| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """vLLM implementation for BailingMoeV3ForCausalLM. |
| 4 | |
| 5 | The HuggingFace reference model mixes MLA full-attention layers with Kimi |
| 6 | Delta Attention linear layers and Bailing MoE blocks. This file keeps the V3 |
| 7 | module/weight names aligned with the reference implementation while reusing |
| 8 | vLLM's parallel linear layers, MLA kernel, KDA kernel and fused MoE loader. |
| 9 | """ |
| 10 | |
| 11 | import copy |
| 12 | from collections.abc import Callable, Iterable |
| 13 | from math import lcm |
| 14 | from typing import TypeGuard |
| 15 | |
| 16 | import regex as re |
| 17 | import torch |
| 18 | import torch.nn as nn |
| 19 | from einops import rearrange |
| 20 | from transformers.configuration_utils import PreTrainedConfig |
| 21 | |
| 22 | from vllm.compilation.decorators import support_torch_compile |
| 23 | from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config |
| 24 | from vllm.distributed import ( |
| 25 | get_pp_group, |
| 26 | get_tensor_model_parallel_rank, |
| 27 | get_tensor_model_parallel_world_size, |
| 28 | ) |
| 29 | from vllm.forward_context import ForwardContext, get_forward_context |
| 30 | from vllm.model_executor.custom_op import PluggableLayer |
| 31 | from vllm.model_executor.layers.activation import SiluAndMul, SwigluStepAndMul |
| 32 | from vllm.model_executor.layers.fused_moe import ( |
| 33 | FusedMoEFactory, |
| 34 | GateLinear, |
| 35 | fused_moe_make_expert_params_mapping, |
| 36 | ) |
| 37 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 38 | from vllm.model_executor.layers.linear import ( |
| 39 | ColumnParallelLinear, |
| 40 | MergedColumnParallelLinear, |
| 41 | ReplicatedLinear, |
| 42 | RowParallelLinear, |
| 43 | ) |
| 44 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 45 | from vllm.model_executor.layers.mamba.abstract import MambaBase |
| 46 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 47 | MambaStateCopyFuncCalculator, |
| 48 | MambaStateDtypeCalculator, |
| 49 | MambaStateShapeCalculator, |
| 50 | is_conv_state_dim_first, |
| 51 | ) |
| 52 | from vllm.model_executor.layers.mamba.ops.causal_conv1d import ( |
| 53 | causal_conv1d_fn, |
| 54 | causal_conv1d_update, |
| 55 | ) |
| 56 | from vllm.model_executor.layers.mla import ( |
| 57 | MLAModules, |
| 58 | MultiHeadLatentAttentionWrapper, |
| 59 | ) |
| 60 | from vllm.model_executor.layers.quantization.base_config import QuantizationConfig |
| 61 | from vllm.model_executor.layers.quantization.fp8 import Fp8Config |
| 62 | from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped |
| 63 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 64 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 65 | ParallelLMHead, |
| 66 | VocabParallelEmbedding, |
| 67 | ) |
| 68 | from vllm.model_executor.model_loader.weight_utils import ( |
| 69 | default_weight_loader, |
| 70 | maybe_remap_kv_scale_name, |
| 71 | sharded_weight_loader, |
| 72 | ) |
| 73 | from vllm.model_executor.utils import set_weight_attrs |
| 74 | from vllm.sequence import IntermediateTensors |
| 75 | from vllm.third_party.flash_linear_attention.ops.kda import ( |
| 76 | FusedRMSNormGated, |
| 77 | chunk_kda, |
| 78 | fused_kda_gate, |
| 79 | fused_recurrent_kda, |
| 80 | fused_recurrent_kda_fwd, |
| 81 | ) |
| 82 | from vllm.utils.torch_utils import direct_register_custom_op |
| 83 | from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata |
| 84 | from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum |
| 85 | |
| 86 | from .interfaces import HasInnerState, IsHybrid, SupportsPP |
| 87 | from .utils import ( |
| 88 | PPMissingLayer, |
| 89 | WeightsMapper, |
| 90 | is_pp_missing_parameter, |
| 91 | make_layers, |
| 92 | maybe_prefix, |
| 93 | ) |
| 94 | |
| 95 | |
| 96 | def bailing_v3_kda_attention( |
| 97 | q_proj_states: torch.Tensor, |
| 98 | k_proj_states: torch.Tensor, |
| 99 | v_proj_states: torch.Tensor, |
| 100 | g1: torch.Tensor, |
| 101 | beta: torch.Tensor, |
| 102 | core_attn_out: torch.Tensor, |
| 103 | layer_name: str, |
| 104 | ) -> None: |
| 105 | """Run Bailing V3's KDA state update outside the compiled graph.""" |
| 106 | forward_context: ForwardContext = get_forward_context() |
