RetroSearch Browse

Home - News ( United States | United Kingdom | Italy | Germany ) - Football scores

Showing content from https://github.com/sgl-project/sglang/issues/6088 below:

[Bug] Qwen 2 MoE crashes on idle batch when DP Attention/EPMoE is enabled · Issue #6088 · sgl-project/sglang · GitHub

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

As described in #5858 (comment), DP Attention (together with the new EPMoE support introduced by #5917) is failing on idle batches.

Reproduction

Server:

python -m sglang.launch_server \
  --model-path path-to-Qwen2-57B-A14B --trust-remote-code \
  --mem-fraction-static 0.8 \
  --max-prefill-tokens 2048 \ 
  --chunked-prefill-size -1 \
  --tp-size 4 \
  --enable-dp-attention \
  --dp-size 4

Client:

import requests

port = 30000
prompt = "The capital of France is"

response = requests.post(
    f"http://localhost:{port}/generate",
    json={
        "text": prompt,
        "sampling_params": {
            "temperature": 0.7,
            "max_new_tokens": 2048,
        },
        "stream": True,
    },
    stream=True,
)

This would cause the server to crash. Error from one TPWorker:

[2025-05-07 12:19:42 DP1 TP1] TpModelWorkerClient hit an exception: Traceback (most recent call last):
  File "/code/sglang/python/sglang/srt/managers/tp_worker_overlap_thread.py", line 118, in forward_thread_func
    self.forward_thread_func_()
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/code/sglang/python/sglang/srt/managers/tp_worker_overlap_thread.py", line 148, in forward_thread_func_
    logits_output, next_token_ids = self.worker.forward_batch_generation(
  File "/code/sglang/python/sglang/srt/managers/tp_worker.py", line 206, in forward_batch_generation
    logits_output = self.model_runner.forward(
  File "/code/sglang/python/sglang/srt/model_executor/model_runner.py", line 1105, in forward
    return self.forward_idle(forward_batch, pp_proxy_tensors=pp_proxy_tensors)
  File "/code/sglang/python/sglang/srt/model_executor/model_runner.py", line 1071, in forward_idle
    return self.model.forward(
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/code/sglang/python/sglang/srt/models/qwen2_moe.py", line 418, in forward
    hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
    return forward_call(*args, **kwargs)
  File "/code/sglang/python/sglang/srt/models/qwen2_moe.py", line 379, in forward
    hidden_states, residual = layer(
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
    return forward_call(*args, **kwargs)
  File "/code/sglang/python/sglang/srt/models/qwen2_moe.py", line 318, in forward
    hidden_states = self.input_layernorm(hidden_states)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
    return forward_call(*args, **kwargs)
  File "/code/sglang/python/sglang/srt/layers/layernorm.py", line 56, in forward
    return self.forward_cuda(*args, **kwargs)
  File "/code/sglang/python/sglang/srt/layers/layernorm.py", line 70, in forward_cuda
    out = rmsnorm(x, self.weight.data, self.variance_epsilon)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/sgl_kernel/elementwise.py", line 41, in rmsnorm
    torch.ops.sgl_kernel.rmsnorm.default(out, input, weight, eps, enable_pdl)
  File "/opt/conda/envs/sglang/lib/python3.10/site-packages/torch/_ops.py", line 723, in __call__
    return self._op(*args, **kwargs)
RuntimeError: RMSNorm failed with error code invalid configuration argument

Environment

$ python3 -m sglang.check_env
Python: 3.10.16 (main, Dec 11 2024, 16:24:50) [GCC 11.2.0]
CUDA available: True
GPU 0,1,2,3,4,5,6,7: NVIDIA A40
GPU 0,1,2,3,4,5,6,7 Compute Capability: 8.6
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.4, V12.4.99
CUDA Driver Version: 535.146.02
PyTorch: 2.6.0+cu124
sglang: 0.4.6.post2
sgl_kernel: 0.1.1
flashinfer_python: 0.2.5
triton: 3.2.0
transformers: 4.51.1
torchao: 0.10.0
numpy: 2.2.5
aiohttp: 3.11.18
fastapi: 0.115.12
hf_transfer: 0.1.9
huggingface_hub: 0.30.2
interegular: 0.3.3
modelscope: 1.25.0
orjson: 3.10.16
outlines: 0.1.11
packaging: 25.0
psutil: 7.0.0
pydantic: 2.11.3
python-multipart: 0.0.20
pyzmq: 26.4.0
uvicorn: 0.34.2
uvloop: 0.21.0
vllm: Module Not Found
xgrammar: 0.1.17
openai: 1.76.0
tiktoken: 0.9.0
anthropic: 0.50.0
litellm: 1.67.4
decord: 0.6.0
NVIDIA Topology:
        GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    NIC0    NIC1    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      PXB     PXB     PXB     SYS     SYS     SYS     SYS     PXB     PXB     0-25,52-77      0               N/A
GPU1    PXB      X      PXB     PXB     SYS     SYS     SYS     SYS     PIX     PIX     0-25,52-77      0               N/A
GPU2    PXB     PXB      X      PIX     SYS     SYS     SYS     SYS     PXB     PXB     0-25,52-77      0               N/A
GPU3    PXB     PXB     PIX      X      SYS     SYS     SYS     SYS     PXB     PXB     0-25,52-77      0               N/A
GPU4    SYS     SYS     SYS     SYS      X      PXB     PXB     PXB     SYS     SYS     26-51,78-103    1               N/A
GPU5    SYS     SYS     SYS     SYS     PXB      X      PXB     PXB     SYS     SYS     26-51,78-103    1               N/A
GPU6    SYS     SYS     SYS     SYS     PXB     PXB      X      PIX     SYS     SYS     26-51,78-103    1               N/A
GPU7    SYS     SYS     SYS     SYS     PXB     PXB     PIX      X      SYS     SYS     26-51,78-103    1               N/A
NIC0    PXB     PIX     PXB     PXB     SYS     SYS     SYS     SYS      X      PIX
NIC1    PXB     PIX     PXB     PXB     SYS     SYS     SYS     SYS     PIX      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_0
  NIC1: mlx5_1


ulimit soft: 1024

RetroSearch is an open source project built by @garambo | Open a GitHub Issue

Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo

HTML: 3.2 | Encoding: UTF-8 | Version: 0.7.4