| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Clean Multi-head Latent Attention for Kimi-K3 (NVIDIA). |
| 4 | |
| 5 | This is a self-contained MLA layer that owns the full attention path: |
| 6 | |
| 7 | hidden_states |
| 8 | -> fused pre-attention ops (fused_qkv_a_proj / norms / q_b_proj; |
| 9 | with an output gate the rows use fused_qkv_a_g_proj instead) |
| 10 | -> explicit prefill / decode split |
| 11 | prefill: fused key-concat + cache-insert kernel -> run_prefill_new_tokens |
| 12 | (+ chunked-context merge, whose per-chunk gather -> kv_b_proj |
| 13 | -> fused K/V pack loop this layer owns); dispatched by cache |
| 14 | dtype (bf16 / plain fp8 / fp8_ds_mla) |
| 15 | decode : W_UK absorb (BMM1) -> fused q-concat + cache-insert kernel |
| 16 | -> impl.forward_mqa -> W_UV up-proj (MQA) |
| 17 | -> optional output gate |
| 18 | -> o_proj |
| 19 | |
| 20 | Unlike ``MultiHeadLatentAttentionWrapper`` (which delegates orchestration to |
| 21 | ``MLAAttention.forward``), this class *is* the ``AttentionLayerBase``: it selects |
| 22 | the backend, builds the impl, registers itself in the forward context, owns the |
| 23 | KV cache, and absorbs ``kv_b_proj`` into ``W_UK_T`` / ``W_UV`` -- mirroring the |
| 24 | ``DeepseekV4Attention`` structure. |
| 25 | |
| 26 | K3 specifics: optional rotary embedding (disabled for the target model's NoPE |
| 27 | layers, enabled for DSpark) and an optional sigmoid output gate (``g_proj``, |
| 28 | merged into ``fused_qkv_a_g_proj`` when the q-LoRA front-end is present). |
| 29 | |
| 30 | Out of scope (extension points, not wired here): prefill context parallelism |
| 31 | (PCP), sparse/indexer MLA, and the ROCm/aiter fp8/fp4 BMM fast paths. |
| 32 | """ |
| 33 | |
| 34 | import math |
| 35 | from typing import TYPE_CHECKING, cast |
| 36 | |
| 37 | import torch |
| 38 | import torch.nn.functional as F |
| 39 | from torch import nn |
| 40 | |
| 41 | from vllm import _custom_ops as ops |
| 42 | from vllm.compilation.breakable_cudagraph import eager_break_during_capture |
| 43 | from vllm.config import ( |
| 44 | CacheConfig, |
| 45 | VllmConfig, |
| 46 | get_current_vllm_config, |
| 47 | ) |
| 48 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 49 | from vllm.forward_context import get_forward_context |
| 50 | from vllm.logger import init_logger |
| 51 | from vllm.model_executor.layers.attention.attention import ( |
| 52 | _init_kv_cache_quant, |
| 53 | set_default_quant_scales, |
| 54 | should_load_quant_weights, |
| 55 | ) |
| 56 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 57 | _get_kv_b_proj_input_dtype, |
| 58 | accumulate_mla_context_chunk, |
| 59 | init_mla_context_partial, |
| 60 | neutralize_empty_context_partials, |
| 61 | split_kv_b_proj, |
| 62 | ) |
| 63 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 64 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 65 | from vllm.model_executor.layers.linear import ( |
| 66 | ColumnParallelLinear, |
| 67 | KimiK3MergedQKVGateLinear, |
| 68 | MergedColumnParallelLinear, |
| 69 | ReplicatedLinear, |
| 70 | RowParallelLinear, |
| 71 | UnquantizedLinearMethod, |
| 72 | ) |
| 73 | from vllm.model_executor.layers.quantization import ( |
| 74 | QuantizationConfig, |
| 75 | resolve_quant_method, |
| 76 | ) |
| 77 | from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope |
| 78 | from vllm.model_executor.utils import replace_parameter |
| 79 | from vllm.models.common.ops import fused_q_kv_rmsnorm |
| 80 | from vllm.models.kimi_k3.nvidia.low_latency_gemm import try_low_latency_gemm |
| 81 | from vllm.models.kimi_k3.nvidia.ops.fused_mla_key_concat_kv_cache import ( |
| 82 | fused_mla_decode_q_concat_kv_cache_insert, |
| 83 | fused_mla_key_concat_ds_mla_insert, |
| 84 | fused_mla_key_concat_kv_cache_insert, |
| 85 | fused_mla_kv_concat, |
| 86 | fused_mla_kv_concat_quant_fp8, |
| 87 | fused_mla_qkv_quant_kv_cache_fp8_insert, |
| 88 | ) |
| 89 | from vllm.platforms import current_platform |
| 90 | from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig |
| 91 | from vllm.utils.multi_stream_utils import maybe_execute_in_parallel |
| 92 | from vllm.utils.torch_utils import ( |
| 93 | is_quantized_kv_cache, |
| 94 | kv_cache_dtype_str_to_dtype, |
| 95 | ) |
| 96 | from vllm.v1.attention.backend import ( |
| 97 | AttentionBackend, |
| 98 | AttentionType, |
| 99 | MLAAttentionImpl, |
| 100 | ) |
| 101 | from vllm.v1.attention.backends.mla.prefill import get_mla_prefill_backend |
| 102 | from vllm.v1.attention.ops.dcp import MLADCPManager |
| 103 | from vllm.v1.attention.ops.merge_attn_states import merge_attn_states |
| 104 | from vllm.v1.attention.selector import get_attn_backend |
| 105 | from vllm.v1.kv_cache_interface import KVCacheSpec, MLAAttentionSpec, get_kv_quant_mode |
| 106 | |
| 107 | if TYPE_CHECKING: |
| 108 | from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadata |
| 109 | |
| 110 | logger = init_logger(__name__) |
| 111 | |
| 112 | # Below this conservative threshold, overlap the gate projection with attention |
| 113 | # on the auxiliary stream. |
| 114 | _GATE_MULTI_STREAM_TOKEN_THRESHOLD = 512 |
| 115 | |
| 116 | |
| 117 | @torch.compile(backend=current_platform.simple_compile_backend) |
| 118 | def _gate_sigmoid_mul(attn_out: torch.Tensor, gate: torch.Tensor) -> torch.Tensor: |
| 119 | """Apply the sigmoid output gate to a precomputed gate projection.""" |
| 120 | return attn_out * gate.sigmoid() |
| 121 | |
| 122 | |
| 123 | class MultiHeadLatentAttention(nn.Module, AttentionLayerBase): |
| 124 | """Kimi-K3 Multi-head Latent Attention with optional RoPE and output gate.""" |
| 125 | |
| 126 | def __init__( |
| 127 | self, |
| 128 | config: KimiLinearConfig, |
| 129 | hidden_size: int, |
