| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | from typing import cast |
| 6 | |
| 7 | import torch |
| 8 | from torch import nn |
| 9 | |
| 10 | from vllm.compilation.breakable_cudagraph import eager_break_during_capture |
| 11 | from vllm.config import VllmConfig, get_current_vllm_config |
| 12 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 13 | from vllm.forward_context import get_forward_context |
| 14 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 15 | from vllm.model_executor.layers.linear import ( |
| 16 | MergedColumnParallelLinear, |
| 17 | RowParallelLinear, |
| 18 | ) |
| 19 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 20 | from vllm.utils.torch_utils import ( |
| 21 | canonicalize_singleton_dim_strides, |
| 22 | kv_cache_dtype_str_to_dtype, |
| 23 | ) |
| 24 | from vllm.v1.attention.backend import AttentionBackend |
| 25 | from vllm.v1.attention.backends.flash_attn import ( |
| 26 | FlashAttentionBackend, |
| 27 | FlashAttentionMetadata, |
| 28 | ) |
| 29 | from vllm.v1.kv_cache_interface import ( |
| 30 | FullAttentionSpec, |
| 31 | KVCacheSpec, |
| 32 | SlidingWindowSpec, |
| 33 | ) |
| 34 | |
| 35 | from ..configs import InklingModelConfig |
| 36 | from .layernorm import InklingRMSNorm |
| 37 | from .ops.fa4_rel_attention import ( |
| 38 | _INKLING_FA4_REL_ATTENTION_KERNEL, |
| 39 | bucket_max_seqlen_q, |
| 40 | inkling_fa4_num_splits, |
| 41 | ) |
| 42 | from .ops.qkvr_prep import fused_qkvr_prep |
| 43 | from .sconv_swa_attn import _K, _V, InklingConvState, InklingSconvMetadata |
| 44 | from .short_conv import InklingShortConv |
| 45 | |
| 46 | |
| 47 | def compute_log_scaling_tau( |
| 48 | positions: torch.Tensor, n_floor: int, alpha: float |
| 49 | ) -> torch.Tensor: |
| 50 | effective_n = (positions + 1).to(torch.float32) |
| 51 | return 1.0 + alpha * torch.log(torch.clamp(effective_n / float(n_floor), min=1.0)) |
| 52 | |
| 53 | |
| 54 | class RelLogitsProj(nn.Module): |
| 55 | """Project the per-head relative branch ``r`` to per-distance logits.""" |
| 56 | |
| 57 | def __init__(self, d_rel: int, rel_extent: int) -> None: |
| 58 | super().__init__() |
| 59 | self.d_rel = d_rel |
| 60 | self.rel_extent = rel_extent |
| 61 | self.proj = nn.Parameter(torch.empty(d_rel, rel_extent), requires_grad=False) |
| 62 | |
| 63 | def forward(self, r_out: torch.Tensor) -> torch.Tensor: |
| 64 | # r_out: (T, num_heads, d_rel) -> (T, num_heads, rel_extent) |
| 65 | return torch.einsum("thd,de->the", r_out, self.proj) |
| 66 | |
| 67 | |
| 68 | class InklingAttention(nn.Module, AttentionLayerBase): |
| 69 | def __init__( |
| 70 | self, |
| 71 | config: InklingModelConfig, |
| 72 | *, |
| 73 | num_heads: int, |
| 74 | num_kv_heads: int, |
| 75 | head_dim: int, |
| 76 | rel_extent: int, |
| 77 | local_extent: int, |
| 78 | is_local: bool, |
| 79 | prefix: str, |
| 80 | quant_config: QuantizationConfig | None = None, |
| 81 | conv_owner: InklingConvState, |
| 82 | ) -> None: |
| 83 | super().__init__() |
| 84 | self.prefix = prefix |
| 85 | self.is_local = is_local |
| 86 | self.hidden_size = config.hidden_size |
