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 9, 16:00 UTC
[BugFix] Add env variable to control PDL in LoRA (#32836)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Jan 25, 2026 at 16:32 UTC
73b243463b159433ff49c65b54c2e3d66a2e434e
2 files changed
+11
-1
vllm/envs.py
+5
@@ -251,6 +251,7 @@ if TYPE_CHECKING:
251
VLLM_LOG_MODEL_INSPECTION: bool = False
252
VLLM_DEBUG_MFU_METRICS: bool = False
253
VLLM_DISABLE_LOG_LOGO: bool = False
254
+ VLLM_LORA_DISABLE_PDL: bool = False
255
256
257
def get_default_cache_root():
@@ -1618,8 +1619,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
1619
),
1620
# Disable logging of vLLM logo at server startup time.
1621
"VLLM_DISABLE_LOG_LOGO": lambda: bool(int(os.getenv("VLLM_DISABLE_LOG_LOGO", "0"))),
1622
+ # Disable PDL for LoRA, as enabling PDL with LoRA on SM100 causes
1623
+ # Triton compilation to fail.
1624
+ "VLLM_LORA_DISABLE_PDL": lambda: bool(int(os.getenv("VLLM_LORA_DISABLE_PDL", "0"))),
1625
}
1626
1627
+
1628
# --8<-- [end:env-vars-definition]
1629
1630
vllm/lora/ops/triton_ops/utils.py
+6
-1
@@ -310,4 +310,9 @@ def supports_pdl(device: torch.device | None = None) -> bool:
310
Refer to: https://github.com/triton-lang/triton/blob/v3.5.0/python/tutorials/11-programmatic-dependent-launch.py
311
"""
312
# PDL requires compute capability SM90 or above
313
- return current_platform.is_cuda() and current_platform.has_device_capability(90)
313
+
314
+ return (
315
+ current_platform.is_cuda()
316
+ and current_platform.has_device_capability(90)
317
+ and not envs.VLLM_LORA_DISABLE_PDL
318
+ )