| 130 | num_heads: int, |
| 131 | qk_nope_head_dim: int, |
| 132 | qk_rope_head_dim: int, |
| 133 | v_head_dim: int, |
| 134 | q_lora_rank: int | None, |
| 135 | kv_lora_rank: int, |
| 136 | use_output_gate: bool = False, |
| 137 | cache_config: CacheConfig | None = None, |
| 138 | quant_config: QuantizationConfig | None = None, |
| 139 | prefix: str = "", |
| 140 | aux_stream: torch.cuda.Stream | None = None, |
| 141 | use_rope: bool = False, |
| 142 | non_causal_multi_token_decode: bool = False, |
| 143 | run_gemm_rs_ar: bool = False, |
| 144 | ) -> None: |
| 145 | super().__init__() |
| 146 | self.hidden_size = hidden_size |
| 147 | self.qk_nope_head_dim = qk_nope_head_dim |
| 148 | self.qk_rope_head_dim = qk_rope_head_dim |
| 149 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 150 | self.v_head_dim = v_head_dim |
| 151 | self.q_lora_rank = q_lora_rank |
| 152 | self.kv_lora_rank = kv_lora_rank |
| 153 | self.use_output_gate = use_output_gate |
| 154 | self.non_causal_multi_token_decode = non_causal_multi_token_decode |
| 155 | # Latent "head" seen by the attention kernel / KV cache. |
| 156 | self.head_size = kv_lora_rank + qk_rope_head_dim |
| 157 | self.scale = self.qk_head_dim**-0.5 |
| 158 | self.rms_norm_eps = config.rms_norm_eps |
| 159 | self.layer_name = prefix |
| 160 | |
| 161 | self.rotary_emb: RotaryEmbedding | None = None |
| 162 | if use_rope: |
| 163 | rope_parameters = dict(config.rope_parameters) |
| 164 | if rope_parameters["rope_type"] != "default": |
| 165 | rope_parameters["rope_type"] = ( |
| 166 | "deepseek_llama_scaling" |
| 167 | if rope_parameters.get("attention_factor") == 1.0 |
| 168 | else "deepseek_yarn" |
| 169 | ) |
| 170 | self.rotary_emb = get_rope( |
| 171 | qk_rope_head_dim, |
| 172 | max_position=config.max_position_embeddings, |
| 173 | rope_parameters=rope_parameters, |
| 174 | is_neox_style=False, |
| 175 | dtype=torch.float32, |
| 176 | ) |
| 177 | if rope_parameters["rope_type"] == "deepseek_yarn": |
| 178 | mscale_all_dim = rope_parameters.get("mscale_all_dim", False) |
| 179 | scaling_factor = rope_parameters["factor"] |
| 180 | mscale = ( |
| 181 | 1.0 |
| 182 | if scaling_factor <= 1 |
| 183 | else 0.1 * float(mscale_all_dim) * math.log(scaling_factor) + 1.0 |
| 184 | ) |
| 185 | self.scale *= mscale * mscale |
| 186 | # The fused epilogues read the cos/sin table directly in fp32 and run |
| 187 | # the RoPE math in fp32, so there is no per-forward dtype cast (and no |
| 188 | # precision loss). deepseek_yarn builds cos_sin_cache in fp32 already; |
| 189 | # dtype=torch.float32 above forces it for the default rope too (the |
| 190 | # DSpark draft, which has no yarn scaling). |
| 191 | assert self.rotary_emb.cos_sin_cache.dtype == torch.float32, ( |
| 192 | "K3 fused MLA RoPE requires an fp32 cos/sin cache; got " |
| 193 | f"{self.rotary_emb.cos_sin_cache.dtype}." |
| 194 | ) |
| 195 | |
| 196 | tp_size = get_tensor_model_parallel_world_size() |
| 197 | assert num_heads % tp_size == 0 |
| 198 | self.num_heads = num_heads |
| 199 | self.num_local_heads = num_heads // tp_size |
| 200 | |
| 201 | # ---- Pre-attention projections (fusable front-end) ---- |
| 202 | # Exactly one query layout is constructed below. Kimi-K3 uses q-LoRA; |
| 203 | # Kimi-Linear uses independent full-rank Q and KV projections. |
| 204 | self.fused_qkv_a_proj: MergedColumnParallelLinear | None = None |
| 205 | self.fused_qkv_a_g_proj: KimiK3MergedQKVGateLinear | None = None |
| 206 | self.q_proj: ColumnParallelLinear | None = None |
| 207 | self.kv_a_proj_with_mqa: ReplicatedLinear | None = None |
| 208 | self.g_proj: ColumnParallelLinear | None = None |
| 209 | self.aux_stream: torch.cuda.Stream | None = None |
| 210 | self._gate_events: tuple[torch.cuda.Event, torch.cuda.Event] | None = None |
| 211 | |
| 212 | # Kimi-K3: low-rank Q and KV projections share one input projection. |
| 213 | if self.q_lora_rank is not None: |
| 214 | # The latent Q and KV shards are replicated across TP ranks. Their |
| 215 | # TP splitting happens in q_b_proj and kv_b_proj instead. |
| 216 | if use_output_gate: |
| 217 | # The output gate is shard 2 of the same input projection. |
| 218 | self.fused_qkv_a_g_proj = KimiK3MergedQKVGateLinear( |
| 219 | hidden_size=self.hidden_size, |
| 220 | q_lora_rank=self.q_lora_rank, |
| 221 | kv_lora_rank=self.kv_lora_rank, |
| 222 | qk_rope_head_dim=self.qk_rope_head_dim, |
| 223 | total_num_heads=num_heads, |
| 224 | v_head_dim=self.v_head_dim, |
| 225 | bias=False, |
| 226 | quant_config=quant_config, |
| 227 | prefix=f"{prefix}.fused_qkv_a_g_proj", |
| 228 | ) |
| 229 | # Only supports unquantized weights for sliced side-stream path |
| 230 | # for now. |
| 231 | if aux_stream is not None and isinstance( |
| 232 | self.fused_qkv_a_g_proj.quant_method, UnquantizedLinearMethod |
| 233 | ): |
| 234 | self.aux_stream = aux_stream |
| 235 | self._gate_events = (torch.cuda.Event(), torch.cuda.Event()) |
| 236 | else: |
| 237 | self.fused_qkv_a_proj = MergedColumnParallelLinear( |
| 238 | self.hidden_size, |
| 239 | [ |
| 240 | self.q_lora_rank, |
| 241 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 242 | ], |
| 243 | bias=False, |
| 244 | quant_config=quant_config, |
| 245 | prefix=f"{prefix}.fused_qkv_a_proj", |
| 246 | disable_tp=True, |
| 247 | ) |
| 248 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 249 | self.q_b_proj = ColumnParallelLinear( |
| 250 | self.q_lora_rank, |
| 251 | self.num_heads * self.qk_head_dim, |
| 252 | bias=False, |
| 253 | quant_config=quant_config, |
| 254 | prefix=f"{prefix}.q_b_proj", |
| 255 | ) |
| 256 | |
| 257 | # Kimi-Linear: Q and KV use independent full-rank input projections. |
| 258 | else: |
| 259 | self.q_proj = ColumnParallelLinear( |
