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, 19:45 UTC
[Bug] Fix torch warning of tf32 usage (#29112)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Nov 20, 2025 at 20:54 UTC
e1eefa4c40fc5b28bd7e83b6596bb5d2f420fd92
1 file changed
+3
-2
vllm/model_executor/layers/batch_invariant.py
+3
-2
@@ -852,5 +852,6 @@ def init_batch_invariance():
852
enable_batch_invariant_mode()
853
854
# Disable TF32 for batch invariance - it causes non-deterministic rounding
855
- torch.backends.cuda.matmul.allow_tf32 = False
856
- torch.backends.cudnn.allow_tf32 = False
855
+ torch.backends.cuda.matmul.fp32_precision = "ieee"
856
+ torch.backends.cudnn.conv.fp32_precision = "ieee"
857
+ torch.backends.cudnn.rnn.fp32_precision = "ieee"