[Bug] generate wrong sequences with higher temperature

[StevenZHB]

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.

Describe the bug

I started a vllm server and a sglang server with the same model. I found sglang server output unreadable tokens with high temeratur while vllm not.
Example:
sampling_params = {"temperature":1,"n":1}
sglang server: 'Expossible! description! description!爽 ipairs soc!!爽爽爽爽'
vllm server: "Let's start by using the given information to set up three equations:\n1."

sampling_params = {"temperature":0.2,"n":1}
sglang server: "Let's start by using the given information to set up three equations:\n1."
vllm server: "Let's start by using the given information to set up three equations:\n1."

maybe it's related to #523 , I don't know how to fix it

Reproduction

CUDA_VISIBLE_DEVICES=0 nohup python3 -m sglang.launch_server --model-path llama3-8B-instruct --port 9554 --disable-cuda-graph --mem-fraction-static 0.75 --max-prefill-tokens 12800

Environment

python3 -m sglang.check_env
Python: 3.9.17 (main, Jul  5 2023, 20:41:20) [GCC 11.2.0]
CUDA available: True
GPU 0,1: NVIDIA A800-SXM4-80GB
CUDA_HOME: /zhanghongbo/CUDA/cuda-11_8
NVCC: Cuda compilation tools, release 11.8, V11.8.89
CUDA Driver Version: 525.105.17
525.105.17
PyTorch: 2.3.0+cu118
flashinfer: 0.1.1+cu118torch2.3
requests: 2.31.0
tqdm: 4.66.1
numpy: 1.25.0
aiohttp: 3.8.5
fastapi: 0.110.0
hf_transfer: Module Not Found
huggingface_hub: 0.23.2
interegular: 0.3.3
packaging: 24.0
pillow: Module Not Found
psutil: 5.9.8
pydantic: 2.5.0
uvicorn: 0.23.2
uvloop: 0.19.0
zmq: 25.1.2
vllm: 0.5.3.post1
openai: 1.30.0
anthropic: Module Not Found
litellm: Module Not Found
NVIDIA Topology: 
        GPU0    GPU1    NIC0    CPU Affinity    NUMA Affinity
GPU0     X      NV8     SYS     48-63   3
GPU1    NV8      X      SYS     48-63   3
NIC0    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_0


ulimit soft: 1048576

[merrymercy]

Thanks for reporting this. I can reproduce the error with llama-3-8b.

One possible reason is that we previously stored the logits in bfloat16. This can cause some problems, but it has been fixed by PR #773. Could you try the main branch again?

After PR #733, temp=1 works great for me. However, I still get random output with temp=3. I guess this is expected because llama-3 has a vocab size of 128K, and temp=3 flattens the logits too much, making softmax/sampling unstable at this point.

[ispobock]

Hi @merrymercy , the issue seems still there when I use the flashinfer sampling kernel. But it's fine when I switch to the fallback torch implementation.

python3 -m sglang.launch_server --model-path /workdir/llm_models/Meta-Llama-3-8B --port 30000 --trust-remote-code

curl -X POST http://0.0.0.0:30000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "/workdir/llm_models/Meta-Llama-3-8B",
"prompt": "Please introduce yourself",
"max_tokens": 100,
"stream": false,
"temperature": 1
}'

env:

cuda 11.8
flashinfer: 0.1.1+cu118torch2.3 & main branch
torch: 2.3.1+cu118
sglang: main branch

cc: @yzh119

[yzh119]

This should be a bug in flashinfer sampling kernel, thanks for reporting.

[merrymercy]

@StevenZHB Please try the main branch first to see if putting logits to float32 resolves your issue (#773).
If you still see wrong outputs, please try to add --disable-flashinfer-sampling when you launch the server (#778).

[StevenZHB]

@merrymercy Thanks, I have tried with the main branch. It still generates wrong outputs. But it can be resolved by setting --disable-flashinfer-sampling. So maybe it's a bug in flashinfer.

[merrymercy]

@yzh119 This seems like a critical bug. Could you prioritize a fix?

[yzh119]

Yes I'll release v0.1.2 tonight.

[yzh119]

I just noticed that the user is using cu118, the bug was already fixed in flashinfer-ai/flashinfer#386

[ispobock]

@yzh119 I tried flashinfer 0.1.2, but still got the issue. This issue is different from flashinfer-ai/flashinfer#384, could you help check?
@StevenZHB could you verify flashinfer 0.1.2 on your environment?

[yzh119]

Sure I can check, could you help dump the logits score so that I can debug on them directly?

[yzh119]

Bug reproduced, I think I have figured out the bug, v0.1.3 will fix the issue.

[yzh119]

Fixed in flashinfer-ai/flashinfer#410

[zhyncs]

Hi @StevenZHB May you try the latest main branch? It should be ok now. Thanks.

git clone https://github.com/sgl-project/sglang.git
cd sglang

pip install --upgrade pip
pip install -e "python[all]"

pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/

[StevenZHB]

Thanks a lot! I tested the latest version, and everything is working perfectly now.