| 260 | self.hidden_size, |
| 261 | self.num_heads * self.qk_head_dim, |
| 262 | bias=False, |
| 263 | quant_config=quant_config, |
| 264 | prefix=f"{prefix}.q_proj", |
| 265 | ) |
| 266 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 267 | self.hidden_size, |
| 268 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 269 | bias=False, |
| 270 | quant_config=quant_config, |
| 271 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 272 | ) |
| 273 | if use_output_gate: |
| 274 | # Keep the gate standalone for the non-LoRA layout. |
| 275 | self.g_proj = ColumnParallelLinear( |
| 276 | self.hidden_size, |
| 277 | self.num_heads * self.v_head_dim, |
| 278 | bias=False, |
| 279 | quant_config=quant_config, |
| 280 | prefix=f"{prefix}.g_proj", |
| 281 | ) |
| 282 | |
| 283 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 284 | self.kv_b_proj = ColumnParallelLinear( |
| 285 | self.kv_lora_rank, |
| 286 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 287 | bias=False, |
| 288 | quant_config=quant_config, |
| 289 | prefix=f"{prefix}.kv_b_proj", |
| 290 | ) |
| 291 | |
| 292 | # ---- Post-attention projection ---- |
| 293 | self.o_proj = RowParallelLinear( |
| 294 | self.num_heads * self.v_head_dim, |
| 295 | self.hidden_size, |
| 296 | bias=False, |
| 297 | quant_config=quant_config, |
| 298 | prefix=f"{prefix}.o_proj", |
| 299 | ) |
| 300 | self.gemm_rs_ar = None |
| 301 | if run_gemm_rs_ar: |
| 302 | from vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar import ( |
| 303 | get_gemm_rs_ar, |
| 304 | ) |
| 305 | |
| 306 | gemm_rs_ar = get_gemm_rs_ar() |
| 307 | if gemm_rs_ar.can_run(self.o_proj): |
| 308 | self.gemm_rs_ar = gemm_rs_ar |
| 309 | else: |
| 310 | gemm_rs_ar.warn_incompatible_projection() |
| 311 | |
| 312 | # ---- Attention backend / impl / KV cache ---- |
| 313 | self.quant_config = quant_config |
| 314 | if cache_config is not None: |
| 315 | self.kv_cache_dtype = cache_config.cache_dtype |
| 316 | else: |
| 317 | self.kv_cache_dtype = "auto" |
| 318 | |
| 319 | dtype = torch.get_default_dtype() |
| 320 | self.attn_backend = get_attn_backend( |
| 321 | self.head_size, |
| 322 | dtype, |
| 323 | self.kv_cache_dtype, |
| 324 | use_mla=True, |
| 325 | use_sparse=False, |
| 326 | num_heads=self.num_local_heads, |
| 327 | ) |
| 328 | _init_kv_cache_quant(self, quant_config, prefix) |
| 329 | # Unit (1.0) scale for the fused fp8 prefill path: q/k/v are cast |
| 330 | # unscaled to match forward_mha (the prefill flash path does not |
| 331 | # dequantize); only the cache uses _k_scale. |
| 332 | self.register_buffer( |
| 333 | "_one_scale", torch.ones(1, dtype=torch.float32), persistent=False |
| 334 | ) |
| 335 | |
| 336 | impl_cls = cast(type[MLAAttentionImpl], self.attn_backend.get_impl_cls()) |
| 337 | self.impl = impl_cls( # type: ignore[assignment] |
| 338 | num_heads=self.num_local_heads, |
| 339 | head_size=self.head_size, |
| 340 | scale=self.scale, |
| 341 | num_kv_heads=1, |
| 342 | alibi_slopes=None, |
| 343 | sliding_window=None, |
| 344 | kv_cache_dtype=self.kv_cache_dtype, |
| 345 | logits_soft_cap=None, |
| 346 | attn_type=AttentionType.DECODER, |
| 347 | kv_sharing_target_layer_name=None, |
| 348 | q_lora_rank=self.q_lora_rank, |
| 349 | kv_lora_rank=self.kv_lora_rank, |
| 350 | qk_nope_head_dim=self.qk_nope_head_dim, |
| 351 | qk_rope_head_dim=self.qk_rope_head_dim, |
| 352 | qk_head_dim=self.qk_head_dim, |
| 353 | v_head_dim=self.v_head_dim, |
| 354 | kv_b_proj=self.kv_b_proj, |
| 355 | indexer=None, |
| 356 | ) |
| 357 | if getattr(self.impl, "dcp_world_size", -1) < 1: |
| 358 | # FlashAttention requires the cp_world_size is positive and the cp_rank |
| 359 | # is non negative; manually set here if not set by caller (-1 is unset) |
| 360 | self.impl.dcp_world_size = 1 |
| 361 | self.impl.dcp_rank = 0 |
| 362 | self.q_pad_num_heads = getattr(self.impl, "q_pad_num_heads", None) |
| 363 | |
| 364 | vllm_config = get_current_vllm_config() |
| 365 | parallel_config = vllm_config.parallel_config |
| 366 | assert parallel_config.prefill_context_parallel_size == 1, ( |
| 367 | "Kimi-K3 MultiHeadLatentAttention does not support prefill context " |
| 368 | "parallelism." |
| 369 | ) |
| 370 | self.dcp_world_size = parallel_config.decode_context_parallel_size |
| 371 | self.dcp_manager: MLADCPManager | None = None |
| 372 | if self.dcp_world_size > 1: |
| 373 | query_dtype = ( |
| 374 | torch.float8_e4m3fn |
| 375 | if is_quantized_kv_cache(self.kv_cache_dtype) |
| 376 | and self.kv_cache_dtype != "fp8_ds_mla" |
| 377 | else dtype |
| 378 | ) |
| 379 | self.dcp_manager = MLADCPManager( |
| 380 | vllm_config=vllm_config, |
| 381 | device=next(self.kv_b_proj.parameters()).device, |
| 382 | num_heads=self.num_local_heads, |
| 383 | query_head_dim=self.head_size, |
| 384 | output_head_dim=self.kv_lora_rank, |
| 385 | query_dtype=query_dtype, |
| 386 | output_dtype=dtype, |
| 387 | padded_num_heads=self.q_pad_num_heads, |
| 388 | is_lse_base_on_e=self.impl.lse_base_on_e, |
| 389 | use_pcp=False, |
| 390 | ) |
| 391 | self.prefill_backend = get_mla_prefill_backend(vllm_config)( |
| 392 | num_heads=self.num_local_heads, |
| 393 | scale=self.scale, |
| 394 | kv_lora_rank=self.kv_lora_rank, |
| 395 | qk_nope_head_dim=self.qk_nope_head_dim, |
| 396 | qk_rope_head_dim=self.qk_rope_head_dim, |
| 397 | v_head_dim=self.v_head_dim, |
| 398 | vllm_config=vllm_config, |
| 399 | ) |
| 400 | |
| 401 | compilation_config = vllm_config.compilation_config |
| 402 | if prefix in compilation_config.static_forward_context: |
| 403 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 404 | compilation_config.static_forward_context[prefix] = self |
| 405 | self.kv_cache = torch.tensor([]) |
| 406 | |