| 87 | self.head_dim = head_dim |
| 88 | self.d_rel = config.d_rel |
| 89 | self.log_scaling_n_floor = config.log_scaling_n_floor |
| 90 | self.log_scaling_alpha = config.log_scaling_alpha |
| 91 | # q/k are per-head RMS-normed (unit norm), so Inkling scales by 1/head_dim. |
| 92 | self.scaling = 1.0 / head_dim |
| 93 | |
| 94 | tp_size = get_tensor_model_parallel_world_size() |
| 95 | self.num_total_heads = num_heads |
| 96 | self.num_total_kv_heads = num_kv_heads |
| 97 | assert self.num_total_heads % tp_size == 0 |
| 98 | self.num_heads = self.num_total_heads // tp_size |
| 99 | if self.num_total_kv_heads >= tp_size: |
| 100 | assert self.num_total_kv_heads % tp_size == 0 |
| 101 | else: |
| 102 | assert tp_size % self.num_total_kv_heads == 0 |
| 103 | self.num_kv_heads = max(1, self.num_total_kv_heads // tp_size) |
| 104 | # When tp_size > num_kv_heads the K/V projections are padded up to |
| 105 | # tp_size heads so each rank gets at least one (GQA replication). |
| 106 | kv_total_for_sizing = max(self.num_total_kv_heads, tp_size) |
| 107 | |
| 108 | self.qkvr = MergedColumnParallelLinear( |
| 109 | input_size=config.hidden_size, |
| 110 | output_sizes=[ |
| 111 | head_dim * self.num_total_heads, |
| 112 | head_dim * kv_total_for_sizing, |
| 113 | head_dim * kv_total_for_sizing, |
| 114 | self.d_rel * self.num_total_heads, |
| 115 | ], |
| 116 | bias=config.q_bias, |
| 117 | quant_config=quant_config, |
| 118 | prefix=f"{prefix}.qkvr", |
| 119 | ) |
| 120 | self.wo_ud = RowParallelLinear( |
| 121 | input_size=head_dim * self.num_total_heads, |
| 122 | output_size=config.hidden_size, |
| 123 | bias=config.o_bias, |
| 124 | quant_config=quant_config, |
| 125 | # reduce_results=False: the partial output is all-reduced below |
| 126 | # (one-shot custom AR) so the attention-output sconv can run on the |
| 127 | # full hidden width fused with the residual add + rmsnorm. |
| 128 | reduce_results=False, |
| 129 | prefix=f"{prefix}.wo_ud", |
| 130 | ) |
| 131 | self.rel_extent = local_extent if is_local else rel_extent |
| 132 | self.local_extent = local_extent if is_local else None |
| 133 | self.rel_logits_proj = RelLogitsProj(self.d_rel, self.rel_extent) |
| 134 | self.q_norm = InklingRMSNorm(head_dim, eps=config.rms_norm_eps) |
| 135 | self.k_norm = InklingRMSNorm(head_dim, eps=config.rms_norm_eps) |
| 136 | |
| 137 | # Short convolution on the K/V streams (per-head-width, TP sharded), |
| 138 | # applied after the qkvr projection and before q/k norm. |
| 139 | kv_conv_dim = self.num_kv_heads * head_dim |
| 140 | self.conv_owner = conv_owner |
| 141 | self.k_sconv = InklingShortConv( |
| 142 | kv_conv_dim, config.sconv_kernel_size, owner=conv_owner, stream_idx=_K |
| 143 | ) |
| 144 | self.v_sconv = InklingShortConv( |
| 145 | kv_conv_dim, config.sconv_kernel_size, owner=conv_owner, stream_idx=_V |
| 146 | ) |
| 147 | |
| 148 | # FA4 left/right window; right=0 keeps it causal. local_extent-1 mirrors |
| 149 | # the source (sliding_window_size - 1). |
| 150 | self.window_size: tuple[int, int] = ( |
