File "/sglang/python/sglang/srt/layers/sampler.py", line 164, in top_k_top_p_min_p_sampling_from_probs_torch
    min_p_thresholds = probs_sort[:, 0] * min_ps
    TypeError: unsupported operand type(s) for *: 'Tensor' and 'NoneType' (where min_ps is None)

root@x:/sglang# VLLM_MOE_PADDING=0 python -m sglang.bench_latency --batch-size 32 --input 1024 --output 8 --model dummy_half_grok1/ --tokenizer-path Xenova/grok-1-tokenizer --load-format dummy --tp 8 --quant fp8  --attention-backend triton --sampling-backend  pytorch --disable-cuda-graph
Warmup ...
Prefill. latency: 25.79838 s, throughput:   1270.16 token/s
Decode.  latency: 0.53607 s, throughput:     59.69 token/s
Decode.  latency: 0.31325 s, throughput:    102.15 token/s
Decode.  latency: 0.04105 s, throughput:    779.47 token/s
Decode.  latency: 0.04075 s, throughput:    785.31 token/s
Decode.  median latency: 0.17715 s, median throughput:    180.64 token/s
Total. latency: 26.730 s, throughput:   1230.70 token/s
Benchmark ...
Prefill. latency: 1.11868 s, throughput:  29291.72 token/s
Decode.  latency: 0.02593 s, throughput:   1234.10 token/s
Decode.  latency: 0.02575 s, throughput:   1242.87 token/s
Decode.  latency: 0.02559 s, throughput:   1250.46 token/s
Decode.  latency: 0.02646 s, throughput:   1209.23 token/s
Decode.  latency: 0.02574 s, throughput:   1243.42 token/s
Decode.  median latency: 0.02574 s, median throughput:   1243.15 token/s
Total. latency:  1.325 s, throughput:  24925.60 token/s
root@x:/sglang#