| 407 | # ------------------------------------------------------------------ |
| 408 | # AttentionLayerBase interface |
| 409 | # ------------------------------------------------------------------ |
| 410 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 411 | # [B, H=1, N, C] -> [B, N, C] |
| 412 | self.kv_cache = kv_cache.squeeze(1) |
| 413 | |
| 414 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 415 | return self.attn_backend |
| 416 | |
| 417 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 418 | kv_cache_dtype = kv_cache_dtype_str_to_dtype( |
| 419 | self.kv_cache_dtype, vllm_config.model_config |
| 420 | ) |
| 421 | # TODO: Remove this mypy workaround once the K3 PR is fully merged. |
| 422 | return MLAAttentionSpec( # type: ignore[call-arg] |
| 423 | block_size=vllm_config.cache_config.block_size, |
| 424 | num_kv_heads=1, |
| 425 | head_size=self.head_size, |
| 426 | dtype=kv_cache_dtype, |
| 427 | cache_dtype_str=self.kv_cache_dtype, |
| 428 | kv_quant_mode=get_kv_quant_mode(self.kv_cache_dtype), |
| 429 | # fp8_ds_mla: 656-byte custom layout; see flashmla_sparse.py. |
| 430 | state_content_bytes=656 if self.kv_cache_dtype == "fp8_ds_mla" else None, |
| 431 | non_causal_multi_token_decode=self.non_causal_multi_token_decode, |
| 432 | ) |
| 433 | |
| 434 | def process_weights_after_loading(self, act_dtype: torch.dtype) -> None: |
| 435 | """Absorb ``kv_b_proj`` into decode-time ``W_UK_T`` / ``W_UV`` bmm weights. |
| 436 | |
| 437 | ``kv_b_proj`` produces ``[k_nope; v]`` per head from the ``kv_lora_rank`` |
| 438 | latent. For the MQA decode path we pre-split it so that queries are |
| 439 | projected into latent space by ``W_UK_T`` and the attention output is |
| 440 | projected back to ``v`` by ``W_UV`` -- avoiding materializing full K/V. |
| 441 | """ |
| 442 | W_UK, W_UV = split_kv_b_proj( |
| 443 | self.kv_b_proj, |
| 444 | act_dtype, |
| 445 | self.kv_lora_rank, |
| 446 | self.num_local_heads, |
| 447 | self.qk_nope_head_dim, |
| 448 | self.v_head_dim, |
| 449 | ) |
| 450 | # (L, N, V) -> (N, L, V) |
| 451 | replace_parameter(self, "W_UV", W_UV.transpose(0, 1), prefer_copy=True) |
| 452 | # (L, N, P) -> (N, P, L) |
| 453 | replace_parameter(self, "W_UK_T", W_UK.permute(1, 2, 0), prefer_copy=True) |
| 454 | |
| 455 | quant_method = ( |
| 456 | resolve_quant_method(self.quant_config, self, prefix=self.layer_name) |
| 457 | if self.quant_config |
| 458 | else None |
| 459 | ) |
| 460 | if not should_load_quant_weights(quant_method): |
| 461 | set_default_quant_scales(self, register_buffer=False) |
| 462 | |
| 463 | # Precompute reciprocal scales once here (scales are final after load; |
| 464 | # K3 has no runtime calculate_kv_scales path) so the fp8 fused kernels |
| 465 | # in the decode/prefill hot path take a ready inverse instead of |
| 466 | # launching a per-step reciprocal kernel. |
| 467 | self.register_buffer( |
| 468 | "_q_scale_inv", self._q_scale.reciprocal().reshape(1), persistent=False |
| 469 | ) |
| 470 | self.register_buffer( |
| 471 | "_k_scale_inv", self._k_scale.reciprocal().reshape(1), persistent=False |
| 472 | ) |
| 473 | |
| 474 | def _v_up_proj(self, x: torch.Tensor, out: torch.Tensor) -> None: |
| 475 | """Project latent attention output back to ``v`` via ``W_UV`` (bmm).""" |
| 476 | # (B, N, L) -> (N, B, L) |
| 477 | x = x.view(-1, self.num_local_heads, self.kv_lora_rank).transpose(0, 1) |
| 478 | out = out.view(-1, self.num_local_heads, self.v_head_dim) |
| 479 | # (N, B, L) x (N, L, V) -> (N, B, V) written transposed into (B, N, V) |
| 480 | torch.bmm(x, self.W_UV, out=out.transpose(0, 1)) |
| 481 | |
| 482 | def _attn_read_kv_cache(self) -> torch.Tensor: |
| 483 | """Latent cache as seen by the attention read kernels (decode / context). |
| 484 | |
| 485 | A plain per-tensor fp8 cache is stored as ``uint8``; view it as fp8 so |
| 486 | the backend reads it as E4M3 rather than fp4/E2M1 -- the latter doubles |
| 487 | the perceived head dim (``head_size * 2``) and fails the kernel's |
| 488 | ``head_dim_k == head_dim_q`` check. Mirrors ``MLAAttention.forward``; |
| 489 | the fp8_ds_mla layout keeps its native uint8 view. |
| 490 | """ |
| 491 | cache = self.kv_cache |
| 492 | if ( |
| 493 | is_quantized_kv_cache(self.kv_cache_dtype) |
| 494 | and self.kv_cache_dtype != "fp8_ds_mla" |
| 495 | ): |
| 496 | return cache.view(current_platform.fp8_dtype()) |
| 497 | return cache |
| 498 | |
| 499 | # ------------------------------------------------------------------ |
| 500 | # Forward |
| 501 | # ------------------------------------------------------------------ |
| 502 | def _unquantized_gemm( |
| 503 | self, hidden_states: torch.Tensor, weight: torch.Tensor |
| 504 | ) -> torch.Tensor: |
| 505 | out = try_low_latency_gemm(hidden_states, weight) |
| 506 | return out if out is not None else F.linear(hidden_states, weight) |
| 507 | |
| 508 | def _apply_q_lora_attention( |
| 509 | self, |
| 510 | positions: torch.Tensor, |
| 511 | hidden_states: torch.Tensor, |
| 512 | qkv_lora: torch.Tensor, |
| 513 | ) -> torch.Tensor: |
| 514 | """Expand q-LoRA output and run attention.""" |
| 515 | q_lora_rank = self.q_lora_rank |
| 516 | assert q_lora_rank is not None |
| 517 | q_c, kv_c, k_pe = qkv_lora.split( |
| 518 | [q_lora_rank, self.kv_lora_rank, self.qk_rope_head_dim], dim=-1 |
| 519 | ) |
| 520 | q_c, kv_c_normed = fused_q_kv_rmsnorm( |
| 521 | q_c, |
| 522 | kv_c, |
| 523 | self.q_a_layernorm.weight.data, |
| 524 | self.kv_a_layernorm.weight.data, |
| 525 | self.rms_norm_eps, |
| 526 | ) |
| 527 | q = self.q_b_proj(q_c)[0].view(-1, self.num_local_heads, self.qk_head_dim) |
| 528 | |
| 529 | attn_out = torch.empty( |
| 530 | (hidden_states.shape[0], self.num_local_heads * self.v_head_dim), |
