| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py |
| 6 | # Copyright 2023 The vLLM team. |
| 7 | # |
| 8 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 9 | # and OPT implementations in this library. It has been modified from its |
| 10 | # original forms to accommodate minor architectural differences compared |
| 11 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 12 | # |
| 13 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 14 | # you may not use this file except in compliance with the License. |
| 15 | # You may obtain a copy of the License at |
| 16 | # |
| 17 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 18 | # |
| 19 | # Unless required by applicable law or agreed to in writing, software |
| 20 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 21 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 22 | # See the License for the specific language governing permissions and |
| 23 | # limitations under the License. |
| 24 | """Inference-only A.X K1 model.""" |
| 25 | |
| 26 | import typing |
| 27 | from collections.abc import Callable, Iterable |
| 28 | from itertools import islice |
| 29 | |
| 30 | import torch |
| 31 | from torch import nn |
| 32 | |
| 33 | from vllm._aiter_ops import rocm_aiter_ops |
| 34 | from vllm.compilation.decorators import support_torch_compile |
| 35 | from vllm.config import CacheConfig, ParallelConfig, VllmConfig |
| 36 | from vllm.distributed import ( |
| 37 | get_ep_group, |
| 38 | get_pp_group, |
| 39 | get_tensor_model_parallel_rank, |
| 40 | get_tensor_model_parallel_world_size, |
| 41 | tensor_model_parallel_all_gather, |
| 42 | ) |
| 43 | from vllm.logger import init_logger |
| 44 | from vllm.model_executor.layers.attention import Attention |
| 45 | from vllm.model_executor.layers.fused_moe import ( |
| 46 | FusedMoEFactory, |
| 47 | GateLinear, |
| 48 | fused_moe_make_expert_params_mapping, |
| 49 | ) |
| 50 | from vllm.model_executor.layers.fused_moe.utils import ( |
| 51 | is_model_fused_shared_expert_compatible, |
| 52 | resolve_layer_fused_shared_expert, |
| 53 | ) |
| 54 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 55 | from vllm.model_executor.layers.linear import ( |
| 56 | ColumnParallelLinear, |
| 57 | MergedColumnParallelLinear, |
| 58 | ReplicatedLinear, |
| 59 | RowParallelLinear, |
| 60 | ) |
| 61 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 62 | from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper |
| 63 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 64 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 65 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 66 | ParallelLMHead, |
| 67 | VocabParallelEmbedding, |
| 68 | ) |
| 69 | from vllm.model_executor.model_loader.weight_utils import ( |
| 70 | default_weight_loader, |
| 71 | maybe_remap_kv_scale_name, |
| 72 | ) |
| 73 | from vllm.model_executor.models.deepseek_v2 import ( |
| 74 | DeepseekAttention, |
| 75 | DeepseekV2MLP, |
| 76 | yarn_get_mscale, |
| 77 | ) |
| 78 | from vllm.model_executor.models.utils import sequence_parallel_chunk |
| 79 | from vllm.platforms import current_platform |
| 80 | from vllm.sequence import IntermediateTensors |
| 81 | from vllm.transformers_utils.configs.AXK1 import AXK1Config |
| 82 | |
| 83 | from .interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP |
| 84 | from .utils import ( |
| 85 | AutoWeightsLoader, |
| 86 | PPMissingLayer, |
| 87 | get_spec_layer_idx_from_weight_name, |
| 88 | is_pp_missing_parameter, |
| 89 | make_empty_intermediate_tensors_factory, |
| 90 | make_layers, |
| 91 | maybe_prefix, |
| 92 | ) |
| 93 | |
| 94 | logger = init_logger(__name__) |
| 95 | |
| 96 | |
| 97 | class AXK1MLP(DeepseekV2MLP): |
| 98 | pass |
| 99 | |
| 100 | |
| 101 | class AXK1MoE(nn.Module): |
| 102 | def __init__( |
| 103 | self, |
| 104 | config: AXK1Config, |
| 105 | parallel_config: ParallelConfig, |
| 106 | quant_config: QuantizationConfig | None = None, |
| 107 | prefix: str = "", |
| 108 | ): |
| 109 | super().__init__() |
| 110 | self.tp_size = get_tensor_model_parallel_world_size() |
| 111 | self.tp_rank = get_tensor_model_parallel_rank() |
| 112 | |
| 113 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 114 | |
| 115 | self.ep_group = get_ep_group().device_group |
| 116 | self.ep_size = self.ep_group.size() |
| 117 | assert config.n_routed_experts is not None |
| 118 | assert config.num_experts_per_tok is not None |
| 119 | assert config.scoring_func is not None |
| 120 | assert config.hidden_act is not None |
| 121 | self.n_routed_experts: int = config.n_routed_experts |
| 122 | self.n_shared_experts: int | None = config.n_shared_experts |
| 123 | |
| 124 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 125 | |
| 126 | if config.hidden_act != "silu": |
| 127 | raise ValueError( |
| 128 | f"Unsupported activation: {config.hidden_act}. " |
| 129 | "Only silu is supported for now." |
| 130 | ) |
| 131 | |
| 132 | self.gate = GateLinear( |
| 133 | config.hidden_size, |
| 134 | config.n_routed_experts, |
| 135 | prefix=f"{prefix}.gate", |
| 136 | ) |
| 137 | if config.topk_method == "noaux_tc": |
| 138 | self.gate.e_score_correction_bias = nn.Parameter( |
| 139 | torch.empty(config.n_routed_experts, dtype=torch.float32) |
| 140 | ) |
| 141 | else: |
| 142 | self.gate.e_score_correction_bias = None |
| 143 | |
| 144 | # Load balancing settings. |
| 145 | eplb_config = parallel_config.eplb_config |
| 146 | self.enable_eplb = parallel_config.enable_eplb |
| 147 | |
| 148 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 149 | self.n_logical_experts = self.n_routed_experts |