| 151 | (local_extent - 1, 0) if is_local else (-1, -1) |
| 152 | ) |
| 153 | # Static per-layer-type KV length bound for the split heuristic: local |
| 154 | # layers never see more than the sliding window. |
| 155 | vllm_config = get_current_vllm_config() |
| 156 | self._max_kv_len = ( |
| 157 | local_extent if is_local else vllm_config.model_config.max_model_len |
| 158 | ) |
| 159 | |
| 160 | # ---- KV-cache wiring (reuse FlashAttentionBackend for metadata) ---- |
| 161 | cache_config = vllm_config.cache_config |
| 162 | self.kv_cache_dtype = ( |
| 163 | cache_config.cache_dtype if cache_config is not None else "auto" |
| 164 | ) |
| 165 | self.kv_cache_torch_dtype = kv_cache_dtype_str_to_dtype( |
| 166 | self.kv_cache_dtype, vllm_config.model_config |
| 167 | ) |
| 168 | self.register_buffer("k_scale", torch.ones((), dtype=torch.float32)) |
| 169 | self.register_buffer("v_scale", torch.ones((), dtype=torch.float32)) |
| 170 | |
| 171 | compilation_config = vllm_config.compilation_config |
| 172 | if prefix in compilation_config.static_forward_context: |
| 173 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 174 | compilation_config.static_forward_context[prefix] = self |
| 175 | self.kv_cache = torch.tensor([]) # replaced by bind_kv_cache |
| 176 | |
| 177 | if vllm_config.kernel_config.enable_jit_warmup: |
| 178 | _INKLING_FA4_REL_ATTENTION_KERNEL.register_warmup() |
| 179 | |
| 180 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 181 | return FlashAttentionBackend |
| 182 | |
| 183 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 184 | block_size = vllm_config.cache_config.block_size |
| 185 | if self.is_local: |
| 186 | assert self.local_extent is not None |
| 187 | return SlidingWindowSpec( |
| 188 | block_size=block_size, |
| 189 | num_kv_heads=self.num_kv_heads, |
| 190 | head_size=self.head_dim, |
| 191 | dtype=self.kv_cache_torch_dtype, |
| 192 | sliding_window=self.local_extent, |
| 193 | ) |
| 194 | return FullAttentionSpec( |
| 195 | block_size=block_size, |
| 196 | num_kv_heads=self.num_kv_heads, |
| 197 | head_size=self.head_dim, |
| 198 | dtype=self.kv_cache_torch_dtype, |
| 199 | ) |
| 200 | |
| 201 | def _split_kv_cache(self) -> tuple[torch.Tensor, torch.Tensor]: |
| 202 | key_cache, value_cache = self.kv_cache.transpose(1, 2).split( |
| 203 | self.head_dim, dim=-1 |
| 204 | ) |
| 205 | return ( |
| 206 | canonicalize_singleton_dim_strides(key_cache), |
| 207 | canonicalize_singleton_dim_strides(value_cache), |
| 208 | ) |
| 209 | |
| 210 | def forward( |
| 211 | self, |
| 212 | positions: torch.Tensor, |
| 213 | hidden_states: torch.Tensor, |
| 214 | log_scaling: torch.Tensor | None = None, |
| 215 | ) -> torch.Tensor: |
| 216 | num_tokens = hidden_states.shape[0] |
| 217 | qkvr, _ = self.qkvr(hidden_states) |
| 218 | |
| 219 | attn_metadata = get_forward_context().attn_metadata |
| 220 | attn_output = torch.empty( |
| 221 | (num_tokens, self.num_heads, self.head_dim), |
| 222 | dtype=qkvr.dtype, |
| 223 | device=qkvr.device, |
| 224 | ) |
| 225 | if not isinstance(attn_metadata, dict): |
| 226 | attn_output.zero_() |
