biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] Remove pad_for_cudagraphs from config (#30143)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-01-20 13:05:48 -07:00
committed by GitHub
parent 86c69dc54c
commit 2261340806
9 changed files with 157 additions and 136 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/v1/cudagraph/test_cudagraph_dispatch.py
View File

@@ -61,9 +61,6 @@ def _create_vllm_config(
)
compilation_config.post_init_cudagraph_sizes()
mock_config.pad_for_cudagraph = (
lambda batch_size: compilation_config.bs_to_padded_graph_size[batch_size]
)
return mock_config
@@ -169,6 +166,7 @@ class TestCudagraphDispatcher:
rt_mode, key = dispatcher.dispatch(
num_tokens=8, uniform_decode=False, has_lora=False, disable_full=True
)
if "PIECEWISE" in cudagraph_mode_str: # string contains check
assert rt_mode == CUDAGraphMode.PIECEWISE
assert key == desc_full_exact.relax_for_mixed_batch_cudagraphs()
@@ -360,7 +358,7 @@ class TestCudagraphIntegration:
):
full_wrapper(input_1)
rt_mode, key = self.dispatcher.dispatch(desc_1)
rt_mode, key = self.dispatcher.dispatch(num_tokens=desc_1.num_tokens)
# 1. Capture first shape
action = self._run_and_monitor_call(full_wrapper, input_1, rt_mode, key)
assert action == "capture_global"
@@ -369,7 +367,7 @@ class TestCudagraphIntegration:
action = self._run_and_monitor_call(full_wrapper, input_1, rt_mode, key)
assert action == "replay"
rt_mode, key = self.dispatcher.dispatch(desc_2)
rt_mode, key = self.dispatcher.dispatch(num_tokens=desc_2.num_tokens)
# 3. Capture second shape
action = self._run_and_monitor_call(full_wrapper, input_2, rt_mode, key)
assert action == "capture_global"
@@ -381,7 +379,7 @@ class TestCudagraphIntegration:
assert action == "replay"
# 5. Bypass if no key match
rt_mode, key = self.dispatcher.dispatch(desc_3_unseen)
rt_mode, key = self.dispatcher.dispatch(num_tokens=desc_3_unseen.num_tokens)
assert rt_mode == CUDAGraphMode.NONE
action = self._run_and_monitor_call(full_wrapper, input_3, rt_mode, key)
assert action == "bypass"