| 150 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 151 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 152 | |
| 153 | self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled() |
| 154 | |
| 155 | self.is_fused_shared_expert_enabled = False |
| 156 | if config.n_shared_experts is not None: |
| 157 | self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert( |
| 158 | quant_config, prefix |
| 159 | ) |
| 160 | |
| 161 | if config.n_shared_experts is None or self.is_fused_shared_expert_enabled: |
| 162 | self.shared_experts = None |
| 163 | else: |
| 164 | intermediate_size = config.moe_intermediate_size * config.n_shared_experts |
| 165 | |
| 166 | self.shared_experts = AXK1MLP( |
| 167 | hidden_size=config.hidden_size, |
| 168 | intermediate_size=intermediate_size, |
| 169 | hidden_act=config.hidden_act, |
| 170 | quant_config=quant_config, |
| 171 | is_sequence_parallel=self.is_sequence_parallel, |
| 172 | reduce_results=False, |
| 173 | prefix=f"{prefix}.shared_experts", |
| 174 | ) |
| 175 | |
| 176 | self.experts = FusedMoEFactory( |
| 177 | shared_experts=self.shared_experts, |
| 178 | gate=self.gate, |
| 179 | num_experts=config.n_routed_experts, |
| 180 | top_k=config.num_experts_per_tok, |
| 181 | hidden_size=config.hidden_size, |
| 182 | intermediate_size=config.moe_intermediate_size, |
| 183 | renormalize=config.norm_topk_prob, |
| 184 | quant_config=quant_config, |
| 185 | use_grouped_topk=True, |
| 186 | num_expert_group=config.n_group, |
| 187 | topk_group=config.topk_group, |
| 188 | prefix=f"{prefix}.experts", |
| 189 | scoring_func=config.scoring_func, |
| 190 | # we do scaling outside, set factor to 1.0 to avoid double mul |
| 191 | # aiter applies routed_scaling_factor internally |
| 192 | routed_scaling_factor=self.routed_scaling_factor, |
| 193 | apply_routed_scale_to_output=not self.is_rocm_aiter_moe_enabled, |
| 194 | e_score_correction_bias=self.gate.e_score_correction_bias, |
| 195 | enable_eplb=self.enable_eplb, |
| 196 | num_redundant_experts=self.n_redundant_experts, |
| 197 | is_sequence_parallel=self.is_sequence_parallel, |
| 198 | n_shared_experts=config.n_shared_experts |
| 199 | if self.is_fused_shared_expert_enabled |
| 200 | else None, |
| 201 | fuse_shared_experts=self.is_fused_shared_expert_enabled, |
| 202 | ) |
| 203 | |
| 204 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 205 | num_tokens, hidden_dim = hidden_states.shape |
| 206 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 207 | |
| 208 | # Chunk the hidden states so they aren't replicated across TP ranks. |
| 209 | # This avoids duplicate computation in self.experts. |
| 210 | # TODO: We can replace the all_reduce at the end of attn with a |
| 211 | # reduce_scatter instead of chunking here. |
| 212 | if self.is_sequence_parallel: |
| 213 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 214 | |
| 215 | final_hidden_states = self.experts( |
| 216 | hidden_states=hidden_states, router_logits=hidden_states |
| 217 | ) |
| 218 | |
| 219 | if self.is_sequence_parallel: |
| 220 | final_hidden_states = tensor_model_parallel_all_gather( |
| 221 | final_hidden_states, 0 |
| 222 | ) |
| 223 | final_hidden_states = final_hidden_states[:num_tokens] |
| 224 | |
| 225 | return final_hidden_states.view(num_tokens, hidden_dim) |
| 226 | |
| 227 | |
| 228 | def _get_llama_4_scaling( |
| 229 | original_max_position_embeddings: int, scaling_beta: float, positions: torch.Tensor |
| 230 | ) -> torch.Tensor: |
| 231 | scaling = 1 + scaling_beta * torch.log( |
| 232 | 1 + torch.floor(positions / original_max_position_embeddings) |
| 233 | ) |
| 234 | # Broadcast over num_heads and head_dim |
| 235 | return scaling[..., None, None] |
| 236 | |
| 237 | |
| 238 | class AXK1Attention(nn.Module): |
| 239 | def __init__( |
| 240 | self, |
| 241 | vllm_config: VllmConfig, |
| 242 | config: AXK1Config, |
| 243 | hidden_size: int, |
| 244 | num_heads: int, |
| 245 | qk_nope_head_dim: int, |
| 246 | qk_rope_head_dim: int, |
| 247 | v_head_dim: int, |
| 248 | q_lora_rank: int | None, |
| 249 | kv_lora_rank: int, |
| 250 | max_position_embeddings: int = 8192, |
| 251 | cache_config: CacheConfig | None = None, |
| 252 | quant_config: QuantizationConfig | None = None, |
| 253 | topk_indices_buffer: torch.Tensor | None = None, |
| 254 | prefix: str = "", |
| 255 | ) -> None: |
| 256 | super().__init__() |
| 257 | self.hidden_size = hidden_size |
| 258 | self.qk_nope_head_dim = qk_nope_head_dim |
| 259 | self.qk_rope_head_dim = qk_rope_head_dim |
| 260 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 261 | self.v_head_dim = v_head_dim |
| 262 | self.q_lora_rank = q_lora_rank |
| 263 | self.kv_lora_rank = kv_lora_rank |
| 264 | self.num_heads = num_heads |
| 265 | tp_size = get_tensor_model_parallel_world_size() |
| 266 | assert num_heads % tp_size == 0 |
| 267 | self.num_local_heads = num_heads // tp_size |
| 268 | self.scaling = self.qk_head_dim**-0.5 |
| 269 | self.max_position_embeddings = max_position_embeddings |
| 270 | assert topk_indices_buffer is None, ( |
| 271 | "topk_indices_buffer is not \ |
| 272 | supported for AXK1Attention" |
| 273 | ) |
| 274 | |
| 275 | if self.q_lora_rank is not None: |
| 276 | self.q_a_proj = ReplicatedLinear( |
| 277 | self.hidden_size, |
| 278 | self.q_lora_rank, |
| 279 | bias=False, |
| 280 | quant_config=quant_config, |
| 281 | prefix=f"{prefix}.q_a_proj", |
| 282 | ) |
| 283 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 284 | self.q_b_proj = ColumnParallelLinear( |
| 285 | self.q_lora_rank, |
| 286 | self.num_heads * self.qk_head_dim, |
