diff --git a/python/sglang/srt/layers/moe/topk.py b/python/sglang/srt/layers/moe/topk.py index 768e03137df1..67742d208adb 100644 --- a/python/sglang/srt/layers/moe/topk.py +++ b/python/sglang/srt/layers/moe/topk.py @@ -149,6 +149,78 @@ def routing( # an accuracy run before becoming the default. _skip_hip_pad_mask = get_bool_env_var("SGLANG_MORI_NO_PAD_MASK", "False") +# ATOM-style shared-expert fusion for the aiter grouped-topk path: keep a +# persistent topk buffer whose shared-expert columns are pre-populated once (NOT +# related to the LLM prefill phase — this applies to both prefill and decode), and +# let the aiter kernel write only the routed columns (via row stride). This removes +# the per-layer _fused_append_shared_experts kernel. +# Auto-enabled (no env) when: non-EP aiter path (moe_ep_size == 1) and +# num_fused_shared_experts > 0. Bit-identical to the plain append: the shared column +# is filled with the same constant scale_factor the aiter append writes. The +# persistent buffer is sized to the max prefill batch (chunked-prefill-size); for +# token counts above it we fall back to the plain path (condition mirrored in +# _post_process_topk_ids). Mirrors ATOM's init_aiter_topK_meta_data. +# Hard upper bound on the persistent buffer (safety cap when chunked prefill is +# disabled / unexpectedly huge); the buffer only costs ~[MAX, topk+n_shared] * 8B. +_AITER_TOPK_FUSE_SHARED_MAX_TOKENS_CAP = 131072 +_aiter_topk_fuse_shared_max_tokens_cache = None +_aiter_topk_fuse_shared_bufs: dict = {} + + +def _get_aiter_topk_fuse_shared_max_tokens() -> int: + """Max per-forward token count the persistent buffer must cover. Sized to the + largest prefill batch (chunked-prefill-size / max-prefill-tokens); decode is + always tiny (bs * num_tokens_per_bs). Above this we fall back to the plain + path, so this only bounds the fast-path coverage, not correctness. Cached + (server args are fixed after startup).""" + global _aiter_topk_fuse_shared_max_tokens_cache + if _aiter_topk_fuse_shared_max_tokens_cache is None: + from sglang.srt.runtime_context import get_server_args + + try: + sa = get_server_args() + except ValueError: + # Global server args not published yet (e.g. a unit test or offline + # init that reaches the aiter grouped-topk path before startup). + # Degrade gracefully instead of crashing -- this value only bounds + # the fast-path coverage, not correctness (see docstring). Return the + # safety cap WITHOUT caching, so a later call (once args are set) + # still computes and caches the real value. + return _AITER_TOPK_FUSE_SHARED_MAX_TOKENS_CAP + cps = getattr(sa, "chunked_prefill_size", None) or 0 + mpt = getattr(sa, "max_prefill_tokens", None) or 0 + m = max(int(cps), int(mpt), 8192) # 8192 floor for tiny configs + if int(cps) <= 0 and int(mpt) <= 0: + # chunked prefill disabled -> use the safety cap + m = _AITER_TOPK_FUSE_SHARED_MAX_TOKENS_CAP + _aiter_topk_fuse_shared_max_tokens_cache = min( + m, _AITER_TOPK_FUSE_SHARED_MAX_TOKENS_CAP + ) + return _aiter_topk_fuse_shared_max_tokens_cache + + +def _get_aiter_topk_fuse_shared_buf( + topk_routed: int, n_shared: int, num_experts: int, shared_weight: float, device +): + """Persistent [MAX, topk_routed + n_shared] weight/id buffers whose shared + columns are pre-filled once (id = num_experts + i, weight = shared_weight). + Fixed max size (>= max prefill batch) so the tensor address is stable across + CUDA-graph replays.""" + key = (topk_routed, n_shared, num_experts, float(shared_weight), str(device)) + buf = _aiter_topk_fuse_shared_bufs.get(key) + if buf is None: + total = topk_routed + n_shared + M = _get_aiter_topk_fuse_shared_max_tokens() + w = torch.empty((M, total), dtype=torch.float32, device=device) + ids = torch.empty((M, total), dtype=torch.int32, device=device) + ids[:, topk_routed:] = torch.arange( + num_experts, num_experts + n_shared, dtype=torch.int32, device=device + ).unsqueeze(0) + w[:, topk_routed:] = shared_weight + buf = (w, ids) + _aiter_topk_fuse_shared_bufs[key] = buf + return buf + if _is_cuda: try: @@ -1418,6 +1490,7 @@ def biased_grouped_topk_gpu( num_fused_shared_experts: int = 0, routed_scaling_factor: Optional[float] = None, apply_routed_scaling_factor_on_output: Optional[bool] = False, + fused_shared_experts_scaling_factor: Optional[float] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: num_tokens = gating_output.shape[0] num_experts = gating_output.shape[1] @@ -1543,8 +1616,34 @@ def