| 107 | layer = forward_context.no_compile_layers[layer_name] |
| 108 | layer._forward( |
| 109 | q_proj_states=q_proj_states, |
| 110 | k_proj_states=k_proj_states, |
| 111 | v_proj_states=v_proj_states, |
| 112 | g1=g1, |
| 113 | beta=beta, |
| 114 | core_attn_out=core_attn_out, |
| 115 | ) |
| 116 | |
| 117 | |
| 118 | direct_register_custom_op( |
| 119 | op_name="bailing_v3_kda_attention", |
| 120 | op_func=bailing_v3_kda_attention, |
| 121 | mutates_args=["core_attn_out"], |
| 122 | ) |
| 123 | |
| 124 | |
| 125 | def _is_kda_layer( |
| 126 | layer_idx: int, layer_group_size: int, num_hidden_layers: int |
| 127 | ) -> bool: |
| 128 | return not ( |
| 129 | (layer_idx + 1) % layer_group_size == 0 |
| 130 | or layer_idx >= num_hidden_layers // layer_group_size * layer_group_size |
| 131 | ) |
| 132 | |
| 133 | |
| 134 | def _get_kda_state_shape_for_config( |
| 135 | vllm_config: VllmConfig, |
| 136 | ) -> tuple[tuple[int, int], tuple[int, int, int]]: |
| 137 | config = vllm_config.model_config.hf_config |
| 138 | num_spec = ( |
| 139 | vllm_config.speculative_config.num_speculative_tokens |
| 140 | if vllm_config.speculative_config |
| 141 | else 0 |
| 142 | ) |
| 143 | shapes = MambaStateShapeCalculator.kda_state_shape( |
| 144 | vllm_config.parallel_config.tensor_parallel_size, |
| 145 | config.num_attention_heads, |
| 146 | config.head_dim, |
| 147 | conv_kernel_size=config.short_conv_kernel_size, |
| 148 | num_spec=num_spec, |
| 149 | ) |
| 150 | |
| 151 | return shapes |
| 152 | |
| 153 | |
| 154 | def _build_rope_parameters(config: PreTrainedConfig) -> dict | None: |
| 155 | rope_parameters = copy.deepcopy(getattr(config, "rope_parameters", None)) or {} |
| 156 | if "rope_theta" not in rope_parameters and hasattr(config, "rope_theta"): |
| 157 | rope_parameters["rope_theta"] = config.rope_theta |
| 158 | # BailingMoeV3 stores the MLA rotary width explicitly in qk_rope_head_dim. |
| 159 | # Some checkpoints also carry partial_rotary_factor=0.5, but applying it |
| 160 | # here would shrink the rotary cache to half of the qk_rope_head_dim. |
| 161 | rope_parameters.pop("partial_rotary_factor", None) |
| 162 | |
| 163 | rope_scaling = getattr(config, "rope_scaling", None) |
| 164 | if isinstance(rope_scaling, dict): |
| 165 | rope_scaling = copy.deepcopy(rope_scaling) |
| 166 | if "type" in rope_scaling and "rope_type" not in rope_scaling: |
| 167 | rope_scaling["rope_type"] = rope_scaling.pop("type") |
| 168 | rope_scaling.pop("partial_rotary_factor", None) |
| 169 | rope_parameters.update(rope_scaling) |
| 170 | |
| 171 | return rope_parameters or None |
| 172 | |
| 173 | |
| 174 | def _build_mla_rotary_embedding( |
| 175 | config: PreTrainedConfig, |
| 176 | head_size: int, |
| 177 | ) -> nn.Module: |
| 178 | rope_parameters = _build_rope_parameters(config) |
| 179 | if rope_parameters is not None and "mrope_section" in rope_parameters: |
| 180 | rope_type = rope_parameters.get("rope_type", "default") |
| 181 | if rope_type != "default": |
| 182 | raise ValueError( |
| 183 | f"Bailing M-RoPE only supports rope_type='default', got {rope_type!r}" |
| 184 | ) |
| 185 | rope_parameters["rope_type"] = "bailing_mrope" |
| 186 | |
| 187 | return get_rope( |
| 188 | head_size=head_size, |
| 189 | max_position=getattr(config, "max_position_embeddings", 8192), |
| 190 | is_neox_style=False, |
| 191 | rope_parameters=rope_parameters, |
| 192 | ) |
| 193 | |
| 194 | |
| 195 | def _get_layer_swiglu_limit(limit_list: list | None, layer_idx: int) -> float | None: |
| 196 | if limit_list is None or layer_idx >= len(limit_list): |
| 197 | return None |
| 198 | limit = limit_list[layer_idx] |
| 199 | if limit in (None, 0): |
| 200 | return None |
| 201 | return float(limit) |
| 202 | |
| 203 | |
| 204 | def _load_a_log(param: torch.nn.Parameter, loaded_weight: torch.Tensor) -> None: |
| 205 | tp_rank = get_tensor_model_parallel_rank() |
| 206 | tp_size = get_tensor_model_parallel_world_size() |
| 207 | if loaded_weight.ndim == 1: |
| 208 | shard_size = loaded_weight.shape[0] // tp_size |
| 209 | shard = loaded_weight.narrow(0, tp_rank * shard_size, shard_size) |
| 210 | param.data.copy_(shard.view_as(param.data)) |
| 211 | else: |
| 212 | sharded_weight_loader(2)(param, loaded_weight) |
