|
| 1 | +import torch |
| 2 | +import torch_npu |
| 3 | +from vllm.distributed import get_tp_group |
| 4 | +from vllm_ascend.ascend_forward_context import _EXTRA_CTX |
| 5 | +from vllm_ascend.attention.attention_v1 import AscendAttentionState |
| 6 | +from vllm_ascend.attention.mla_v1 import ( |
| 7 | + MAX_O_PROJ_PREFETCH_SIZE, |
| 8 | + MLAPO_MAX_SUPPORTED_TOKENS, |
| 9 | +) |
| 10 | +from vllm_ascend.attention.utils import ( |
| 11 | + maybe_save_kv_layer_to_connector, |
| 12 | + wait_for_kv_layer_from_connector, |
| 13 | +) |
| 14 | +from vllm_ascend.device.device_op import DeviceOperator |
| 15 | +from vllm_ascend.distributed.utils import all_gather_async |
| 16 | +from vllm_ascend.ops.layer_shard_linear import ( |
| 17 | + is_hidden_layer, |
| 18 | + reach_layer_for_shard_weight_series, |
| 19 | +) |
| 20 | +from vllm_ascend.utils import get_weight_prefetch_method |
| 21 | + |
| 22 | + |
| 23 | +class AscendSFAImpl: |
| 24 | + def forward( |
| 25 | + self, |
| 26 | + layer_name, |
| 27 | + hidden_states: torch.Tensor, # query in unified attn |
| 28 | + kv_cache: tuple[torch.Tensor, torch.Tensor, torch.Tensor], |
| 29 | + attn_metadata, |
| 30 | + need_gather_q_kv: bool = False, |
| 31 | + output: torch.Tensor | None = None, |
| 32 | + ) -> torch.Tensor: |
| 33 | + assert output is not None, "Output tensor must be provided." |
| 34 | + if attn_metadata is None: |
| 35 | + # Profiling run. |
| 36 | + if self.enable_dsa_cp_with_layer_shard and not _EXTRA_CTX.in_profile_run: |
| 37 | + for layer in self.layer_sharding_kwargs or []: |
| 38 | + if is_hidden_layer(layer): |
| 39 | + reach_layer_for_shard_weight_series(layer) |
| 40 | + return output.fill_(0) |
| 41 | + |
| 42 | + cos = attn_metadata.cos |
| 43 | + sin = attn_metadata.sin |
| 44 | + slot_mapping = attn_metadata.slot_mapping |
| 45 | + slot_mapping_cp = None |
| 46 | + if self.enable_dsa_cp: |
| 47 | + assert attn_metadata.dsa_cp_context is not None |
| 48 | + slot_mapping_cp = attn_metadata.dsa_cp_context.slot_mapping_cp |
| 49 | + actual_seq_lengths_query = ( |
| 50 | + attn_metadata.dsa_cp_context.actual_seq_lengths_query |
| 51 | + ) |
| 52 | + actual_seq_lengths_key = attn_metadata.dsa_cp_context.actual_seq_lengths_key |
| 53 | + else: |
| 54 | + actual_seq_lengths_query = attn_metadata.cum_query_lens |
| 55 | + actual_seq_lengths_key = attn_metadata.seq_lens |
| 56 | + |
| 57 | + # Inputs and outputs may be padded for CUDA graphs |
| 58 | + num_input_tokens = attn_metadata.num_input_tokens |
| 59 | + output_padded = output |
| 60 | + |
| 61 | + # all-gather o_proj weight for prefill stage of PD mix node |
| 62 | + o_proj_full_handle = None |
| 63 | + # if is PD mix stage, using original TP o_proj weight, and also need to full gather for o_proj |
| 64 | + # weight for prefill stage. |
| 65 | + full_gather_o_proj_enabled = ( |
| 66 | + self.enable_dsa_cp_with_o_proj_tp |
| 67 | + and attn_metadata.attn_state |
| 68 | + not in { |
| 69 | + AscendAttentionState.DecodeOnly, |
| 70 | + AscendAttentionState.SpecDecoding, |
| 71 | + } |
| 72 | + ) |
| 73 | + |
| 74 | + # run mlapo ops when dsa-cp is disabled, and ensure that num_tokens satisfies the count limitation |
| 75 | + if self.enable_mlapo and num_input_tokens <= MLAPO_MAX_SUPPORTED_TOKENS: |
| 76 | + hidden_states, ql_nope, q_pe, q_c = self._sfa_preprocess_with_mlapo( |
| 77 | + hidden_states=hidden_states, |
| 78 | + kv_cache=kv_cache, |
| 79 | + cos=cos, |
| 80 | + sin=sin, |
| 81 | + slot_mapping=slot_mapping, |
| 82 | + num_input_tokens=num_input_tokens, |
| 83 | + ) |
| 84 | + k_li, k_li_scale = self.indexer_select_pre_process( |
| 85 | + x=hidden_states, cos=cos, sin=sin |
