[Bug] Running DeepSeek R1 error when enable flashinfer-mla and torch-compile #4964
Description
Checklist
- 1. I have searched related issues but cannot get the expected help.
- 2. The bug has not been fixed in the latest version.
- 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
- 4. If the issue you raised is not a bug but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
- 5. Please use English, otherwise it will be closed.
Describe the bug
I use the latest main (1c63e797560405b9ec5149d93f02d6147383c627), start the DeepSeek R1 model, and when using the --flashinfer-mla --enable-torch-compile parameter, an error is report.
Reproduction
node 1:
python3 -m sglang.launch_server --model /path/to/DeepSeek-R1-Channel-INT8 --trust-remote --host 0.0.0.0 --port 30000 --enable-flashinfer-mla --speculative-algo EAGLE --speculative-draft /path/to/DeepSeek-R1-NextN-Channel-INT8 --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --enable-torch-compile --torch-compile-max-bs 4 --quantization w8a8_int8 --tp 16 --dist-init-addr ip:20000 --nnodes 2 --node-rank 0
node 2:
python3 -m sglang.launch_server --model /path/to/DeepSeek-R1-Channel-INT8 --trust-remote --host 0.0.0.0 --port 30000 --enable-flashinfer-mla --speculative-algo EAGLE --speculative-draft /path/to/DeepSeek-R1-NextN-Channel-INT8 --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --enable-torch-compile --torch-compile-max-bs 4 --quantization w8a8_int8 --tp 16 --dist-init-addr ip:20000 --nnodes 2 --node-rank 1
The exception information is as follows:
[2025-04-01 04:49:35 TP3] Scheduler hit an exception: Traceback (most recent call last):
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 267, in __init__
self.capture()
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 351, in capture
) = self.capture_one_batch_size(bs, forward)
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 443, in capture_one_batch_size
run_once()
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 436, in run_once
logits_output = forward(input_ids, forward_batch.positions, forward_batch)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/eval_frame.py", line 465, in _fn
return fn(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/external_utils.py", line 40, in inner
return fn(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 1323, in forward
hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 1277, in forward
hidden_states, residual = layer(
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 1084, in forward
return self.forward_normal(
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 1110, in forward_normal
hidden_states = self.self_attn(
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 1140, in torch_dynamo_resume_in_forward_normal_at_1110
hidden_states = self.mlp(hidden_states)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 256, in forward
return self.forward_normal(hidden_states)
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 262, in forward_normal
shared_output = self.shared_experts(hidden_states)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1747, in _call_impl
return forward_call(*args, **kwargs)
File "/sgl-workspace/sglang/python/sglang/srt/models/deepseek_v2.py", line 131, in forward
gate_up, _ = self.gate_up_proj(x)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 1269, in __call__
return self._torchdynamo_orig_callable(
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 1064, in __call__
result = self._inner_convert(
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 526, in __call__
return _compile(
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 952, in _compile
raise InternalTorchDynamoError(
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 924, in _compile
guarded_code = compile_inner(code, one_graph, hooks, transform)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 666, in compile_inner
return _compile_inner(code, one_graph, hooks, transform)
File "/usr/local/lib/python3.10/dist-packages/torch/_utils_internal.py", line 87, in wrapper_function
return function(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 699, in _compile_inner
out_code = transform_code_object(code, transform)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/bytecode_transformation.py", line 1322, in transform_code_object
transformations(instructions, code_options)
File "/usr/local/lib/python3.10/dist-packages/torch/_dynamo/convert_frame.py", line 231, in _fn
