[BugFix] Add env variable to control PDL in LoRA (#32836)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/envs.py

@@ -251,6 +251,7 @@ if TYPE_CHECKING:
     VLLM_LOG_MODEL_INSPECTION: bool = False
     VLLM_DEBUG_MFU_METRICS: bool = False
     VLLM_DISABLE_LOG_LOGO: bool = False
+    VLLM_LORA_DISABLE_PDL: bool = False
 
 
 def get_default_cache_root():
@@ -1618,8 +1619,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
     ),
     # Disable logging of vLLM logo at server startup time.
     "VLLM_DISABLE_LOG_LOGO": lambda: bool(int(os.getenv("VLLM_DISABLE_LOG_LOGO", "0"))),
+    # Disable PDL for LoRA, as enabling PDL with LoRA on SM100 causes
+    # Triton compilation to fail.
+    "VLLM_LORA_DISABLE_PDL": lambda: bool(int(os.getenv("VLLM_LORA_DISABLE_PDL", "0"))),
 }
 
+
 # --8<-- [end:env-vars-definition]
 
 

vllm/lora/ops/triton_ops/utils.py

@@ -310,4 +310,9 @@ def supports_pdl(device: torch.device | None = None) -> bool:
     Refer to: https://github.com/triton-lang/triton/blob/v3.5.0/python/tutorials/11-programmatic-dependent-launch.py
     """
     # PDL requires compute capability SM90 or above
-    return current_platform.is_cuda() and current_platform.has_device_capability(90)
+
+    return (
+        current_platform.is_cuda()
+        and current_platform.has_device_capability(90)
+        and not envs.VLLM_LORA_DISABLE_PDL
+    )