| 287 | bias=False, |
| 288 | quant_config=quant_config, |
| 289 | prefix=f"{prefix}.q_b_proj", |
| 290 | ) |
| 291 | else: |
| 292 | self.q_proj = ColumnParallelLinear( |
| 293 | self.hidden_size, |
| 294 | self.num_heads * self.qk_head_dim, |
| 295 | bias=False, |
| 296 | quant_config=quant_config, |
| 297 | prefix=f"{prefix}.q_proj", |
| 298 | ) |
| 299 | |
| 300 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 301 | self.hidden_size, |
| 302 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 303 | bias=False, |
| 304 | quant_config=quant_config, |
| 305 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 306 | ) |
| 307 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 308 | self.kv_b_proj = ColumnParallelLinear( |
| 309 | self.kv_lora_rank, |
| 310 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 311 | bias=False, |
| 312 | quant_config=quant_config, |
| 313 | prefix=f"{prefix}.kv_b_proj", |
| 314 | ) |
| 315 | # O projection. |
| 316 | self.o_proj = RowParallelLinear( |
| 317 | self.num_heads * self.v_head_dim, |
| 318 | self.hidden_size, |
| 319 | bias=False, |
| 320 | quant_config=quant_config, |
| 321 | prefix=f"{prefix}.o_proj", |
| 322 | ) |
| 323 | assert config.rope_parameters is not None |
| 324 | if config.rope_parameters["rope_type"] != "default": |
| 325 | config.rope_parameters["rope_type"] = ( |
| 326 | "deepseek_llama_scaling" |
| 327 | if config.rope_parameters.get("attention_factor") == 1.0 |
| 328 | else "deepseek_yarn" |
| 329 | ) |
| 330 | |
| 331 | self.rotary_emb = get_rope( |
| 332 | qk_rope_head_dim, |
| 333 | max_position=max_position_embeddings, |
| 334 | rope_parameters=config.rope_parameters, |
| 335 | is_neox_style=False, |
| 336 | ) |
| 337 | |
| 338 | if config.rope_parameters["rope_type"] == "deepseek_yarn": |
| 339 | mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) |
| 340 | scaling_factor = config.rope_parameters["factor"] |
| 341 | mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) |
| 342 | self.scaling = self.scaling * mscale * mscale |
| 343 | |
| 344 | self.attn = Attention( |
| 345 | self.num_local_heads, |
| 346 | self.qk_head_dim, |
| 347 | self.scaling, |
| 348 | num_kv_heads=self.num_local_heads, |
| 349 | cache_config=cache_config, |
| 350 | quant_config=quant_config, |
| 351 | prefix=f"{prefix}.attn", |
| 352 | ) |
| 353 | |
| 354 | def forward( |
| 355 | self, |
| 356 | positions: torch.Tensor, |
| 357 | hidden_states: torch.Tensor, |
| 358 | llama_4_scaling: torch.Tensor | None, |
| 359 | ) -> torch.Tensor: |
| 360 | if self.q_lora_rank is not None: |
| 361 | q = self.q_a_proj(hidden_states)[0] |
| 362 | q = self.q_a_layernorm(q) |
| 363 | q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim) |
| 364 | else: |
| 365 | q = self.q_proj(hidden_states)[0].view( |
| 366 | -1, self.num_local_heads, self.qk_head_dim |
| 367 | ) |
| 368 | q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) |
| 369 | latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0] |
| 370 | kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) |
| 371 | latent_cache = latent_cache.unsqueeze(1) |
| 372 | kv_a = self.kv_a_layernorm(kv_a) |
| 373 | kv = self.kv_b_proj(kv_a)[0] |
| 374 | kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim) |
| 375 | k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) |
| 376 | k_pe = latent_cache[:, :, self.kv_lora_rank :] |
| 377 | |
| 378 | q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe) |
| 379 | |
| 380 | q[..., self.qk_nope_head_dim :] = q_pe |
| 381 | k = torch.empty_like(q) |
| 382 | k[..., : self.qk_nope_head_dim] = k_nope |
| 383 | k[..., self.qk_nope_head_dim :] = k_pe |
| 384 | |
| 385 | # Apply llama 4 scaling if provided |
| 386 | if llama_4_scaling is not None: |
| 387 | q *= llama_4_scaling |
| 388 | |
| 389 | # padding value to qk_head_dim for alignment |
| 390 | v = torch.nn.functional.pad( |
| 391 | v, [0, self.qk_head_dim - self.v_head_dim], value=0 |
| 392 | ).view(-1, self.num_local_heads * self.qk_head_dim) |
| 393 | attn_output = self.attn(q, k, v) |
| 394 | attn_output = attn_output.view(-1, self.num_local_heads, self.qk_head_dim)[ |
| 395 | ..., : self.v_head_dim |
| 396 | ].reshape(-1, self.num_local_heads * self.v_head_dim) |
| 397 | output, _ = self.o_proj(attn_output) |
| 398 | return output |
| 399 | |
| 400 | |
| 401 | class AXK1MLAAttention(nn.Module): |
| 402 | """Main reference: DeepseekV2 paper, and FlashInfer Implementation |
| 403 | (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551). |
| 404 | |
| 405 | For more info see MLACommonImpl in: |
| 406 | vllm/v1/attention/backends/mla/utils.py |
| 407 | """ |
| 408 | |
| 409 | def __init__( |
| 410 | self, |
| 411 | vllm_config: VllmConfig, |
| 412 | config: AXK1Config, |
| 413 | hidden_size: int, |
| 414 | num_heads: int, |
| 415 | qk_nope_head_dim: int, |
| 416 | qk_rope_head_dim: int, |
| 417 | v_head_dim: int, |
| 418 | q_lora_rank: int | None, |
| 419 | kv_lora_rank: int, |
| 420 | max_position_embeddings: int = 8192, |
| 421 | cache_config: CacheConfig | None = None, |
| 422 | quant_config: QuantizationConfig | None = None, |
| 423 | prefix: str = "", |
| 424 | topk_indices_buffer: torch.Tensor | None = None, |
| 425 | ) -> None: |
| 426 | super().__init__() |
| 427 | self.hidden_size = hidden_size |
| 428 | self.qk_nope_head_dim = qk_nope_head_dim |
| 429 | self.qk_rope_head_dim = qk_rope_head_dim |
| 430 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 431 | self.v_head_dim = v_head_dim |
| 432 | |
| 433 | self.q_lora_rank = q_lora_rank |
| 434 | self.kv_lora_rank = kv_lora_rank |
