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 11, 00:00 UTC
[Chore] Update more locations to use `attention_config.backend` (#31153)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Dec 23, 2025 at 11:19 UTC
8cef137689170338fdd573838414fc7f9e3ec84a
2 files changed
+3
-2
benchmarks/benchmark_batch_invariance.py
+1
-1
@@ -104,7 +104,6 @@ def run_benchmark_with_batch_invariant(
104
random.seed(seed)
105
106
# Set environment variables
107
- os.environ["VLLM_ATTENTION_BACKEND"] = backend
107
if batch_invariant:
108
os.environ["VLLM_BATCH_INVARIANT"] = "1"
109
else:
@@ -140,6 +139,7 @@ def run_benchmark_with_batch_invariant(
139
max_model_len=max_model_len,
140
dtype="bfloat16",
141
tensor_parallel_size=tp_size,
142
+ attention_config={"backend": backend},
143
enable_prefix_caching=False,
144
)
145
init_time = time.perf_counter() - start_init
tests/compile/distributed/test_fusions_e2e.py
+2
-1
@@ -557,7 +557,8 @@ def test_rms_group_quant(
557
# To capture subprocess logs, we need to know whether spawn or fork is used.
558
# Force spawn as it is more general.
559
monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
560
- monkeypatch.setenv("VLLM_ATTENTION_BACKEND", backend.name)
560
+
561
+ model_kwargs["attention_config"] = {"backend": backend.name}
562
563
compilation_config = CompilationConfig(
564
# Testing properties