| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2026 MBZUAI-IFM. |
| 5 | # Copyright 2024 The Qwen team. |
| 6 | # Copyright 2023 The vLLM team. |
| 7 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 8 | # |
| 9 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 10 | # and OPT implementations in this library. It has been modified from its |
| 11 | # original forms to accommodate minor architectural differences compared |
| 12 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 13 | # |
| 14 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 15 | # you may not use this file except in compliance with the License. |
| 16 | # You may obtain a copy of the License at |
| 17 | # |
| 18 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 19 | # |
| 20 | # Unless required by applicable law or agreed to in writing, software |
| 21 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 22 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 23 | # See the License for the specific language governing permissions and |
| 24 | # limitations under the License. |
| 25 | """Inference-only K2Horizon model compatible with HuggingFace weights.""" |
| 26 | |
| 27 | import math |
| 28 | import typing |
| 29 | from collections.abc import Callable, Iterable |
| 30 | from itertools import islice |
| 31 | from typing import Any |
| 32 | |
| 33 | import torch |
| 34 | import torch.nn.functional as F |
| 35 | from torch import nn |
| 36 | |
| 37 | from vllm import _custom_ops as ops |
| 38 | from vllm.compilation.decorators import support_torch_compile |
| 39 | from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config |
| 40 | from vllm.distributed import ( |
| 41 | get_ep_group, |
| 42 | get_pp_group, |
| 43 | get_tensor_model_parallel_rank, |
| 44 | get_tensor_model_parallel_world_size, |
| 45 | tensor_model_parallel_all_gather, |
| 46 | ) |
| 47 | from vllm.logger import init_logger |
| 48 | from vllm.model_executor.layers.activation import SiluAndMul |
| 49 | from vllm.model_executor.layers.attention import Attention |
| 50 | from vllm.model_executor.layers.fused_moe import ( |
| 51 | FusedMoEFactory, |
| 52 | GateLinear, |
| 53 | fused_moe_make_expert_params_mapping, |
| 54 | ) |
| 55 | from vllm.model_executor.layers.fused_moe.config import ( |
| 56 | _get_config_dtype_str, |
| 57 | ) |
| 58 | from vllm.model_executor.layers.fused_moe.fused_moe import ( |
| 59 | _get_config_quant_dtype, |
| 60 | _prepare_expert_assignment, |
| 61 | dispatch_fused_moe_kernel, |
| 62 | try_get_optimal_moe_config, |
| 63 | ) |
| 64 | from vllm.model_executor.layers.fused_moe.utils import ( |
| 65 | moe_kernel_quantize_input, |
| 66 | ) |
| 67 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 68 | from vllm.model_executor.layers.linear import ( |
| 69 | ColumnParallelLinear, |
| 70 | MergedColumnParallelLinear, |
| 71 | QKVParallelLinear, |
| 72 | ReplicatedLinear, |
| 73 | RowParallelLinear, |
| 74 | UnquantizedLinearMethod, |
| 75 | ) |
| 76 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 77 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 78 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 79 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 80 | ParallelLMHead, |
| 81 | VocabParallelEmbedding, |
| 82 | ) |
| 83 | from vllm.model_executor.model_loader.weight_utils import ( |
| 84 | default_weight_loader, |
| 85 | maybe_remap_kv_scale_name, |
| 86 | ) |
| 87 | from vllm.model_executor.models.interfaces import ( |
| 88 | EagleModelMixin, |
| 89 | MixtureOfExperts, |
| 90 | SupportsEagle, |
| 91 | SupportsEagle3, |
| 92 | SupportsLoRA, |
| 93 | SupportsPP, |
| 94 | ) |
| 95 | from vllm.model_executor.models.utils import ( |
| 96 | AutoWeightsLoader, |
| 97 | PPMissingLayer, |
| 98 | extract_layer_index, |
| 99 | is_pp_missing_parameter, |
| 100 | make_empty_intermediate_tensors_factory, |
| 101 | make_layers, |
| 102 | maybe_prefix, |
| 103 | sequence_parallel_chunk, |
| 104 | ) |
| 105 | from vllm.sequence import IntermediateTensors |
| 106 | from vllm.triton_utils import tl |
| 107 | |
| 108 | logger = init_logger(__name__) |
| 109 | |
| 110 | |
| 111 | def split_to_interleaved(x): |
| 112 | # Split halves: x0 x1 x2 x3 ... y0 y1 y2 y3 ... |
| 113 | # Interleaved: x0 y0 x1 y1 x2 y2 x3 y3 ... |
| 114 | return x.reshape(*x.shape[:-1], 2, -1).transpose(-1, -2).reshape(*x.shape[:-1], -1) |
| 115 | |
| 116 | |
| 117 | def interleaved_to_split(x): |
| 118 | # Interleaved: x0 y0 x1 y1 x2 y2 x3 y3 ... |
| 119 | # Split halves: x0 x1 x2 x3 ... y0 y1 y2 y3 ... |
| 120 | return x.reshape(*x.shape[:-1], -1, 2).transpose(-1, -2).reshape(*x.shape[:-1], -1) |
| 121 | |
| 122 | |
| 123 | def is_rope_weights_folding_supported( |
| 124 | qk_proj: nn.Module, |
| 125 | dual_chunk_attention_config: dict[str, Any] | None, |
| 126 | ) -> bool: |
| 127 | """Whether the partial-RoPE weight-fold path is safe for this attention. |
| 128 | |
| 129 | Folding permutes the q/k projection's output channels at load time, which |
| 130 | only produces correct numerics when: |
| 131 | |
| 132 | * the q/k projection is dense/unquantized and |
| 133 | * dual-chunk attention is disabled -- the P.RoPE equivalence has not been |
| 134 | validated with dual-chunk attention. |
| 135 | |
| 136 | When either condition fails we fall back to the runtime channel-permutation |
| 137 | RoPE path instead of folding. |
| 138 | """ |
| 139 | is_dense_qk_proj = isinstance( |
| 140 | getattr(qk_proj, "quant_method", None), UnquantizedLinearMethod |
| 141 | ) |
| 142 | return is_dense_qk_proj and dual_chunk_attention_config is None |
| 143 | |
| 144 | |
| 145 | def apply_partial_rope( |
| 146 | rotary_emb, |
| 147 | positions: torch.Tensor, |
| 148 | q: torch.Tensor, |
| 149 | k: torch.Tensor, |
| 150 | num_heads: int, |
| 151 | num_kv_heads: int, |
| 152 | head_dim: int, |
| 153 | rope_head_dim: int, |
| 154 | fold_rope_weights: bool, |
| 155 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 156 | if fold_rope_weights or rope_head_dim == head_dim: |
| 157 | return rotary_emb(positions, q, k) |
| 158 | |
| 159 | q = q.reshape(*q.shape[:-1], num_heads, head_dim) |
| 160 | k = k.reshape(*k.shape[:-1], num_kv_heads, head_dim) |
| 161 | |
| 162 | q_rope, q_nope = torch.split( |
| 163 | split_to_interleaved(q), |
| 164 | split_size_or_sections=[rope_head_dim, head_dim - rope_head_dim], |
| 165 | dim=-1, |
| 166 | ) |
| 167 | k_rope, k_nope = torch.split( |
| 168 | split_to_interleaved(k), |
| 169 | split_size_or_sections=[rope_head_dim, head_dim - rope_head_dim], |
| 170 | dim=-1, |
| 171 | ) |
| 172 | |
| 173 | q_rope, k_rope = rotary_emb( |
| 174 | positions, interleaved_to_split(q_rope), interleaved_to_split(k_rope) |
| 175 | ) |
| 176 | |
| 177 | q = interleaved_to_split( |
| 178 | torch.cat([split_to_interleaved(q_rope), q_nope], dim=-1) |
| 179 | ).reshape(*q.shape[:-2], -1) |
| 180 | k = interleaved_to_split( |
| 181 | torch.cat([split_to_interleaved(k_rope), k_nope], dim=-1) |
| 182 | ).reshape(*k.shape[:-2], -1) |
| 183 | |
| 184 | return q, k |
| 185 | |
| 186 | |
| 187 | def _rope_weight_perm(head_dim: int, rope_head_dim: int) -> torch.Tensor: |
| 188 | """Per-head gather index that folds K2Horizon's partial-RoPE channel reordering. |
| 189 | |
| 190 | The K2Horizon checkpoint stores each head's channels in a GPT-J-interleaved |
| 191 | convention with rope/nope channels interleaved. Applying this fixed |
| 192 | (position-independent) permutation to the q/k weight rows lets the runtime |