| 435 | |
| 436 | self.num_heads = num_heads |
| 437 | tp_size = get_tensor_model_parallel_world_size() |
| 438 | assert num_heads % tp_size == 0 |
| 439 | self.num_local_heads = num_heads // tp_size |
| 440 | |
| 441 | self.scaling = self.qk_head_dim**-0.5 |
| 442 | self.max_position_embeddings = max_position_embeddings |
| 443 | |
| 444 | if self.q_lora_rank is not None: |
| 445 | self.fused_qkv_a_proj = MergedColumnParallelLinear( |
| 446 | self.hidden_size, |
| 447 | [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], |
| 448 | bias=False, |
| 449 | quant_config=quant_config, |
| 450 | prefix=f"{prefix}.fused_qkv_a_proj", |
| 451 | disable_tp=True, |
| 452 | ) |
| 453 | else: |
| 454 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 455 | self.hidden_size, |
| 456 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 457 | bias=False, |
| 458 | quant_config=quant_config, |
| 459 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 460 | ) |
| 461 | |
| 462 | if self.q_lora_rank is not None: |
| 463 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 464 | self.q_b_proj = ColumnParallelLinear( |
| 465 | self.q_lora_rank, |
| 466 | self.num_heads * self.qk_head_dim, |
| 467 | bias=False, |
| 468 | quant_config=quant_config, |
| 469 | prefix=f"{prefix}.q_b_proj", |
| 470 | ) |
| 471 | else: |
| 472 | self.q_proj = ColumnParallelLinear( |
| 473 | self.hidden_size, |
| 474 | self.num_heads * self.qk_head_dim, |
| 475 | bias=False, |
| 476 | quant_config=quant_config, |
| 477 | prefix=f"{prefix}.q_proj", |
| 478 | ) |
| 479 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 480 | self.kv_b_proj = ColumnParallelLinear( |
| 481 | self.kv_lora_rank, |
| 482 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 483 | bias=False, |
| 484 | quant_config=quant_config, |
| 485 | prefix=f"{prefix}.kv_b_proj", |
| 486 | ) |
| 487 | self.o_proj = RowParallelLinear( |
| 488 | self.num_heads * self.v_head_dim, |
| 489 | self.hidden_size, |
| 490 | bias=False, |
| 491 | quant_config=quant_config, |
| 492 | prefix=f"{prefix}.o_proj", |
| 493 | ) |
| 494 | |
| 495 | assert config.rope_parameters is not None |
| 496 | if config.rope_parameters["rope_type"] != "default": |
| 497 | config.rope_parameters["rope_type"] = ( |
| 498 | "deepseek_llama_scaling" |
| 499 | if config.rope_parameters.get("attention_factor") == 1.0 |
| 500 | else "deepseek_yarn" |
| 501 | ) |
| 502 | |
| 503 | self.rotary_emb = get_rope( |
| 504 | qk_rope_head_dim, |
| 505 | max_position=max_position_embeddings, |
| 506 | rope_parameters=config.rope_parameters, |
| 507 | is_neox_style=False, |
| 508 | ) |
| 509 | |
| 510 | if config.rope_parameters["rope_type"] == "deepseek_yarn": |
| 511 | mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) |
| 512 | scaling_factor = config.rope_parameters["factor"] |
| 513 | mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) |
| 514 | self.scaling = self.scaling * mscale * mscale |
| 515 | |
| 516 | mla_modules = MLAModules( |
| 517 | kv_a_layernorm=self.kv_a_layernorm, |
| 518 | kv_b_proj=self.kv_b_proj, |
| 519 | rotary_emb=self.rotary_emb, |
| 520 | o_proj=self.o_proj, |
| 521 | fused_qkv_a_proj=self.fused_qkv_a_proj |
| 522 | if self.q_lora_rank is not None |
| 523 | else None, |
| 524 | kv_a_proj_with_mqa=self.kv_a_proj_with_mqa |
| 525 | if self.q_lora_rank is None |
| 526 | else None, |
| 527 | q_a_layernorm=self.q_a_layernorm if self.q_lora_rank is not None else None, |
| 528 | q_b_proj=self.q_b_proj if self.q_lora_rank is not None else None, |
| 529 | q_proj=self.q_proj if self.q_lora_rank is None else None, |
| 530 | indexer=None, |
| 531 | indexer_rotary_emb=None, |
| 532 | is_sparse=False, |
| 533 | topk_indices_buffer=topk_indices_buffer, |
| 534 | ) |
| 535 | |
| 536 | self.mla_attn = MultiHeadLatentAttentionWrapper( |
| 537 | self.hidden_size, |
| 538 | self.num_local_heads, |
| 539 | self.scaling, |
| 540 | self.qk_nope_head_dim, |
| 541 | self.qk_rope_head_dim, |
| 542 | self.v_head_dim, |
| 543 | self.q_lora_rank, |
| 544 | self.kv_lora_rank, |
| 545 | mla_modules, |
| 546 | cache_config, |
| 547 | quant_config, |
| 548 | prefix, |
| 549 | ) |
| 550 | |
| 551 | def forward( |
| 552 | self, |
| 553 | positions: torch.Tensor, |
| 554 | hidden_states: torch.Tensor, |
| 555 | llama_4_scaling: torch.Tensor | None, |
| 556 | ) -> torch.Tensor: |
| 557 | return self.mla_attn(positions, hidden_states, llama_4_scaling) |
| 558 | |
| 559 | |
| 560 | class AXK1DecoderLayer(nn.Module): |
| 561 | def __init__( |
| 562 | self, |
| 563 | vllm_config: VllmConfig, |
| 564 | prefix: str, |
| 565 | config: AXK1Config | None = None, |
| 566 | ) -> None: |
| 567 | super().__init__() |
| 568 | |
| 569 | if config is None: |
| 570 | config = vllm_config.model_config.hf_config |
| 571 | model_config = vllm_config.model_config |
| 572 | cache_config = vllm_config.cache_config |
| 573 | quant_config = vllm_config.quant_config |
| 574 | parallel_config = vllm_config.parallel_config |
| 575 | self.config = config |
| 576 | |
| 577 | assert config.max_position_embeddings is not None |
| 578 | assert config.hidden_act is not None |
| 579 | assert config.routed_scaling_factor is not None |
| 580 | assert config.qk_nope_head_dim is not None |
| 581 | assert config.qk_rope_head_dim is not None |
| 582 | assert config.v_head_dim is not None |
| 583 | assert config.kv_lora_rank is not None |
| 584 | self.hidden_size = config.hidden_size |
| 585 | max_position_embeddings = config.max_position_embeddings |
| 586 | # DecoderLayers are created with `make_layers` which passes the prefix |
