| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterator |
| 5 | from contextlib import contextmanager |
| 6 | from typing import TYPE_CHECKING, Any |
| 7 | |
| 8 | import torch |
| 9 | |
| 10 | from vllm.config.mamba import MambaBackendEnum |
| 11 | from vllm.logger import init_logger |
| 12 | from vllm.model_executor.layers.mamba.ops.ssu_dispatch import ( |
| 13 | flashinfer_replayssm_autotune_supported, |
| 14 | ) |
| 15 | |
| 16 | if TYPE_CHECKING: |
| 17 | from vllm.v1.worker.gpu_model_runner import GPUModelRunner |
| 18 | |
| 19 | logger = init_logger(__name__) |
| 20 | |
| 21 | |
| 22 | def _replayssm_autotune_kwargs( |
| 23 | runner: "GPUModelRunner", |
| 24 | max_token_prefill_kwargs: dict[str, Any], |
| 25 | ) -> tuple[int, dict[str, Any]] | None: |
| 26 | config = runner.vllm_config |
| 27 | if not ( |
| 28 | config.cache_config.use_replayssm |
| 29 | and config.mamba_config.backend == MambaBackendEnum.FLASHINFER |
| 30 | ): |
| 31 | return None |
| 32 | if not flashinfer_replayssm_autotune_supported(): |
| 33 | logger.info_once( |
| 34 | "Skipping FlashInfer ReplaySSM autotuning because " |
| 35 | "flashinfer.mamba.checkpointing_ssu.CheckpointingSSURunner " |
| 36 | "is unavailable." |
| 37 | ) |
| 38 | return None |
| 39 | v2_runner: Any = runner |
| 40 | query_len = ( |
| 41 | v2_runner.decode_query_len |
| 42 | if config.use_v2_model_runner |
| 43 | else runner.uniform_decode_query_len |
| 44 | ) |
| 45 | max_num_reqs = min( |
| 46 | runner.scheduler_config.max_num_seqs, |
| 47 | runner.max_num_tokens // query_len, |
| 48 | runner.kv_cache_config.num_blocks - 1, |
| 49 | ) |
| 50 | if max_num_reqs <= 0: |
| 51 | logger.warning_once( |
| 52 | "Skipping FlashInfer ReplaySSM autotuning because no non-padding " |
| 53 | "state slot is available." |
| 54 | ) |
| 55 | return None |
| 56 | |
| 57 | decode_kwargs = { |
| 58 | **max_token_prefill_kwargs, |
| 59 | "num_tokens": max_num_reqs * query_len, |
| 60 | "uniform_decode": True, |
| 61 | } |
| 62 | if config.use_v2_model_runner: |
| 63 | decode_kwargs["valid_dummy_state_slots"] = True |
| 64 | else: |
| 65 | decode_kwargs.update( |
| 66 | allow_microbatching=False, |
| 67 | force_attention=True, |
| 68 | profile_seq_lens=query_len + 1, |
| 69 | ) |
| 70 | return max_num_reqs, decode_kwargs |
| 71 | |
| 72 | |
| 73 | @contextmanager |
| 74 | def _temporary_replayssm_autotune_state( |
| 75 | runner: "GPUModelRunner", max_num_reqs: int |
| 76 | ) -> Iterator[None]: |
| 77 | from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2 |
| 78 | from vllm.model_executor.layers.mamba.ops.ssu_dispatch import ( |
| 79 | reset_replayssm_ring_trackers, |
| 80 | update_replayssm_ring_trackers, |
| 81 | ) |
| 82 | |
| 83 | reset_tensors: dict[int, torch.Tensor] = {} |
| 84 | tracker_specs: dict[ |
| 85 | int, tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, int] |
| 86 | ] = {} |
| 87 | for module in runner.get_model().modules(): |
| 88 | if not isinstance(module, MambaMixer2) or not module.use_replayssm: |
| 89 | continue |
| 90 | assert module.replayssm_buffer_len is not None |
| 91 | ring_start = module._replayssm_ring_start |
| 92 | prev_num_accepted = module._replayssm_prev_num_accepted |
| 93 | prev_query_len = module._replayssm_prev_query_len |
| 94 | tracker_specs.setdefault( |
| 95 | ring_start.data_ptr(), |