| 86 | + ) |
| 87 | + # [patch] Add 'wait_for_kv_layer_from_connector' call for mlapo path |
| 88 | + wait_for_kv_layer_from_connector(layer_name) |
| 89 | + # native |
| 90 | + else: |
| 91 | + assert self.fused_qkv_a_proj is not None, "q lora is required for DSA." |
| 92 | + weight_prefetch_method = get_weight_prefetch_method() |
| 93 | + weight_prefetch_method.maybe_prefetch_mla_or_sla_weight_in_current_stream( |
| 94 | + inputs=self.fused_qkv_a_proj.weight, dependency=hidden_states |
| 95 | + ) |
| 96 | + qkv_lora = self.fused_qkv_a_proj(hidden_states)[0] |
| 97 | + q_c, kv_no_split = qkv_lora.split( |
| 98 | + [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], |
| 99 | + dim=-1, |
| 100 | + ) |
| 101 | + assert self.q_a_layernorm is not None, "q_a_layernorm must be initialized" |
| 102 | + q_c = self.q_a_layernorm(q_c) |
| 103 | + |
| 104 | + k_li, k_li_scale = self.indexer_select_pre_process( |
| 105 | + x=hidden_states, cos=cos, sin=sin |
| 106 | + ) |
| 107 | + |
| 108 | + wait_for_kv_layer_from_connector(layer_name) |
| 109 | + |
| 110 | + if self.enable_dsa_cp: |
| 111 | + assert slot_mapping_cp is not None |
| 112 | + k_pe, k_nope = self.exec_kv( |
| 113 | + kv_no_split, cos, sin, kv_cache, slot_mapping_cp, attn_metadata |
| 114 | + ) |
| 115 | + else: |
| 116 | + k_pe, k_nope = self.exec_kv( |
| 117 | + kv_no_split, cos, sin, kv_cache, slot_mapping, attn_metadata |
| 118 | + ) |
| 119 | + |
| 120 | + if self.enable_dsa_cp: |
| 121 | + assert k_pe is not None |
| 122 | + assert k_nope is not None |
| 123 | + assert k_li is not None |
| 124 | + async_op = ( |
| 125 | + self.enable_dsa_cp_with_layer_shard or full_gather_o_proj_enabled |
| 126 | + ) |
| 127 | + # support all_gather kv async for communication calculation overlap |
| 128 | + if not self.use_sparse_c8_indexer: |
| 129 | + fused_kv_no_split, kv_ag_handle = all_gather_async( |
| 130 | + torch.cat( |
| 131 | + [ |
| 132 | + k_pe.view(-1, k_pe.shape[-1]), |
| 133 | + k_nope.view(-1, k_nope.shape[-1]), |
| 134 | + k_li.view(-1, k_li.shape[-1]), |
| 135 | + ], |
| 136 | + dim=1, |
| 137 | + ), |
| 138 | + get_tp_group(), |
| 139 | + async_op=async_op, |
| 140 | + ) |
| 141 | + else: |
| 142 | + # due to different dtypes, we have to split commu pass |
| 143 | + assert k_li_scale is not None |
| 144 | + fused_kv_no_split, _ = all_gather_async( |
| 145 | + torch.cat( |
| 146 | + [ |
| 147 | + k_pe.view(-1, k_pe.shape[-1]), |
| 148 | + k_nope.view(-1, k_nope.shape[-1]), |
| 149 | + ], |
| 150 | + dim=1, |
| 151 | + ), |
| 152 | + get_tp_group(), |
| 153 | + async_op=async_op, |
| 154 | + ) |
| 155 | + k_li, _ = all_gather_async( |
| 156 | + k_li, |
| 157 | + get_tp_group(), |
| 158 | + async_op=async_op, |
| 159 | + ) |
| 160 | + k_li_scale, kv_ag_handle = all_gather_async( |
| 161 | + k_li_scale, |
| 162 | + get_tp_group(), |
| 163 | + async_op=async_op, |
| 164 | + ) |
| 165 | + |
| 166 | + ql_nope, q_pe = self._q_proj_and_k_up_proj(q_c) |
| 167 | + q_pe = self.rope_single(q_pe, cos, sin) |
| 168 | + |
| 169 | + if self.enable_dsa_cp: |
| 170 | + if kv_ag_handle is not None: |
| 171 | + kv_ag_handle.wait() |
| 172 | + |
| 173 | + if self.enable_dsa_cp_with_layer_shard: |
| 174 | + for layer in self.layer_sharding_kwargs or []: |
| 175 | + if is_hidden_layer(layer): |
| 176 | + reach_layer_for_shard_weight_series(layer) |
| 177 | + elif full_gather_o_proj_enabled: |
| 178 | + _, o_proj_full_handle = all_gather_async( |
| 179 | + self.o_proj_tp_weight, |
| 180 | + get_tp_group(), |
| 181 | + output=type(self).o_proj_full_pool, |
| 182 | + ) |
| 183 | + |
| 184 | + if kv_cache is not None: |