| 587 | # with the layer's index. |
| 588 | layer_idx = int(prefix.split(sep=".")[-1]) |
| 589 | self.layer_idx = layer_idx |
| 590 | |
| 591 | # verify MLA attention specific fields |
| 592 | qk_nope_head_dim = config.qk_nope_head_dim |
| 593 | qk_rope_head_dim = config.qk_rope_head_dim |
| 594 | v_head_dim = config.v_head_dim |
| 595 | kv_lora_rank = config.kv_lora_rank |
| 596 | use_mha = all(dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)) |
| 597 | self.use_mha = use_mha |
| 598 | |
| 599 | attn_cls: type[nn.Module] |
| 600 | if use_mha: |
| 601 | attn_cls = DeepseekAttention |
| 602 | elif model_config.use_mla: |
| 603 | attn_cls = AXK1MLAAttention |
| 604 | else: |
| 605 | attn_cls = AXK1Attention |
| 606 | self.self_attn = attn_cls( |
| 607 | vllm_config=vllm_config, |
| 608 | config=config, |
| 609 | hidden_size=self.hidden_size, |
| 610 | num_heads=config.num_attention_heads, |
| 611 | qk_nope_head_dim=qk_nope_head_dim, |
| 612 | qk_rope_head_dim=qk_rope_head_dim, |
| 613 | v_head_dim=v_head_dim, |
| 614 | q_lora_rank=config.q_lora_rank, |
| 615 | kv_lora_rank=kv_lora_rank, |
| 616 | max_position_embeddings=max_position_embeddings, |
| 617 | cache_config=cache_config, |
| 618 | quant_config=quant_config, |
| 619 | prefix=f"{prefix}.self_attn", |
| 620 | topk_indices_buffer=None, |
| 621 | ) |
| 622 | |
| 623 | self.is_layer_sparse = self._is_layer_sparse() |
| 624 | if self.is_layer_sparse: |
| 625 | self.mlp = AXK1MoE( |
| 626 | config=config, |
| 627 | parallel_config=parallel_config, |
| 628 | quant_config=quant_config, |
| 629 | prefix=f"{prefix}.mlp", |
| 630 | ) |
| 631 | else: |
| 632 | self.mlp = AXK1MLP( |
| 633 | hidden_size=config.hidden_size, |
| 634 | intermediate_size=config.intermediate_size, |
| 635 | hidden_act=config.hidden_act, |
| 636 | quant_config=quant_config, |
| 637 | prefix=f"{prefix}.mlp", |
| 638 | ) |
| 639 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 640 | self.post_attention_layernorm = RMSNorm( |
| 641 | config.hidden_size, eps=config.rms_norm_eps |
| 642 | ) |
| 643 | self.post_mlp_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 644 | self.routed_scaling_factor = config.routed_scaling_factor |
| 645 | |
| 646 | def _is_layer_sparse(self) -> bool: |
| 647 | assert self.config.moe_layer_freq is not None |
| 648 | return ( |
| 649 | self.config.n_routed_experts is not None |
| 650 | and self.layer_idx >= self.config.first_k_dense_replace |
| 651 | and self.layer_idx % self.config.moe_layer_freq == 0 |
| 652 | ) |
| 653 | |
| 654 | def forward( |
| 655 | self, |
| 656 | positions: torch.Tensor, |
| 657 | hidden_states: torch.Tensor, |
| 658 | residual: torch.Tensor | None, |
| 659 | llama_4_scaling: torch.Tensor | None = None, |
| 660 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 661 | # Self Attention |
| 662 | if residual is None: |
| 663 | residual = hidden_states |
| 664 | hidden_states = self.input_layernorm(hidden_states) |
| 665 | else: |
| 666 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 667 | |
| 668 | attn_kwargs = { |
| 669 | "positions": positions, |
| 670 | "hidden_states": hidden_states, |
| 671 | } |
| 672 | if not self.use_mha: |
| 673 | attn_kwargs["llama_4_scaling"] = llama_4_scaling |
| 674 | hidden_states = self.self_attn(**attn_kwargs) |
| 675 | |
| 676 | if ( |
| 677 | not isinstance(self.self_attn, DeepseekAttention) |
| 678 | and hidden_states.dtype == torch.float16 |
| 679 | ): |
| 680 | # Fix FP16 overflow |
| 681 | # We scale both hidden_states and residual before |
| 682 | # rmsnorm, and rmsnorm result would not affect by scale. |
| 683 | hidden_states *= 1.0 / self.routed_scaling_factor |
| 684 | if self.layer_idx == 0: |
| 685 | # The residual is shared by all layers, we only scale it on |
| 686 | # first layer. |
| 687 | residual *= 1.0 / self.routed_scaling_factor |
| 688 | |
| 689 | # Fully Connected |
| 690 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 691 | hidden_states = self.mlp(hidden_states) |
| 692 | |
| 693 | if self.is_layer_sparse: |
| 694 | hidden_states = self.post_mlp_layernorm(hidden_states) |
| 695 | |
| 696 | if isinstance(self.mlp, AXK1MLP) and hidden_states.dtype == torch.float16: |
| 697 | # Fix FP16 overflow |
| 698 | # Scaling the AXK1MLP output, it is the input of |
| 699 | # input_layernorm of next decoder layer. |
| 700 | # The scaling of AXK1MOE output would be done in the forward |
| 701 | # of AXK1MOE |
| 702 | hidden_states *= 1.0 / self.routed_scaling_factor |
| 703 | |
| 704 | return hidden_states, residual |
| 705 | |
| 706 | |
| 707 | @support_torch_compile |
| 708 | class AXK1Model(nn.Module): |
| 709 | fall_back_to_pt_during_load = False |
| 710 | |
| 711 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 712 | super().__init__() |
| 713 | |
| 714 | config: AXK1Config = vllm_config.model_config.hf_config |
| 715 | quant_config = vllm_config.quant_config |
| 716 | self.config = config |
| 717 | self.device = current_platform.device_type |
| 718 | self.vocab_size = config.vocab_size |
| 719 | |
| 720 | if get_pp_group().is_first_rank: |
| 721 | self.embed_tokens = VocabParallelEmbedding( |
| 722 | config.vocab_size, |
| 723 | config.hidden_size, |
| 724 | quant_config=quant_config, |
| 725 | prefix=f"{prefix}.embed_tokens", |
| 726 | ) |
| 727 | else: |
| 728 | self.embed_tokens = PPMissingLayer() |
| 729 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 730 | config.num_hidden_layers, |
| 731 | lambda prefix: AXK1DecoderLayer(vllm_config, prefix), |
| 732 | prefix=f"{prefix}.layers", |
| 733 | ) |
| 734 | |
