[Model Runner V2] Fix lora Triton Error [CUDA]: device-side assert triggered (#43139)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Wentao Ye
2026-05-21 01:00:30 +00:00
committed by GitHub
co-authored by Nick Hill
parent 63ea11709b
commit 9640970de2
+14 -1
View File
@@ -37,6 +37,7 @@ from vllm.distributed.parallel_state import (
)
from vllm.forward_context import BatchDescriptor, set_forward_context
from vllm.logger import init_logger
from vllm.lora.layers import LoRAMapping
from vllm.model_executor.layers.mamba.ops.ssu_dispatch import (
initialize_mamba_ssu_backend,
)
@@ -1084,7 +1085,19 @@ class GPUModelRunner(LoRAModelRunnerMixin):
)
block_tables = None
slot_mappings = None
# FIXME(woosuk): Fix warmup for LoRA.
if self.lora_config:
# program a no-LoRA mapping here so kernels early-exit instead of
# reading uninitialized metadata during dummy runs.
# FIXME: Replace this with LoRA warmup:
# https://github.com/vllm-project/vllm/pull/35536
assert hasattr(self, "lora_manager")
self.lora_manager._adapter_manager.set_adapter_mapping(
LoRAMapping(
index_mapping=(0,) * input_batch.num_tokens_after_padding,
prompt_mapping=(0,) * input_batch.num_reqs,
is_prefill=True,
)
)
attn_metadata = None
slot_mappings_by_layer = None