| 193 | use vLLM's *native* NeoX partial-RoPE path (``head_size = head_dim``, |
| 194 | ``rotary_dim = rope_head_dim``) with no per-forward permutes/all-gather. |
| 195 | """ |
| 196 | D, R = head_dim, rope_head_dim |
| 197 | assert R % 2 == 0 and R <= D, ( |
| 198 | f"partial NeoX RoPE requires an even rope_head_dim <= head_dim, " |
| 199 | f"got R={R}, D={D}" |
| 200 | ) |
| 201 | h = D // 2 |
| 202 | idx = torch.cat( |
| 203 | [ |
| 204 | torch.arange(0, R // 2), # rope first-half channels |
| 205 | torch.arange(h, h + R // 2), # rope second-half channels |
| 206 | torch.arange(R // 2, h), # nope remainder (first half) |
| 207 | torch.arange(h + R // 2, D), # nope remainder (second half) |
| 208 | ] |
| 209 | ) |
| 210 | return idx |
| 211 | |
| 212 | |
| 213 | def calc_router_weights( |
| 214 | router_logits: torch.Tensor, |
| 215 | e_score_correction_bias: torch.Tensor | None, |
| 216 | score_func: str, |
| 217 | top_k: int, |
| 218 | scaling_factor: float | None, |
| 219 | renormalize: bool | None = None, |
| 220 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 221 | if score_func == "softmax": |
| 222 | routing_scores = F.softmax(router_logits, dim=-1, dtype=torch.float32) |
| 223 | else: |
| 224 | assert score_func == "sigmoid", ( |
| 225 | f"Unsupported router score function: {score_func}" |
| 226 | ) |
| 227 | routing_scores = torch.sigmoid(router_logits.to(torch.float32)) |
| 228 | |
| 229 | selection_scores = routing_scores |
| 230 | if e_score_correction_bias is not None: |
| 231 | selection_scores = selection_scores + e_score_correction_bias.to( |
| 232 | selection_scores |
| 233 | ) |
| 234 | |
| 235 | selected_indices = torch.topk(selection_scores, top_k, dim=-1).indices |
| 236 | routing_weights = torch.gather(routing_scores, dim=-1, index=selected_indices) |
| 237 | if renormalize is None: |
| 238 | renormalize = top_k > 1 |
| 239 | if renormalize: |
| 240 | routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True) |
| 241 | if scaling_factor is not None: |
| 242 | routing_weights = routing_weights * scaling_factor |
| 243 | return routing_weights.to(router_logits.dtype), selected_indices |
| 244 | |
| 245 | |
| 246 | def fused_mova_impl( |
| 247 | config, |
| 248 | hidden_states: torch.Tensor, |
| 249 | w1: torch.Tensor, |
| 250 | topk_weights: torch.Tensor, |
| 251 | topk_ids: torch.Tensor, |
| 252 | use_fp8_w8a8: bool = False, |
| 253 | use_int8_w8a8: bool = False, |
| 254 | use_int8_w8a16: bool = False, |
| 255 | use_int4_w4a16: bool = False, |
| 256 | ocp_mx_scheme: str | None = None, |
| 257 | per_channel_quant: bool = False, |
| 258 | global_num_experts: int = -1, |
| 259 | expert_map: torch.Tensor | None = None, |
| 260 | w1_scale: torch.Tensor | None = None, |
| 261 | w1_zp: torch.Tensor | None = None, |
| 262 | a1_scale: torch.Tensor | None = None, |
| 263 | block_shape: list[int] | None = None, |
| 264 | w1_bias: torch.Tensor | None = None, |
| 265 | ) -> torch.Tensor: |
| 266 | if ocp_mx_scheme is not None: |
| 267 | raise NotImplementedError( |
| 268 | f"Using ocp_mx_scheme={ocp_mx_scheme} in functional fused_experts call is " |
| 269 | "deprecated. Please use OCP_MXQuantizationEmulationTritonExperts." |
| 270 | ) |
| 271 | |
| 272 | # Check constraints. |
| 273 | if use_int4_w4a16: |
| 274 | assert hidden_states.size(1) // 2 == w1.size(2), "Hidden size mismatch" |
| 275 | else: |
| 276 | assert hidden_states.size(1) == w1.size(2), ( |
| 277 | f"Hidden size mismatch {hidden_states.size(1)} != {w1.size(2)}" |
| 278 | ) |
| 279 | |
| 280 | assert topk_weights.size() == topk_ids.size(), "topk shape mismatch" |
| 281 | assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" |
| 282 | assert w1.stride(-1) == 1, "Stride of last dimension must be 1" |
| 283 | assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16] |
| 284 | |
| 285 | num_tokens = hidden_states.size(0) |
| 286 | E, N, K = w1.size() |
| 287 | |
| 288 | if global_num_experts == -1: |
| 289 | global_num_experts = E |
| 290 | top_k_num = topk_ids.size(1) |
| 291 | M = num_tokens |
| 292 | |
| 293 | # Note: for use_int8_w8a16 or use_int4_w4a16, the activations are |
| 294 | # quantized prior to calling fused_experts. |
| 295 | quant_dtype = _get_config_quant_dtype( |
| 296 | use_fp8_w8a8=use_fp8_w8a8, |
| 297 | use_int8_w8a8=use_int8_w8a8, |
| 298 | ) |
| 299 | |
| 300 | intermediate_cache1 = torch.empty( |
| 301 | (M, top_k_num, N), device=hidden_states.device, dtype=hidden_states.dtype |
| 302 | ) |
| 303 | |
| 304 | if hidden_states.dtype == torch.bfloat16: |
| 305 | compute_type = tl.bfloat16 |
| 306 | elif hidden_states.dtype == torch.float16: |
| 307 | compute_type = tl.float16 |
| 308 | elif hidden_states.dtype == torch.float32: |
| 309 | compute_type = tl.float32 |
| 310 | else: |
| 311 | raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}") |
| 312 | |
| 313 | out_hidden_states = torch.empty( |
| 314 | (M, N), |
| 315 | device=hidden_states.device, |
| 316 | dtype=hidden_states.dtype, |
| 317 | ) |
| 318 | |
| 319 | qhidden_states, a1q_scale = moe_kernel_quantize_input( |
| 320 | A=hidden_states, |
| 321 | A_scale=a1_scale, |
| 322 | quant_dtype=quant_dtype, |
| 323 | per_act_token_quant=per_channel_quant, |
| 324 | block_shape=block_shape, |
| 325 | ) |
| 326 | |
| 327 | sorted_token_ids, expert_ids, num_tokens_post_padded = _prepare_expert_assignment( |
| 328 | topk_ids, |
| 329 | config, |
| 330 | num_tokens, |
| 331 | top_k_num, |
| 332 | global_num_experts, |
| 333 | expert_map, |
| 334 | use_int8_w8a16=use_int8_w8a16, |
| 335 | use_int4_w4a16=use_int4_w4a16, |
| 336 | block_shape=block_shape, |
| 337 | ignore_invalid_experts=True, |
| 338 | ) |
| 339 | |
| 340 | dispatch_fused_moe_kernel( |
| 341 | qhidden_states, |
| 342 | w1, |
| 343 | intermediate_cache1, |
| 344 | a1q_scale, |
| 345 | w1_scale, |
| 346 | w1_zp, |
| 347 | topk_weights, |
| 348 | sorted_token_ids, |
| 349 | expert_ids, |
| 350 | num_tokens_post_padded, |
| 351 | False, # apply_router_weight_on_input, |
| 352 | top_k_num, |
| 353 | config, |
| 354 | compute_type=compute_type, |
| 355 | use_fp8_w8a8=use_fp8_w8a8, |
| 356 | use_int8_w8a8=use_int8_w8a8, |
| 357 | use_int8_w8a16=use_int8_w8a16, |
| 358 | use_int4_w4a16=use_int4_w4a16, |
| 359 | per_channel_quant=per_channel_quant, |
| 360 | block_shape=block_shape, |
| 361 | B_bias=w1_bias, |
| 362 | ) |
| 363 | |
| 364 | intermediate_cache1 = F.silu(intermediate_cache1) |
| 365 | intermediate_cache1 *= topk_weights[:, :, None].to(intermediate_cache1.dtype) |
| 366 | |
| 367 | ops.moe_sum(intermediate_cache1, out_hidden_states) |
| 368 | |
| 369 | return out_hidden_states |
| 370 | |
| 371 | |
| 372 | class K2HorizonRMSNorm(RMSNorm): |
| 373 | def __init__(self, hidden_size: int, n_groups: int, eps: float = 1e-6): |
| 374 | assert hidden_size % n_groups == 0 |
| 375 | super().__init__(hidden_size, eps=eps) |
| 376 | self.hidden_size = hidden_size |
| 377 | self.n_groups = n_groups |
| 378 | self.group_size = hidden_size // n_groups |
| 379 | |
| 380 | def forward(self, x, residual=None): |
| 381 | if self.n_groups == 1: |
| 382 | return super().forward(x, residual) |
| 383 | |
| 384 | if residual is not None: |
| 385 | x = x + residual |
| 386 | residual = x |
| 387 | |
| 388 | orig_shape = x.shape |
| 389 | x_grouped = x.reshape(-1, self.n_groups, self.group_size).reshape( |
| 390 | -1, self.group_size |
| 391 | ) |
| 392 | y_grouped = torch.empty_like(x_grouped) |
| 393 | |
| 394 | # Reuse vLLM's RMSNorm CUDA kernel on each group. |