| 735 | self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible( |
| 736 | self.layers, |
| 737 | AXK1MoE, |
| 738 | "mlp", |
| 739 | ) |
| 740 | |
| 741 | if get_pp_group().is_last_rank: |
| 742 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 743 | else: |
| 744 | self.norm = PPMissingLayer() |
| 745 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 746 | ["hidden_states", "residual"], config.hidden_size |
| 747 | ) |
| 748 | self.use_mha = all( |
| 749 | dim == 0 for dim in (config.qk_nope_head_dim, config.qk_rope_head_dim) |
| 750 | ) |
| 751 | self.fuse_qkv_a_proj = config.q_lora_rank is not None |
| 752 | self.num_redundant_experts = ( |
| 753 | vllm_config.parallel_config.eplb_config.num_redundant_experts |
| 754 | ) |
| 755 | |
| 756 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 757 | return self.embed_tokens(input_ids) |
| 758 | |
| 759 | def forward( |
| 760 | self, |
| 761 | input_ids: torch.Tensor | None, |
| 762 | positions: torch.Tensor, |
| 763 | intermediate_tensors: IntermediateTensors | None, |
| 764 | inputs_embeds: torch.Tensor | None = None, |
| 765 | ) -> torch.Tensor | IntermediateTensors: |
| 766 | if get_pp_group().is_first_rank: |
| 767 | if inputs_embeds is not None: |
| 768 | hidden_states = inputs_embeds |
| 769 | else: |
| 770 | hidden_states = self.embed_input_ids(input_ids) |
| 771 | residual = None |
| 772 | else: |
| 773 | assert intermediate_tensors is not None |
| 774 | hidden_states = intermediate_tensors["hidden_states"] |
| 775 | residual = intermediate_tensors["residual"] |
| 776 | |
| 777 | # Compute llama 4 scaling once per forward pass if enabled |
| 778 | llama_4_scaling_config = getattr(self.config, "llama_4_scaling", None) |
| 779 | llama_4_scaling: torch.Tensor | None |
| 780 | if llama_4_scaling_config is not None: |
| 781 | llama_4_scaling = _get_llama_4_scaling( |
| 782 | original_max_position_embeddings=llama_4_scaling_config[ |
| 783 | "original_max_position_embeddings" |
| 784 | ], |
| 785 | scaling_beta=llama_4_scaling_config["beta"], |
| 786 | positions=positions, |
| 787 | ) |
| 788 | else: |
| 789 | llama_4_scaling = None |
| 790 | |
| 791 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 792 | hidden_states, residual = layer( |
| 793 | positions, hidden_states, residual, llama_4_scaling |
| 794 | ) |
| 795 | |
| 796 | if not get_pp_group().is_last_rank: |
| 797 | return IntermediateTensors( |
| 798 | {"hidden_states": hidden_states, "residual": residual} |
| 799 | ) |
| 800 | |
| 801 | hidden_states, _ = self.norm(hidden_states, residual) |
| 802 | return hidden_states |
| 803 | |
| 804 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 805 | stacked_params_mapping: list[tuple[str, str, int | str]] = [ |
| 806 | # (param_name, shard_name, shard_id) |
| 807 | ("gate_up_proj", "gate_proj", 0), |
| 808 | ("gate_up_proj", "up_proj", 1), |
| 809 | ] |
| 810 | mla_params_mapping = [ |
| 811 | ("fused_qkv_a_proj", "q_a_proj", 0), |
| 812 | ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), |
| 813 | ] |
| 814 | mha_params_mapping = [ |
| 815 | ("qkv_proj", "q_proj", "q"), |
| 816 | ("qkv_proj", "k_proj", "k"), |
| 817 | ("qkv_proj", "v_proj", "v"), |
| 818 | ] |
| 819 | if self.use_mha: |
| 820 | stacked_params_mapping.extend(mha_params_mapping) |
| 821 | else: |
| 822 | stacked_params_mapping.extend(mla_params_mapping) |
| 823 | |
| 824 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 825 | # (param_name, weight_name, expert_id, shard_id) |
| 826 | assert self.config.n_routed_experts is not None |
| 827 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 828 | self, |
| 829 | ckpt_gate_proj_name="gate_proj", |
| 830 | ckpt_down_proj_name="down_proj", |
| 831 | ckpt_up_proj_name="up_proj", |
| 832 | num_experts=self.config.n_routed_experts |
| 833 | + ( |
| 834 | (self.config.n_shared_experts or 0) |
| 835 | if self.is_fused_shared_expert_enabled |
| 836 | else 0 |
| 837 | ), |
| 838 | num_redundant_experts=self.num_redundant_experts, |
| 839 | ) |
| 840 | |
| 841 | params_dict = dict(self.named_parameters()) |
| 842 | loaded_params: set[str] = set() |
| 843 | for name, loaded_weight in weights: |
| 844 | if "rotary_emb.inv_freq" in name: |
| 845 | continue |
| 846 | |
| 847 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) |
| 848 | if spec_layer is not None: |
| 849 | continue # skip spec decode layers for main model |
| 850 | |
| 851 | is_fusion_moe_shared_experts_layer = ( |
| 852 | self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name) |
| 853 | ) |
| 854 | |
| 855 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 856 | # Skip non-stacked layers and experts (experts handled below). |
| 857 | if weight_name not in name: |
| 858 | continue |
| 859 | # We have mlp.experts[0].gate_proj in the checkpoint. |
| 860 | # Since we handle the experts below in expert_params_mapping, |
| 861 | # we need to skip here BEFORE we update the name, otherwise |
| 862 | # name will be updated to mlp.experts[0].gate_up_proj, which |
| 863 | # will then be updated below in expert_params_mapping |
| 864 | # for mlp.experts[0].gate_gate_up_proj, which breaks load. |
| 865 | if ("mlp.experts." in name) and name not in params_dict: |
| 866 | continue |
| 867 | if is_fusion_moe_shared_experts_layer: |
| 868 | continue |
| 869 | name_mapped = name.replace(weight_name, param_name) |
| 870 | |
| 871 | # QKV fusion is optional, fall back to normal |
| 872 | # weight loading if it's not enabled |
| 873 | # if go with fusion option, then update name |
| 874 | if ( |
| 875 | param_name == "fused_qkv_a_proj" |