| 185 | + assert fused_kv_no_split is not None |
| 186 | + if not self.use_sparse_c8_indexer: |
| 187 | + k_pe, k_nope, k_li = fused_kv_no_split.split( |
| 188 | + [self.qk_rope_head_dim, self.kv_lora_rank, self.head_dim], |
| 189 | + dim=-1, |
| 190 | + ) |
| 191 | + else: |
| 192 | + k_pe, k_nope = fused_kv_no_split.split( |
| 193 | + [self.qk_rope_head_dim, self.kv_lora_rank], dim=-1 |
| 194 | + ) |
| 195 | + k_nope = k_nope.view(k_nope.shape[0], 1, -1) |
| 196 | + k_pe = k_pe.view(k_pe.shape[0], 1, -1) |
| 197 | + DeviceOperator.reshape_and_cache( |
| 198 | + key=k_nope[: attn_metadata.num_actual_tokens], |
| 199 | + value=k_pe[: attn_metadata.num_actual_tokens], |
| 200 | + key_cache=kv_cache[0], |
| 201 | + value_cache=kv_cache[1], |
| 202 | + slot_mapping=slot_mapping[: attn_metadata.num_actual_tokens], |
| 203 | + ) |
| 204 | + |
| 205 | + k_li = self._get_full_kv(k_li, attn_metadata) |
| 206 | + |
| 207 | + if kv_cache is not None: |
| 208 | + if self.is_kv_producer: |
| 209 | + attn_metadata.reshape_cache_event = torch.npu.Event() |
| 210 | + torch_npu.npu_scatter_nd_update_( |
| 211 | + kv_cache[2].view(-1, k_li.shape[-1]), |
| 212 | + slot_mapping.view(-1, 1), |
| 213 | + k_li.view(-1, k_li.shape[-1]), |
| 214 | + ) # b, s, n, d |
| 215 | + if self.use_sparse_c8_indexer: |
| 216 | + assert len(kv_cache) == 4 |
| 217 | + assert k_li_scale is not None |
| 218 | + torch_npu.npu_scatter_nd_update_( |
| 219 | + kv_cache[3].view(-1, k_li_scale.shape[-1]), |
| 220 | + slot_mapping.view(-1, 1), |
| 221 | + k_li_scale.view(-1, k_li_scale.shape[-1]), |
| 222 | + ) |
| 223 | + if self.is_kv_producer: |
| 224 | + attn_metadata.reshape_cache_event.record() |
| 225 | + |
| 226 | + topk_indices = self.indexer_select_post_process( |
| 227 | + x=hidden_states, |
| 228 | + q_c=q_c, |
| 229 | + kv_cache=kv_cache, |
| 230 | + attn_metadata=attn_metadata, |
| 231 | + cos=cos, |
| 232 | + sin=sin, |
| 233 | + actual_seq_lengths_query=actual_seq_lengths_query, |
| 234 | + actual_seq_lengths_key=actual_seq_lengths_key, |
| 235 | + ) |
| 236 | + |
| 237 | + attn_output = self._execute_sparse_flash_attention_process( |
| 238 | + ql_nope, |
| 239 | + q_pe, |
| 240 | + kv_cache, |
| 241 | + topk_indices, |
| 242 | + attn_metadata, |
| 243 | + actual_seq_lengths_query, |
| 244 | + actual_seq_lengths_key, |
| 245 | + ) |
| 246 | + |
| 247 | + attn_output = self._v_up_proj(attn_output) |
| 248 | + weight_prefetch_method = get_weight_prefetch_method() |
| 249 | + weight_prefetch_method.maybe_prefetch_mla_or_sla_weight_in_current_stream( |
| 250 | + inputs=self.o_proj.weight, |
| 251 | + dependency=attn_output, |
| 252 | + max_size=MAX_O_PROJ_PREFETCH_SIZE, |
| 253 | + linear_layer=self.o_proj, |
| 254 | + ) |
| 255 | + |
| 256 | + if self.enable_dsa_cp_with_o_proj_tp: |
| 257 | + # When using SFA-CP with pd mixed, o_proj has two cases: |
| 258 | + # 1. prefill: o_proj is a TP weight, we need to all-gather o_proj weight to switch TP=1. |
| 259 | + # 2. decode: all-to-all the hidden_state before the o_proj forward. |
| 260 | + result, require_o_proj_forward = ( |
| 261 | + self._handle_o_proj_weight_switch_and_forward( |
| 262 | + attn_output=attn_output, |
| 263 | + output=output, |
| 264 | + o_proj_full_handle=o_proj_full_handle, |
| 265 | + should_shard_weight=full_gather_o_proj_enabled, |
| 266 | + ) |
| 267 | + ) |
| 268 | + if not require_o_proj_forward: |
| 269 | + return result |
| 270 | + attn_output = result |
| 271 | + |
| 272 | + output[...] = self.o_proj(attn_output)[0] |
| 273 | + |
| 274 | + maybe_save_kv_layer_to_connector(layer_name, list(kv_cache)) |
| 275 | + |
| 276 | + return output_padded |
0 commit comments