| 213 | |
| 214 | |
| 215 | def _is_block_fp8_config( |
| 216 | quant_config: QuantizationConfig | None, |
| 217 | ) -> TypeGuard[Fp8Config]: |
| 218 | return ( |
| 219 | isinstance(quant_config, Fp8Config) |
| 220 | and quant_config.weight_block_size is not None |
| 221 | ) |
| 222 | |
| 223 | |
| 224 | def _configure_ling_fp8_quant_config( |
| 225 | quant_config: QuantizationConfig | None, |
| 226 | config: PreTrainedConfig, |
| 227 | ) -> None: |
| 228 | if not _is_block_fp8_config(quant_config): |
| 229 | return |
| 230 | |
| 231 | quant_config.ignored_layers_match_mode = "suffix" |
| 232 | hf_quant_config = getattr(config, "quantization_config", None) |
| 233 | if not isinstance(hf_quant_config, dict): |
| 234 | return |
| 235 | |
| 236 | quant_config.is_scale_e8m0 = ( # type: ignore[attr-defined] |
| 237 | hf_quant_config.get("scale_fmt") == "ue8m0" |
| 238 | ) |
| 239 | |
| 240 | routed_quant_method = hf_quant_config.get("routed_experts_quant_method") |
| 241 | if routed_quant_method is None: |
| 242 | return |
| 243 | if routed_quant_method != "mxfp4": |
| 244 | raise ValueError( |
| 245 | f"Unsupported routed experts quantization: {routed_quant_method!r}" |
| 246 | ) |
| 247 | |
| 248 | quant_config.store_dtype = "mxfp4" |
| 249 | |
| 250 | |
| 251 | _LING_MXFP4_WEIGHTS_MAPPER = WeightsMapper( |
| 252 | orig_to_new_regex={ |
| 253 | re.compile( |
| 254 | r"(\.mlp\.experts\.\d+\." |
| 255 | r"(?:gate_proj|up_proj|down_proj)\.weight_scale)_inv$" |
| 256 | ): r"\1" |
| 257 | } |
| 258 | ) |
| 259 | |
| 260 | |
| 261 | def _maybe_remap_ling_mxfp4_weight_names( |
| 262 | weights: Iterable[tuple[str, torch.Tensor]], |
| 263 | quant_config: QuantizationConfig | None, |
| 264 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 265 | """Map Ling's MXFP4 expert scales to Mxfp4MoEMethod parameters.""" |
| 266 | if isinstance(quant_config, Fp8Config) and quant_config.store_dtype == "mxfp4": |
| 267 | return _LING_MXFP4_WEIGHTS_MAPPER.apply(weights) |
| 268 | return weights |
| 269 | |
| 270 | |
| 271 | def _is_fp8_module_excluded( |
| 272 | quant_config: QuantizationConfig | None, |
| 273 | prefix: str, |
| 274 | ) -> bool: |
| 275 | """Match Ling's abbreviated FP8 exclusions against mapped vLLM prefixes.""" |
| 276 | if not _is_block_fp8_config(quant_config): |
| 277 | return False |
| 278 | |
| 279 | return is_layer_skipped( |
| 280 | prefix=prefix, |
| 281 | ignored_layers=quant_config.ignored_layers, |
| 282 | fused_mapping=quant_config.packed_modules_mapping, |
| 283 | match_mode=quant_config.ignored_layers_match_mode, |
| 284 | ) |
| 285 | |
| 286 | |
| 287 | def _get_block_fp8_mlp_padded_intermediate_size( |
| 288 | quant_config: QuantizationConfig | None, |
| 289 | intermediate_size: int, |
| 290 | prefix: str, |
| 291 | ) -> int: |
| 292 | """Pad a block-FP8 MLP so each TP shard contains whole quant blocks.""" |
| 293 | if not _is_block_fp8_config(quant_config): |
| 294 | return intermediate_size |
| 295 | block_size = quant_config.weight_block_size |
| 296 | assert block_size is not None |
| 297 | |
| 298 | alignments: list[int] = [] |
| 299 | if not _is_fp8_module_excluded(quant_config, f"{prefix}.gate_up_proj"): |
| 300 | alignments.append(int(block_size[0])) |
| 301 | if not _is_fp8_module_excluded(quant_config, f"{prefix}.down_proj"): |
| 302 | alignments.append(int(block_size[1])) |
| 303 | if not alignments: |
| 304 | return intermediate_size |
| 305 | |
| 306 | tp_size = get_tensor_model_parallel_world_size() |
| 307 | tp_alignment = tp_size * lcm(*alignments) |
| 308 | return (intermediate_size + tp_alignment - 1) // tp_alignment * tp_alignment |
| 309 | |
| 310 | |
| 311 | def _pad_block_fp8_mlp_checkpoint_tensor( |
| 312 | quant_config: QuantizationConfig, |
| 313 | name: str, |
| 314 | loaded_weight: torch.Tensor, |
| 315 | intermediate_size: int, |
| 316 | padded_intermediate_size: int, |
| 317 | ) -> torch.Tensor: |
| 318 | """Zero-pad an MLP checkpoint tensor on its intermediate dimension.""" |
| 319 | if padded_intermediate_size == intermediate_size: |
| 320 | return loaded_weight |
| 321 | |
| 322 | block_size = getattr(quant_config, "weight_block_size", None) |
| 323 | assert block_size is not None |
| 324 | |
| 325 | if ".down_proj." in name: |