| 876 | ) and name_mapped not in params_dict: |
| 877 | continue |
| 878 | else: |
| 879 | name = name_mapped |
| 880 | # Skip loading extra bias for GPTQ models. |
| 881 | if name.endswith(".bias") and name not in params_dict: |
| 882 | continue |
| 883 | |
| 884 | if is_pp_missing_parameter(name, self): |
| 885 | continue |
| 886 | |
| 887 | param = params_dict[name] |
| 888 | weight_loader = param.weight_loader |
| 889 | weight_loader(param, loaded_weight, shard_id) |
| 890 | break |
| 891 | else: |
| 892 | is_expert_weight = False |
| 893 | |
| 894 | # Special handling: when AITER fusion_shared_experts is enabled, |
| 895 | # checkpoints may provide a single widened shared_experts tensor |
| 896 | # without explicit expert indices |
| 897 | # (e.g. ...mlp.shared_experts.gate_proj.weight). |
| 898 | # For models with multiple shared experts, split that tensor |
| 899 | # evenly into per-shared-expert slices and load them into |
| 900 | # appended expert slots mlp.experts.{n_routed_experts + j}.* |
| 901 | # accordingly. |
| 902 | num_chunks = 1 |
| 903 | if is_fusion_moe_shared_experts_layer: |
| 904 | num_chunks = getattr(self.config, "n_shared_experts", 1) or 1 |
| 905 | # Determine split axis based on op type |
| 906 | # gate/up: ColumnParallel → split along dim 0 |
| 907 | # down: RowParallel → split along dim 1 |
| 908 | split_dim = ( |
| 909 | 1 |
| 910 | if ("down_proj.weight" in name and loaded_weight.ndim > 1) |
| 911 | else 0 |
| 912 | ) |
| 913 | total = loaded_weight.shape[split_dim] |
| 914 | assert total % num_chunks == 0, ( |
| 915 | f"Shared expert weight dim {total} " |
| 916 | f"not divisible by num_chunks {num_chunks}" |
| 917 | ) |
| 918 | chunk_size = total // num_chunks |
| 919 | |
| 920 | for j in range(num_chunks): |
| 921 | chunk_name = name |
| 922 | weight_to_load = loaded_weight |
| 923 | |
| 924 | if is_fusion_moe_shared_experts_layer: |
| 925 | chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size) |
| 926 | if loaded_weight.ndim == 1: |
| 927 | weight_to_load = loaded_weight[chunk_slice] |
| 928 | elif split_dim == 0: |
| 929 | weight_to_load = loaded_weight[chunk_slice, :] |
| 930 | else: |
| 931 | weight_to_load = loaded_weight[:, chunk_slice] |
| 932 | # Synthesize an expert-style name so expert mapping |
| 933 | # can route it |
| 934 | chunk_name = name.replace( |
| 935 | "mlp.shared_experts", |
| 936 | f"mlp.experts.{self.config.n_routed_experts + j}", |
| 937 | ) |
| 938 | |
| 939 | # Use expert_params_mapping to locate the destination |
| 940 | # param and delegate to its expert-aware weight_loader |
| 941 | # with expert_id. |
| 942 | for mapping in expert_params_mapping: |
| 943 | param_name, weight_name, expert_id, expert_shard_id = mapping |
| 944 | if weight_name not in chunk_name: |
| 945 | continue |
| 946 | |
| 947 | # Anyway, this is an expert weight and should not be |
| 948 | # attempted to load as other weights later |
| 949 | is_expert_weight = True |
| 950 | |
| 951 | # Do not modify `name` since the loop may continue here |
| 952 | # Instead, create a new variable |
| 953 | name_mapped = chunk_name.replace(weight_name, param_name) |
| 954 | |
| 955 | if is_pp_missing_parameter(name_mapped, self): |
| 956 | continue |
| 957 | |
| 958 | param = params_dict[name_mapped] |
| 959 | # We should ask the weight loader to return success or |
| 960 | # not here since otherwise we may skip experts with |
| 961 | # other available replicas. |
| 962 | weight_loader = typing.cast( |
| 963 | Callable[..., bool], param.weight_loader |
| 964 | ) |
| 965 | success = weight_loader( |
| 966 | param, |
| 967 | weight_to_load, |
| 968 | name_mapped, |
| 969 | shard_id=expert_shard_id, |
| 970 | expert_id=expert_id, |
| 971 | return_success=True, |
| 972 | ) |
| 973 | if success: |
| 974 | if not is_fusion_moe_shared_experts_layer: |
| 975 | name = name_mapped |
| 976 | else: |
| 977 | loaded_params.add(name_mapped) |
| 978 | break |
| 979 | else: |
| 980 | if is_expert_weight: |
| 981 | # We've checked that this is an expert weight |
| 982 | # However it's not mapped locally to this rank |
| 983 | # So we simply skip it |
| 984 | continue |
| 985 | |
| 986 | # Skip loading extra bias for GPTQ models. |
| 987 | if name.endswith(".bias") and name not in params_dict: |
| 988 | continue |
| 989 | |
| 990 | # Remapping the name of FP8 kv-scale. |
| 991 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 992 | if remapped_name is None: |
| 993 | continue |
| 994 | name = remapped_name |
| 995 | |
| 996 | if is_pp_missing_parameter(name, self): |
| 997 | continue |
| 998 | |
| 999 | param = params_dict[name] |
| 1000 | weight_loader = getattr( |
| 1001 | param, "weight_loader", default_weight_loader |
| 1002 | ) |
| 1003 | weight_loader(param, loaded_weight) |
| 1004 | if not is_fusion_moe_shared_experts_layer: |
| 1005 | loaded_params.add(name) |
| 1006 | |
| 1007 | return loaded_params |
| 1008 | |
| 1009 | |
| 1010 | class AXK1MixtureOfExperts(MixtureOfExperts): |
| 1011 | moe_mlp_layers: list[AXK1MoE] |
| 1012 | """ |
| 1013 | List of MoE MLP layers in the model. |
| 1014 | """ |
| 1015 | |
| 1016 | def extract_moe_parameters(self, example_moe: AXK1MoE | None): |
| 1017 | if example_moe is None: |
| 1018 | self.num_moe_layers = 0 |
| 1019 | self.num_expert_groups = 0 |
| 1020 | self.num_logical_experts = 0 |
| 1021 | self.num_physical_experts = 0 |
| 1022 | self.num_local_physical_experts = 0 |
| 1023 | self.num_routed_experts = 0 |
| 1024 | self.num_shared_experts = 0 |
| 1025 | self.num_redundant_experts = 0 |
| 1026 | logger.warning("AXK1: No AXK1MoE layer found in model.layers.") |
