大模型推理引擎vLLM(30): 参考sglang代码，重构vllm021中EP高吞吐代码，消除空泡问题:400us减小到25us-北京尧图网络

目录1 问题描述2 sglang的这个过程一件事做完再干下一件2.1 代码第 1 步 —— dispatch 只交作业不拷名单2.2 代码第 2 步 —— 同一个函数里拷名单 scatter2.2.1 拆开 dispatch 的结果2.2.2 纯 CPU算总长度开输出 buffer2.2.3 立刻 HtoD2.2.4 下一行就是 scatter里面先 scan 再 scatter3 vLLM同样两件事但拆成两个房间做3.1 代码第 1 步 —— _receiver 里就把名单拷了3.1.1 等通信3.1.2 改 topksglang 这条 DeepGEMM 路径基本不做3.1.3 立刻 HtoD也就是memcpy3.2 代码中间走廊 —— modular_kernel3.3 代码第 2 步 —— 很晚才 scatter3.3.1 用之前拷上来的 meta 算对齐长度CPU3.3.2 两个 native fill你 profile 里看到的3.3.3 再数一遍 counts不用刚才 Memcpy 上去的那份做 scatter3.3.4 这才 scan scatter4 消除空泡方法15 消除空泡方法26 消除空泡方法37 消除空泡方法48 总结abstract:其实这里消除空泡的核心方法就是看dispatch和scatter之间有哪些cpu调用消耗了时间然后看看这些cpu调用能不能替换成更省时间的或者直接删掉最终效果就是空泡从400us减小成了25us效果显著。1 问题描述上面的这个是vllm的prof图这个是sglang的prof可以看到sglang是没有空泡的那么把 sglang和vllm的这块代码看懂然后借鉴sglang的代码消除下vllm的空泡问题。2 sglang的这个过程一件事做完再干下一件假设 DeepEP 通信刚结束本 rank 手里有GPU 上的 token 数据hiddenGPU 上的topk_idsCPU 上的一份名单counts [3, 5, 2, ...]每个 expert 分到几个 token后面要做的事本质一样把这份名单拷到 GPU再按名单做 scan scatter。差别只在于这两步中间夹了没有别的事。sglang的大体过程如下时间 →[1] DeepEP dispatch 结束手里有 counts还在 CPU 的 List[2] 马上进 pre_permute 这一个函数CPU: sum(counts) → 算要开多大 bufferGPU: empty 开几块内存GPU: 把 counts 拷上去 ← profile 里的 MemcpyGPU: 立刻 ep_scatter ← 紧接着 scan scatter[3] 去做 grouped gemmMemcpy 和 scatter 写在同一个函数里前后两行所以中间几乎没空泡。2.1 代码第 1 步 —— dispatch 只交作业不拷名单sglang/python/sglang/srt/layers/moe/token_dispatcher/deepep.pydef dispatch_b(self, hidden_states, topk_ids, topk_weights, previous_event): ( hidden_states, topk_ids, topk_weights, num_recv_tokens_per_expert, event, ) self._dispatch_core(hidden_states, topk_ids, topk_weights, previous_event) event.current_stream_wait() if self.async_finish else () if isinstance(hidden_states, tuple): hidden_states, hidden_states_scale hidden_states else: hidden_states_scale None return DeepEPNormalDispatchOutput( hidden_states, hidden_states_scale, topk_ids, topk_weights, num_recv_tokens_per_expert, )DeepEP 跑完了通信结束。num_recv_tokens_per_expert仍然是 CPU 上的List[int]。这里没有.cuda()所以这里不会出现你盯的那次 Memcpy。输出是这样的DeepEPNormalDispatchOutput( hidden_states..., # GPU hidden_states_scale..., # GPU topk_ids..., # GPU topk_weights..., # GPU num_recv_tokens_per_expert[3, 5, 2, ...], # CPU list )2.2 代码第 2 步 —— 同一个函数里拷名单 scattersglang/python/sglang/srt/layers/moe/moe_runner/deep_gemm.pyregister_pre_permute(deepep_normal, deep_gemm) def pre_permute_deepep_normal_to_deep_gemm( dispatch_output: DeepEPNormalDispatchOutput, quant_info: DeepGemmMoeQuantInfo, runner_config: MoeRunnerConfig, running_state: dict, ) - DeepGemmRunnerInput: from sglang.srt.layers.moe.ep_moe.kernels import ep_scatter ( hidden_states, hidden_states_scale, topk_ids, topk_weights, num_recv_tokens_per_expert, ) dispatch_output assert runner_config.activation silu all_tokens sum(num_recv_tokens_per_expert) running_state[all_tokens] all_tokens K hidden_states.shape[1] hidden_states_shape hidden_states.shape hidden_states_device hidden_states.device hidden_states_dtype hidden_states.dtype running_state[hidden_states_shape] hidden_states_shape running_state[hidden_states_device] hidden_states_device running_state[hidden_states_dtype] hidden_states_dtype running_state[topk_ids] topk_ids running_state[topk_weights] topk_weights input_tensor torch.empty( (all_tokens, K), devicehidden_states.device, dtypehidden_states.dtype, ) if deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0: # TODO check whether need zeros input_tensor_scale torch.zeros( (ceil_div(K // 128, 4), all_tokens), devicehidden_states.device, dtypetorch.int, ).transpose(0, 1) else: input_tensor_scale torch.empty( (all_tokens, K // 128), devicehidden_states.device, dtypetorch.float32, ) m_indices torch.empty(all_tokens, devicehidden_states.device, dtypetorch.int32) output_index torch.empty_like(topk_ids) if get_offloader().forbid_copy_engine_usage: num_recv_tokens_per_expert_gpu copy_list_to_gpu_no_ce( num_recv_tokens_per_expert ) else: num_recv_tokens_per_expert_gpu torch.tensor( num_recv_tokens_per_expert, dtypetorch.int32, pin_memoryTrue, devicecpu, ).cuda(non_blockingTrue) expert_start_loc