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 12, 21:00 UTC
[BugFix] [FEAT] Enable fastsafetensors for ROCm platform (#28225)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
TJian committed
Nov 20, 2025 at 23:34 UTC
82b05b15e61badfd0c5912d4c3eebc88043c9ef8
3 files changed
+5
-2
requirements/rocm.txt
+1
@@ -15,3 +15,4 @@ setuptools-scm>=8
15
runai-model-streamer[s3,gcs]==0.15.0
16
conch-triton-kernels==1.2.1
17
timm>=1.0.17
18
+fastsafetensors @ git+https://github.com/foundation-model-stack/fastsafetensors.git@d6f998a03432b2452f8de2bb5cefb5af9795d459
tests/model_executor/model_loader/fastsafetensors_loader/test_fastsafetensors_loader.py
+2
-1
@@ -19,7 +19,8 @@ sampling_params = SamplingParams(temperature=0.8, top_p=0.95, seed=0)
19
20
21
@pytest.mark.skipif(
22
- not current_platform.is_cuda(), reason="fastsafetensors requires CUDA/NVIDIA GPUs"
22
+ not current_platform.is_cuda_alike(),
23
+ reason="fastsafetensors requires NVIDIA/AMD GPUs",
24
)
25
def test_model_loader_download_files(vllm_runner):
26
with vllm_runner(test_model, load_format="fastsafetensors") as llm:
tests/model_executor/model_loader/fastsafetensors_loader/test_weight_utils.py
+2
-1
@@ -17,7 +17,8 @@ from vllm.platforms import current_platform
17
18
19
@pytest.mark.skipif(
20
- not current_platform.is_cuda(), reason="fastsafetensors requires CUDA/NVIDIA GPUs"
20
+ not current_platform.is_cuda_alike(),
21
+ reason="fastsafetensors requires NVIDIA/AMD GPUs",
22
)
23
def test_fastsafetensors_model_loader():
24
with tempfile.TemporaryDirectory() as tmpdir: