| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """TRT-LLM Ragged backend for MLA prefill.""" |
| 4 | |
| 5 | from typing import TYPE_CHECKING, ClassVar |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | import vllm.envs as envs |
| 10 | from vllm.v1.attention.backends.mla.prefill.base import ( |
| 11 | MLADimensions, |
| 12 | MLAPrefillBackend, |
| 13 | ) |
| 14 | from vllm.v1.attention.backends.utils import log2_lse_to_ln |
| 15 | from vllm.v1.worker.workspace import current_workspace_manager |
| 16 | |
| 17 | if TYPE_CHECKING: |
| 18 | from vllm.config import VllmConfig |
| 19 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 20 | MLACommonPrefillMetadata, |
| 21 | ) |
| 22 | from vllm.platforms.interface import DeviceCapability |
| 23 | |
| 24 | |
| 25 | class TrtllmRaggedPrefillBackend(MLAPrefillBackend): |
| 26 | """TRT-LLM Ragged backend for MLA prefill.""" |
| 27 | |
| 28 | supported_mla_dimensions: ClassVar[list[MLADimensions]] = [ |
| 29 | MLADimensions( |
| 30 | qk_nope_head_dim=128, |
| 31 | qk_rope_head_dim=64, |
| 32 | v_head_dim=128, |
| 33 | ), |
| 34 | MLADimensions( |
| 35 | qk_nope_head_dim=192, |
| 36 | qk_rope_head_dim=64, |
| 37 | v_head_dim=256, |
| 38 | ), |
| 39 | ] |
| 40 | |
| 41 | @staticmethod |
| 42 | def get_name() -> str: |
| 43 | return "TRTLLM_RAGGED" |
| 44 | |
| 45 | @classmethod |
| 46 | def supports_compute_capability(cls, device_capability: "DeviceCapability") -> bool: |
| 47 | return device_capability.major == 10 |
| 48 | |
| 49 | @classmethod |
| 50 | def is_available(cls) -> bool: |
| 51 | try: |
| 52 | from flashinfer.prefill import ( |
| 53 | trtllm_ragged_attention_deepseek, # noqa: F401 |
| 54 | ) |
| 55 | |
| 56 | return True |
| 57 | except ImportError: |
| 58 | return False |
| 59 | |
| 60 | def __init__( |
| 61 | self, |
| 62 | num_heads: int, |
| 63 | scale: float, |
| 64 | kv_lora_rank: int, |
| 65 | qk_nope_head_dim: int, |
| 66 | qk_rope_head_dim: int, |
| 67 | v_head_dim: int, |
| 68 | vllm_config: "VllmConfig", |
| 69 | ) -> None: |
| 70 | super().__init__( |
| 71 | num_heads=num_heads, |
| 72 | scale=scale, |
| 73 | kv_lora_rank=kv_lora_rank, |
| 74 | qk_nope_head_dim=qk_nope_head_dim, |
| 75 | qk_rope_head_dim=qk_rope_head_dim, |
| 76 | v_head_dim=v_head_dim, |
| 77 | vllm_config=vllm_config, |
| 78 | ) |
| 79 | (self._workspace_buffer,) = current_workspace_manager().get_simultaneous( |
| 80 | ( |
| 81 | (envs.VLLM_FLASHINFER_WORKSPACE_BUFFER_SIZE,), |
| 82 | torch.uint8, |
| 83 | ), |
| 84 | ) |
| 85 | |
| 86 | def prepare_metadata( |
| 87 | self, |
| 88 | prefill_metadata: "MLACommonPrefillMetadata", |
| 89 | ) -> None: |
| 90 | super().prepare_metadata(prefill_metadata) |
| 91 | self._query_seq_lens = ( |
| 92 | prefill_metadata.query_start_loc[1:] - prefill_metadata.query_start_loc[:-1] |
| 93 | ) |
| 94 | query_lens_cpu = prefill_metadata.query_lens_cpu |
| 95 | if query_lens_cpu is None: |
| 96 | raise ValueError("TRTLLM ragged prefill requires CPU query lengths") |
| 97 | if query_lens_cpu.device.type != "cpu": |
| 98 | raise ValueError("TRTLLM ragged prefill query lengths must be on CPU") |
| 99 | if query_lens_cpu.ndim != 1 or query_lens_cpu.numel() == 0: |
| 100 | raise ValueError( |
| 101 | "TRTLLM ragged prefill requires a non-empty 1D query-length tensor" |
| 102 | ) |
| 103 | min_query_len = int(torch.min(query_lens_cpu).item()) |
| 104 | if min_query_len < 0: |
| 105 | raise ValueError("TRTLLM ragged prefill query lengths must be non-negative") |
| 106 | self._has_active_rows = min_query_len > 0 |
| 107 | if not self._has_active_rows: |
| 108 | has_mixed_rows = bool(torch.any(query_lens_cpu > 0).item()) |
| 109 | if has_mixed_rows: |
| 110 | raise ValueError( |
| 111 | "TRTLLM ragged prefill contains mixed active and empty query " |
| 112 | f"rows: query_lens={query_lens_cpu.tolist()}" |
| 113 | ) |
| 114 | |
| 115 | def supports_out(self) -> bool: |
| 116 | # Output head dim is v.shape[-1] == v_head_dim, so `out` is unpadded. |
| 117 | return True |
| 118 | |
| 119 | def run_prefill_new_tokens( |
| 120 | self, |
| 121 | q: torch.Tensor, |
| 122 | k: torch.Tensor, |
| 123 | v: torch.Tensor, |
| 124 | return_softmax_lse: bool, |
| 125 | out: torch.Tensor | None = None, |
| 126 | output_scale: torch.Tensor | None = None, |
| 127 | ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: |
| 128 | from flashinfer.prefill import trtllm_ragged_attention_deepseek |
| 129 | |
| 130 | if out is None: |
| 131 | out = torch.empty( |
| 132 | q.shape[0], |
| 133 | q.shape[1], |
| 134 | v.shape[2], |
| 135 | device=q.device, |
| 136 | dtype=self._prefill_metadata.output_dtype, |
| 137 | ) |
| 138 | |
| 139 | if not self._has_active_rows: |
| 140 | out.zero_() |
| 141 | if return_softmax_lse: |
| 142 | lse = torch.empty( |
| 143 | self.num_heads, |
| 144 | q.shape[0], |
| 145 | dtype=torch.float32, |
| 146 | device=q.device, |
| 147 | ) |
| 148 | return out, lse.fill_(-float("inf")) |
| 149 | return out |
| 150 | |
| 151 | ret = trtllm_ragged_attention_deepseek( |
| 152 | query=q, |
| 153 | key=k, |
| 154 | value=v, |
| 155 | workspace_buffer=self._workspace_buffer, |
| 156 | seq_lens=self._query_seq_lens, |
| 157 | max_q_len=self._prefill_metadata.max_query_len, |
| 158 | max_kv_len=self._prefill_metadata.max_query_len, |
| 159 | bmm1_scale=self.scale, |
| 160 | bmm2_scale=1.0, |
| 161 | o_sf_scale=1.0, |
| 162 | batch_size=self._query_seq_lens.shape[0], |
| 163 | window_left=-1, |
| 164 | cum_seq_lens_q=self._prefill_metadata.query_start_loc, |
| 165 | cum_seq_lens_kv=self._prefill_metadata.query_start_loc, |
| 166 | enable_pdl=False, |
| 167 | is_causal=True, |
| 168 | return_lse=return_softmax_lse, |
| 169 | out=out, |
| 170 | skip_all_rows_active_check=True, |
| 171 | ) |
| 172 | |
| 173 | if isinstance(ret, tuple): |
| 174 | # Convert from (q_len, num_heads) to (num_heads, q_len) |
| 175 | return ret[0], log2_lse_to_ln(ret[1].transpose(0, 1)) |
| 176 | return ret |
| 177 | |
| 178 | def run_prefill_context_chunk( |
| 179 | self, |
| 180 | chunk: "MLACommonPrefillMetadata.ContextChunk", |
| 181 | q: torch.Tensor, |
| 182 | k: torch.Tensor, |
| 183 | v: torch.Tensor, |
| 184 | out: torch.Tensor | None = None, |
| 185 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 186 | from flashinfer.prefill import trtllm_ragged_attention_deepseek |
| 187 | |
| 188 | if not chunk.all_rows_active: |
| 189 | raise ValueError( |
| 190 | "TRTLLM ragged context prefill contains an empty query or KV row" |
| 191 | ) |
| 192 | |
| 193 | if out is None: |
| 194 | out = torch.empty( |
| 195 | q.shape[0], |
| 196 | q.shape[1], |
| 197 | v.shape[2], |
| 198 | device=q.device, |
| 199 | dtype=self._prefill_metadata.output_dtype, |
| 200 | ) |
| 201 | |
| 202 | attn_out, lse = trtllm_ragged_attention_deepseek( |
| 203 | query=q, |
| 204 | key=k, |
| 205 | value=v, |
| 206 | workspace_buffer=self._workspace_buffer, |
| 207 | seq_lens=chunk.seq_lens, |
| 208 | max_q_len=chunk.max_query_len, |
| 209 | max_kv_len=chunk.max_seq_len, |
| 210 | bmm1_scale=self.scale, |
| 211 | bmm2_scale=1.0, |
| 212 | o_sf_scale=1.0, |
| 213 | batch_size=chunk.num_requests, |
| 214 | window_left=-1, |
| 215 | cum_seq_lens_q=chunk.query_start_loc, |
| 216 | cum_seq_lens_kv=chunk.cu_seq_lens, |
| 217 | enable_pdl=False, |
| 218 | is_causal=False, |
| 219 | return_lse=True, |
| 220 | out=out, |
| 221 | skip_all_rows_active_check=True, |
| 222 | ) |
| 223 | |
| 224 | # Convert from (q_len, num_heads) to (num_heads, q_len) |
| 225 | return attn_out, log2_lse_to_ln(lse.transpose(0, 1)) |