| 326 | if name.endswith(".bias"): |
| 327 | return loaded_weight |
| 328 | dim = 1 |
| 329 | block = int(block_size[1]) |
| 330 | logical_shards = 1 |
| 331 | elif any( |
| 332 | projection in name |
| 333 | for projection in (".gate_proj.", ".up_proj.", ".gate_up_proj.") |
| 334 | ): |
| 335 | dim = 0 |
| 336 | block = int(block_size[0]) |
| 337 | logical_shards = 2 if ".gate_up_proj." in name else 1 |
| 338 | else: |
| 339 | return loaded_weight |
| 340 | |
| 341 | if name.endswith(".weight_scale_inv"): |
| 342 | expected_shard_size = (intermediate_size + block - 1) // block |
| 343 | target_shard_size = (padded_intermediate_size + block - 1) // block |
| 344 | elif name.endswith((".weight", ".bias")): |
| 345 | expected_shard_size = intermediate_size |
| 346 | target_shard_size = padded_intermediate_size |
| 347 | else: |
| 348 | return loaded_weight |
| 349 | |
| 350 | current_size = loaded_weight.shape[dim] |
| 351 | if current_size % logical_shards != 0: |
| 352 | raise ValueError( |
| 353 | f"Cannot split {name} dimension {current_size} into " |
| 354 | f"{logical_shards} logical shards." |
| 355 | ) |
| 356 | current_shard_size = current_size // logical_shards |
| 357 | if current_shard_size == target_shard_size: |
| 358 | return loaded_weight |
| 359 | if current_shard_size != expected_shard_size: |
| 360 | raise ValueError( |
| 361 | f"Cannot pad {name}: expected each logical intermediate dimension " |
| 362 | f"to be {expected_shard_size}, but got {current_shard_size}." |
| 363 | ) |
| 364 | |
| 365 | padded_shards: list[torch.Tensor] = [] |
| 366 | for shard in loaded_weight.split(current_shard_size, dim=dim): |
| 367 | pad_shape = list(shard.shape) |
| 368 | pad_shape[dim] = target_shard_size - current_shard_size |
| 369 | padded_shards.extend([shard, shard.new_zeros(pad_shape)]) |
| 370 | return torch.cat(padded_shards, dim=dim) |
| 371 | |
| 372 | |
| 373 | def _maybe_pad_block_fp8_shared_expert_checkpoint_tensor( |
| 374 | quant_config: QuantizationConfig | None, |
| 375 | config: PreTrainedConfig, |
| 376 | name: str, |
| 377 | loaded_weight: torch.Tensor, |
| 378 | ) -> torch.Tensor: |
| 379 | if ".mlp.shared_experts." not in name: |
| 380 | return loaded_weight |
| 381 | |
| 382 | shared_prefix = name.split(".shared_experts.", 1)[0] + ".shared_experts" |
| 383 | shared_intermediate = ( |
| 384 | config.moe_shared_expert_intermediate_size * config.num_shared_experts |
| 385 | ) |
| 386 | padded_shared_intermediate = _get_block_fp8_mlp_padded_intermediate_size( |
| 387 | quant_config, |
| 388 | shared_intermediate, |
| 389 | shared_prefix, |
| 390 | ) |
| 391 | if padded_shared_intermediate == shared_intermediate: |
| 392 | return loaded_weight |
| 393 | |
| 394 | assert quant_config is not None |
| 395 | return _pad_block_fp8_mlp_checkpoint_tensor( |
| 396 | quant_config, |
| 397 | name, |
| 398 | loaded_weight, |
| 399 | shared_intermediate, |
| 400 | padded_shared_intermediate, |
| 401 | ) |
| 402 | |
| 403 | |
| 404 | class _IdentityProjection(nn.Module): |
| 405 | def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, None]: |
| 406 | return x, None |
| 407 | |
| 408 | |
| 409 | class BailingMoeV3MLP(nn.Module): |
| 410 | def __init__( |
| 411 | self, |
| 412 | intermediate_size: int, |
| 413 | config: PreTrainedConfig, |
| 414 | quant_config: QuantizationConfig | None = None, |
| 415 | reduce_results: bool = True, |
| 416 | prefix: str = "", |
| 417 | swiglu_limit: float | None = None, |
| 418 | ) -> None: |
| 419 | super().__init__() |
| 420 | self.gate_up_proj = MergedColumnParallelLinear( |
| 421 | config.hidden_size, |
| 422 | [intermediate_size] * 2, |
| 423 | bias=config.use_bias, |
| 424 | quant_config=quant_config, |
| 425 | prefix=f"{prefix}.gate_up_proj", |
| 426 | ) |
| 427 | self.down_proj = RowParallelLinear( |
| 428 | intermediate_size, |
| 429 | config.hidden_size, |
| 430 | bias=config.use_bias, |
| 431 | quant_config=quant_config, |
| 432 | reduce_results=reduce_results, |
| 433 | prefix=f"{prefix}.down_proj", |
| 434 | ) |
| 435 | self.act_fn = ( |
| 436 | SwigluStepAndMul(limit=swiglu_limit) if swiglu_limit else SiluAndMul() |
| 437 | ) |
| 438 | |
| 439 | def forward(self, x): |
| 440 | x, _ = self.gate_up_proj(x) |
