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
[CI] fix mamba kernel test (#26250)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Oct 6, 2025 at 02:26 UTC
9c3c21c519fc0dc7b7b49c86058749c9af202ad9
2 files changed
+12
-1
.buildkite/test-pipeline.yaml
+1
@@ -477,6 +477,7 @@ steps:
477
source_file_dependencies:
478
- csrc/mamba/
479
- tests/kernels/mamba
480
+ - vllm/model_executor/layers/mamba/ops
481
commands:
482
- pytest -v -s kernels/mamba
483
tests/kernels/mamba/test_causal_conv1d.py
+11
-1
@@ -165,7 +165,17 @@ def test_causal_conv1d_update(dim, width, seqlen, has_bias, silu_activation, ity
165
bias = torch.randn(dim, device=device, dtype=itype) if has_bias else None
166
conv_state_ref = conv_state.detach().clone()
167
activation = None if not silu_activation else "silu"
168
- out = causal_conv1d_update(x, conv_state, weight, bias, activation=activation)
168
+
169
+ conv_state_indices = torch.arange(batch, dtype=torch.int32, device=device)
170
+
171
+ out = causal_conv1d_update(
172
+ x,
173
+ conv_state,
174
+ weight,
175
+ bias,
176
+ activation=activation,
177
+ conv_state_indices=conv_state_indices,
178
+ )
179
out_ref = causal_conv1d_update_ref(
180
x_ref, conv_state_ref, weight, bias, activation=activation
181
)