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, 04:15 UTC
[Bugfix] Fix non-contiguous tensor error in `rocm_unquantized_gemm_impl` (#27605)
Signed-off-by: zhewenli <zhewenli@meta.com>
Zhewen Li committed
Oct 29, 2025 at 00:00 UTC
8b62495076fde4a73c4a397f5307f49c93dd7c6e
2 files changed
+4
-4
.buildkite/test-amd.yaml
+1
-1
@@ -286,7 +286,7 @@ steps:
286
287
- label: Engine Test # 25min
288
timeout_in_minutes: 40
289
- mirror_hardwares: [amdexperimental]
289
+ mirror_hardwares: [amdexperimental, amdproduction]
290
agent_pool: mi325_1
291
#grade: Blocking
292
source_file_dependencies:
vllm/model_executor/layers/utils.py
+3
-3
@@ -119,17 +119,17 @@ def rocm_unquantized_gemm_impl(
119
if use_skinny is not True:
120
return torch.nn.functional.linear(x, weight, bias)
121
122
- x_view = x.view(-1, x.size(-1))
122
+ x_view = x.reshape(-1, x.size(-1))
123
n = x_view.shape[0]
124
m = weight.shape[0]
125
cu_count = current_platform.get_cu_count()
126
127
if m > 8 and 0 < n <= 4:
128
out = ops.wvSplitK(weight, x_view, cu_count, bias)
129
- return out.view(*x.shape[:-1], weight.shape[0])
129
+ return out.reshape(*x.shape[:-1], weight.shape[0])
130
elif m % 4 == 0 and n == 1 and k <= 8192 and bias is None:
131
out = ops.LLMM1(weight, x_view, 4)
132
- return out.view(*x.shape[:-1], weight.shape[0])
132
+ return out.reshape(*x.shape[:-1], weight.shape[0])
133
return torch.nn.functional.linear(x, weight, bias)
134
135