| 441 | x = self.act_fn(x) |
| 442 | x, _ = self.down_proj(x) |
| 443 | return x |
| 444 | |
| 445 | |
| 446 | class BailingMoeV3MLAAttention(nn.Module): |
| 447 | def __init__( |
| 448 | self, |
| 449 | config: PreTrainedConfig, |
| 450 | quant_config: QuantizationConfig | None = None, |
| 451 | layer_id: int = 0, |
| 452 | prefix: str = "attention", |
| 453 | cache_config: CacheConfig | None = None, |
| 454 | ) -> None: |
| 455 | super().__init__() |
| 456 | self.hidden_size = config.hidden_size |
| 457 | self.num_heads = config.num_attention_heads |
| 458 | self.layer_id = layer_id |
| 459 | |
| 460 | self.qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 128) |
| 461 | self.qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 64) |
| 462 | self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim |
| 463 | self.v_head_dim = getattr(config, "v_head_dim", 128) |
| 464 | self.q_lora_rank = getattr(config, "q_lora_rank", None) |
| 465 | self.kv_lora_rank = getattr(config, "kv_lora_rank", 512) |
| 466 | |
| 467 | tp_size = get_tensor_model_parallel_world_size() |
| 468 | assert self.num_heads % tp_size == 0 |
| 469 | self.num_local_heads = self.num_heads // tp_size |
| 470 | self.scaling = self.qk_head_dim**-0.5 |
| 471 | |
| 472 | self.q_proj: ColumnParallelLinear | None |
| 473 | self.kv_a_proj_with_mqa: ReplicatedLinear | None |
| 474 | if self.q_lora_rank is None: |
| 475 | self.q_proj = ColumnParallelLinear( |
| 476 | self.hidden_size, |
| 477 | self.num_heads * self.qk_head_dim, |
| 478 | bias=False, |
| 479 | quant_config=quant_config, |
| 480 | prefix=f"{prefix}.q_proj", |
| 481 | ) |
| 482 | self.fused_qkv_a_proj = None |
| 483 | self.q_a_layernorm = None |
| 484 | self.q_b_proj = None |
| 485 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 486 | self.hidden_size, |
| 487 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 488 | bias=False, |
| 489 | quant_config=quant_config, |
| 490 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 491 | ) |
| 492 | else: |
| 493 | self.fused_qkv_a_proj = MergedColumnParallelLinear( |
| 494 | self.hidden_size, |
| 495 | [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], |
| 496 | bias=False, |
| 497 | quant_config=quant_config, |
| 498 | prefix=f"{prefix}.fused_qkv_a_proj", |
| 499 | disable_tp=True, |
| 500 | ) |
| 501 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 502 | self.q_b_proj = ColumnParallelLinear( |
| 503 | self.q_lora_rank, |
| 504 | self.num_heads * self.qk_head_dim, |
| 505 | bias=False, |
| 506 | quant_config=quant_config, |
| 507 | prefix=f"{prefix}.q_b_proj", |
| 508 | ) |
| 509 | self.q_proj = None |
| 510 | self.kv_a_proj_with_mqa = None |
| 511 | |
| 512 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 513 | self.kv_b_proj = ColumnParallelLinear( |
| 514 | self.kv_lora_rank, |
| 515 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 516 | bias=False, |
| 517 | quant_config=quant_config, |
| 518 | prefix=f"{prefix}.kv_b_proj", |
| 519 | ) |
| 520 | self.gated_attention_proj_granularity_type = getattr( |
| 521 | config, "gated_attention_proj_granularity_type", None |
| 522 | ) |
| 523 | if self.gated_attention_proj_granularity_type == "head_wise": |
| 524 | g_out = self.num_heads |
| 525 | elif self.gated_attention_proj_granularity_type == "element_wise": |
| 526 | g_out = self.num_heads * self.v_head_dim |
| 527 | else: |
| 528 | g_out = 0 |
| 529 | self.g_proj = ( |
| 530 | ColumnParallelLinear( |
| 531 | self.hidden_size, |
| 532 | g_out, |
| 533 | bias=False, |
| 534 | quant_config=quant_config, |
| 535 | prefix=f"{prefix}.g_proj", |
| 536 | ) |
| 537 | if g_out > 0 |
| 538 | else None |
| 539 | ) |
| 540 | self.dense = RowParallelLinear( |
| 541 | self.num_heads * self.v_head_dim, |
| 542 | self.hidden_size, |
| 543 | bias=False, |
| 544 | quant_config=quant_config, |
| 545 | prefix=f"{prefix}.dense", |
| 546 | ) |
| 547 | |
| 548 | self.rotary_emb = _build_mla_rotary_embedding( |
| 549 | config, |
| 550 | head_size=self.qk_rope_head_dim, |
| 551 | ) |
| 552 | mla_modules = MLAModules( |
| 553 | kv_a_layernorm=self.kv_a_layernorm, |
| 554 | kv_b_proj=self.kv_b_proj, |
| 555 | rotary_emb=self.rotary_emb, |
| 556 | o_proj=_IdentityProjection(), |
