[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

• scaled_fp8_quant @HandH1998
  

  sglang/python/sglang/srt/layers/quantization/fp8.py

  Line 10 in 2add697

  from vllm import _custom_ops as ops

  
  

  sglang/python/sglang/srt/layers/quantization/fp8.py

  Line 297 in 2add697

  qweight, weight_scale = ops.scaled_fp8_quant(layer.weight, scale=None)

  
  

  sglang/python/sglang/srt/layers/moe/ep_moe/layer.py

  Line 6 in 2add697

  from vllm import _custom_ops as ops

  
  

  sglang/python/sglang/srt/layers/moe/ep_moe/layer.py

  Line 603 in 2add697

  ops.scaled_fp8_quant(layer.w13_weight.data[expert, :, :])

• rotary_embedding @ByronHsu
  

  sglang/python/sglang/srt/layers/rotary_embedding.py

  Line 142 in 2add697

  from vllm import _custom_ops as ops

  
  

  sglang/python/sglang/srt/layers/rotary_embedding.py

  Lines 159 to 166 in 2add697

  	ops.rotary_embedding(
  	positions,
  	query,
  	key,
  	self.head_size,
  	self.cos_sin_cache,
  	self.is_neox_style,
  	)

  
  BTW we don't need batched_rotary_embedding

• topk_softmax @zhyncs
  

  sglang/python/sglang/srt/layers/moe/topk.py

  Line 48 in 2add697

  from vllm import _custom_ops as ops

  
  

  sglang/python/sglang/srt/layers/moe/topk.py

  Lines 62 to 67 in 2add697

  	ops.topk_softmax(
  	topk_weights,
  	topk_ids,
  	token_expert_indicies,
  	gating_output.float(),
  	)

• moe_align_block_size moe_sum @BBuf @zhyncs
  

  sglang/python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

  Line 14 in 2add697

  from vllm import _custom_ops as ops

  
  

  sglang/python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

  Lines 448 to 455 in 2add697

  	ops.moe_align_block_size(
  	topk_ids,
  	num_experts,
  	block_size,
  	sorted_ids,
  	expert_ids,
  	num_tokens_post_pad,
  	)

  
  

  sglang/python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

  Lines 1012 to 1015 in 2add697

  	ops.moe_sum(
  	intermediate_cache3.view(*intermediate_cache3.shape),
  	out_hidden_states[begin_chunk_idx:end_chunk_idx],
  	)

• awq_dequantize @bjmsong @zhyncs
  

  sglang/python/sglang/srt/models/deepseek_v2.py

  Line 25 in 2add697

  from vllm import _custom_ops as ops

  
  

  sglang/python/sglang/srt/models/deepseek_v2.py

  Lines 946 to 952 in 2add697

  	w = ops.awq_dequantize(
  	self_attn.kv_b_proj.qweight,
  	self_attn.kv_b_proj.scales,
  	self_attn.kv_b_proj.qzeros,
  	0,
  	0,
  	0,

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