| 531 | dtype=hidden_states.dtype, |
| 532 | device=hidden_states.device, |
| 533 | ) |
| 534 | self._attention(positions, q, kv_c_normed, k_pe.unsqueeze(1), attn_out) |
| 535 | return attn_out |
| 536 | |
| 537 | def _forward_q_lora( |
| 538 | self, |
| 539 | positions: torch.Tensor, |
| 540 | hidden_states: torch.Tensor, |
| 541 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 542 | """Run the q-LoRA path, including gate overlap when eligible.""" |
| 543 | q_lora_rank = self.q_lora_rank |
| 544 | assert q_lora_rank is not None |
| 545 | qkv_a_rows = q_lora_rank + self.kv_lora_rank + self.qk_rope_head_dim |
| 546 | |
| 547 | # run g_proj in aux_stream |
| 548 | if ( |
| 549 | self._gate_events is not None |
| 550 | and hidden_states.shape[0] < _GATE_MULTI_STREAM_TOKEN_THRESHOLD |
| 551 | ): |
| 552 | fused_qkv_a_g_proj = self.fused_qkv_a_g_proj |
| 553 | assert fused_qkv_a_g_proj is not None |
| 554 | qkv_a_weight = fused_qkv_a_g_proj.weight[:qkv_a_rows] |
| 555 | gate_weight = fused_qkv_a_g_proj.weight[qkv_a_rows:] |
| 556 | |
| 557 | start_event, end_event = self._gate_events |
| 558 | return maybe_execute_in_parallel( |
| 559 | lambda: self._apply_q_lora_attention( |
| 560 | positions, |
| 561 | hidden_states, |
| 562 | self._unquantized_gemm(hidden_states, qkv_a_weight), |
| 563 | ), |
| 564 | lambda: self._unquantized_gemm(hidden_states, gate_weight), |
| 565 | start_event, |
| 566 | end_event, |
| 567 | self.aux_stream, |
| 568 | ) |
| 569 | |
| 570 | # run g_proj together with qkv_a |
| 571 | if self.use_output_gate: |
| 572 | qkv_a_proj = self.fused_qkv_a_g_proj |
| 573 | assert qkv_a_proj is not None |
| 574 | qkv_lora, gate = qkv_a_proj(hidden_states)[0].split( |
| 575 | [qkv_a_rows, self.num_local_heads * self.v_head_dim], dim=-1 |
| 576 | ) |
| 577 | else: |
| 578 | qkv_a_proj = self.fused_qkv_a_proj |
| 579 | assert qkv_a_proj is not None |
| 580 | qkv_lora = qkv_a_proj(hidden_states)[0] |
| 581 | gate = None |
| 582 | |
| 583 | attn_out = self._apply_q_lora_attention(positions, hidden_states, qkv_lora) |
| 584 | return attn_out, gate |
| 585 | |
| 586 | def _forward_full_rank_q( |
| 587 | self, |
| 588 | positions: torch.Tensor, |
| 589 | hidden_states: torch.Tensor, |
| 590 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 591 | """Run the full-rank Q path with an optional sequential gate.""" |
| 592 | q_proj = self.q_proj |
| 593 | kv_a_proj_with_mqa = self.kv_a_proj_with_mqa |
| 594 | assert q_proj is not None |
| 595 | assert kv_a_proj_with_mqa is not None |
| 596 | |
| 597 | q = q_proj(hidden_states)[0].view(-1, self.num_local_heads, self.qk_head_dim) |
| 598 | kv_lora = kv_a_proj_with_mqa(hidden_states)[0] |
| 599 | kv_c, k_pe = kv_lora.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) |
| 600 | kv_c_normed = self.kv_a_layernorm(kv_c) |
| 601 | attn_out = torch.empty( |
| 602 | (hidden_states.shape[0], self.num_local_heads * self.v_head_dim), |
| 603 | dtype=hidden_states.dtype, |
| 604 | device=hidden_states.device, |
| 605 | ) |
| 606 | self._attention(positions, q, kv_c_normed, k_pe.unsqueeze(1), attn_out) |
| 607 | |
| 608 | gate = self.g_proj(hidden_states)[0] if self.g_proj is not None else None |
| 609 | return attn_out, gate |
| 610 | |
| 611 | def forward( |
| 612 | self, |
| 613 | positions: torch.Tensor, |
| 614 | hidden_states: torch.Tensor, |
| 615 | ) -> torch.Tensor: |
| 616 | if self.q_lora_rank is None: |
| 617 | attn_out, gate = self._forward_full_rank_q(positions, hidden_states) |
| 618 | else: |
| 619 | attn_out, gate = self._forward_q_lora(positions, hidden_states) |
| 620 | |
| 621 | if gate is not None: |
| 622 | attn_out = _gate_sigmoid_mul(attn_out, gate) |
| 623 | |
| 624 | if self.gemm_rs_ar is not None and self.gemm_rs_ar.should_run(attn_out): |
| 625 | return self.gemm_rs_ar.apply(attn_out, self.o_proj) |
| 626 | |
| 627 | return self.o_proj(attn_out)[0] |
| 628 | |
| 629 | @eager_break_during_capture |
| 630 | def _attention( |
| 631 | self, |
| 632 | positions: torch.Tensor, |
| 633 | q: torch.Tensor, |
| 634 | kv_c_normed: torch.Tensor, |
| 635 | k_pe: torch.Tensor, |
| 636 | attn_out: torch.Tensor, |
| 637 | ) -> None: |
| 638 | forward_context = get_forward_context() |
| 639 | attn_metadata_by_layer = forward_context.attn_metadata |
| 640 | if attn_metadata_by_layer is None: |
| 641 | attn_out.zero_() |
| 642 | return |
| 643 | assert isinstance(attn_metadata_by_layer, dict) |
| 644 | attn_metadata = cast( |
| 645 | "MLACommonMetadata", attn_metadata_by_layer[self.layer_name] |
| 646 | ) |
| 647 | |
| 648 | num_actual_toks = attn_metadata.num_actual_tokens |
| 649 | slot_mapping_by_layer = forward_context.slot_mapping |
| 650 | assert isinstance(slot_mapping_by_layer, dict) |
| 651 | slot_mapping = slot_mapping_by_layer[self.layer_name] |
| 652 | |
| 653 | q = q[:num_actual_toks] |
| 654 | kv_c_normed = kv_c_normed[:num_actual_toks] |
| 655 | k_pe = k_pe[:num_actual_toks] |
| 656 | positions = positions[:num_actual_toks] |
| 657 | attn_out = attn_out[:num_actual_toks] |
| 658 | |
| 659 | cos_sin_cache = None |
| 660 | rope_positions = None |
| 661 | if self.rotary_emb is not None: |
| 662 | # Pass the fp32 cos/sin table straight to the fused epilogue (it reads |
| 663 | # fp32 and does the RoPE math in fp32) -- no per-forward dtype cast. |
| 664 | cos_sin_cache = self.rotary_emb.cos_sin_cache |
| 665 | rope_positions = positions |
| 666 | |
| 667 | # Decode tokens are laid out first, prefill tokens after. The fused |
| 668 | # prefill covers every supported config (bf16 / plain-fp8 / |
| 669 | # fp8_ds_mla), so there is no dense-MHA (forward_mha) fallback. |
| 670 | num_mqa_tokens = attn_metadata.num_decode_tokens |
| 671 | num_mha_tokens = q.size(0) - num_mqa_tokens |
| 672 | |
