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

[Attention] Update tests to remove deprecated env vars (#30563)

Browse Source 
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2025-12-17 12:49:59 -05:00
committed by GitHub
parent 9ca8cb38fd
commit 7eb6cb6c18
34 changed files with 580 additions and 447 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
tests/v1/e2e/test_async_scheduling.py
View File

@@ -142,16 +142,17 @@ def run_tests(
"""Test consistency of combos of async scheduling, preemption,
uni/multiproc executor with spec decoding."""
with monkeypatch.context() as m:
# avoid precision errors
if current_platform.is_rocm():
if is_testing_with_spec_decoding:
# Use TRITON_ATTN for spec decoding test for consistency
m.setenv("VLLM_ATTENTION_BACKEND", "TRITON_ATTN")
else:
m.setenv("VLLM_ATTENTION_BACKEND", "ROCM_AITER_FA")
# Determine attention config based on platform
if current_platform.is_rocm():
if is_testing_with_spec_decoding:
# Use TRITON_ATTN for spec decoding test for consistency
attention_config = {"backend": "TRITON_ATTN"}
else:
m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
attention_config = {"backend": "ROCM_AITER_FA"}
else:
attention_config = {"backend": "FLEX_ATTENTION"}
with monkeypatch.context() as m:
# lock matmul precision to full FP32 (IEEE)
m.setenv("VLLM_FLOAT32_MATMUL_PRECISION", "ieee")
# m.setenv("VLLM_BATCH_INVARIANT", "1")
@@ -174,6 +175,7 @@ def run_tests(
spec_config,
test_prefill_chunking=test_prefill_chunking,
is_testing_with_spec_decoding=is_testing_with_spec_decoding,
attention_config=attention_config,
)
outputs.append(test_results)
@@ -262,6 +264,7 @@ def run_test(
spec_config: dict[str, Any] | None,
test_prefill_chunking: bool,
is_testing_with_spec_decoding: bool = False,
attention_config: dict[str, Any] | None = None,
):
spec_decoding = spec_config is not None
cache_arg: dict[str, Any] = (
@@ -301,6 +304,7 @@ def run_test(
dtype=dtype,
speculative_config=spec_config,
disable_log_stats=False,
attention_config=attention_config,
**cache_arg,
) as vllm_model:
results = []