| 557 | fused_qkv_a_proj=self.fused_qkv_a_proj, |
| 558 | kv_a_proj_with_mqa=self.kv_a_proj_with_mqa, |
| 559 | q_a_layernorm=self.q_a_layernorm, |
| 560 | q_b_proj=self.q_b_proj, |
| 561 | q_proj=self.q_proj, |
| 562 | indexer=None, |
| 563 | is_sparse=False, |
| 564 | topk_indices_buffer=None, |
| 565 | ) |
| 566 | self.mla_attn = MultiHeadLatentAttentionWrapper( |
| 567 | self.hidden_size, |
| 568 | self.num_local_heads, |
| 569 | self.scaling, |
| 570 | self.qk_nope_head_dim, |
| 571 | self.qk_rope_head_dim, |
| 572 | self.v_head_dim, |
| 573 | self.q_lora_rank, |
| 574 | self.kv_lora_rank, |
| 575 | mla_modules, |
| 576 | cache_config, |
| 577 | quant_config, |
| 578 | prefix, |
| 579 | ) |
| 580 | |
| 581 | def forward(self, hidden_states: torch.Tensor, positions: torch.Tensor): |
| 582 | attn_out = self.mla_attn(positions, hidden_states) |
| 583 | if self.g_proj is not None: |
| 584 | gate = torch.sigmoid(self.g_proj(hidden_states)[0].float()).to( |
| 585 | hidden_states.dtype |
| 586 | ) |
| 587 | if self.gated_attention_proj_granularity_type == "head_wise": |
| 588 | attn_out = attn_out.view(-1, self.num_local_heads, self.v_head_dim) |
| 589 | attn_out = attn_out * gate.unsqueeze(-1) |
| 590 | attn_out = attn_out.reshape(hidden_states.shape[0], -1) |
| 591 | else: |
| 592 | attn_out = attn_out * gate |
| 593 | |
| 594 | return self.dense(attn_out)[0] |
| 595 | |
| 596 | |
| 597 | # --8<-- [start:bailing_moe_v3_kimi_delta_attention] |
| 598 | @PluggableLayer.register("bailing_moe_v3_kimi_delta_attention") |
| 599 | class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase): |
| 600 | # --8<-- [end:bailing_moe_v3_kimi_delta_attention] |
| 601 | |
| 602 | @property |
| 603 | def mamba_type(self) -> MambaAttentionBackendEnum: |
| 604 | return MambaAttentionBackendEnum.GDN_ATTN |
| 605 | |
| 606 | def get_state_dtype( |
| 607 | self, |
| 608 | ) -> tuple[torch.dtype, torch.dtype]: |
| 609 | if self.model_config is None or self.cache_config is None: |
| 610 | raise ValueError("model_config and cache_config must be set") |
| 611 | return MambaStateDtypeCalculator.kda_state_dtype( |
| 612 | self.model_config.dtype, self.cache_config.mamba_cache_dtype |
| 613 | ) |
| 614 | |
| 615 | def get_state_shape( |
| 616 | self, |
| 617 | ) -> tuple[tuple[int, int], tuple[int, int, int]]: |
| 618 | return MambaStateShapeCalculator.kda_state_shape( |
| 619 | self.tp_size, |
| 620 | self.num_heads, |
| 621 | self.head_dim, |
| 622 | conv_kernel_size=self.conv_size, |
| 623 | num_spec=self.num_speculative_tokens, |
| 624 | ) |
| 625 | |
| 626 | def __init__( |
| 627 | self, |
| 628 | config: PreTrainedConfig, |
| 629 | quant_config: QuantizationConfig | None = None, |
| 630 | layer_id: int = 0, |
| 631 | prefix: str = "attention", |
| 632 | model_config: ModelConfig | None = None, |
| 633 | cache_config: CacheConfig | None = None, |
| 634 | num_speculative_tokens: int = 0, |
| 635 | ) -> None: |
| 636 | super().__init__() |
| 637 | self.tp_size = get_tensor_model_parallel_world_size() |
| 638 | self.hidden_size = config.hidden_size |
| 639 | self.head_dim = config.head_dim |
| 640 | self.num_heads = config.num_attention_heads |
| 641 | self.local_num_heads = self.num_heads // self.tp_size |
| 642 | self.conv_size = config.short_conv_kernel_size |
| 643 | self.layer_id = layer_id |
| 644 | self.prefix = prefix |
| 645 | self.model_config = model_config |
| 646 | self.cache_config = cache_config |
| 647 | self.num_speculative_tokens = num_speculative_tokens |
| 648 | self.safe_gate = getattr(config, "kda_safe_gate", True) |
| 649 | self.lower_bound = getattr(config, "kda_lower_bound", -5.0) |
| 650 | if not getattr(config, "no_kda_lora", False): |
| 651 | raise ValueError("BailingMoeV3 KDA currently expects no_kda_lora=True") |
| 652 | |
| 653 | projection_size = self.head_dim * self.num_heads |
| 654 | self.projection_size_per_partition = projection_size // self.tp_size |
| 655 | self.separate_b_proj = _is_fp8_module_excluded(quant_config, f"{prefix}.b_proj") |
| 656 | self.qkv_proj: MergedColumnParallelLinear | None |
| 657 | self.b_proj: ColumnParallelLinear | None |
| 658 | self.qkvb_proj: MergedColumnParallelLinear | None |
| 659 | if self.separate_b_proj: |