| 673 | # Both the prefill and decode fused epilogues write their own cache |
| 674 | # slice, so there is no separate do_kv_cache_update. |
| 675 | |
| 676 | # ---- Prefill: fused key-concat + cache-insert + attention ---- |
| 677 | if num_mha_tokens > 0: |
| 678 | self._forward_prefill_fused( |
| 679 | q[num_mqa_tokens:], |
| 680 | kv_c_normed[num_mqa_tokens:], |
| 681 | k_pe[num_mqa_tokens:], |
| 682 | rope_positions[num_mqa_tokens:] if rope_positions is not None else None, |
| 683 | cos_sin_cache, |
| 684 | slot_mapping[num_mqa_tokens:num_actual_toks], |
| 685 | attn_metadata, |
| 686 | attn_out[num_mqa_tokens:], |
| 687 | ) |
| 688 | |
| 689 | # ---- Decode: latent multi-query attention ---- |
| 690 | if num_mqa_tokens > 0: |
| 691 | mqa_q_nope, mqa_q_pe = q[:num_mqa_tokens].split( |
| 692 | [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1 |
| 693 | ) |
| 694 | # BMM1: absorb q_nope into latent space. (N,B,P) x (N,P,L) -> (B,N,L) |
| 695 | ql_nope = torch.bmm(mqa_q_nope.transpose(0, 1), self.W_UK_T).transpose(0, 1) |
| 696 | # Fused: concat mqa_q = [ql_nope | q_pe] and insert the decode-token |
| 697 | # latent into the paged cache (one launch, right before forward_mqa). |
| 698 | mqa_q = self._decode_concat_cache( |
| 699 | ql_nope, |
| 700 | mqa_q_pe, |
| 701 | kv_c_normed[:num_mqa_tokens], |
| 702 | k_pe[:num_mqa_tokens], |
| 703 | rope_positions[:num_mqa_tokens] if rope_positions is not None else None, |
| 704 | cos_sin_cache, |
| 705 | slot_mapping[:num_mqa_tokens], |
| 706 | ) |
| 707 | if self.dcp_world_size > 1: |
| 708 | assert self.dcp_manager is not None |
| 709 | assert self.dcp_manager.query_gather is not None |
| 710 | mqa_q = self.dcp_manager.query_gather(mqa_q) |
| 711 | latent_out, lse = self.impl.forward_mqa( # type: ignore[attr-defined] |
| 712 | mqa_q, self._attn_read_kv_cache(), attn_metadata, self |
| 713 | ) |
| 714 | if self.dcp_world_size > 1: |
| 715 | assert lse is not None |
| 716 | assert self.dcp_manager is not None |
| 717 | assert attn_metadata.decode is not None |
| 718 | latent_out = self.dcp_manager.combine( |
| 719 | latent_out, |
| 720 | lse, |
| 721 | seq_lens=attn_metadata.decode.seq_lens, |
| 722 | query_start_loc=attn_metadata.query_start_loc[ |
| 723 | : attn_metadata.num_decodes + 1 |
| 724 | ], |
| 725 | ) |
| 726 | self._v_up_proj(latent_out, out=attn_out[:num_mqa_tokens]) |
| 727 | |
| 728 | def _decode_concat_cache( |
| 729 | self, |
| 730 | ql_nope: torch.Tensor, |
| 731 | q_pe: torch.Tensor, |
| 732 | kv_c_normed: torch.Tensor, |
| 733 | k_pe: torch.Tensor, |
| 734 | positions: torch.Tensor | None, |
| 735 | cos_sin_cache: torch.Tensor | None, |
| 736 | slot_mapping: torch.Tensor, |
| 737 | ) -> torch.Tensor: |
| 738 | """Fused decode query-concat + latent cache insert, dispatched by cache |
| 739 | dtype (same policy as prefill: fp8 cache -> fp8 query).""" |
| 740 | if self.kv_cache_dtype == "fp8_ds_mla": |
| 741 | cache = self.kv_cache |
| 742 | if cache.dtype != torch.uint8: |
| 743 | cache = cache.view(torch.uint8) |
| 744 | return fused_mla_decode_q_concat_kv_cache_insert( |
| 745 | ql_nope, |
| 746 | q_pe, |
| 747 | kv_c_normed, |
| 748 | k_pe, |
| 749 | cache, |
| 750 | slot_mapping, |
| 751 | ds_mla=True, |
| 752 | positions=positions, |
| 753 | cos_sin_cache=cos_sin_cache, |
| 754 | ) |
| 755 | if is_quantized_kv_cache(self.kv_cache_dtype): |
| 756 | cache = self.kv_cache |
| 757 | if cache.dtype != torch.float8_e4m3fn: |
| 758 | cache = cache.view(torch.float8_e4m3fn) |
| 759 | mqa_q = fused_mla_decode_q_concat_kv_cache_insert( |
| 760 | ql_nope, |
| 761 | q_pe, |
| 762 | kv_c_normed, |
| 763 | k_pe, |
| 764 | cache, |
| 765 | slot_mapping, |
| 766 | q_scale_inv=self._q_scale_inv, |
| 767 | cache_scale_inv=self._k_scale_inv, |
| 768 | positions=positions, |
| 769 | cos_sin_cache=cos_sin_cache, |
| 770 | ) |
| 771 | if not self.impl.supports_quant_query_input: # type: ignore[attr-defined] |
| 772 | # Backend dequantizes fp8 KV on load and takes a bf16 query |
| 773 | # (e.g. TRITON_MLA, the DSpark draft); undo the query |
| 774 | # quantization. |
| 775 | mqa_q = (mqa_q.to(torch.float32) * self._q_scale).to(ql_nope.dtype) |
| 776 | return mqa_q |
| 777 | return fused_mla_decode_q_concat_kv_cache_insert( |
| 778 | ql_nope, |
| 779 | q_pe, |
| 780 | kv_c_normed, |
| 781 | k_pe, |
| 782 | self.kv_cache, |
| 783 | slot_mapping, |
| 784 | positions=positions, |
| 785 | cos_sin_cache=cos_sin_cache, |
| 786 | ) |
| 787 | |
| 788 | def _compute_prefill_context( |
| 789 | self, |
| 790 | q: torch.Tensor, |
| 791 | attn_metadata: "MLACommonMetadata", |
| 792 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 793 | """Chunked-context prefill, K3-fused. Replaces the impl's version. |
| 794 | |
| 795 | Per chunk the impl gathers the paged latent, up-projects it, then casts |
| 796 | and concatenates K (and casts V) in two or three more launches. Here that |
| 797 | tail is one fused kernel per chunk -- ``fused_mla_kv_concat`` for a bf16 |
| 798 | query, ``fused_mla_kv_concat_quant_fp8`` when the query is fp8 -- reading |
| 799 | the strided ``kv_b_proj`` output in place and writing a contiguous key, so |
| 800 | only the gather and ``kv_b_proj`` remain. |
| 801 | |
| 802 | The impl's query cast is gone as well: ``q`` already carries |
| 803 | ``prefill.q_data_type`` because the new-token epilogue quantized it. The |
| 804 | gathered latent still gets the impl's cast to whatever ``kv_b_proj`` |
| 805 | consumes -- free (a no-op ``.to``) for a checkpoint whose ``kv_b_proj`` |
| 806 | takes the fp8 latent directly, and required for a bf16 one, which is |
| 807 | what a stock K3 checkpoint carries. Its output is bf16 either way. |
| 808 | |