| 395 | ops.rms_norm( |
| 396 | y_grouped, |
| 397 | x_grouped, |
| 398 | torch.ones(self.group_size, dtype=x.dtype, device=x.device), |
| 399 | self.variance_epsilon, |
| 400 | ) |
| 401 | |
| 402 | y = y_grouped.reshape(orig_shape) * self.weight |
| 403 | return (y, residual) if residual is not None else y |
| 404 | |
| 405 | |
| 406 | class K2HorizonMLP(nn.Module): |
| 407 | def __init__( |
| 408 | self, |
| 409 | hidden_size: int, |
| 410 | intermediate_size: int, |
| 411 | hidden_act: str, |
| 412 | quant_config: QuantizationConfig | None = None, |
| 413 | reduce_results: bool = True, |
| 414 | expert_gate: torch.nn.Linear | None = None, |
| 415 | prefix: str = "", |
| 416 | ) -> None: |
| 417 | super().__init__() |
| 418 | self.gate_up_proj = MergedColumnParallelLinear( |
| 419 | hidden_size, |
| 420 | [intermediate_size] * 2, |
| 421 | bias=False, |
| 422 | quant_config=quant_config, |
| 423 | prefix=f"{prefix}.gate_up_proj", |
| 424 | ) |
| 425 | self.down_proj = RowParallelLinear( |
| 426 | intermediate_size, |
| 427 | hidden_size, |
| 428 | bias=False, |
| 429 | quant_config=quant_config, |
| 430 | reduce_results=reduce_results, |
| 431 | prefix=f"{prefix}.down_proj", |
| 432 | ) |
| 433 | if hidden_act != "silu": |
| 434 | raise ValueError( |
| 435 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 436 | ) |
| 437 | self.act_fn = SiluAndMul() |
| 438 | self.expert_gate = expert_gate |
| 439 | |
| 440 | def forward(self, x): |
| 441 | gate_up, _ = self.gate_up_proj(x) |
| 442 | out = self.act_fn(gate_up) |
| 443 | out, _ = self.down_proj(out) |
| 444 | |
| 445 | if self.expert_gate is not None: |
| 446 | out = F.sigmoid(self.expert_gate(x)[0]) * out |
| 447 | |
| 448 | return out |
| 449 | |
| 450 | |
| 451 | class K2HorizonSparseMoeBlock(nn.Module): |
| 452 | def __init__( |
| 453 | self, |
| 454 | vllm_config: VllmConfig, |
| 455 | prefix: str = "", |
| 456 | ): |
| 457 | super().__init__() |
| 458 | |
| 459 | config = vllm_config.model_config.hf_text_config |
| 460 | parallel_config = vllm_config.parallel_config |
| 461 | quant_config = vllm_config.quant_config |
| 462 | |
| 463 | self.tp_size = get_tensor_model_parallel_world_size() |
| 464 | |
| 465 | self.ep_group = get_ep_group().device_group |
| 466 | self.ep_rank = get_ep_group().rank_in_group |
| 467 | self.ep_size = self.ep_group.size() |
| 468 | self.n_routed_experts = config.num_experts |
| 469 | |
| 470 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 471 | |
| 472 | if self.tp_size > config.num_experts: |
| 473 | raise ValueError( |
| 474 | f"Tensor parallel size {self.tp_size} is greater than " |
| 475 | f"the number of experts {config.num_experts}." |
| 476 | ) |
| 477 | |
| 478 | # Load balancing settings. |
| 479 | vllm_config = get_current_vllm_config() |
| 480 | eplb_config = vllm_config.parallel_config.eplb_config |
| 481 | self.enable_eplb = parallel_config.enable_eplb |
| 482 | |
| 483 | self.n_logical_experts = self.n_routed_experts |
| 484 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 485 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 486 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 487 | |
| 488 | self.physical_expert_start = self.ep_rank * self.n_local_physical_experts |
| 489 | self.physical_expert_end = ( |
| 490 | self.physical_expert_start + self.n_local_physical_experts |
| 491 | ) |
| 492 | |
| 493 | self.gate = GateLinear( |
| 494 | config.hidden_size, |
| 495 | config.num_experts, |
| 496 | bias=config.moe_gate_bias, |
| 497 | skip_bias_add=True, |
| 498 | quant_config=quant_config, |
| 499 | prefix=f"{prefix}.gate", |
| 500 | ) |
| 501 | if self.gate.bias is not None: |
| 502 | self.gate.bias = nn.Parameter(self.gate.bias.float(), requires_grad=False) |
| 503 | |
| 504 | self.num_shared_experts = config.num_shared_experts |
| 505 | self.shared_experts: K2HorizonMLP | None |
| 506 | if config.num_shared_experts > 0: |
| 507 | self.shared_experts = K2HorizonMLP( |
| 508 | hidden_size=config.hidden_size, |
| 509 | intermediate_size=config.moe_intermediate_size |
| 510 | * config.num_shared_experts, |
| 511 | hidden_act=config.hidden_act, |
| 512 | quant_config=quant_config, |
| 513 | reduce_results=False, |
| 514 | expert_gate=None, |
| 515 | prefix=f"{prefix}.shared_experts", |
| 516 | ) |
| 517 | else: |
| 518 | self.shared_experts = None |
| 519 | |
| 520 | self.experts = FusedMoEFactory( |
| 521 | shared_experts=self.shared_experts, |
| 522 | gate=self.gate, |
| 523 | num_experts=self.n_routed_experts, |
| 524 | top_k=config.num_experts_per_tok, |
| 525 | hidden_size=config.hidden_size, |
| 526 | intermediate_size=config.moe_intermediate_size, |
| 527 | renormalize=config.norm_topk_prob, |
| 528 | quant_config=quant_config, |
| 529 | prefix=f"{prefix}.experts", |
| 530 | enable_eplb=self.enable_eplb, |
| 531 | num_redundant_experts=self.n_redundant_experts, |
| 532 | is_sequence_parallel=self.is_sequence_parallel, |
| 533 | scoring_func=config.router_score_func, |
| 534 | e_score_correction_bias=self.gate.bias, |
| 535 | routed_scaling_factor=config.router_scaling_factor, |
| 536 | ) |
| 537 | |
| 538 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 539 | assert hidden_states.dim() <= 2, ( |
| 540 | "K2HorizonSparseMoeBlock only supports 1D or 2D inputs" |
| 541 | ) |
| 542 | is_input_1d = hidden_states.dim() == 1 |
| 543 | num_tokens, hidden_dim = hidden_states.shape |
| 544 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 545 | |
| 546 | if self.is_sequence_parallel: |
| 547 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 548 | |
| 549 | final_hidden_states = self.experts( |
| 550 | hidden_states=hidden_states, router_logits=hidden_states |
| 551 | ) |
| 552 | |
| 553 | if self.is_sequence_parallel: |
| 554 | final_hidden_states = tensor_model_parallel_all_gather( |
| 555 | final_hidden_states, 0 |
| 556 | ) |
| 557 | final_hidden_states = final_hidden_states[:num_tokens] |
| 558 | |
| 559 | # return to 1d if input is 1d |
| 560 | return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states |
| 561 | |
| 562 | |
| 563 | class K2HorizonAttention(nn.Module): |
| 564 | def __init__( |
| 565 | self, |
| 566 | hidden_size: int, |
| 567 | num_heads: int, |
| 568 | num_kv_heads: int, |
| 569 | rope_parameters: dict[str, Any], |
| 570 | rope_head_dim: int, |
| 571 | query_key_norm: bool, |
| 572 | max_position_embeddings: int = 8192, |
| 573 | head_dim: int | None = None, |
| 574 | rms_norm_eps: float = 1e-06, |
| 575 | qkv_bias: bool = False, |
| 576 | cache_config: CacheConfig | None = None, |
| 577 | quant_config: QuantizationConfig | None = None, |
| 578 | prefix: str = "", |
| 579 | dual_chunk_attention_config: dict[str, Any] | None = None, |
| 580 | gate_func: str | None = None, |
| 581 | ) -> None: |
| 582 | super().__init__() |
| 583 | self.hidden_size = hidden_size |
| 584 | tp_size = get_tensor_model_parallel_world_size() |
| 585 | self.total_num_heads = num_heads |
| 586 | assert self.total_num_heads % tp_size == 0 |
| 587 | self.num_heads = self.total_num_heads // tp_size |
| 588 | self.total_num_kv_heads = num_kv_heads |
| 589 | if self.total_num_kv_heads >= tp_size: |
| 590 | # Number of KV heads is greater than TP size, so we partition |
| 591 | # the KV heads across multiple tensor parallel GPUs. |
| 592 | assert self.total_num_kv_heads % tp_size == 0 |
| 593 | else: |
| 594 | # Number of KV heads is less than TP size, so we replicate |
| 595 | # the KV heads across multiple tensor parallel GPUs. |
