| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Adapted from |
| 4 | # https://github.com/sgl-project/sglang/blob/bed301a5acaa9577c9aa706468bdf242f6a43051/python/sglang/srt/layers/moe/routed_experts_capturer.py |
| 5 | |
| 6 | from __future__ import annotations |
| 7 | |
| 8 | import logging |
| 9 | from collections.abc import Callable |
| 10 | from functools import partial |
| 11 | from typing import Protocol, runtime_checkable |
| 12 | |
| 13 | import torch |
| 14 | |
| 15 | from vllm.config import VllmConfig |
| 16 | from vllm.distributed.parallel_state import get_tp_group |
| 17 | from vllm.forward_context import get_forward_context |
| 18 | from vllm.platforms import current_platform |
| 19 | |
| 20 | logger = logging.getLogger(__name__) |
| 21 | |
| 22 | |
| 23 | @runtime_checkable |
| 24 | class RoutedExpertsCaptureSource(Protocol): |
| 25 | layer_id: int |
| 26 | capture_fn: Callable[[torch.Tensor], None] | None |
| 27 | |
| 28 | |
| 29 | def _get_routed_experts_shape(vllm_config: VllmConfig) -> tuple[int, int, int]: |
| 30 | model_config = vllm_config.model_config |
| 31 | num_layers = model_config.get_total_num_hidden_layers() |
| 32 | num_experts = model_config.get_num_experts() |
| 33 | num_experts_per_tok = model_config.get_num_experts_per_tok() |
| 34 | if num_layers <= 0 or num_experts <= 0 or num_experts_per_tok <= 0: |
| 35 | raise ValueError( |
| 36 | "Routed-experts capture requires positive layer, expert, and " |
| 37 | "experts-per-token counts, got " |
| 38 | f"{num_layers=}, {num_experts=}, {num_experts_per_tok=}." |
| 39 | ) |
| 40 | return num_layers, num_experts, num_experts_per_tok |
| 41 | |
| 42 | |
| 43 | class RoutedExpertsCapturer: |
| 44 | """Worker-side capturer for routed experts, lives on GPU. |
| 45 | |
| 46 | Layer-level hooks call :meth:`capture` inside the forward pass. Routing |
| 47 | rows owned by this DP rank are written into a preallocated device buffer. |
| 48 | |
| 49 | The device buffer uses ``int32``. Stable snapshots use the narrowest dtype |
| 50 | that can represent every logical expert ID. |
| 51 | |
| 52 | Invariants: |
| 53 | - One instance per worker; shape is fixed at init and covers the |
| 54 | worst-case step (``max_num_batched_tokens`` tokens). |
| 55 | - Every routed layer overwrites the current step's token rows. |
| 56 | """ |
| 57 | |
| 58 | def __init__( |
| 59 | self, |
| 60 | max_num_batched_tokens: int, |
| 61 | vllm_config: VllmConfig, |
| 62 | ) -> None: |
| 63 | num_layers, num_experts, num_experts_per_tok = _get_routed_experts_shape( |
| 64 | vllm_config |
| 65 | ) |
| 66 | self.shape_per_token = (num_layers, num_experts_per_tok) |
| 67 | self.output_dtype_name = "uint8" if num_experts <= 256 else "uint16" |
| 68 | self.output_dtype = getattr(torch, self.output_dtype_name) |
| 69 | logger.info( |
| 70 | "RoutedExpertsCapturer: allocating buffer with " |
| 71 | "max_tokens=%d, num_layers=%d, num_experts_per_tok=%d " |
| 72 | "(hf_config.model_type=%s)", |
| 73 | max_num_batched_tokens, |
| 74 | num_layers, |
| 75 | num_experts_per_tok, |
| 76 | vllm_config.model_config.hf_text_config.model_type, |
| 77 | ) |
| 78 | self.device_buffer = torch.zeros( |
| 79 | (max_num_batched_tokens, *self.shape_per_token), |
| 80 | dtype=torch.int32, |
| 81 | device=current_platform.device_type, |
| 82 | ) |
| 83 | self.dp_rank = vllm_config.parallel_config.data_parallel_rank |
| 84 | self.tp_size = vllm_config.parallel_config.tensor_parallel_size |
| 85 | |
| 86 | def capture(self, layer_id: int, topk_ids: torch.Tensor) -> None: |
| 87 | """Capture expert routing decisions for a specific layer. |