torch.empty_like(num_recv_tokens_per_expert_gpu) ep_scatter( hidden_states, hidden_states_scale, topk_ids, num_recv_tokens_per_expert_gpu, expert_start_loc, input_tensor, input_tensor_scale, m_indices, output_index, scale_ue8m0deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0, ) dispose_tensor(hidden_states) dispose_tensor(hidden_states_scale) running_state[output_index] output_index return DeepGemmRunnerInput( hidden_statesinput_tensor, hidden_states_scaleinput_tensor_scale, use_masked_gemmFalse, m_indicesm_indices, )把上面的代码逐段读一下2.2.1 拆开 dispatch 的结果( hidden_states, hidden_states_scale, topk_ids, topk_weights, num_recv_tokens_per_expert, # 还是 list ) dispatch_output2.2.2 纯 CPU算总长度开输出 bufferall_tokens sum(num_recv_tokens_per_expert) # CPU 加法不上 GPU input_tensor torch.empty((all_tokens, K), ...) # 开输出 input_tensor_scale torch.empty(...) m_indices torch.empty(...) # 注意empty不是 full(-1) output_index torch.empty_like(topk_ids)这些是在准备 scatter 要用的空盒子。还没拷 counts。2.2.3 立刻 HtoDnum_recv_tokens_per_expert_gpu torch.tensor( num_recv_tokens_per_expert, # CPU list dtypetorch.int32, pin_memoryTrue, devicecpu, ).cuda(non_blockingTrue) # ← 这里出现 Memcpy2.2.4 下一行就是 scatter里面先 scan 再 scatterep_scatter( hidden_states, hidden_states_scale, topk_ids, num_recv_tokens_per_expert_gpu, # 刚拷上去的 counts expert_start_loc, input_tensor, ... )所以 sglang 的 GPU 时间线就是... dispatch 通信 ... | Memcpy(counts) | scan | scatter | gemm ...↑________________↑几乎贴在一起3 vLLM同样两件事但拆成两个房间做时间 → [1] DeepEP dispatch 结束和 sglang 一样手里也有 counts: List[int] [2] 进 _receiverprepare 收尾 ← 「第一个房间」 torch.where 改 topk_ids 立刻 make_from_list把 counts 拷到 GPU ← Memcpy 出现在这里 return带着 meta 离开这个房间 [3] 回到 modular_kernel ← 「走廊」 _prepare 结束再调 _fused_experts 再进 DeepGemmExperts.apply 再算 workspace / M_sum ... 这段 GPU 往往没事干 → 空泡 [4] 终于进 deepgemm_moe_permute ← 「第二个房间」 torch.full(-1) × 2 count_expert再数一遍才 ep_scatterscan scatter3.1 代码第 1 步 ——_receiver里就把名单拷了vllm021/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_ht.py3.1.1 等通信if event.event is not None: event.current_stream_wait()和 sglangdispatch_b里 wait 一样通信结束。3.1.2 改 topksglang 这条 DeepGEMM 路径基本不做expert_topk_ids torch.where( expert_topk_ids -1, ..., expert_topk_ids self.rank_expert_offset, # local → global )3.1.3 立刻 HtoD也就是memcpyexpert_tokens_meta mk.ExpertTokensMetadata.make_from_list( expert_num_tokens_per_expert_list, deviceexpert_x.device )make_from_list实际干的事expert_num_tokens_cpu torch.tensor(list, devicecpu, pin_memoryTrue) return ExpertTokensMetadata( expert_num_tokensexpert_num_tokens_cpu.to(device, non_blockingTrue), # ↑ 这里就是 Memcpy expert_num_tokens_cpuexpert_num_tokens_cpu, )注意到这里还没有调用ep_scatter。函数直接return了 token、scale、meta、topk。3.2 代码中间走廊 —— modular_kernela1q, a1q_scale, expert_tokens_meta, topk_ids, topk_weights self._prepare(...) # ↑ 里面已经跑完 _receiver → Memcpy 已经发生 fused_out self._fused_experts(..., expert_tokens_metaexpert_tokens_meta, ...) # ↑ 这里面很晚才调到 deepgemm_moe_permute → 才 scatter_prepare和_fused_experts之间CPU 还在调 Python、进 experts、算 workspace。GPU 上 counts 已经拷完了但 scan/scatter 还没 enqueue → profile 里就是白的。3.3 代码第 2 步 —— 很晚才 scattervllmhcu021/vllm_hcu/model_executor/layers/fused_moe/deep_gemm_utils.pydef deepgemm_moe_permute( aq: torch.Tensor, aq_scale: torch.Tensor, topk_ids: torch.Tensor, local_num_experts: int, expert_map: torch.Tensor | None, expert_tokens_meta: mk.ExpertTokensMetadata | None, aq_out: torch.Tensor | None None, ): assert aq.ndim 2 assert topk_ids.dtype.is_signed, The kernel uses -1 to represent invalid topk_ids H aq.size(1) device aq.device # block_m, block_k get_mk_alignment_for_contiguous_layout() block_m 256 M_sum compute_aligned_M( Mtopk_ids.size(0), num_topktopk_ids.size(1), local_num_expertslocal_num_experts, alignmentblock_m, expert_tokens_metaexpert_tokens_meta, ) expert_start_loc torch.empty( (local_num_experts), devicedevice, dtypetorch.int32 ) assert aq_out is None or aq_out.shape (M_sum, H) if aq_out is None: aq_out torch.empty((M_sum, H), devicedevice, dtypeaq.dtype) # aq_scale_out