utils.import_utils import has_humming
from .base import NvFp4LinearKernel, NvFp4LinearLayerConfig
@@ -71,7 +72,7 @@ class HummingNvFp4LinearKernel(NvFp4LinearKernel):
layer, quant_config, input_schema=input_schema
)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,
@@ -11,6 +11,7 @@ from vllm.model_executor.layers.quantization.utils.humming import (
prepare_humming_linear_layer_config,
)
from vllm.platforms import current_platform
from vllm.utils.humming import create_humming_locks
from vllm.utils.import_utils import has_humming
from .ScaledMMLinearKernel import (
@@ -81,7 +82,7 @@ class HummingFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,
@@ -148,7 +149,7 @@ class HummingInt8ScaledMMLinearKernel(Int8ScaledMMLinearKernel):
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
self.locks = create_humming_locks(layer.weight.device)
def apply_weights(
self,