| 596 | assert tp_size % self.total_num_kv_heads == 0 |
| 597 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 598 | self.head_dim = head_dim or (hidden_size // self.total_num_heads) |
| 599 | self.q_size = self.num_heads * self.head_dim |
| 600 | self.kv_size = self.num_kv_heads * self.head_dim |
| 601 | self.scaling = self.head_dim**-0.5 |
| 602 | self.max_position_embeddings = max_position_embeddings |
| 603 | self.dual_chunk_attention_config = dual_chunk_attention_config |
| 604 | |
| 605 | self.qkv_proj = QKVParallelLinear( |
| 606 | hidden_size, |
| 607 | self.head_dim, |
| 608 | self.total_num_heads, |
| 609 | self.total_num_kv_heads, |
| 610 | bias=qkv_bias, |
| 611 | quant_config=quant_config, |
| 612 | prefix=f"{prefix}.qkv_proj", |
| 613 | ) |
| 614 | |
| 615 | self.o_proj = RowParallelLinear( |
| 616 | self.total_num_heads * self.head_dim, |
| 617 | hidden_size, |
| 618 | bias=qkv_bias, |
| 619 | quant_config=quant_config, |
| 620 | prefix=f"{prefix}.o_proj", |
| 621 | ) |
| 622 | |
| 623 | self.rope_head_dim = rope_head_dim or self.head_dim |
| 624 | # Fold the partial-RoPE channel permutation into the q/k weights only |
| 625 | # when supported (dense projections + no dual-chunk attention); |
| 626 | # otherwise keep the runtime channel-permutation path |
| 627 | # (see ``is_rope_weights_folding_supported``). |
| 628 | self.fold_rope_weights = is_rope_weights_folding_supported( |
| 629 | self.qkv_proj, dual_chunk_attention_config |
| 630 | ) |
| 631 | if self.fold_rope_weights: |
| 632 | rope_parameters = dict(rope_parameters) |
| 633 | rope_parameters["rope_dim"] = self.rope_head_dim |
| 634 | self.rotary_emb = get_rope( |
| 635 | self.head_dim, |
| 636 | max_position=max_position_embeddings, |
| 637 | is_neox_style=True, |
| 638 | rope_parameters=rope_parameters, |
| 639 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 640 | ) |
| 641 | else: |
| 642 | self.rotary_emb = get_rope( |
| 643 | self.rope_head_dim, |
| 644 | max_position=max_position_embeddings, |
| 645 | rope_parameters=rope_parameters, |
| 646 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 647 | ) |
| 648 | attn_kwargs: dict[str, Any] = ( |
| 649 | { |
| 650 | "layer_idx": extract_layer_index(prefix), |
| 651 | "dual_chunk_attention_config": dual_chunk_attention_config, |
| 652 | } |
| 653 | if dual_chunk_attention_config |
| 654 | else {} |
| 655 | ) |
| 656 | self.attn = Attention( |
| 657 | self.num_heads, |
| 658 | self.head_dim, |
| 659 | self.scaling, |
| 660 | num_kv_heads=self.num_kv_heads, |
| 661 | cache_config=cache_config, |
| 662 | quant_config=quant_config, |
| 663 | prefix=f"{prefix}.attn", |
| 664 | **attn_kwargs, |
| 665 | ) |
| 666 | |
| 667 | self.query_key_norm = query_key_norm |
| 668 | if self.query_key_norm: |
| 669 | self.q_norm = K2HorizonRMSNorm( |
| 670 | hidden_size=self.q_size, n_groups=self.num_heads, eps=rms_norm_eps |
| 671 | ) |
| 672 | self.k_norm = K2HorizonRMSNorm( |
| 673 | hidden_size=self.kv_size, n_groups=self.num_kv_heads, eps=rms_norm_eps |
| 674 | ) |
| 675 | |
| 676 | self.gate_func = gate_func |
| 677 | if self.gate_func is not None: |
| 678 | self.gate_proj = ColumnParallelLinear( |
| 679 | hidden_size, |
| 680 | self.total_num_heads * self.head_dim, |
| 681 | bias=False, |
| 682 | quant_config=quant_config, |
| 683 | prefix=f"{prefix}.gate_proj", |
| 684 | ) |
| 685 | |
| 686 | self.tp_size = get_tensor_model_parallel_world_size() |
| 687 | self.tp_rank = get_tensor_model_parallel_rank() |
| 688 | |
| 689 | def forward( |
| 690 | self, |
| 691 | positions: torch.Tensor, |
| 692 | hidden_states: torch.Tensor, |
| 693 | ) -> torch.Tensor: |
| 694 | qkv, _ = self.qkv_proj(hidden_states) |
| 695 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 696 | if self.query_key_norm: |
| 697 | q = self.q_norm(q) |
| 698 | k = self.k_norm(k) |
| 699 | |
| 700 | q, k = apply_partial_rope( |
| 701 | self.rotary_emb, |
| 702 | positions, |
| 703 | q, |
| 704 | k, |
| 705 | self.num_heads, |
| 706 | self.num_kv_heads, |
| 707 | self.head_dim, |
| 708 | self.rope_head_dim, |
| 709 | self.fold_rope_weights, |
| 710 | ) |
| 711 | |
| 712 | attn_output = self.attn(q, k, v) |
| 713 | |
| 714 | if self.gate_func is not None: |
| 715 | gate, _ = self.gate_proj(hidden_states) |
| 716 | if self.gate_func == "silu": |
| 717 | gate = F.silu(gate) |
| 718 | else: |
| 719 | assert self.gate_func == "softplus" |
| 720 | gate = F.softplus(gate, beta=math.log(2)) |
| 721 | |
| 722 | attn_output = attn_output * gate |
| 723 | |
| 724 | output, _ = self.o_proj(attn_output) |
| 725 | return output |
| 726 | |
| 727 | |
| 728 | class K2HorizonMoVAAttention(nn.Module): |
| 729 | def __init__( |
| 730 | self, |
| 731 | vllm_config: VllmConfig, |
| 732 | hidden_size: int, |
| 733 | num_heads: int, |
| 734 | num_kv_heads: int, |
| 735 | rope_parameters: dict[str, Any], |
| 736 | rope_head_dim: int, |
| 737 | query_key_norm: bool, |
| 738 | num_experts: int, |
| 739 | num_experts_per_tok: int, |
| 740 | moe_gate_bias: bool, |
| 741 | router_score_func: str, |
| 742 | router_scaling_factor: float, |
| 743 | max_position_embeddings: int = 8192, |
| 744 | head_dim: int | None = None, |
| 745 | rms_norm_eps: float = 1e-06, |
| 746 | qkv_bias: bool = False, |
| 747 | cache_config: CacheConfig | None = None, |
| 748 | quant_config: QuantizationConfig | None = None, |
| 749 | prefix: str = "", |
| 750 | dual_chunk_attention_config: dict[str, Any] | None = None, |
| 751 | gate_func: str | None = None, |
| 752 | ) -> None: |
| 753 | super().__init__() |
| 754 | self.hidden_size = hidden_size |
| 755 | tp_size = get_tensor_model_parallel_world_size() |
| 756 | self.total_num_heads = num_heads |
| 757 | assert self.total_num_heads % tp_size == 0 |
| 758 | self.num_heads = self.total_num_heads // tp_size |
| 759 | self.total_num_kv_heads = num_kv_heads |
| 760 | if self.total_num_kv_heads >= tp_size: |
| 761 | # Number of KV heads is greater than TP size, so we partition |
| 762 | # the KV heads across multiple tensor parallel GPUs. |
| 763 | assert self.total_num_kv_heads % tp_size == 0 |
| 764 | else: |
| 765 | # Number of KV heads is less than TP size, so we replicate |
| 766 | # the KV heads across multiple tensor parallel GPUs. |
| 767 | assert tp_size % self.total_num_kv_heads == 0 |
| 768 | |
| 769 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 770 | self.head_dim = head_dim or (hidden_size // self.total_num_heads) |
| 771 | self.q_size = self.num_heads * self.head_dim |
| 772 | self.kv_size = self.num_kv_heads * self.head_dim |
| 773 | self.scaling = self.head_dim**-0.5 |
| 774 | self.max_position_embeddings = max_position_embeddings |
| 775 | self.dual_chunk_attention_config = dual_chunk_attention_config |
| 776 | |
| 777 | self.num_experts = num_experts |
| 778 | self.num_experts_per_tok = num_experts_per_tok |
| 779 | self.router_score_func = router_score_func |
| 780 | self.router_scaling_factor = router_scaling_factor |
| 781 | |
| 782 | self.qk_proj = MergedColumnParallelLinear( |
| 783 | hidden_size, |
| 784 | [self.total_num_heads * self.head_dim, self.kv_size * tp_size], |
| 785 | bias=qkv_bias, |
| 786 | quant_config=quant_config, |
| 787 | prefix=f"{prefix}.qk_proj", |
| 788 | ) |
| 789 | |
| 790 | self.v_router = ReplicatedLinear( |
| 791 | hidden_size, |
| 792 | num_experts, |
| 793 | bias=moe_gate_bias, |
| 794 | skip_bias_add=True, |
| 795 | quant_config=quant_config, |
| 796 | prefix=f"{prefix}.v_router", |