| 96 | ( |
| 97 | ring_start, |
| 98 | prev_num_accepted, |
| 99 | prev_query_len, |
| 100 | module.replayssm_buffer_len, |
| 101 | module.kv_cache[2].size(2), |
| 102 | ), |
| 103 | ) |
| 104 | tensors = ( |
| 105 | *module.kv_cache, |
| 106 | ring_start, |
| 107 | prev_num_accepted, |
| 108 | prev_query_len, |
| 109 | ) |
| 110 | for tensor in tensors: |
| 111 | if tensor.numel(): |
| 112 | reset_tensors.setdefault(tensor.data_ptr(), tensor) |
| 113 | |
| 114 | v2_runner: Any = runner |
| 115 | block_tables = saved_block_ids = None |
| 116 | if not runner.vllm_config.use_v2_model_runner: |
| 117 | block_tables = runner.input_batch.block_table.block_tables |
| 118 | saved_block_ids = tuple( |
| 119 | block_table.block_table.np[:max_num_reqs, 0].copy() |
| 120 | for block_table in block_tables |
| 121 | ) |
| 122 | dummy_block_ids = range(1, max_num_reqs + 1) |
| 123 | for block_table in block_tables: |
| 124 | block_table.block_table.np[:max_num_reqs, 0] = dummy_block_ids |
| 125 | runner.input_batch.block_table.commit_block_table(max_num_reqs) |
| 126 | |
| 127 | first_tracker = next(iter(tracker_specs.values()), None) |
| 128 | if first_tracker is not None and first_tracker[0].is_cuda: |
| 129 | state_slots = torch.arange( |
| 130 | 1, max_num_reqs + 1, dtype=torch.int32, device=first_tracker[0].device |
| 131 | ) |
| 132 | for ( |
| 133 | ring_start, |
| 134 | prev_num_accepted, |
| 135 | prev_query_len, |
| 136 | logical_window, |
| 137 | ring_buffer_len, |
| 138 | ) in tracker_specs.values(): |
| 139 | # Compile reset (prefill) and advance (decode) before inference. |
| 140 | # The final reset leaves the decode tuning run in a clean state. |
| 141 | reset_replayssm_ring_trackers( |
| 142 | ring_start, prev_num_accepted, prev_query_len, state_slots |
| 143 | ) |
| 144 | update_replayssm_ring_trackers( |
| 145 | ring_start, |
| 146 | prev_num_accepted, |
| 147 | prev_query_len, |
| 148 | state_slots, |
| 149 | logical_window, |
| 150 | ring_buffer_len, |
| 151 | ) |
| 152 | reset_replayssm_ring_trackers( |
| 153 | ring_start, prev_num_accepted, prev_query_len, state_slots |
| 154 | ) |
| 155 | |
| 156 | try: |
| 157 | yield |
| 158 | finally: |
| 159 | if runner.vllm_config.use_v2_model_runner: |
| 160 | v2_runner.block_tables.get_dummy_block_tables(max_num_reqs) |
| 161 | else: |
| 162 | assert block_tables is not None and saved_block_ids is not None |
| 163 | for block_table, block_ids in zip(block_tables, saved_block_ids): |
| 164 | block_table.block_table.np[:max_num_reqs, 0] = block_ids |
| 165 | runner.input_batch.block_table.commit_block_table(max_num_reqs) |
| 166 | for tensor in reset_tensors.values(): |
| 167 | tensor[1 : max_num_reqs + 1].zero_() |
| 168 | |
| 169 | |
| 170 | def replayssm_autotune_warmup(runner: "GPUModelRunner") -> None: |
| 171 | max_token_prefill_kwargs = { |
| 172 | "num_tokens": runner.scheduler_config.max_num_batched_tokens, |
| 173 | "skip_eplb": True, |
| 174 | "is_profile": True, |
| 175 | "randomize_inputs": True, |
| 176 | } |
| 177 | autotune = _replayssm_autotune_kwargs(runner, max_token_prefill_kwargs) |
| 178 | if autotune is None: |
| 179 | return |
| 180 | max_num_reqs, decode_kwargs = autotune |
| 181 | with _temporary_replayssm_autotune_state(runner, max_num_reqs): |
| 182 | runner._dummy_run(**decode_kwargs) |