[Feature] remove vllm _custom_ops

[zhyncs]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

• Support for silu_and_mul and gelu_and_mul in AMD, remove the current dependencies on vllm ops.silu_and_mul and ops.gelu_and_muli. Used in fused_moe_triton.py. [ROCm] Enable silu_and_mul, gelu_and_mul, gelu_tanh_and_mul in amd platform #4150 @yiakwy-xpu-ml-framework-team
• remove from vllm.model_executor.layers.activation import GeluAndMul, SiluAndMul in sglang/python/sglang/srt/layers/activation.py.
• Support GemmaRMSNorm and RMSNorm in AMD.
• remove from vllm.model_executor.layers.layernorm import GemmaRMSNorm, RMSNorm in sglang/python/sglang/srt/layers/layernorm.py.
• Support rotary_embedding kernel in AMD.
• Support for ops.moe_sum in AMD, remove the dependency on vllm ops.moe_sum. Used in fused_moe_triton.py.
• Benchmark vllm ops.moe_align_block_size, moe_align_block_size_triton, and sgl_moe_align_block_size, and remove the num_experts=256 limitation in sgl_moe_align_block_size. After this, directly select the kernel from moe_align_block_size_triton and sgl_moe_align_block_size, and remove the dependency on vllm ops.moe_align_block_size. Used in fused_moe_triton.py. remove moe_align vllm dep #4249 & refine sgl_moe_align_block_size_benchmark #4327
• Implement scaled_int8_quant in sgl-kernel and remove the current dependency on vllm ops.scaled_int8_quant. Used in fused_moe_triton.py. @zcnrex
• Implement per_token_group_quant_int8 in CUDA, replacing the current per_token_group_quant_int8 triton implementation. Used in fused_moe_triton.py. @zcnrex
• Support sglang_per_token_group_quant_fp8 in AMD. Used in fused_moe_triton.py. [tools] add fp8 max/min constant in utils #3959 [ROCm] Enable per token group quant fp8 in amd #3702 @yiakwy-xpu-ml-framework-team
• Implement scaled_fp8_quant kernel and remove the current dependency on vllm ops.scaled_fp8_quant. (This is in progress, 50% complete—see [link]([quant kernel] sgl-kernel support per_tensor_quant fp8 #3786) for per-tensor support, and @hebiao064 is working on per-token support. vllm ops.scaled_fp8_quant will support both per-tensor and per-token.). Used in fused_moe_triton.py, layer.py and fp8.py. @BBuf @hebiao064 Add sgl_per_token_quant_fp8 #4089 [Refactor] Reducing code duplication across FP8 CUDA quantization kernels #4163. https://github.com/sgl-project/sglang/pull/4231。https://github.com/sgl-project/sglang/pull/4215
• Support for apply_rope_with_cos_sin_cache_inplace kernel in AMD, remove the current dependencies on vllm os.rotary_embedding . Used in rotary_embedding.py.
• Implement topk_softmax kernel and remove the current dependency on vllm.ops.topk_softmax. Used in topk.py. Add moe topk softmax templated from vllm #4302
• Support topk_softmax in amd. [ROCm] enable moe topk softmax in amd #4448
• Remove vllm ops.topk_softmax in python/sglang/srt/layers/moe/topk.py. remove vllm ops.topk_softmax dependency #4498
• Implement awq_dequantize kernel and remove the current dependency on vllm ops.awq_dequantize. Used in deepseek_nextn.py and deepseek_v2.py. Add awq dequantize kernel to sgl with 1x to 3x speedup #4104 @zcnrex

Related resources

No response

[zhyncs]

We need to ensure that

1. the functionality is consistent with the original
2. without precision issues
3. and that performance is as good as or better than before