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 11, 05:00 UTC
[Quantization] enable compressed-tensors marlin support for turing (2) (#31008)
Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Jinzhen Lin committed
Dec 19, 2025 at 16:56 UTC
9187de9fac94996120ae29afc43d5a34aa86cf18
3 files changed
+3
-3
vllm/model_executor/layers/quantization/utils/marlin_utils.py
+1
-1
@@ -48,7 +48,7 @@ def query_marlin_supported_quant_types(
48
-1 if capability_tuple is None else capability_tuple.to_int()
49
)
50
51
- if device_capability < 80:
51
+ if device_capability < 75:
52
return []
53
54
# - has_zp is True: return quant_types that has zero points
vllm/model_executor/layers/quantization/utils/marlin_utils_fp4.py
+1
-1
@@ -23,7 +23,7 @@ logger = init_logger(__name__)
23
24
25
def is_fp4_marlin_supported():
26
- return current_platform.has_device_capability(80)
26
+ return current_platform.has_device_capability(75)
27
28
29
def nvfp4_marlin_process_scales(marlin_scales):
vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py
+1
-1
@@ -22,7 +22,7 @@ logger = init_logger(__name__)
22
23
24
def is_fp8_marlin_supported():
25
- return current_platform.has_device_capability(80)
25
+ return current_platform.has_device_capability(75)
26
27
28
def fp8_fused_exponent_bias_into_scales(scales):