| 227 | else: |
| 228 | conv_meta = attn_metadata[self.conv_owner.prefix] |
| 229 | md = attn_metadata[self.prefix] |
| 230 | assert isinstance(conv_meta, InklingSconvMetadata) |
| 231 | fa_md = cast(FlashAttentionMetadata, md) |
| 232 | assert self.kv_cache.numel() > 0 |
| 233 | assert self.conv_owner.kv_cache.numel() > 0 |
| 234 | # One launch: K/V sconv (conv-cache insert + conv + residual), |
| 235 | # Q/K per-head rmsnorm, and the attention KV-cache write. K/V are |
| 236 | # consumed via the KV cache; only normed q is materialized. |
| 237 | key_cache, value_cache = self._split_kv_cache() |
| 238 | off_k, _ = self.conv_owner.stream_ranges[_K] |
| 239 | off_v, _ = self.conv_owner.stream_ranges[_V] |
| 240 | q, rel_logits = fused_qkvr_prep( |
| 241 | qkvr, |
| 242 | self.k_sconv.weight.squeeze(1), |
| 243 | self.v_sconv.weight.squeeze(1), |
| 244 | self.q_norm.weight, |
| 245 | self.k_norm.weight, |
| 246 | self.rel_logits_proj.proj, |
| 247 | self.q_norm.variance_epsilon, |
| 248 | self.num_heads, |
| 249 | self.num_kv_heads, |
| 250 | self.head_dim, |
| 251 | self.d_rel, |
| 252 | self.conv_owner.kv_cache, |
| 253 | key_cache, |
| 254 | value_cache, |
| 255 | positions, |
| 256 | conv_meta.block_table, |
| 257 | conv_meta.seq_idx, |
| 258 | conv_meta.slot_mapping, |
| 259 | conv_meta.query_start, |
| 260 | fa_md.slot_mapping, |
| 261 | off_k, |
| 262 | off_v, |
| 263 | self.conv_owner.block_size, |
| 264 | log_scaling if not self.is_local else None, |
| 265 | ) |
| 266 | q = q.view(num_tokens, self.num_heads, self.head_dim) |
| 267 | self._attention(q, rel_logits, attn_output) |
| 268 | |
| 269 | flat = attn_output.view(num_tokens, -1) |
| 270 | output, _ = self.wo_ud(flat) |
| 271 | return output |
| 272 | |
| 273 | @eager_break_during_capture |
| 274 | def _attention( |
| 275 | self, |
| 276 | q: torch.Tensor, |
| 277 | rel_logits: torch.Tensor, |
| 278 | output: torch.Tensor, |
| 279 | ) -> None: |
| 280 | attn_metadata = get_forward_context().attn_metadata |
| 281 | assert isinstance(attn_metadata, dict) |
| 282 | md = cast(FlashAttentionMetadata, attn_metadata[self.prefix]) |
| 283 | |
| 284 | nt = md.num_actual_tokens |
| 285 | key_cache, value_cache = self._split_kv_cache() |
| 286 | max_seqlen_q = bucket_max_seqlen_q(md.max_query_len) |
| 287 | num_splits = inkling_fa4_num_splits( |
| 288 | is_local=self.is_local, |
| 289 | batch_size=md.seq_lens.shape[0], |
| 290 | max_query_len=max_seqlen_q, |
| 291 | num_heads=self.num_heads, |
| 292 | num_kv_heads=self.num_kv_heads, |
| 293 | max_kv_len=self._max_kv_len, |
| 294 | ) |
| 295 | _INKLING_FA4_REL_ATTENTION_KERNEL( |
| 296 | q[:nt], |
| 297 | key_cache, |
| 298 | value_cache, |
| 299 | block_table=md.block_table, |
| 300 | cache_seqlens=md.seq_lens, |
| 301 | cu_seqlens_q=md.query_start_loc, |
| 302 | max_seqlen_q=max_seqlen_q, |
| 303 | softmax_scale=self.scaling, |
| 304 | causal=True, |
| 305 | window_size=self.window_size, |
| 306 | rel_extent=self.rel_extent, |
| 307 | rel_logits=rel_logits[:nt], |
| 308 | num_splits=num_splits, |
| 309 | out=output[:nt], |
| 310 | ) |