| 809 | The gathered ``k_pe`` is likewise used as-is (fp8 for a plain fp8 cache) |
| 810 | and needs no RoPE: it was rotated on the way in. |
| 811 | |
| 812 | Chunk partials are written straight into the accumulating context partial |
| 813 | when the prefill backend honors ``out``, so only the (64x smaller) lse is |
| 814 | copied per chunk. |
| 815 | |
| 816 | Decode context parallelism keeps using |
| 817 | ``impl._context_parallel_compute_prefill_context``; its extra allgather |
| 818 | and reorg are not fused here. |
| 819 | """ |
| 820 | prefill = attn_metadata.prefill |
| 821 | assert prefill is not None |
| 822 | prefill_backend = prefill.prefill_backend |
| 823 | assert prefill_backend is not None |
| 824 | chunked_context = prefill.chunked_context |
| 825 | assert chunked_context is not None |
| 826 | assert q.dtype == prefill.q_data_type, ( |
| 827 | "Kimi-K3 chunked context expects the new-token epilogue to have " |
| 828 | f"produced a {prefill.q_data_type} query; got {q.dtype}." |
| 829 | ) |
| 830 | |
| 831 | fp8_prefill = q.dtype == current_platform.fp8_dtype() |
| 832 | workspace = chunked_context.workspace |
| 833 | kv_cache = self._attn_read_kv_cache() |
| 834 | kv_b_proj_input_dtype = _get_kv_b_proj_input_dtype(self.kv_b_proj, fp8_prefill) |
| 835 | |
| 836 | def run_chunk( |
| 837 | chunk, out: torch.Tensor | None = None |
| 838 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 839 | self._gather_context_latent(chunk, kv_cache, prefill, fp8_prefill) |
| 840 | gathered = workspace[: chunk.num_context_tokens] |
| 841 | kv_c_normed = gathered[..., : self.kv_lora_rank] |
| 842 | if kv_b_proj_input_dtype is not None: |
| 843 | kv_c_normed = kv_c_normed.to(kv_b_proj_input_dtype) |
| 844 | kv_nope = self.kv_b_proj(kv_c_normed)[0].view( |
| 845 | -1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim |
| 846 | ) |
| 847 | k_nope, v = kv_nope.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) |
| 848 | k_pe = gathered[..., self.kv_lora_rank :] |
| 849 | if fp8_prefill: |
| 850 | k, v = fused_mla_kv_concat_quant_fp8(k_nope, k_pe, v) |
| 851 | else: |
| 852 | k = fused_mla_kv_concat(k_nope, k_pe) |
| 853 | attn_output, attn_lse = prefill_backend.run_prefill_context_chunk( |
| 854 | chunk=chunk, q=q[chunk.token_slice], k=k, v=v, out=out |
| 855 | ) |
| 856 | assert out is None or attn_output.data_ptr() == out.data_ptr(), ( |
| 857 | f"{prefill_backend.get_name()} reports supports_out() but did not " |
| 858 | "write the context chunk into the `out` it was given." |
| 859 | ) |
| 860 | return attn_output, attn_lse |
| 861 | |
| 862 | chunks = chunked_context.chunks |
| 863 | if len(chunks) == 1 and not chunked_context.empty_token_slices: |
| 864 | # One chunk covering every prefill token: its partial *is* the context |
| 865 | # partial, so it needs neither an accumulator nor a copy. |
| 866 | return run_chunk(chunks[0]) |
| 867 | |
| 868 | # A backend honoring `out` writes each chunk's partial straight into the |
| 869 | # accumulator, so the per-chunk output copy disappears -- and because that |
| 870 | # contract fixes the trailing shape, the accumulator can be sized before |
| 871 | # any chunk runs. Otherwise the shape is only knowable from a real partial, |
| 872 | # so the first chunk runs ahead of the loop and is copied in. |
| 873 | writes_out = prefill_backend.supports_out() |
| 874 | if writes_out: |
| 875 | assert prefill.output_dtype is not None |
| 876 | output = torch.empty( |
| 877 | (q.shape[0], self.num_local_heads, self.v_head_dim), |
| 878 | dtype=prefill.output_dtype, |
| 879 | device=q.device, |
| 880 | ) |
| 881 | output_lse = torch.empty( |
| 882 | (self.num_local_heads, q.shape[0]), |
| 883 | dtype=torch.float32, |
| 884 | device=q.device, |
| 885 | ) |
| 886 | neutralize_empty_context_partials(chunked_context, output, output_lse) |
| 887 | else: |
| 888 | attn_output, attn_lse = run_chunk(chunks[0]) |
| 889 | output, output_lse = init_mla_context_partial( |
| 890 | chunked_context, attn_output, attn_lse, num_tokens=q.shape[0] |
| 891 | ) |
| 892 | accumulate_mla_context_chunk( |
| 893 | chunks[0], attn_output, attn_lse, output, output_lse |
| 894 | ) |
| 895 | chunks = chunks[1:] |
| 896 | |
| 897 | for chunk in chunks: |
| 898 | # A continuation chunk's leading tokens have to be merged with the |
| 899 | # partial already sitting there, so it cannot write in place. |
| 900 | out = ( |
| 901 | output[chunk.token_slice] |
| 902 | if writes_out and not chunk.is_continuation |
| 903 | else None |
| 904 | ) |
| 905 | attn_output, attn_lse = run_chunk(chunk, out=out) |
| 906 | accumulate_mla_context_chunk( |
| 907 | chunk, |
| 908 | attn_output, |
| 909 | attn_lse, |
| 910 | output, |
| 911 | output_lse, |
| 912 | output_written=out is not None, |
| 913 | ) |
| 914 | return output, output_lse |
| 915 | |
| 916 | def _gather_context_latent( |
| 917 | self, |
| 918 | chunk, |
| 919 | kv_cache: torch.Tensor, |
| 920 | prefill, |
| 921 | fp8_prefill: bool, |
| 922 | ) -> None: |
| 923 | """Gather one chunk's paged context latent into the workspace. |
| 924 | |
| 925 | Dispatched exactly as in ``impl._compute_prefill_context``: an fp8 query |
| 926 | reads the plain fp8 cache in its stored layout, anything else lands in the |
| 927 | workspace as the model dtype. |
| 928 | """ |
| 929 | workspace = prefill.chunked_context.workspace |
| 930 | toks = chunk.num_context_tokens |
| 931 | block_table = prefill.block_table[chunk.request_slice] |
| 932 | if self.kv_cache_dtype == "fp8_ds_mla": |
| 933 | ops.cp_gather_and_upconvert_fp8_kv_cache( |
| 934 | src_cache=kv_cache, |
| 935 | dst=workspace[:toks], |
| 936 | block_table=block_table, |
| 937 | workspace_starts=chunk.cu_seq_lens, |
| 938 | batch_size=chunk.num_requests, |
