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 4, 18:45 UTC
Fix test_startup.py for torch 2.12 (#40636)
Signed-off-by: Angela Yi <yiangela7@gmail.com>
Angela Yi committed
Apr 22, 2026 at 12:31 UTC
eb6661d52260a3a43e40dae6c808db49487e87c4
3 files changed
+8
-10
tests/compile/h100/test_startup.py
+6
-8
@@ -135,10 +135,9 @@ MODEL_SPECS = [
135
model="deepseek-ai/DeepSeek-V3.2",
136
hf_overrides=_SMALL_MOE_OVERRIDES,
137
cold_artifacts_saved=4,
138
- # TODO: https://github.com/vllm-project/vllm/issues/38051
139
- # We shouldn't be saving any artifacts on warm start.
140
- warm_artifacts_saved=4,
141
- warm_artifacts_loaded=0,
138
+ # https://github.com/vllm-project/vllm/issues/38051
139
+ warm_artifacts_saved=0 if is_torch_equal_or_newer("2.12.0") else 4,
140
+ warm_artifacts_loaded=4 if is_torch_equal_or_newer("2.12.0") else 0,
141
),
142
id="deepseek_v3.2",
143
),
@@ -147,10 +146,9 @@ MODEL_SPECS = [
146
model="moonshotai/Kimi-K2.5",
147
hf_overrides={"text_config": _SMALL_MOE_OVERRIDES},
148
cold_artifacts_saved=4,
150
- # TODO: https://github.com/vllm-project/vllm/issues/38051
151
- # We shouldn't be saving any artifacts on warm start.
152
- warm_artifacts_saved=4,
153
- warm_artifacts_loaded=0,
149
+ # https://github.com/vllm-project/vllm/issues/38051
150
+ warm_artifacts_saved=0 if is_torch_equal_or_newer("2.12.0") else 4,
151
+ warm_artifacts_loaded=4 if is_torch_equal_or_newer("2.12.0") else 0,
152
),
153
id="kimi_k2.5",
154
),
tests/compile/test_dynamic_shapes_compilation.py
+1
-1
@@ -28,7 +28,7 @@ def get_test_models():
28
"Qwen/Qwen2-7B-Instruct",
29
"meta-llama/Llama-3.1-8B",
30
]
31
- if is_torch_equal_or_newer("2.12.0"):
31
+ if is_torch_equal_or_newer("2.12.0.dev"):
32
models.append("Qwen/Qwen3-4B-Instruct-2507")
33
return models
34
vllm/env_override.py
+1
-1
@@ -553,7 +553,7 @@ def _apply_constrain_to_fx_strides_patch():
553
_lowering.constrain_to_fx_strides = _patched
554
555
556
-if is_torch_equal_or_newer("2.10.0") and not is_torch_equal_or_newer("2.12.0"):
556
+if is_torch_equal_or_newer("2.10.0") and not is_torch_equal_or_newer("2.12.0.dev"):
557
import builtins as _builtins
558
import pickle
559