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 14, 05:30 UTC
[BugFix] Fix DeepGEMM warmup, no m.weight_scale_inv (#25206)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Sep 18, 2025 at 17:26 UTC
9fac6aa30b669de75d8718164cd99676d3530e7d
1 file changed
+1
-1
vllm/model_executor/warmup/deep_gemm_warmup.py
+1
-1
@@ -36,7 +36,7 @@ def _extract_data_from_linear_base_module(
36
assert m.quant_method.quant_config is not None
37
38
w = m.weight
39
- ws = m.weight_scale_inv
39
+ ws = m.weight_scale
40
quant_block_size = m.quant_method.quant_config.weight_block_size
41
42
assert isinstance(w, torch.Tensor)