torch.empty( # (M_sum, H // block_k), devicedevice, dtypetorch.float32 # ) aq_scale_out torch.empty( (M_sum, aq_scale.shape[-1]), devicedevice, dtypetorch.float32 ) # DeepGEMM uses negative values in m_indices (here expert_ids) to mark # completely invalid / padded blocks that should be skipped. We always # initialize expert_ids to -1 so any row that is not explicitly written # by the scatter kernel will be treated as invalid and skipped by # DeepGEMMs scheduler. expert_ids torch.full( (M_sum,), fill_value-1, devicedevice, dtypetorch.int32, ) inv_perm torch.full( topk_ids.shape, fill_value-1, devicedevice, dtypetorch.int32 ) # Derive per-expert counts from topk_ids so ep_scatter layout matches the # indices written into inv_perm (dispatch meta can diverge after remap). expert_num_tokens count_expert_num_tokens( topk_ids, local_num_experts, expert_map ) ep_scatter( recv_xaq, recv_x_scaleaq_scale, recv_topktopk_ids, num_recv_tokens_per_expertexpert_num_tokens, expert_start_locexpert_start_loc, expert_mapexpert_map, output_tensoraq_out, output_tensor_scaleaq_scale_out, m_indicesexpert_ids, output_indexinv_perm, ) return aq_out, aq_scale_out, expert_ids, inv_perm3.3.1 用之前拷上来的 meta 算对齐长度CPUM_sum compute_aligned_M(..., expert_tokens_metaexpert_tokens_meta)3.3.2 两个 native fill你 profile 里看到的expert_ids torch.full((M_sum,), fill_value-1, ...) inv_perm torch.full(topk_ids.shape, fill_value-1, ...)3.3.3 再数一遍 counts不用刚才 Memcpy 上去的那份做 scatterexpert_num_tokens count_expert_num_tokens(topk_ids, local_num_experts, expert_map)3.3.4 这才 scan scatterep_scatter(..., num_recv_tokens_per_expertexpert_num_tokens, ...)所以 vLLM 的 GPU 时间线是... dispatch ... | Memcpy | ........空白........ | fill | fill | count | scan | scatter | gemm↑ ↑_receiver 里 permute 里才到4 消除空泡方法1通过prof发现在memcpy之后还有很多cpu调用于是要想办法减少这些cpu调用def compute_aligned_M( M: int, num_topk: int, local_num_experts: int, alignment: int, expert_tokens_meta: mk.ExpertTokensMetadata | None, ): # Conservative upper bound on permuted rows (M_sum). Safe even when # dispatch meta under-counts vs post-dispatch topk_ids after DeepEP remap. M_sum_upper (M * num_topk) local_num_experts * (alignment - 1) M_sum_upper round_up(M_sum_upper, alignment) # Fast path: reuse cached sum(list) from make_from_list (no aten round_up), # but still take max with upper bound for safety. if expert_tokens_meta is not None and expert_tokens_meta.m_sum is not None: return max(expert_tokens_meta.m_sum, M_sum_upper) if (expert_tokens_meta is not None) and ( expert_tokens_meta.expert_num_tokens_cpu is not None ): M_sum_meta expert_num_tokens_round_up_and_sum( expert_tokens_meta.expert_num_tokens_cpu, alignmentalignment ) return max(M_sum_meta, M_sum_upper) return M_sum_upper通过分析prof发现其中一个函数被调用了很多次而通过sglang代码以及添加打印发现其实这里不需要这么复杂因为dispatch接口已经传入了256对齐了所以之类计算的时候只需要简单的一个sum函数就可以解决dataclass class ExpertTokensMetadata: Metadata regarding expert-token routing. expert_num_tokens: torch.Tensor expert_num_tokens_cpu: torch.Tensor | None m_sum: int | None None staticmethod def make_from_list( expert_num_tokens_list: list[int], device: str ) - ExpertTokensMetadata: expert_num_tokens_cpu torch.tensor( expert_num_tokens_list, devicecpu, dtypetorch.int32, pin_memoryTrue ) return ExpertTokensMetadata( expert_num_tokensexpert_num_tokens_cpu.to(device, non_blockingTrue), expert_num_tokens_cpuexpert_num_tokens_cpu, m_sumsum(expert_num_tokens_list), )这样修改之后空泡有所减小但还是不够需要继续修改。5 消除空泡方法2那么继续看还有什么那么接下来去看vllm在memcpy之后cpu在干什么那么vllm中间的cpu调用是哪些东西这里把allocate_buffer里面的这个替换了一下6 消除空泡方法3刚才从prof看到这里的import也占用了时间于是这里加个判断只有ep的时候才走下面的代码7 消除空泡方法4刚才有个误区老是看memcpy之后的cpu调用其实应该再往前看看memcpy之前的有哪些调用可以优化发现了一个这个torchwhere在deepep_ht.py文件中这里给他删掉现在新路径不用全局的了不用expertmap了探后topkids里面就是局部的然后scatter也是直接用局部的就是本来吧这个topk_ids在distapch之后收到的里面的是本地局部的专家并且里面是带有负一的然后这个torch.where给他加上了偏置把局部的都给转成了全局的然后scatter里面到时候还要根据expertmap给把这个topk_ids给再转成局部的才做scatter以前的路径多此一举去掉torch.where之后这四个算子都没了8 其他消除空泡方法其实就是和上面一样还是看dispatch和scatter之间有哪些cpu调用消耗了时间然后看看这些cpu调用能不能替换成更省时间的或者直接删掉就这样一步步来。9 总结下面是最终消除空泡前后的对比图