| 797 | ) |
| 798 | if self.v_router.bias is not None: |
| 799 | self.v_router.bias = nn.Parameter( |
| 800 | self.v_router.bias.float(), requires_grad=False |
| 801 | ) |
| 802 | |
| 803 | self.v_experts = nn.ModuleList( |
| 804 | [ |
| 805 | ColumnParallelLinear( |
| 806 | hidden_size, |
| 807 | self.kv_size * tp_size, |
| 808 | bias=False, |
| 809 | quant_config=quant_config, |
| 810 | prefix=f"{prefix}.v_experts.{value}", |
| 811 | ) |
| 812 | for value in range(num_experts) |
| 813 | ] |
| 814 | ) |
| 815 | |
| 816 | self.o_proj = RowParallelLinear( |
| 817 | self.total_num_heads * self.head_dim, |
| 818 | hidden_size, |
| 819 | bias=qkv_bias, |
| 820 | quant_config=quant_config, |
| 821 | prefix=f"{prefix}.o_proj", |
| 822 | ) |
| 823 | |
| 824 | self.rope_head_dim = rope_head_dim or self.head_dim |
| 825 | # Fold the partial-RoPE channel permutation into the q/k weights only |
| 826 | # when supported (dense projections + no dual-chunk attention); |
| 827 | # otherwise keep the runtime channel-permutation path |
| 828 | # (see ``is_rope_weights_folding_supported``). |
| 829 | self.fold_rope_weights = is_rope_weights_folding_supported( |
| 830 | self.qk_proj, dual_chunk_attention_config |
| 831 | ) |
| 832 | if self.fold_rope_weights: |
| 833 | rope_parameters = dict(rope_parameters) |
| 834 | rope_parameters["rope_dim"] = self.rope_head_dim |
| 835 | self.rotary_emb = get_rope( |
| 836 | self.head_dim, |
| 837 | max_position=max_position_embeddings, |
| 838 | is_neox_style=True, |
| 839 | rope_parameters=rope_parameters, |
| 840 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 841 | ) |
| 842 | else: |
| 843 | self.rotary_emb = get_rope( |
| 844 | self.rope_head_dim, |
| 845 | max_position=max_position_embeddings, |
| 846 | rope_parameters=rope_parameters, |
| 847 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 848 | ) |
| 849 | attn_kwargs: dict[str, Any] = ( |
| 850 | { |
| 851 | "layer_idx": extract_layer_index(prefix), |
| 852 | "dual_chunk_attention_config": dual_chunk_attention_config, |
| 853 | } |
| 854 | if dual_chunk_attention_config |
| 855 | else {} |
| 856 | ) |
| 857 | self.attn = Attention( |
| 858 | self.num_heads, |
| 859 | self.head_dim, |
| 860 | self.scaling, |
| 861 | num_kv_heads=self.num_kv_heads, |
| 862 | cache_config=cache_config, |
| 863 | quant_config=quant_config, |
| 864 | prefix=f"{prefix}.attn", |
| 865 | **attn_kwargs, |
| 866 | ) |
| 867 | |
| 868 | self.query_key_norm = query_key_norm |
| 869 | if self.query_key_norm: |
| 870 | self.q_norm = K2HorizonRMSNorm( |
| 871 | hidden_size=self.q_size, n_groups=self.num_heads, eps=rms_norm_eps |
| 872 | ) |
| 873 | self.k_norm = K2HorizonRMSNorm( |
| 874 | hidden_size=self.kv_size, n_groups=self.num_kv_heads, eps=rms_norm_eps |
| 875 | ) |
| 876 | |
| 877 | self.gate_func = gate_func |
| 878 | if self.gate_func is not None: |
| 879 | self.gate_proj = ColumnParallelLinear( |
| 880 | hidden_size, |
| 881 | self.total_num_heads * self.head_dim, |
| 882 | bias=False, |
| 883 | quant_config=quant_config, |
| 884 | prefix=f"{prefix}.gate_proj", |
| 885 | ) |
| 886 | |
| 887 | self.tp_size = get_tensor_model_parallel_world_size() |
| 888 | self.tp_rank = get_tensor_model_parallel_rank() |
| 889 | |
| 890 | w1_shape = torch.Size( |
| 891 | [ |
| 892 | self.num_experts, |
| 893 | self.total_num_kv_heads * self.head_dim, |
| 894 | hidden_size, |
| 895 | ] |
| 896 | ) |
| 897 | config_dtype = _get_config_dtype_str( |
| 898 | use_fp8_w8a8=False, |
| 899 | use_int8_w8a16=False, |
| 900 | use_int4_w4a16=False, |
| 901 | dtype=vllm_config.model_config.dtype, |
| 902 | ) |
| 903 | self.fused_mova_config = try_get_optimal_moe_config( |
| 904 | w1_shape=w1_shape, |
| 905 | w2_shape=torch.Size([w1_shape[0], w1_shape[2], w1_shape[1]]), |
| 906 | top_k=self.num_experts_per_tok, |
| 907 | dtype=config_dtype, |
| 908 | M=vllm_config.scheduler_config.max_num_batched_tokens, |
| 909 | block_shape=None, |
| 910 | ) |
| 911 | |
| 912 | def compute_mova_v_sparse(self, hidden_states): |
| 913 | router_logits, _ = self.v_router(hidden_states) |
| 914 | |
| 915 | routing_weights, selected_values = calc_router_weights( |
| 916 | router_logits=router_logits, |
| 917 | e_score_correction_bias=self.v_router.bias, |
| 918 | score_func=self.router_score_func, |
| 919 | top_k=self.num_experts_per_tok, |
| 920 | scaling_factor=self.router_scaling_factor, |
| 921 | ) |
| 922 | |
| 923 | w1 = torch.stack( |
| 924 | [expert.weight for expert in self.v_experts], dim=0 |
| 925 | ).contiguous() |
| 926 | |
| 927 | v = fused_mova_impl( |
| 928 | config=self.fused_mova_config, |
| 929 | hidden_states=hidden_states.contiguous(), |
| 930 | w1=w1, |
| 931 | topk_weights=routing_weights.contiguous(), |
| 932 | topk_ids=selected_values.contiguous(), |
| 933 | global_num_experts=self.num_experts, |
| 934 | expert_map=None, |
| 935 | ) |
| 936 | |
| 937 | return v |
| 938 | |
| 939 | def forward( |
| 940 | self, |
| 941 | positions: torch.Tensor, |
| 942 | hidden_states: torch.Tensor, |
| 943 | ) -> torch.Tensor: |
| 944 | qk, _ = self.qk_proj(hidden_states) |
| 945 | q, k = qk.split([self.q_size, self.kv_size], dim=-1) |
| 946 | v = self.compute_mova_v_sparse(hidden_states) |
| 947 | |
| 948 | if self.query_key_norm: |
| 949 | q = self.q_norm(q) |
| 950 | k = self.k_norm(k) |
| 951 | |
| 952 | q, k = apply_partial_rope( |
| 953 | self.rotary_emb, |
| 954 | positions, |
| 955 | q, |
| 956 | k, |
| 957 | self.num_heads, |
| 958 | self.num_kv_heads, |
| 959 | self.head_dim, |
| 960 | self.rope_head_dim, |
| 961 | self.fold_rope_weights, |
| 962 | ) |
| 963 | |
| 964 | attn_output = self.attn(q, k, v) |
| 965 | |
| 966 | if self.gate_func is not None: |
| 967 | gate, _ = self.gate_proj(hidden_states) |
| 968 | if self.gate_func == "silu": |
| 969 | gate = F.silu(gate) |
| 970 | else: |
| 971 | assert self.gate_func == "softplus" |
| 972 | gate = F.softplus(gate, beta=math.log(2)) |
| 973 | |
| 974 | attn_output = attn_output * gate |
| 975 | |
| 976 | output, _ = self.o_proj(attn_output) |
| 977 | return output |
| 978 | |
| 979 | |
| 980 | class K2HorizonDecoderLayer(nn.Module): |
| 981 | def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 982 | super().__init__() |
| 983 | |
| 984 | config = vllm_config.model_config.hf_text_config |
| 985 | cache_config = vllm_config.cache_config |
| 986 | quant_config = vllm_config.quant_config |
| 987 | |
| 988 | self.hidden_size = config.hidden_size |
| 989 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 990 | dual_chunk_attention_config = getattr( |
| 991 | config, "dual_chunk_attention_config", None |
| 992 | ) |
| 993 | |
| 994 | # `mlp_only_layers` in the config. |
| 995 | layer_idx = extract_layer_index(prefix) |
| 996 | mlp_only_layers = ( |
| 997 | [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers |
| 998 | ) |
| 999 | is_sparse_layer = (layer_idx not in mlp_only_layers) and ( |
| 1000 | config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0 |
| 1001 | ) |
| 1002 | |
| 1003 | if is_sparse_layer and config.mova_num_experts > 0: |
| 1004 | self.self_attn = K2HorizonMoVAAttention( |
| 1005 | vllm_config=vllm_config, |
| 1006 | hidden_size=self.hidden_size, |
| 1007 | num_heads=config.num_attention_heads, |
| 1008 | num_kv_heads=config.num_key_value_heads, |
| 1009 | rope_parameters=config.rope_parameters, |
| 1010 | rope_head_dim=config.rope_head_dim, |
| 1011 | query_key_norm=config.query_key_norm, |
| 1012 | max_position_embeddings=max_position_embeddings, |
