Enable Model Config Input via a Centralized Parser in utils.py

[DiweiSun]

This PR proposes to enhance the model configuration system by introducing a flexible and reusable mechanism to load model parameters — either from a Hugging Face model config (--model-name) or via manual CLI overrides — instead of relying on hardcoded values. The implementation will be centralized in a new utils.py module to enable consistent configuration handling across multiple benchmark and inference scripts.

This change enables users to:
✅ Enable native pytorch op path for topk-softmax

✅ Benchmark models using real-world configurations (e.g., deepseek-ai/DeepSeek-R1)
✅ Override specific parameters (e.g., --num-experts 128)
✅ Avoid hardcoded test configurations
✅ Reuse config logic across tools

[DiweiSun]

Top-K Softmax Performance Benchmark

Case 1

python bench_moe_topk_softmax.py  --model-name Qwen/Qwen3-VL-8B-Instruct --top-k 2

📡 Loading config from Hugging Face: Qwen/Qwen3-VL-8B-Instruct
`torch_dtype` is deprecated! Use `dtype` instead!
⚙️ Overriding top_k = [2] (from CLI)
sweep_params {'num_tokens': [100], 'num_experts': [64], 'top_k': [2], 'dtype': [torch.bfloat16], 'renormalize': [False]}
Testing 1 configurations...
Config: num_tokens=100, num_experts=64, topk=2, dtype=torch.bfloat16, renormalize=False
Starting performance benchmark...
topk-softmax-performance:
   num_tokens  num_experts  topk           dtype  renormalize  SGLang  native
0         100           64     2  torch.bfloat16        False   15.16   86.32

Case 2 python bench_moe_topk_softmax.py

python bench_moe_topk_softmax.py  

🔧 No --model-name provided. Using CLI args or defaults.
💡 Using default num_experts = 64
💡 Using default top_k = 2
💡 Using default num_layers = 32
💡 Using default hidden_size = 4096
💡 Using default ffn_hidden_size = 11008
💡 Using default num_heads = 32
💡 Using default num_kv_heads = 8
💡 Using default head_dim = 128
💡 Using default vocab_size = 32000
💡 Using default max_seq_len = 32768
💡 Using default norm_eps = 1e-06
💡 Using default architectures = ['LlamaForCausalLM']
💡 Using default dtype = float16
sweep_params {'num_tokens': [100], 'num_experts': [64], 'top_k': [2], 'dtype': [torch.bfloat16], 'renormalize': [False]}
Testing 1 configurations...
Config: num_tokens=100, num_experts=64, topk=2, dtype=torch.bfloat16, renormalize=False
Starting performance benchmark...
topk-softmax-performance:
   num_tokens  num_experts  topk           dtype  renormalize  SGLang  native
0         100           64     2  torch.bfloat16        False    15.2    86.8

[chunyuan-w]

This native function does not support renormalize. You may follow this one in SGLang:

https://github.com/sgl-project/sglang/blob/36942660513213f3ecd4a39ad64d6cf3127328a9/python/sglang/srt/layers/moe/topk.py#L411-L417

[airMeng]

so bench_moe_topk_softmax.py not needed in

sgl-kernel-xpu/.github/workflows/pr-test-xpu.yml

Line 58 in 688c0b8

/bin/bash -c "cd /root/sglang/sgl-kernel-xpu/benchmark && python3 bench_flash_attn.py "

?

[DiweiSun]

so bench_moe_topk_softmax.py not needed in

sgl-kernel-xpu/.github/workflows/pr-test-xpu.yml

Line 58 in 688c0b8

/bin/bash -c "cd /root/sglang/sgl-kernel-xpu/benchmark && python3 bench_flash_attn.py "

?

Tested in CI? For performance test? But We dont have performance check in benchmark_kernel scripts, like criteria, scenarios, fail/pass conditions. I suppose functionality test has already been enabled in UT cases, no need to track with kernel bench

[airMeng]

better to add bench_moe_topk_softmax.py first even we haven't tracked it