| 660 | self.qkv_proj = MergedColumnParallelLinear( |
| 661 | self.hidden_size, |
| 662 | [projection_size, projection_size, projection_size], |
| 663 | bias=False, |
| 664 | quant_config=quant_config, |
| 665 | prefix=f"{prefix}.qkv_proj", |
| 666 | ) |
| 667 | self.b_proj = ColumnParallelLinear( |
| 668 | self.hidden_size, |
| 669 | self.num_heads, |
| 670 | bias=False, |
| 671 | quant_config=quant_config, |
| 672 | prefix=f"{prefix}.b_proj", |
| 673 | ) |
| 674 | self.qkvb_proj = None |
| 675 | else: |
| 676 | self.qkvb_proj = MergedColumnParallelLinear( |
| 677 | self.hidden_size, |
| 678 | [projection_size, projection_size, projection_size, self.num_heads], |
| 679 | bias=False, |
| 680 | quant_config=quant_config, |
| 681 | prefix=f"{prefix}.qkvb_proj", |
| 682 | ) |
| 683 | self.qkv_proj = None |
| 684 | self.b_proj = None |
| 685 | self.f_proj = ColumnParallelLinear( |
| 686 | self.hidden_size, |
| 687 | projection_size, |
| 688 | bias=False, |
| 689 | quant_config=quant_config, |
| 690 | prefix=f"{prefix}.f_proj", |
| 691 | ) |
| 692 | self.dt_bias = nn.Parameter( |
| 693 | torch.empty(projection_size // self.tp_size, dtype=torch.float32) |
| 694 | ) |
| 695 | set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)}) |
| 696 | |
| 697 | self.q_conv1d = ColumnParallelLinear( |
| 698 | input_size=self.conv_size, |
| 699 | output_size=projection_size, |
| 700 | bias=False, |
| 701 | params_dtype=torch.float32, |
| 702 | prefix=f"{prefix}.q_conv1d", |
| 703 | ) |
| 704 | self.k_conv1d = ColumnParallelLinear( |
| 705 | input_size=self.conv_size, |
| 706 | output_size=projection_size, |
| 707 | bias=False, |
| 708 | params_dtype=torch.float32, |
| 709 | prefix=f"{prefix}.k_conv1d", |
| 710 | ) |
| 711 | self.v_conv1d = ColumnParallelLinear( |
| 712 | input_size=self.conv_size, |
| 713 | output_size=projection_size, |
| 714 | bias=False, |
| 715 | params_dtype=torch.float32, |
| 716 | prefix=f"{prefix}.v_conv1d", |
| 717 | ) |
| 718 | self.q_conv1d.weight.data = self.q_conv1d.weight.data.unsqueeze(1) |
| 719 | self.k_conv1d.weight.data = self.k_conv1d.weight.data.unsqueeze(1) |
| 720 | self.v_conv1d.weight.data = self.v_conv1d.weight.data.unsqueeze(1) |
| 721 | |
| 722 | self.A_log = nn.Parameter( |
| 723 | torch.empty(1, 1, self.local_num_heads, 1, dtype=torch.float32) |
| 724 | ) |
| 725 | set_weight_attrs(self.A_log, {"weight_loader": _load_a_log}) |
| 726 | |
| 727 | self.g_proj = ColumnParallelLinear( |
| 728 | self.hidden_size, |
| 729 | projection_size, |
| 730 | bias=False, |
| 731 | quant_config=quant_config, |
| 732 | prefix=f"{prefix}.g_proj", |
| 733 | ) |
| 734 | self.o_norm = FusedRMSNormGated( |
| 735 | self.head_dim, eps=config.rms_norm_eps, activation="sigmoid" |
| 736 | ) |
| 737 | self.o_proj = RowParallelLinear( |
| 738 | projection_size, |
| 739 | self.hidden_size, |
| 740 | bias=False, |
| 741 | quant_config=quant_config, |
| 742 | prefix=f"{prefix}.o_proj", |
| 743 | ) |
| 744 | |
| 745 | compilation_config = get_current_vllm_config().compilation_config |
| 746 | if prefix in compilation_config.static_forward_context: |
| 747 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 748 | compilation_config.static_forward_context[prefix] = self |
| 749 | splitting_op = "vllm::bailing_v3_kda_attention" |
| 750 | if ( |
| 751 | compilation_config.splitting_ops is not None |
| 752 | and splitting_op not in compilation_config.splitting_ops |
| 753 | ): |
| 754 | compilation_config.splitting_ops.append(splitting_op) |
| 755 | |
| 756 | def forward( |
| 757 | self, |
| 758 | hidden_states: torch.Tensor, |
| 759 | positions: torch.Tensor, |
| 760 | output: torch.Tensor, |
| 761 | ) -> None: |
| 762 | del positions |
| 763 | num_tokens = hidden_states.size(0) |
| 764 | if self.separate_b_proj: |
| 765 | assert self.qkv_proj is not None and self.b_proj is not None |
| 766 | qkv = self.qkv_proj(hidden_states)[0] |
| 767 | q, k, v = qkv.split(self.projection_size_per_partition, dim=-1) |
| 768 | beta_logits = self.b_proj(hidden_states)[0] |
| 769 | else: |
| 770 | assert self.qkvb_proj is not None |
| 771 | qkvb = self.qkvb_proj(hidden_states)[0] |
| 772 | q, k, v, beta_logits = qkvb.split( |
| 773 | [ |
| 774 | self.projection_size_per_partition, |
| 775 | self.projection_size_per_partition, |
