[Bug] AWQ scalar type error

[zjp-shadow]

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

When I run the Deepseek-R1-AWQ, I met a scalar type bug same as pr #3450 . @hnyls2002

Loading safetensors checkpoint shards:  97% Completed | 72/74 [00:44<00:01,  1.55it/s]
Loading safetensors checkpoint shards:  99% Completed | 73/74 [00:45<00:00,  1.72it/s]
[2025-02-22 12:19:24 TP3] Scheduler hit an exception: Traceback (most recent call last):
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/managers/scheduler.py", line 1816, in run_scheduler_process
    scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, dp_rank)
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/managers/scheduler.py", line 240, in __init__
    self.tp_worker = TpWorkerClass(
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/managers/tp_worker_overlap_thread.py", line 63, in __init__
    self.worker = TpModelWorker(server_args, gpu_id, tp_rank, dp_rank, nccl_port)
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/managers/tp_worker.py", line 68, in __init__
    self.model_runner = ModelRunner(
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/model_executor/model_runner.py", line 195, in __init__
    self.load_model()
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/model_executor/model_runner.py", line 318, in load_model
    self.model = get_model(
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/model_loader/__init__.py", line 22, in get_model
    return loader.load_model(
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/model_loader/loader.py", line 362, in load_model
    model.load_weights(self._get_all_weights(model_config, model))
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/sglang/srt/models/deepseek_v2.py", line 962, in load_weights
    w = ops.awq_dequantize(
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/vllm/_custom_ops.py", line 222, in awq_dequantize
    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
  File "/mnt/Shared_h0/zjp/anaconda3/envs/deepseek/lib/python3.9/site-packages/torch/_ops.py", line 1116, in __call__
    return self._op(*args, **(kwargs or {}))
RuntimeError: expected scalar type Half but found BFloat16

Reproduction

I use the command recommended by instructions

python3 -m sglang.launch_server --model cognitivecomputations/DeepSeek-R1-AWQ --tp 8 --trust-remote-code

But it produces the mistake same as preview.

RuntimeError: expected scalar type Half but found BFloat16

Environment

INFO 02-22 12:35:34 __init__.py:190] Automatically detected platform cuda.
Python: 3.9.21 (main, Dec 11 2024, 16:24:11) [GCC 11.2.0]
CUDA available: True
GPU 0,1,2,3,4,5,6,7: NVIDIA H100 80GB HBM3
GPU 0,1,2,3,4,5,6,7 Compute Capability: 9.0
CUDA_HOME: /usr/local/cuda
NVCC: Cuda compilation tools, release 12.4, V12.4.131
CUDA Driver Version: 570.86.15
PyTorch: 2.5.1+cu124
sglang: 0.4.3.post2
sgl_kernel: 0.0.3.post6
flashinfer: 0.2.1.post2+cu124torch2.5
triton: 3.1.0
transformers: 4.48.3
torchao: 0.8.0
numpy: 1.26.4
aiohttp: 3.11.12
fastapi: 0.115.8
hf_transfer: 0.1.9
huggingface_hub: 0.29.1
interegular: 0.3.3
modelscope: 1.23.0
orjson: 3.10.15
packaging: 24.2
psutil: 7.0.0
pydantic: 2.10.6
multipart: 0.0.20
zmq: 26.2.1
uvicorn: 0.34.0
uvloop: 0.21.0
vllm: 0.7.2
openai: 1.63.2
tiktoken: 0.9.0
anthropic: 0.46.0
decord: 0.6.0
NVIDIA Topology: 
        GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV18    NODE    NODE    SYS     SYS     SYS     SYS     0-127,256-383   0               N/A
GPU1    NV18     X      NODE    NODE    SYS     SYS     SYS     SYS     0-127,256-383   0               N/A
GPU2    NODE    NODE     X      NV17    SYS     SYS     SYS     SYS     0-127,256-383   0               N/A
GPU3    NODE    NODE    NV17     X      SYS     SYS     SYS     SYS     0-127,256-383   0               N/A
GPU4    SYS     SYS     SYS     SYS      X      NV18    NODE    NODE    128-255,384-511 1               N/A
GPU5    SYS     SYS     SYS     SYS     NV18     X      NODE    NODE    128-255,384-511 1               N/A
GPU6    SYS     SYS     SYS     SYS     NODE    NODE     X      NV18    128-255,384-511 1               N/A
GPU7    SYS     SYS     SYS     SYS     NODE    NODE    NV18     X      128-255,384-511 1               N/A

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

ulimit soft: 1024

[FrankLeeeee]

Hi let me try to reproduce it.

[FrankLeeeee]

I managed to run this model with the following command by adding the argument --dtype half.

python3 -m sglang.launch_server --model cognitivecomputations/DeepSeek-R1-AWQ --tp 8 --trust-remote-code --dtype half

[zjp-shadow]

I managed to run this model with the following command by adding the argument --dtype half.

python3 -m sglang.launch_server --model cognitivecomputations/DeepSeek-R1-AWQ --tp 8 --trust-remote-code --dtype half

This can indeed be started, but I'm not sure if the issue is caused by the precision of different dtypes. Using R1-AWQ in fp16 tends to result in gibberish, which I didn't encounter in the quantization like 1.58bit of Ollama I deployed earlier. The answer of R1-AWQ is shown in the picture.

[fridayL]

I managed to run this model with the following command by adding the argument --dtype half.
python3 -m sglang.launch_server --model cognitivecomputations/DeepSeek-R1-AWQ --tp 8 --trust-remote-code --dtype half

This can indeed be started, but I'm not sure if the issue is caused by the precision of different dtypes. Using R1-AWQ in fp16 tends to result in gibberish, which I didn't encounter in the quantization like 1.58bit of Ollama I deployed earlier. The answer of R1-AWQ is shown in the picture.

I get same problem, how did you solve it

[FrankLeeeee]

@zjp-shadow @fridayL

I figured it out, AWQ does not work with MLA yet. with this command, you can get the model running and generating the expected output.

python3 -m sglang.launch_server --model cognitivecomputations/DeepSeek-R1-AWQ --tp 8 --trust-remote-code --dtype half --disable-mla

My input is like

messages=[
        {"role": "user", "content": "List 3 countries and their capitals."},
    ],
    temper

This is my output:

"<think>\nOkay, let's see. The user is asking for a list of three countries and their capitals. Hmm, I need to make sure I pick countries that are well-known so the answer is useful. Maybe start with some obvious ones. United States, Canada, Mexico? Wait, but maybe some people might"