| 939 | seq_starts=chunk.starts, |
| 940 | ) |
| 941 | elif not fp8_prefill: |
| 942 | ops.gather_and_maybe_dequant_cache( |
| 943 | src_cache=kv_cache, |
| 944 | dst=workspace, |
| 945 | block_table=block_table, |
| 946 | cu_seq_lens=chunk.cu_seq_lens, |
| 947 | token_to_seq=chunk.token_to_seq, |
| 948 | num_tokens=toks, |
| 949 | kv_cache_dtype=self.kv_cache_dtype, |
| 950 | scale=self._k_scale, |
| 951 | seq_starts=chunk.starts, |
| 952 | ) |
| 953 | else: |
| 954 | ops.cp_gather_cache( |
| 955 | src_cache=kv_cache, |
| 956 | dst=workspace[:toks], |
| 957 | block_table=block_table, |
| 958 | cu_seq_lens=chunk.cu_seq_lens, |
| 959 | batch_size=chunk.num_requests, |
| 960 | seq_starts=chunk.starts, |
| 961 | ) |
| 962 | |
| 963 | def _forward_prefill_fused( |
| 964 | self, |
| 965 | q: torch.Tensor, |
| 966 | kv_c_normed: torch.Tensor, |
| 967 | k_pe: torch.Tensor, |
| 968 | positions: torch.Tensor | None, |
| 969 | cos_sin_cache: torch.Tensor | None, |
| 970 | slot_mapping: torch.Tensor, |
| 971 | attn_metadata, |
| 972 | out: torch.Tensor, |
| 973 | ) -> None: |
| 974 | """Prefill using the fused key-concat + cache-insert kernel. |
| 975 | |
| 976 | Replaces ``_concat_k_nope_k_pe`` and the prefill cache write with one |
| 977 | fused kernel launch, dispatched by cache dtype. Chunked context runs |
| 978 | through this layer's ``_compute_prefill_context``, except under DCP where |
| 979 | it is delegated to the impl. |
| 980 | |
| 981 | Supported configs (K3 fp8 policy): |
| 982 | - bf16 cache -> bf16 prefill query |
| 983 | - plain fp8 cache -> fp8 prefill query (unscaled q/k/v; cache _k_scale) |
| 984 | - fp8_ds_mla cache -> bf16 prefill query (656B per-tile self-scaled) |
| 985 | """ |
| 986 | prefill = attn_metadata.prefill |
| 987 | has_context = prefill.chunked_context is not None |
| 988 | fp8_prefill = prefill.q_data_type == current_platform.fp8_dtype() |
| 989 | |
| 990 | kv_nope = self.kv_b_proj(kv_c_normed)[0].view( |
| 991 | -1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim |
| 992 | ) |
| 993 | k_nope, v = kv_nope.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) |
| 994 | |
| 995 | if self.kv_cache_dtype == "fp8_ds_mla": |
| 996 | # fp8_ds_mla cache (656B, per-tile self-scaled); bf16 attention. |
| 997 | assert not fp8_prefill, ( |
| 998 | "Kimi-K3 fp8_ds_mla uses a bf16 prefill query; fp8 prefill " |
| 999 | "query is not supported with fp8_ds_mla." |
| 1000 | ) |
| 1001 | kv_cache = self.kv_cache |
| 1002 | if kv_cache.dtype != torch.uint8: |
| 1003 | kv_cache = kv_cache.view(torch.uint8) |
| 1004 | k = fused_mla_key_concat_ds_mla_insert( |
| 1005 | q, |
| 1006 | k_nope, |
| 1007 | k_pe, |
| 1008 | kv_c_normed, |
| 1009 | kv_cache, |
| 1010 | slot_mapping, |
| 1011 | positions, |
| 1012 | cos_sin_cache, |
| 1013 | ) |
| 1014 | elif is_quantized_kv_cache(self.kv_cache_dtype): |
| 1015 | assert fp8_prefill, ( |
| 1016 | "Kimi-K3 fp8 KV cache requires an fp8 prefill query; enable " |
| 1017 | "--attention-config '{\"use_prefill_query_quantization\": true}'." |
| 1018 | ) |
| 1019 | # Plain per-tensor fp8: quant q/k/v (unscaled, matching forward_mha's |
| 1020 | # unscaled `.to(fp8)`) and insert the fp8 latent (scaled by _k_scale). |
| 1021 | kv_cache = self.kv_cache |
| 1022 | if kv_cache.dtype != torch.float8_e4m3fn: |
| 1023 | kv_cache = kv_cache.view(torch.float8_e4m3fn) |
| 1024 | q, k, v = fused_mla_qkv_quant_kv_cache_fp8_insert( |
| 1025 | q, |
| 1026 | k_nope, |
| 1027 | k_pe, |
| 1028 | kv_c_normed, |
| 1029 | v, |
| 1030 | kv_cache, |
| 1031 | slot_mapping, |
| 1032 | self._one_scale, |
| 1033 | self._one_scale, |
| 1034 | self._one_scale, |
| 1035 | self._k_scale_inv, |
| 1036 | positions, |
| 1037 | cos_sin_cache, |
| 1038 | ) |
| 1039 | else: |
| 1040 | # Concat full K = [k_nope | k_pe] and insert [kv_c_normed | k_pe] |
| 1041 | # into the paged cache for these prefill tokens, in one launch. |
| 1042 | k = fused_mla_key_concat_kv_cache_insert( |
| 1043 | q, |
| 1044 | k_nope, |
| 1045 | k_pe, |
| 1046 | kv_c_normed, |
| 1047 | self.kv_cache, |
| 1048 | slot_mapping, |
| 1049 | positions, |
| 1050 | cos_sin_cache, |
| 1051 | ) |
| 1052 | |
| 1053 | # When there is no chunked context, backends that honor `out` write the |
| 1054 | # attention result straight into it, avoiding a slice+flatten+copy. |
| 1055 | writes_out = not has_context and prefill.prefill_backend.supports_out() |
| 1056 | output_prefill = prefill.prefill_backend.run_prefill_new_tokens( |
| 1057 | q=q, |
| 1058 | k=k, |
| 1059 | v=v, |
| 1060 | return_softmax_lse=has_context, |
| 1061 | out=( |
| 1062 | out.view(-1, self.num_local_heads, self.v_head_dim) |
| 1063 | if writes_out |
| 1064 | else None |
| 1065 | ), |
| 1066 | ) |
| 1067 | |
| 1068 | if has_context: |
| 1069 | if self.dcp_world_size > 1: |
| 1070 | context_output, context_lse = ( |
| 1071 | self.impl._context_parallel_compute_prefill_context( # type: ignore[attr-defined] |
| 1072 | q, |
| 1073 | self._attn_read_kv_cache(), |
| 1074 | attn_metadata, |
| 1075 | k_scale=self._k_scale, |
| 1076 | dcp_world_size=self.dcp_world_size, |
| 1077 | ) |
| 1078 | ) |
| 1079 | else: |
| 1080 | context_output, context_lse = self._compute_prefill_context( |
| 1081 | q, attn_metadata |
| 1082 | ) |
| 1083 | suffix_output, suffix_lse = output_prefill |
| 1084 | out = out.view(-1, self.num_local_heads, self.v_head_dim) |
| 1085 | merge_attn_states( |
| 1086 | output=out, |
| 1087 | prefix_output=context_output[..., : self.v_head_dim], |
| 1088 | prefix_lse=context_lse, |
| 1089 | suffix_output=suffix_output[..., : self.v_head_dim], |
| 1090 | suffix_lse=suffix_lse, |
| 1091 | ) |
| 1092 | elif not writes_out: |
| 1093 | out.copy_(output_prefill[..., : self.v_head_dim].flatten(start_dim=-2)) |