Skip to content

Commit 622e1b1

Browse files
authored
[Test] Enable default graph mode and set max cudagraph capture size to fix test case (vllm-project#10863)
### What this PR does / why we need it? For the test case `test_suffix_acceptance`: 1. Switches it to use the default graph mode `FULL_AND_PIECEWISE`. 2. Sets `max_cudagraph_capture_size=16` to fix a previous failure due to oversized graph captures. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? Ran `test_suffix_acceptance` locally with the changes applied and verified it passes. - vLLM version: v0.22.1 - vLLM main: vllm-project/vllm@967c5c3 Signed-off-by: zhiyu-wa <1959864813@qq.com>
1 parent 0d1d5bf commit 622e1b1

1 file changed

Lines changed: 1 addition & 4 deletions

File tree

tests/e2e/pull_request/one_card/spec_decode/test_suffix.py

Lines changed: 1 addition & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -23,10 +23,7 @@ def test_suffix_acceptance(
2323
"num_speculative_tokens": 10,
2424
},
2525
max_model_len=1024,
26-
compilation_config={
27-
"cudagraph_mode": "PIECEWISE",
28-
"cudagraph_capture_sizes": [1, 2, 4, 8],
29-
},
26+
max_cudagraph_capture_size=16,
3027
disable_log_stats=False,
3128
) as runner:
3229
for i in range(10):

0 commit comments

Comments
 (0)