| 88 | |
| 89 | Under data parallelism, ``topk_ids`` may have four different batch |
| 90 | layouts depending on where the DP combine happens and whether |
| 91 | Expert Parallelism (EP) or Sequence Parallelism (SP) is active for the |
| 92 | MoE layer: |
| 93 | - ``n == total`` (naive dispatch): all DP ranks' tokens are |
| 94 | concatenated before routing; we slice out this rank's span |
| 95 | using the cumulative per-rank counts. |
| 96 | - ``n == token_num_per_dp`` (modular-kernel path): DP combine |
| 97 | happens inside ``quant_method.apply``; ``select_experts`` only |
| 98 | ever sees this rank's tokens, so we take the whole tensor. |
| 99 | - ``n == sum(dp_metadata.local_sizes)`` (naive DP+EP dispatch): |
| 100 | sequence-parallel shards from every DP rank are gathered through |
| 101 | the flattened EP group. The shard sizes include CUDA-graph / SP |
| 102 | padding, so we use them to locate this DP rank's unpadded rows. |
| 103 | - ``n == ceil(token_num_per_dp / tp_size)`` (SP + modular-kernel |
| 104 | path): tokens were split along dim=0 across the TP group by |
| 105 | ``_sequence_parallel_context`` |
| 106 | (``moe_runner_base.py:_sequence_parallel_context``), so each |
| 107 | TP rank only sees its shard. We all-gather along dim=0 to |
| 108 | reconstruct this DP rank's full routing tensor. SP pads with |
| 109 | ceil-div (see ``_compute_sp_num_tokens`` in |
| 110 | ``forward_context.py``), so the gathered tensor may contain a |
| 111 | few trailing padding rows which are trimmed by the downstream |
| 112 | ``[:token_num_per_dp]`` slice. |
| 113 | |
| 114 | Args: |
| 115 | layer_id: The layer index. |
| 116 | topk_ids: Tensor of shape (batch_size, num_routed_experts). |
| 117 | |
| 118 | """ |
| 119 | ctx = get_forward_context() |
| 120 | if ctx.dp_metadata is None: |
| 121 | local_topk_ids = topk_ids |
| 122 | else: |
| 123 | num_tokens_dp = ctx.dp_metadata.num_tokens_across_dp_cpu |
| 124 | num_local_tokens = int(num_tokens_dp[self.dp_rank].item()) |
| 125 | total = int(num_tokens_dp.sum().item()) |
| 126 | n = topk_ids.shape[0] |
| 127 | shard_sizes = getattr(ctx.dp_metadata, "local_sizes", None) |
| 128 | gathered_size = sum(shard_sizes) if shard_sizes is not None else -1 |
| 129 | |
| 130 | if n == total: |
| 131 | end = int(num_tokens_dp[: self.dp_rank + 1].sum().item()) |
| 132 | local_topk_ids = topk_ids[end - num_local_tokens : end] |
| 133 | elif n == num_local_tokens: |
| 134 | local_topk_ids = topk_ids |
| 135 | elif shard_sizes is not None and n == gathered_size: |
| 136 | # Naive DP+EP dispatch gathers the sequence-parallel shards in |
| 137 | # flattened DP-then-TP order. ``local_sizes`` is the exact |
| 138 | # all-gatherv layout, including padding. Select this DP rank's |
| 139 | # contiguous shard group, then trim its trailing padding. |
| 140 | num_dp_ranks = len(num_tokens_dp) |
| 141 | assert len(shard_sizes) % num_dp_ranks == 0 |
| 142 | shards_per_dp_rank = len(shard_sizes) // num_dp_ranks |
| 143 | first_shard = self.dp_rank * shards_per_dp_rank |
| 144 | start = sum(shard_sizes[:first_shard]) |
| 145 | local_topk_ids = topk_ids[start : start + num_local_tokens] |
| 146 | elif ( |
| 147 | self.tp_size > 1 |
| 148 | and n == (num_local_tokens + self.tp_size - 1) // self.tp_size |
| 149 | ): |
| 150 | # SP + modular-kernel path. All-gather across the TP |
| 151 | # group along dim=0 to reconstruct the full per-DP-rank |
| 152 | # tensor; keep only the first ``token_num_per_dp`` rows |
| 153 | # (trailing rows are SP ceil-div padding). The TP group |