| 1013 | rms_norm_eps=config.rms_norm_eps, |
| 1014 | qkv_bias=config.attention_bias, |
| 1015 | head_dim=config.head_dim, |
| 1016 | gate_func=config.attention_gate_func, |
| 1017 | cache_config=cache_config, |
| 1018 | quant_config=quant_config, |
| 1019 | prefix=f"{prefix}.self_attn", |
| 1020 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 1021 | num_experts=config.mova_num_experts, |
| 1022 | num_experts_per_tok=config.mova_num_experts_per_tok, |
| 1023 | moe_gate_bias=config.moe_gate_bias, |
| 1024 | router_score_func=config.router_score_func, |
| 1025 | router_scaling_factor=config.router_scaling_factor, |
| 1026 | ) |
| 1027 | else: |
| 1028 | self.self_attn = K2HorizonAttention( |
| 1029 | hidden_size=self.hidden_size, |
| 1030 | num_heads=config.num_attention_heads, |
| 1031 | num_kv_heads=config.num_key_value_heads, |
| 1032 | rope_parameters=config.rope_parameters, |
| 1033 | rope_head_dim=config.rope_head_dim, |
| 1034 | query_key_norm=config.query_key_norm, |
| 1035 | max_position_embeddings=max_position_embeddings, |
| 1036 | rms_norm_eps=config.rms_norm_eps, |
| 1037 | qkv_bias=config.attention_bias, |
| 1038 | head_dim=config.head_dim, |
| 1039 | gate_func=config.attention_gate_func, |
| 1040 | cache_config=cache_config, |
| 1041 | quant_config=quant_config, |
| 1042 | prefix=f"{prefix}.self_attn", |
| 1043 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 1044 | ) |
| 1045 | |
| 1046 | if is_sparse_layer: |
| 1047 | self.mlp = K2HorizonSparseMoeBlock( |
| 1048 | vllm_config=vllm_config, prefix=f"{prefix}.mlp" |
| 1049 | ) |
| 1050 | else: |
| 1051 | self.mlp = K2HorizonMLP( |
| 1052 | hidden_size=config.hidden_size, |
| 1053 | intermediate_size=config.intermediate_size, |
| 1054 | hidden_act=config.hidden_act, |
| 1055 | quant_config=quant_config, |
| 1056 | prefix=f"{prefix}.mlp", |
| 1057 | ) |
| 1058 | |
| 1059 | self.input_layernorm = K2HorizonRMSNorm( |
| 1060 | hidden_size=config.hidden_size, |
| 1061 | n_groups=config.layernorm_num_groups, |
| 1062 | eps=config.rms_norm_eps, |
| 1063 | ) |
| 1064 | |
| 1065 | self.post_attention_layernorm = K2HorizonRMSNorm( |
| 1066 | hidden_size=config.hidden_size, |
| 1067 | n_groups=config.layernorm_num_groups, |
| 1068 | eps=config.rms_norm_eps, |
| 1069 | ) |
| 1070 | |
| 1071 | def forward( |
| 1072 | self, |
| 1073 | positions: torch.Tensor, |
| 1074 | hidden_states: torch.Tensor, |
| 1075 | residual: torch.Tensor | None, |
| 1076 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 1077 | # Self Attention |
| 1078 | if residual is None: |
| 1079 | residual = hidden_states |
| 1080 | hidden_states = self.input_layernorm(hidden_states) |
| 1081 | else: |
| 1082 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 1083 | |
| 1084 | hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states) |
| 1085 | |
| 1086 | # Fully Connected |
| 1087 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 1088 | hidden_states = self.mlp(hidden_states) |
| 1089 | return hidden_states, residual |
| 1090 | |
| 1091 | |
| 1092 | @support_torch_compile |
| 1093 | class K2HorizonModel(nn.Module, EagleModelMixin): |
| 1094 | def __init__( |
| 1095 | self, |
| 1096 | *, |
| 1097 | vllm_config: VllmConfig, |
| 1098 | prefix: str = "", |
| 1099 | decoder_layer_type: type[torch.nn.Module] = K2HorizonDecoderLayer, |
| 1100 | ): |
| 1101 | super().__init__() |
| 1102 | |
| 1103 | config = vllm_config.model_config.hf_text_config |
| 1104 | quant_config = vllm_config.quant_config |
| 1105 | parallel_config = vllm_config.parallel_config |
| 1106 | eplb_config = parallel_config.eplb_config |
| 1107 | self.num_redundant_experts = eplb_config.num_redundant_experts |
| 1108 | |
| 1109 | self.vocab_size = config.vocab_size |
| 1110 | self.config = config |
| 1111 | self.quant_config = quant_config |
| 1112 | self.embed_tokens = VocabParallelEmbedding( |
| 1113 | config.vocab_size, |
| 1114 | config.hidden_size, |
| 1115 | quant_config=quant_config, |
| 1116 | prefix=f"{prefix}.embed_tokens", |
| 1117 | ) |
| 1118 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 1119 | config.num_hidden_layers, |
| 1120 | lambda prefix: decoder_layer_type(vllm_config=vllm_config, prefix=prefix), |
| 1121 | prefix=f"{prefix}.layers", |
| 1122 | ) |
| 1123 | |
| 1124 | self.norm = K2HorizonRMSNorm( |
| 1125 | hidden_size=config.hidden_size, |
| 1126 | n_groups=config.layernorm_num_groups, |
| 1127 | eps=config.rms_norm_eps, |
| 1128 | ) |
| 1129 | |
| 1130 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 1131 | ["hidden_states", "residual"], config.hidden_size |
| 1132 | ) |
| 1133 | |
| 1134 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1135 | return self.embed_tokens(input_ids) |
| 1136 | |
| 1137 | def forward( |
| 1138 | self, |
| 1139 | input_ids: torch.Tensor | None, |
| 1140 | positions: torch.Tensor, |
| 1141 | intermediate_tensors: IntermediateTensors | None = None, |
| 1142 | inputs_embeds: torch.Tensor | None = None, |
| 1143 | ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]: |
| 1144 | if get_pp_group().is_first_rank: |
| 1145 | if inputs_embeds is not None: |
| 1146 | hidden_states = inputs_embeds |
| 1147 | else: |
| 1148 | hidden_states = self.embed_input_ids(input_ids) |
| 1149 | residual = None |
| 1150 | else: |
| 1151 | assert intermediate_tensors is not None |
| 1152 | hidden_states = intermediate_tensors["hidden_states"] |
| 1153 | residual = intermediate_tensors["residual"] |
| 1154 | |
| 1155 | aux_hidden_states = self._maybe_add_hidden_state( |
| 1156 | [], self.start_layer, hidden_states, residual |
| 1157 | ) |
| 1158 | for layer_idx, layer in enumerate( |
| 1159 | islice(self.layers, self.start_layer, self.end_layer), |
| 1160 | start=self.start_layer, |
| 1161 | ): |
| 1162 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 1163 | self._maybe_add_hidden_state( |
| 1164 | aux_hidden_states, layer_idx + 1, hidden_states, residual |
| 1165 | ) |
| 1166 | |
| 1167 | if not get_pp_group().is_last_rank: |
| 1168 | return IntermediateTensors( |
| 1169 | {"hidden_states": hidden_states, "residual": residual} |
| 1170 | ) |
| 1171 | hidden_states, _ = self.norm(hidden_states, residual) |
| 1172 | |
| 1173 | # Return auxiliary hidden states if collected |
| 1174 | if len(aux_hidden_states) > 0: |
| 1175 | return hidden_states, aux_hidden_states |
| 1176 | return hidden_states |
| 1177 | |
| 1178 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1179 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1180 | # (param_name, weight_name, expert_id, shard_id) |
| 1181 | return fused_moe_make_expert_params_mapping( |
| 1182 | self, |
| 1183 | ckpt_gate_proj_name="gate_proj", |
| 1184 | ckpt_down_proj_name="down_proj", |
| 1185 | ckpt_up_proj_name="up_proj", |
| 1186 | num_experts=self.config.num_experts, |
| 1187 | num_redundant_experts=self.num_redundant_experts, |
| 1188 | ) |
| 1189 | |
| 1190 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1191 | stacked_params_mapping = [ |
| 1192 | # (param_name, shard_name, shard_id) |
| 1193 | ("qkv_proj", "q_proj", "q"), |
| 1194 | ("qkv_proj", "k_proj", "k"), |
| 1195 | ("qkv_proj", "v_proj", "v"), |
| 1196 | ("gate_up_proj", "gate_proj", 0), |
| 1197 | ("gate_up_proj", "up_proj", 1), |
| 1198 | ] |
| 1199 | |
| 1200 | # Per-head gather index used to fold K2Horizon's partial-RoPE channel |
| 1201 | # reordering into the q_proj/k_proj rows and the q_norm/k_norm scales, |
| 1202 | # so the runtime can use vLLM's native partial NeoX RoPE. |