大模型推理引擎vLLM(30): 参考sglang代码，重构vllm021中EP高吞吐代码，消除空泡问题:400us减小到25us

相关推荐

企业大脑到底是什么：别再把它当成一个更大的知识库

语义鸿沟是企业AI最后一道坎，本体语义平台是怎么填平的

学生事务管理系统

更多资讯

为什么升级到 v11 后所有 Dozzle 用户都会被登出一次？

CesiumJS开发中解决WebGL纹理尺寸错误指南

Ryujinx 上手指南：在 PC 上跑起 Switch 游戏只需这几步

Linux服务器突然变慢？从load average到cron定时任务的完整排障复盘

JeecgBoot AI自然语言生成BPMN流程实战指南

VeraCrypt 救援盘完整指南：崩溃前 30 分钟建好你的应急通道

今日更新

基于高阶循环谱的MPSK载波频率估计与MATLAB实现

MiniPdf方：.NET开源Office转PDF的高效解决方案

Solon AI v3.10：Java全栈智能体开发框架解析

本周精选

AI驱动的自动化专利分析平台技术解析

LeetCode-Go 题解：1252. Cells with Odd Values in a Matrix（奇数值单元格计数）双解法剖析

在线答疑系统Java毕设实战：Spring Boot前后端分离与状态流转

本月热门

基于YOLO的焊缝缺陷检测：从数据标注到工业部署全流程实践

分布式服务的决策记录

创业产品的首版范围

读完文章，想聊聊您的网站？