| 154 | # is always initialized on real rollout workers, and |
| 155 | # every rank in the group reaches this branch in |
| 156 | # lockstep (bind is per-FusedMoEFactory layer, SP is a global |
| 157 | # condition), so a bare all_gather here will not |
| 158 | # deadlock -- let it raise if the precondition is |
| 159 | # violated rather than skip silently. |
| 160 | # |
| 161 | # ``topk_ids`` is already whatever the router produced |
| 162 | # (typically int32/int64, both supported by NCCL); the |
| 163 | # downstream buffer assignment narrows to the capture dtype. |
| 164 | local_topk_ids = get_tp_group().all_gather(topk_ids, dim=0)[ |
| 165 | :num_local_tokens |
| 166 | ] |
| 167 | else: |
| 168 | sp_expected = (num_local_tokens + self.tp_size - 1) // self.tp_size |
| 169 | raise AssertionError( |
| 170 | "RoutedExpertsCapturer: unexpected topk_ids batch " |
| 171 | f"dim {n} (expected {total}, {num_local_tokens}, " |
| 172 | f"{gathered_size}, or {sp_expected} for " |
| 173 | f"dp_rank={self.dp_rank}, tp_size={self.tp_size})" |
| 174 | ) |
| 175 | |
| 176 | if not 0 <= layer_id < self.device_buffer.shape[1]: |
| 177 | raise IndexError( |
| 178 | f"routed-experts layer {layer_id} exceeds capture buffer " |
| 179 | f"layer count {self.device_buffer.shape[1]}" |
| 180 | ) |
| 181 | |
| 182 | self.device_buffer[: len(local_topk_ids), layer_id] = local_topk_ids |
| 183 | |
| 184 | def snapshot_routing_data(self, num_tokens: int) -> torch.Tensor: |
| 185 | """Return a stable snapshot of the current routing data.""" |
| 186 | return self.device_buffer[:num_tokens].to(self.output_dtype) |
| 187 | |
| 188 | |
| 189 | def bind_routed_experts_capturer( |
| 190 | model: torch.nn.Module, |
| 191 | capturer: RoutedExpertsCapturer, |
| 192 | ) -> None: |
| 193 | """Attach capture callbacks to the target model's MoE routers.""" |
| 194 | from vllm.model_executor.layers.fused_moe.layer import MoERunner |
| 195 | from vllm.model_executor.layers.fused_moe.modular_kernel import ( |
| 196 | FusedMoEExpertsMonolithic, |
| 197 | ) |
| 198 | from vllm.model_executor.layers.fused_moe.router.base_router import BaseRouter |
| 199 | |
| 200 | num_bound = 0 |
| 201 | for module in model.modules(): |
| 202 | if isinstance(module, RoutedExpertsCaptureSource): |
| 203 | module.capture_fn = partial(capturer.capture, module.layer_id) |
| 204 | num_bound += 1 |
| 205 | continue |
| 206 | if not isinstance(module, MoERunner): |
| 207 | continue |
| 208 | capture_fn = partial(capturer.capture, module.layer_id) |
| 209 | quant_method = module._quant_method |
| 210 | if quant_method.is_monolithic: |
| 211 | moe_kernel = getattr(quant_method, "moe_kernel", None) |
| 212 | impl = getattr(moe_kernel, "impl", None) |
| 213 | fused_experts = getattr(impl, "fused_experts", None) |
| 214 | if not ( |
| 215 | isinstance(fused_experts, FusedMoEExpertsMonolithic) |
| 216 | and fused_experts.supports_routing_replay_capture() |
| 217 | ): |
| 218 | raise ValueError( |
| 219 | "Routed-experts capture is not supported with monolithic " |
| 220 | f"MoE kernel {type(fused_experts).__name__}." |
| 221 | ) |
| 222 | fused_experts.set_capture_fn(capture_fn) |
| 223 | num_bound += 1 |
| 224 | elif isinstance(module.router, BaseRouter): |
| 225 | module.router.set_capture_fn(capture_fn) |
| 226 | num_bound += 1 |
| 227 | else: |
| 228 | raise ValueError( |
| 229 | "Routed-experts capture is not supported with router " |
| 230 | f"{type(module.router).__name__}." |
| 231 | ) |
| 232 | |
| 233 | if num_bound == 0: |
| 234 | raise ValueError("No supported MoE router found for routed-experts capture.") |