| 1203 | head_dim = self.config.head_dim or ( |
| 1204 | self.config.hidden_size // self.config.num_attention_heads |
| 1205 | ) |
| 1206 | rope_head_dim = self.config.rope_head_dim or head_dim |
| 1207 | rope_perm_idx = _rope_weight_perm(head_dim, rope_head_dim) |
| 1208 | |
| 1209 | # Attention modules whose q/k projections are dense: only these fold the |
| 1210 | # RoPE channel permutation into their q/k weights, biases and q/k norm |
| 1211 | # scales at load time. Attentions with packed/quantized q/k projections |
| 1212 | # keep the runtime channel-permutation path, so their weights must be |
| 1213 | # loaded verbatim (folding would misalign per-channel scales). |
| 1214 | rope_fold_prefixes = tuple( |
| 1215 | f"{module_name}." |
| 1216 | for module_name, module in self.named_modules() |
| 1217 | if isinstance(module, (K2HorizonAttention, K2HorizonMoVAAttention)) |
| 1218 | and getattr(module, "fold_rope_weights", False) |
| 1219 | ) |
| 1220 | |
| 1221 | # Skip loading extra parameters for GPTQ/modelopt models. |
| 1222 | # ignore_suffixes = ( |
| 1223 | # ".bias", |
| 1224 | # "_bias", |
| 1225 | # ".weight_scale", |
| 1226 | # "_weight_scale", |
| 1227 | # ".input_scale", |
| 1228 | # "_input_scale", |
| 1229 | # ) |
| 1230 | ignore_suffixes = () |
| 1231 | |
| 1232 | params_dict = dict(self.named_parameters()) |
| 1233 | |
| 1234 | loaded_params: set[str] = set() |
| 1235 | expert_params_mapping = self.get_expert_mapping() |
| 1236 | for name, loaded_weight in weights: |
| 1237 | if "scale" in name or "zero_point" in name: |
| 1238 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 1239 | if remapped_name is None: |
| 1240 | continue |
| 1241 | name = remapped_name |
| 1242 | |
| 1243 | # QK norm weights |
| 1244 | if name.endswith(".self_attn.q_norm.weight") or name.endswith( |
| 1245 | ".self_attn.k_norm.weight" |
| 1246 | ): |
| 1247 | if is_pp_missing_parameter(name, self): |
| 1248 | continue |
| 1249 | if name not in params_dict: |
| 1250 | continue |
| 1251 | |
| 1252 | # Fold the per-head RoPE channel permutation into the norm scales. |
| 1253 | if self.config.query_key_norm and name.startswith(rope_fold_prefixes): |
| 1254 | loaded_weight = ( |
| 1255 | loaded_weight.view(-1, head_dim)[:, rope_perm_idx] |
| 1256 | .reshape(-1) |
| 1257 | .contiguous() |
| 1258 | ) |
| 1259 | |
| 1260 | param = params_dict[name] |
| 1261 | tp_rank = get_tensor_model_parallel_rank() |
| 1262 | tp_size = get_tensor_model_parallel_world_size() |
| 1263 | |
| 1264 | if loaded_weight.shape != param.shape: |
| 1265 | if name.endswith(".self_attn.q_norm.weight"): |
| 1266 | loaded_weight = loaded_weight.chunk(tp_size, dim=0)[tp_rank] |
| 1267 | else: |
| 1268 | num_kv_heads = self.config.num_key_value_heads |
| 1269 | if num_kv_heads >= tp_size: |
| 1270 | loaded_weight = loaded_weight.chunk(tp_size, dim=0)[tp_rank] |
| 1271 | else: |
| 1272 | num_kv_head_replicas = tp_size // num_kv_heads |
| 1273 | kv_rank = tp_rank // num_kv_head_replicas |
| 1274 | loaded_weight = loaded_weight.chunk(num_kv_heads, dim=0)[ |
| 1275 | kv_rank |
| 1276 | ] |
| 1277 | |
| 1278 | weight_loader: Callable[..., Any] = getattr( |
| 1279 | param, "weight_loader", default_weight_loader |
| 1280 | ) |
| 1281 | weight_loader(param, loaded_weight) |
| 1282 | loaded_params.add(name) |
| 1283 | continue |
| 1284 | |
| 1285 | # MoVA v experts |
| 1286 | if ".self_attn.v_experts." in name: |
| 1287 | if is_pp_missing_parameter(name, self): |
| 1288 | continue |
| 1289 | if name not in params_dict: |
| 1290 | continue |
| 1291 | |
| 1292 | param = params_dict[name] |
| 1293 | tp_rank = get_tensor_model_parallel_rank() |
| 1294 | tp_size = get_tensor_model_parallel_world_size() |
| 1295 | num_kv_heads = self.config.num_key_value_heads |
| 1296 | |
| 1297 | if loaded_weight.shape != param.shape: |
| 1298 | if num_kv_heads >= tp_size: |
| 1299 | loaded_weight = loaded_weight.chunk(tp_size, dim=0)[tp_rank] |
| 1300 | else: |
| 1301 | num_kv_head_replicas = tp_size // num_kv_heads |
| 1302 | kv_rank = tp_rank // num_kv_head_replicas |
| 1303 | loaded_weight = loaded_weight.chunk(num_kv_heads, dim=0)[ |
| 1304 | kv_rank |
| 1305 | ] |
| 1306 | |
| 1307 | default_weight_loader(param, loaded_weight) |
| 1308 | loaded_params.add(name) |
| 1309 | continue |
| 1310 | |
| 1311 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 1312 | # Skip non-stacked layers and experts (experts handled below). |
| 1313 | if weight_name not in name: |
| 1314 | continue |
| 1315 | # We have mlp.experts[0].gate_proj in the checkpoint. |
| 1316 | # Since we handle the experts below in expert_params_mapping, |
| 1317 | # we need to skip here BEFORE we update the name, otherwise |
| 1318 | # name will be updated to mlp.experts[0].gate_up_proj, which |
| 1319 | # will then be updated below in expert_params_mapping |
| 1320 | # for mlp.experts[0].gate_gate_up_proj, which breaks load. |
| 1321 | if "mlp.experts" in name: |
| 1322 | continue |
| 1323 | if ".self_attn." in name and param_name == "gate_up_proj": |
| 1324 | assert weight_name == "gate_proj" |
| 1325 | continue |
| 1326 | |
| 1327 | # Fold the per-head RoPE channel permutation into the dense |
| 1328 | # q_proj / k_proj rows (W' = P W), and biases, before TP |
| 1329 | # sharding. |
| 1330 | if shard_id in ("q", "k") and name.startswith(rope_fold_prefixes): |
| 1331 | if name.endswith(".weight"): |
| 1332 | hidden = loaded_weight.shape[-1] |
| 1333 | loaded_weight = ( |
| 1334 | loaded_weight.view(-1, head_dim, hidden)[ |
| 1335 | :, rope_perm_idx, : |
| 1336 | ] |
| 1337 | .reshape(-1, hidden) |
| 1338 | .contiguous() |
| 1339 | ) |
| 1340 | elif self.config.attention_bias and name.endswith(".bias"): |
| 1341 | loaded_weight = ( |
| 1342 | loaded_weight.view(-1, head_dim)[:, rope_perm_idx] |
| 1343 | .reshape(-1) |
| 1344 | .contiguous() |
| 1345 | ) |
| 1346 | |
| 1347 | if ( |
| 1348 | ".self_attn." in name |
| 1349 | and param_name == "qkv_proj" |
| 1350 | and name.replace(weight_name, param_name) not in params_dict |
| 1351 | ): |
| 1352 | qk_name = name.replace(weight_name, "qk_proj") |
| 1353 | if is_pp_missing_parameter(qk_name, self): |
| 1354 | continue |
| 1355 | if qk_name not in params_dict: |
| 1356 | continue |
| 1357 | |
| 1358 | assert weight_name in ["q_proj", "k_proj"] |
| 1359 | if weight_name == "k_proj": |
| 1360 | tp_size = get_tensor_model_parallel_world_size() |
| 1361 | num_kv_heads = self.config.num_key_value_heads |
| 1362 | if num_kv_heads < tp_size: |
| 1363 | num_kv_head_replicas = tp_size // num_kv_heads |
| 1364 | loaded_weight = ( |
| 1365 | loaded_weight.reshape( |
| 1366 | num_kv_heads, -1, *loaded_weight.shape[1:] |
| 1367 | ) |
| 1368 | .repeat_interleave(num_kv_head_replicas, dim=0) |
| 1369 | .reshape(-1, *loaded_weight.shape[1:]) |
| 1370 | ) |
| 1371 | |
| 1372 | param = params_dict[qk_name] |
| 1373 | weight_loader = getattr( |
| 1374 | param, "weight_loader", default_weight_loader |
| 1375 | ) |
| 1376 | weight_loader( |
| 1377 | param, loaded_weight, 0 if weight_name == "q_proj" else 1 |
| 1378 | ) |
| 1379 | loaded_params.add(qk_name) |
| 1380 | break |
| 1381 | |
| 1382 | name = name.replace(weight_name, param_name) |
| 1383 | |
| 1384 | # Skip loading extra parameters for GPTQ/modelopt models. |
| 1385 | if name.endswith(ignore_suffixes) and name not in params_dict: |
| 1386 | continue |
| 1387 | |
| 1388 | # Skip layers on other devices. |
| 1389 | if is_pp_missing_parameter(name, self): |