biased_grouped_topk_gpu( assert ( hidden_states.shape[0] == gating_output.shape[0] ), f"Number of tokens mismatch: hidden_states.shape[0] = {hidden_states.shape[0]}, gating_output.shape[0] = {gating_output.shape[0]}" - topk_weights = torch.empty((token, topk), dtype=torch.float32, device=device) - topk_ids = torch.empty((token, topk), dtype=torch.int32, device=device) + _shared_fuse = ( + num_fused_shared_experts > 0 + and get_parallel().moe_ep_size == 1 + and token <= _get_aiter_topk_fuse_shared_max_tokens() + ) + if _shared_fuse: + # Persistent buffer with pre-populated shared columns. The weight is the + # constant scale_factor the aiter _post_process append writes (default + # 1.0), so this is bit-identical for any shared-expert scaling. aiter + # writes only the routed columns [:, :topk] via row stride; the shared + # column stays intact. If token exceeds the buffer size we drop to the + # plain path below (and _post_process appends shared experts as usual) — + # the same condition is mirrored there so the two paths stay consistent. + _shared_w = ( + 1.0 + if fused_shared_experts_scaling_factor is None + else fused_shared_experts_scaling_factor + ) + full_w, full_ids = _get_aiter_topk_fuse_shared_buf( + topk, num_fused_shared_experts, num_experts, _shared_w, device + ) + topk_weights = full_w[:token, :topk] + topk_ids = full_ids[:token, :topk] + else: + topk_weights = torch.empty( + (token, topk), dtype=torch.float32, device=device + ) + topk_ids = torch.empty((token, topk), dtype=torch.int32, device=device) aiter_biased_grouped_topk( gating_output, correction_bias.to(dtype=gating_output.dtype), @@ -1555,6 +1654,10 @@ def biased_grouped_topk_gpu( renormalize, routed_scaling_factor if routed_scaling_factor is not None else 1.0, ) + if _shared_fuse: + # Return the full [token, topk + n_shared] view (routed just written, + # shared pre-populated). _post_process_topk_ids skips its append. + return full_w[:token], full_ids[:token] return topk_weights, topk_ids elif _is_musa and ( gating_output.shape[1] // num_expert_group <= 32 @@ -1966,25 +2069,43 @@ def _post_process_topk_ids( num_local_routed, ) elif _aiter_append: - M, N = router_logits.shape - scale_factor = ( - 1.0 - if fused_shared_experts_scaling_factor is None - else fused_shared_experts_scaling_factor + # Detect whether the shared experts were already fused/appended in + # biased_grouped_topk_gpu via the persistent pre-populated topk buffer. + # When fused, topk_ids already has the full width (routed + shared), i.e. + # topk_ids.shape[1] == topk_config.top_k; when the fast path fell back to + # the plain buffer (e.g. token > MAX during a large prefill), only the + # routed columns are present and we must append the shared experts here. + # Checking the tensor shape is robust to the exact fast-path conditions + # (no fragile mirroring of biased_grouped_topk_gpu's _shared_fuse check). + _shared_fused_in_topk = ( + num_fused_shared_experts > 0 + and get_parallel().moe_ep_size == 1 + and topk_ids.shape[1] == topk_config.top_k ) + if _shared_fused_in_topk: + # Shared experts were already appended in biased_grouped_topk_gpu via + # the persistent pre-populated topk buffer; nothing to do here. + pass + else: + M, N = router_logits.shape + scale_factor = ( + 1.0 + if fused_shared_experts_scaling_factor is None + else fused_shared_experts_scaling_factor + ) - # Lazy import to avoid circular-import issues - from sglang.kernels.ops.moe.fused_moe_triton_kernels import ( - fused_append_shared_experts, - ) + # Lazy import to avoid circular-import issues + from sglang.kernels.ops.moe.fused_moe_triton_kernels import ( + fused_append_shared_experts, + ) - topk_ids, topk_weights = fused_append_shared_experts( - topk_ids, - topk_weights, - num_fused_shared_experts, - scale_factor, - N, # base id for shared experts - ) + topk_ids, topk_weights = fused_append_shared_experts( + topk_ids, + topk_weights, + num_fused_shared_experts, + scale_factor, + N, # base id for shared experts + ) elif use_per_rank_shared_slots: # DeepEP/MegaMOE: remap to per-rank shared-slot layout where each @@ -2079,6 +2200,7 @@ def select_experts( num_fused_shared_experts=num_fused_shared_experts, routed_scaling_factor=routed_scaling_factor, apply_routed_scaling_factor_on_output=apply_routed_scaling_factor_on_output, + fused_shared_experts_scaling_factor=topk_config.fused_shared_experts_scaling_factor, ) elif torch_native and custom_routing_function is None: assert (