A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 2, 22:00 UTC
Revert "[Misc] add humming to dependencies" (#43492)
Michael Goin committed
May 23, 2026 at 17:21 UTC
10d264a2b917e3e8b26cdf155d47f37df6d3faa0
3 files changed
+14
-16
requirements/cuda.txt
+1
-4
@@ -25,7 +25,4 @@ nvidia-cutlass-dsl[cu13]==4.5.0
25
quack-kernels>=0.3.3
26
27
# Tokenspeed_MLA for faster mla with spec decode
28
-tokenspeed-mla==0.1.2
29
-
30
-# Humming kernels for quantization gemm
31
-humming-kernels[cu13]==0.1.0
28
+tokenspeed-mla==0.1.2
\ No newline at end of file
setup.py
-2
@@ -1017,8 +1017,6 @@ def get_requirements() -> list[str]:
1017
if "nvidia-cutlass-dsl[cu13]" in req and cuda_major == "12":
1018
# [cu13] extra is the default; strip it on CUDA 12 builds.
1019
req = req.replace("nvidia-cutlass-dsl[cu13]", "nvidia-cutlass-dsl")
1020
- if "humming-kernels[cu13]" in req and cuda_major == "12":
1021
- req = req.replace("humming-kernels[cu13]", "humming-kernels[cu12]")
1020
modified_requirements.append(req)
1021
requirements = modified_requirements
1022
elif _is_hip():
vllm/model_executor/layers/quantization/humming.py
+13
-10
@@ -43,9 +43,12 @@ from vllm.model_executor.parameter import (
43
RowvLLMParameter,
44
)
45
from vllm.model_executor.utils import set_weight_attrs
46
-from vllm.platforms import current_platform
46
48
-if current_platform.is_cuda():
47
+if TYPE_CHECKING:
48
+ from vllm.model_executor.models.utils import WeightsMapper
49
+
50
+
51
+try:
52
from humming.dtypes import DataType
53
from humming.layer import HummingMethod
54
from humming.schema import (
@@ -62,16 +65,15 @@ if current_platform.is_cuda():
65
HummingIndexedExperts,
66
get_humming_moe_gemm_type,
67
)
68
+except ModuleNotFoundError:
69
+ HummingMethod = None
70
66
-if TYPE_CHECKING:
67
- from humming.schema import (
68
- BaseInputSchema,
69
- BaseWeightSchema,
70
- HummingInputSchema,
71
- HummingWeightSchema,
72
- )
71
74
- from vllm.model_executor.models.utils import WeightsMapper
72
+def assert_humming_available():
73
+ assert HummingMethod is not None, (
74
+ "humming is not available, please run "
75
+ "'pip install git+https://github.com/inclusionAI/humming' to install it."
76
+ )
77
78
79
def prepare_padded_shape(shape, x):
@@ -184,6 +186,7 @@ class HummingConfig(QuantizationConfig):
186
packed_modules_mapping: dict[str, list[str]] = {}
187
188
def __init__(self, full_config: dict[str, Any] | None = None):
189
+ assert_humming_available()
190
self.full_config: dict[str, Any] = full_config or {}
191
192
@classmethod