| 1390 | continue |
| 1391 | if name.endswith("scale"): |
| 1392 | # Remapping the name of FP8 kv-scale. |
| 1393 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 1394 | if remapped_name is None: |
| 1395 | continue |
| 1396 | name = remapped_name |
| 1397 | if name not in params_dict: |
| 1398 | continue |
| 1399 | |
| 1400 | param = params_dict[name] |
| 1401 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 1402 | if weight_loader == default_weight_loader: |
| 1403 | weight_loader(param, loaded_weight) |
| 1404 | else: |
| 1405 | weight_loader(param, loaded_weight, shard_id) |
| 1406 | break |
| 1407 | else: |
| 1408 | is_expert_weight = False |
| 1409 | for mapping in expert_params_mapping: |
| 1410 | param_name, weight_name, expert_id, shard_id = mapping |
| 1411 | if weight_name not in name: |
| 1412 | continue |
| 1413 | |
| 1414 | # Anyway, this is an expert weight and should not be |
| 1415 | # attempted to load as other weights later |
| 1416 | is_expert_weight = True |
| 1417 | |
| 1418 | # Do not modify `name` since the loop may continue here |
| 1419 | # Instead, create a new variable |
| 1420 | name_mapped = name.replace(weight_name, param_name) |
| 1421 | |
| 1422 | if is_pp_missing_parameter(name_mapped, self): |
| 1423 | continue |
| 1424 | |
| 1425 | # Skip loading extra parameters for GPTQ/modelopt models. |
| 1426 | if ( |
| 1427 | name_mapped.endswith(ignore_suffixes) |
| 1428 | and name_mapped not in params_dict |
| 1429 | ): |
| 1430 | continue |
| 1431 | |
| 1432 | param = params_dict[name_mapped] |
| 1433 | # We should ask the weight loader to return success or not |
| 1434 | # here since otherwise we may skip experts with other |
| 1435 | # available replicas. |
| 1436 | weight_loader = typing.cast( |
| 1437 | Callable[..., bool], param.weight_loader |
| 1438 | ) |
| 1439 | success = weight_loader( |
| 1440 | param, |
| 1441 | loaded_weight, |
| 1442 | name_mapped, |
| 1443 | shard_id=shard_id, |
| 1444 | expert_id=expert_id, |
| 1445 | return_success=True, |
| 1446 | ) |
| 1447 | if success: |
| 1448 | name = name_mapped |
| 1449 | break |
| 1450 | else: |
| 1451 | if is_expert_weight: |
| 1452 | # We've checked that this is an expert weight |
| 1453 | # However it's not mapped locally to this rank |
| 1454 | # So we simply skip it |
| 1455 | continue |
| 1456 | |
| 1457 | # Skip loading extra parameters for GPTQ/modelopt models. |
| 1458 | if name.endswith(ignore_suffixes) and name not in params_dict: |
| 1459 | continue |
| 1460 | # Skip layers on other devices. |
| 1461 | if is_pp_missing_parameter(name, self): |
| 1462 | continue |
| 1463 | if name not in params_dict: |
| 1464 | continue |
| 1465 | param = params_dict[name] |
| 1466 | weight_loader = getattr( |
| 1467 | param, "weight_loader", default_weight_loader |
| 1468 | ) |
| 1469 | weight_loader(param, loaded_weight) |
| 1470 | loaded_params.add(name) |
| 1471 | |
| 1472 | return loaded_params |
| 1473 | |
| 1474 | |
| 1475 | class K2HorizonForCausalLM( |
| 1476 | nn.Module, SupportsPP, SupportsLoRA, SupportsEagle, SupportsEagle3, MixtureOfExperts |
| 1477 | ): |
| 1478 | packed_modules_mapping = { |
| 1479 | "qkv_proj": [ |
| 1480 | "q_proj", |
| 1481 | "k_proj", |
| 1482 | "v_proj", |
| 1483 | ], |
| 1484 | "qk_proj": [ |
| 1485 | "q_proj", |
| 1486 | "k_proj", |
| 1487 | ], |
| 1488 | } |
| 1489 | |
| 1490 | embedding_modules = { |
| 1491 | "embed_tokens": "input_embeddings", |
| 1492 | "lm_head": "output_embeddings", |
| 1493 | } |
| 1494 | |
| 1495 | fall_back_to_pt_during_load = False |
| 1496 | |
| 1497 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1498 | super().__init__() |
| 1499 | config = vllm_config.model_config.hf_text_config |
| 1500 | quant_config = vllm_config.quant_config |
| 1501 | self.config = config |
| 1502 | self.quant_config = quant_config |
| 1503 | # Only perform the following mapping when K2HorizonMLP exists |
| 1504 | if getattr(config, "mlp_only_layers", []): |
| 1505 | self.packed_modules_mapping["gate_up_proj"] = ["gate_proj", "up_proj"] |
| 1506 | self.model = K2HorizonModel( |
| 1507 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1508 | ) |
| 1509 | self.lm_head = ParallelLMHead( |
| 1510 | config.vocab_size, |
| 1511 | config.hidden_size, |
| 1512 | quant_config=quant_config, |
| 1513 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1514 | ) |
| 1515 | if self.config.tie_word_embeddings: |
| 1516 | self.lm_head.weight = self.model.embed_tokens.weight |
| 1517 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1518 | self.make_empty_intermediate_tensors = ( |
| 1519 | self.model.make_empty_intermediate_tensors |
| 1520 | ) |
| 1521 | |
| 1522 | # Set MoE hyperparameters |
| 1523 | self.expert_weights = [] |
| 1524 | |
| 1525 | self.moe_layers = [] |
| 1526 | example_layer = None |
| 1527 | for layer in self.model.layers: |
| 1528 | if isinstance(layer, PPMissingLayer): |
| 1529 | continue |
| 1530 | |
| 1531 | assert isinstance(layer, K2HorizonDecoderLayer) |
| 1532 | if isinstance(layer.mlp, K2HorizonSparseMoeBlock): |
| 1533 | example_layer = layer.mlp |
| 1534 | self.moe_layers.append(layer.mlp.experts) |
| 1535 | |
| 1536 | self.num_moe_layers = len(self.moe_layers) |
| 1537 | if example_layer is not None: |
| 1538 | self.num_logical_experts = example_layer.n_logical_experts |
| 1539 | self.num_physical_experts = example_layer.n_physical_experts |
| 1540 | self.num_local_physical_experts = example_layer.n_local_physical_experts |
| 1541 | self.num_routed_experts = example_layer.n_routed_experts |
| 1542 | self.num_redundant_experts = example_layer.n_redundant_experts |
| 1543 | |
| 1544 | def update_physical_experts_metadata( |
| 1545 | self, |
| 1546 | num_physical_experts: int, |
| 1547 | num_local_physical_experts: int, |
| 1548 | ) -> None: |
| 1549 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1550 | self.num_physical_experts = num_physical_experts |
| 1551 | self.num_local_physical_experts = num_local_physical_experts |
| 1552 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1553 | for layer in self.model.layers: |
| 1554 | if isinstance(layer.mlp, K2HorizonSparseMoeBlock): |
| 1555 | moe = layer.mlp |
| 1556 | moe.n_local_physical_experts = num_local_physical_experts |
| 1557 | moe.n_physical_experts = num_physical_experts |
| 1558 | moe.n_redundant_experts = self.num_redundant_experts |
| 1559 | moe.experts.update_expert_map() |
| 1560 | |
| 1561 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1562 | return self.model.embed_input_ids(input_ids) |
| 1563 | |
| 1564 | def forward( |
| 1565 | self, |
| 1566 | input_ids: torch.Tensor | None, |
| 1567 | positions: torch.Tensor, |
| 1568 | intermediate_tensors: IntermediateTensors | None = None, |
| 1569 | inputs_embeds: torch.Tensor | None = None, |
| 1570 | ) -> torch.Tensor | IntermediateTensors: |
| 1571 | hidden_states = self.model( |
| 1572 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 1573 | ) |
| 1574 | return hidden_states |
| 1575 | |
| 1576 | def compute_logits( |
| 1577 | self, |
| 1578 | hidden_states: torch.Tensor, |
| 1579 | ) -> torch.Tensor | None: |
| 1580 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1581 | return logits |
| 1582 | |
| 1583 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1584 | loader = AutoWeightsLoader(self) |
| 1585 | return loader.load_weights(weights) |
| 1586 | |
| 1587 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1588 | return self.model.get_expert_mapping() |