torch.cuda.set_rng_state(cuda_rng_state)
File "/usr/local/lib/python3.10/dist-packages/torch/cuda/random.py", line 77, in set_rng_state
_lazy_call(cb)
File "/usr/local/lib/python3.10/dist-packages/torch/cuda/__init__.py", line 249, in _lazy_call
callable()
File "/usr/local/lib/python3.10/dist-packages/torch/cuda/random.py", line 75, in cb
default_generator.set_state(new_state_copy)
torch._dynamo.exc.InternalTorchDynamoError: RuntimeError: CUDA error: an illegal memory access was encountered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 1999, in run_scheduler_process
scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, dp_rank)
File "/sgl-workspace/sglang/python/sglang/srt/managers/scheduler.py", line 249, in __init__
self.tp_worker = TpWorkerClass(
File "/sgl-workspace/sglang/python/sglang/srt/managers/tp_worker.py", line 74, in __init__
self.model_runner = ModelRunner(
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 169, in __init__
self.initialize(min_per_gpu_memory)
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 207, in initialize
self.init_cuda_graphs()
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/model_runner.py", line 919, in init_cuda_graphs
self.cuda_graph_runner = CudaGraphRunner(self)
File "/sgl-workspace/sglang/python/sglang/srt/model_executor/cuda_graph_runner.py", line 269, in __init__
raise Exception(
Exception: Capture cuda graph failed: RuntimeError: CUDA error: an illegal memory access was encountered
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True
Possible solutions:
1. disable cuda graph by --disable-cuda-graph
2. set --mem-fraction-static to a smaller value (e.g., 0.8 or 0.7)
3. disable torch compile by not using --enable-torch-compile
4. set --cuda-graph-max-bs to a smaller value (e.g., 32)
Open an issue on GitHub https://github.com/sgl-project/sglang/issues/new/choose
Environment
Python: 3.10.12 (main, Jan 17 2025, 14:35:34) [GCC 11.4.0]
CUDA available: True
GPU 0,1,2,3,4,5,6,7: NVIDIA A800-SXM4-80GB
GPU 0,1,2,3,4,5,6,7 Compute Capability: 8.0
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.4, V12.4.131
CUDA Driver Version: 550.54.14
PyTorch: 2.5.1+cu124
sglang: 0.4.4.post3+build1
sgl_kernel: 0.0.6
flashinfer: Module Not Found
triton: 3.1.0
transformers: 4.50.0
torchao: 0.9.0
numpy: 1.26.4
aiohttp: 3.11.13
fastapi: 0.115.11
hf_transfer: 0.1.9
huggingface_hub: 0.29.3
interegular: 0.3.3
modelscope: 1.23.2
orjson: 3.10.15
outlines: 0.1.11
packaging: 24.2
psutil: 7.0.0
pydantic: 2.10.6
multipart: Module Not Found
zmq: Module Not Found
uvicorn: 0.34.0
uvloop: 0.21.0
vllm: 0.7.2
xgrammar: 0.1.17
openai: 1.66.3
tiktoken: 0.9.0
anthropic: 0.49.0
litellm: 1.63.7
decord: 0.6.0
NVIDIA Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 NIC7 NIC8 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NV8 NV8 NV8 NV8 NV8 NV8 NV8 PXB PXB SYS SYS SYSSYS SYS SYS SYS 0-31,64-95 0 N/A
GPU1 NV8 X NV8 NV8 NV8 NV8 NV8 NV8 PXB PXB SYS SYS SYSSYS SYS SYS SYS 0-31,64-95 0 N/A
GPU2 NV8 NV8 X NV8 NV8 NV8 NV8 NV8 SYS SYS PXB PXB SYSSYS SYS SYS SYS 0-31,64-95 0 N/A
GPU3 NV8 NV8 NV8 X NV8 NV8 NV8 NV8 SYS SYS PXB PXB SYSSYS SYS SYS SYS 0-31,64-95 0 N/A
GPU4 NV8 NV8 NV8 NV8 X NV8 NV8 NV8 SYS SYS SYS SYS PXBPXB SYS SYS SYS 32-63,96-127 1 N/A
GPU5 NV8 NV8 NV8 NV8 NV8 X NV8 NV8 SYS SYS SYS SYS PXBPXB SYS SYS SYS 32-63,96-127 1 N/A
GPU6 NV8 NV8 NV8 NV8 NV8 NV8 X NV8 SYS SYS SYS SYS SYSSYS PXB PXB SYS 32-63,96-127 1 N/A
GPU7 NV8 NV8 NV8 NV8 NV8 NV8 NV8 X SYS SYS SYS SYS SYSSYS PXB PXB SYS 32-63,96-127 1 N/A
NIC0 PXB PXB SYS SYS SYS SYS SYS SYS X PIX SYS SYS SYSSYS SYS SYS SYS
NIC1 PXB PXB SYS SYS SYS SYS SYS SYS PIX X SYS SYS SYSSYS SYS SYS SYS
NIC2 SYS SYS PXB PXB SYS SYS SYS SYS SYS SYS X PIX SYSSYS SYS SYS SYS
NIC3 SYS SYS PXB PXB SYS SYS SYS SYS SYS SYS PIX X SYSSYS SYS SYS SYS
NIC4 SYS SYS SYS SYS PXB PXB SYS SYS SYS SYS SYS SYS X PIX SYS SYS SYS
NIC5 SYS SYS SYS SYS PXB PXB SYS SYS SYS SYS SYS SYS PIX X SYS SYS SYS
NIC6 SYS SYS SYS SYS SYS SYS PXB PXB SYS SYS SYS SYS SYSSYS X PIX SYS
NIC7 SYS SYS SYS SYS SYS SYS PXB PXB SYS SYS SYS SYS SYSSYS PIX X SYS
NIC8 SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYS SYSSYS SYS SYS X
Legend:
X = Self
SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
PIX = Connection traversing at most a single PCIe bridge
NV# = Connection traversing a bonded set of # NVLinks
NIC Legend:
NIC0: mlx5_2
NIC1: mlx5_3
NIC2: mlx5_4
NIC3: mlx5_5
NIC4: mlx5_6
NIC5: mlx5_7
NIC6: mlx5_8
NIC7: mlx5_9
NIC8: mlx5_bond_0
ulimit soft: 1048576