| 1027 | else: |
| 1028 | self.num_logical_experts = example_moe.n_logical_experts |
| 1029 | self.num_physical_experts = example_moe.n_physical_experts |
| 1030 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 1031 | self.num_routed_experts = example_moe.n_routed_experts |
| 1032 | self.num_shared_experts = example_moe.n_shared_experts or 0 |
| 1033 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 1034 | |
| 1035 | def update_physical_experts_metadata( |
| 1036 | self, |
| 1037 | num_physical_experts: int, |
| 1038 | num_local_physical_experts: int, |
| 1039 | ) -> None: |
| 1040 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1041 | self.num_physical_experts = num_physical_experts |
| 1042 | self.num_local_physical_experts = num_local_physical_experts |
| 1043 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1044 | for moe in self.moe_mlp_layers: |
| 1045 | moe.n_local_physical_experts = num_local_physical_experts |
| 1046 | moe.n_physical_experts = num_physical_experts |
| 1047 | moe.n_redundant_experts = self.num_redundant_experts |
| 1048 | moe.experts.update_expert_map() |
| 1049 | |
| 1050 | |
| 1051 | class AXK1ForCausalLM( |
| 1052 | nn.Module, SupportsPP, AXK1MixtureOfExperts, SupportsLoRA, SupportsEagle |
| 1053 | ): |
| 1054 | packed_modules_mapping = { |
| 1055 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 1056 | } |
| 1057 | model_cls = AXK1Model |
| 1058 | |
| 1059 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1060 | super().__init__() |
| 1061 | config: AXK1Config = vllm_config.model_config.hf_config |
| 1062 | quant_config = vllm_config.quant_config |
| 1063 | self.config = config |
| 1064 | self.quant_config = quant_config |
| 1065 | |
| 1066 | qk_nope_head_dim = config.qk_nope_head_dim |
| 1067 | qk_rope_head_dim = config.qk_rope_head_dim |
| 1068 | self.use_mha = all(dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)) |
| 1069 | |
| 1070 | if self.use_mha: |
| 1071 | self.packed_modules_mapping["qkv_proj"] = ["q_proj", "k_proj", "v_proj"] |
| 1072 | |
| 1073 | # `packed_modules_mapping` needs to be modified before |
| 1074 | # initializing AXK1Model, as it is passed inplace to |
| 1075 | # quantization config init and may be used to select the |
| 1076 | # quant_method for relevant layers during initialization. |
| 1077 | self.fuse_qkv_a_proj = config.q_lora_rank is not None |
| 1078 | if self.fuse_qkv_a_proj: |
| 1079 | self.packed_modules_mapping["fused_qkv_a_proj"] = [ |
| 1080 | "q_a_proj", |
| 1081 | "kv_a_proj_with_mqa", |
| 1082 | ] |
| 1083 | |
| 1084 | self.model = self.model_cls( |
| 1085 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1086 | ) |
| 1087 | if get_pp_group().is_last_rank: |
| 1088 | self.lm_head = ParallelLMHead( |
| 1089 | config.vocab_size, |
| 1090 | config.hidden_size, |
| 1091 | quant_config=quant_config, |
| 1092 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1093 | ) |
| 1094 | else: |
| 1095 | self.lm_head = PPMissingLayer() |
| 1096 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1097 | self.make_empty_intermediate_tensors = ( |
| 1098 | self.model.make_empty_intermediate_tensors |
| 1099 | ) |
| 1100 | # Set MoE hyperparameters |
| 1101 | self.num_moe_layers = ( |
| 1102 | self.config.num_hidden_layers - self.config.first_k_dense_replace |
| 1103 | ) |
| 1104 | self.set_moe_parameters() |
| 1105 | |
| 1106 | def set_moe_parameters(self): |
| 1107 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 1108 | |
| 1109 | self.moe_layers = [] |
| 1110 | self.moe_mlp_layers = [] |
| 1111 | example_moe = None |
| 1112 | for layer in self.model.layers: |
| 1113 | if isinstance(layer, PPMissingLayer): |
| 1114 | continue |
| 1115 | |
| 1116 | assert isinstance(layer, AXK1DecoderLayer) |
| 1117 | if isinstance(layer.mlp, AXK1MoE): |
| 1118 | # Pick last one layer since the first ones may be dense layers. |
| 1119 | example_moe = layer.mlp |
| 1120 | self.moe_mlp_layers.append(layer.mlp) |
| 1121 | self.moe_layers.append(layer.mlp.experts) |
| 1122 | |
| 1123 | self.extract_moe_parameters(example_moe) |
| 1124 | |
| 1125 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1126 | return self.model.embed_input_ids(input_ids) |
| 1127 | |
| 1128 | def forward( |
| 1129 | self, |
| 1130 | input_ids: torch.Tensor | None, |
| 1131 | positions: torch.Tensor, |
| 1132 | intermediate_tensors: IntermediateTensors | None = None, |
| 1133 | inputs_embeds: torch.Tensor | None = None, |
| 1134 | ) -> torch.Tensor | IntermediateTensors: |
| 1135 | hidden_states = self.model( |
| 1136 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 1137 | ) |
| 1138 | return hidden_states |
| 1139 | |
| 1140 | def compute_logits( |
| 1141 | self, |
| 1142 | hidden_states: torch.Tensor, |
| 1143 | ) -> torch.Tensor | None: |
| 1144 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1145 | return logits |
| 1146 | |
| 1147 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1148 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1149 | # (param_name, weight_name, expert_id, shard_id) |
| 1150 | assert self.config.n_routed_experts is not None |
| 1151 | return fused_moe_make_expert_params_mapping( |
| 1152 | self, |
| 1153 | ckpt_gate_proj_name="gate_proj", |
| 1154 | ckpt_down_proj_name="down_proj", |
| 1155 | ckpt_up_proj_name="up_proj", |
| 1156 | num_experts=self.config.n_routed_experts, |
| 1157 | num_redundant_experts=0, |
| 1158 | ) |
| 1159 | |
| 1160 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1161 | loader = AutoWeightsLoader(self) |
| 1162 | return loader.load_weights(weights) |