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
output type conversion fix (#27159)
Jianyu Huang committed
Oct 19, 2025 at 01:10 UTC
221bf725772b891db27e1b9200c36a721ba9a304
1 file changed
+1
-4
vllm/model_executor/layers/batch_invariant.py
+1
-4
@@ -134,10 +134,7 @@ def matmul_kernel_persistent(
134
bias_ptrs = bias_ptr + offs_cn
135
bias = tl.load(bias_ptrs, mask=offs_cn < N, other=0.0).to(tl.float32)
136
accumulator += bias
137
- if c_ptr.dtype.element_ty == tl.float8e4nv:
138
- c = accumulator.to(tl.float8e4nv)
139
- else:
140
- c = accumulator.to(tl.float16)
137
+ c = accumulator.to(c_ptr.dtype.element_ty)
138
tl.store(c_ptrs, c, mask=c_mask)
139
140