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 15, 12:15 UTC
[Bugfix] Temporarily disable gptq_bitblas on ROCm (#17411)
Signed-off-by: Yan Cangang <nalanzeyu@gmail.com>
NaLan ZeYu committed
May 1, 2025 at 10:51 UTC
1144a8efe715d0274412d215a3ad1f941a8469f1
2 files changed
+6
-1
docs/source/features/quantization/supported_hardware.md
+1
-1
@@ -80,7 +80,7 @@ The table below shows the compatibility of various quantization implementations
80
* ✅︎
81
* ✅︎
82
* ✅︎
83
- * ✅︎
83
+ * ❌
84
* ❌
85
* ❌
86
* ❌
vllm/model_executor/layers/quantization/gptq_bitblas.py
+5
@@ -25,6 +25,7 @@ from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
25
PackedColumnParameter,
26
PackedvLLMParameter,
27
RowvLLMParameter)
28
+from vllm.platforms import current_platform
29
from vllm.scalar_type import scalar_types
30
31
logger = init_logger(__name__)
@@ -191,6 +192,10 @@ class GPTQBitBLASConfig(QuantizationConfig):
192
sym = quant_config.get("sym")
193
desc_act = quant_config.get("desc_act")
194
195
+ # temporarily disable on ROCm platform
196
+ if not current_platform.is_cuda():
197
+ return False
198
+
199
# If we cannot find the info needed in the config, cannot convert.
200
if (num_bits is None or group_size is None or sym is None
201
or desc_act is None):