| 776 | self.projection_size_per_partition, |
| 777 | self.local_num_heads, |
| 778 | ], |
| 779 | dim=-1, |
| 780 | ) |
| 781 | beta = beta_logits.float().sigmoid().unsqueeze(0) |
| 782 | g1 = self.f_proj(hidden_states)[0] |
| 783 | g1 = fused_kda_gate( |
| 784 | g1, |
| 785 | self.A_log, |
| 786 | self.head_dim, |
| 787 | g_bias=self.dt_bias, |
| 788 | lower_bound=self.lower_bound if self.safe_gate else None, |
| 789 | ) |
| 790 | g1 = g1.unsqueeze(0) |
| 791 | g2 = rearrange( |
| 792 | self.g_proj(hidden_states)[0], "... (h d) -> ... h d", d=self.head_dim |
| 793 | ) |
| 794 | |
| 795 | core_attn_out = torch.zeros( |
| 796 | (1, num_tokens, self.local_num_heads, self.head_dim), |
| 797 | dtype=hidden_states.dtype, |
| 798 | device=hidden_states.device, |
| 799 | ) |
| 800 | torch.ops.vllm.bailing_v3_kda_attention( |
| 801 | q, k, v, g1, beta, core_attn_out, self.prefix |
| 802 | ) |
| 803 | core_attn_out = self.o_norm(core_attn_out, g2) |
| 804 | core_attn_out = rearrange(core_attn_out, "1 n h d -> n (h d)") |
| 805 | output[:] = self.o_proj(core_attn_out)[0] |
| 806 | |
| 807 | def _forward( |
| 808 | self, |
| 809 | q_proj_states: torch.Tensor, |
| 810 | k_proj_states: torch.Tensor, |
| 811 | v_proj_states: torch.Tensor, |
| 812 | g1: torch.Tensor, |
| 813 | beta: torch.Tensor, |
| 814 | core_attn_out: torch.Tensor, |
| 815 | ) -> None: |
| 816 | forward_context = get_forward_context() |
| 817 | attn_metadata_map = forward_context.attn_metadata |
| 818 | if attn_metadata_map is None: |
| 819 | return |
| 820 | |
| 821 | assert isinstance(attn_metadata_map, dict) |
| 822 | attn_metadata = attn_metadata_map.get(self.prefix) |
| 823 | if attn_metadata is None: |
| 824 | # Profile/warmup dummy runs skip mamba-family metadata. |
| 825 | return |
| 826 | assert isinstance(attn_metadata, GDNAttentionMetadata) |
| 827 | has_initial_state = attn_metadata.has_initial_state |
| 828 | spec_query_start_loc = attn_metadata.spec_query_start_loc |
| 829 | query_start_loc = attn_metadata.non_spec_query_start_loc |
| 830 | spec_state_indices = attn_metadata.spec_state_indices_tensor |
| 831 | state_indices = attn_metadata.non_spec_state_indices_tensor |
| 832 | spec_sequence_masks = attn_metadata.spec_sequence_masks |
| 833 | spec_token_indx = attn_metadata.spec_token_indx |
| 834 | non_spec_token_indx = attn_metadata.non_spec_token_indx |
| 835 | num_accepted_tokens = attn_metadata.num_accepted_tokens |
| 836 | num_actual_tokens = attn_metadata.num_actual_tokens |
| 837 | conv_state, recurrent_state = self.kv_cache |
| 838 | recurrent_state_active = recurrent_state[..., : self.head_dim] |
| 839 | if not is_conv_state_dim_first(): |
| 840 | conv_state = conv_state.transpose(-1, -2) |
| 841 | conv_state_q, conv_state_k, conv_state_v = conv_state.chunk(3, dim=-2) |
| 842 | |
| 843 | q_proj_states = q_proj_states[:num_actual_tokens] |
| 844 | k_proj_states = k_proj_states[:num_actual_tokens] |
| 845 | v_proj_states = v_proj_states[:num_actual_tokens] |
| 846 | g1 = g1[:, :num_actual_tokens] |
| 847 | beta = beta[:, :num_actual_tokens] |
| 848 | |
| 849 | spec_only = ( |
| 850 | spec_sequence_masks is not None |
| 851 | and attn_metadata.num_prefills == 0 |
| 852 | and attn_metadata.num_decodes == 0 |
| 853 | ) |
| 854 | if spec_sequence_masks is not None: |
| 855 | assert spec_token_indx is not None |
| 856 | assert non_spec_token_indx is not None |
| 857 | if spec_only: |
| 858 | q_proj_states_spec = q_proj_states |
| 859 | k_proj_states_spec = k_proj_states |
| 860 | v_proj_states_spec = v_proj_states |
| 861 | g1_spec = g1 |
| 862 | beta_spec = beta |
| 863 | q_proj_states_non_spec = None |
| 864 | k_proj_states_non_spec = None |
| 865 | v_proj_states_non_spec = None |
| 866 | g1_non_spec = None |
| 867 | beta_non_spec = None |
| 868 | else: |
| 869 | q_proj_states_spec = q_proj_states.index_select(0, spec_token_indx) |
| 870 | k_proj_states_spec = k_proj_states.index_select(0, spec_token_indx) |
| 871 | v_proj_states_spec = v_proj_states.index_select(0, spec_token_indx) |
| 872 | g1_spec = g1.index_select(1, spec_token_indx) |
| 873 | beta_spec = beta.index_select(1, spec_token_indx) |
| 874 | q_proj_states_non_spec = q_proj_states.index_select( |
| 875 | 0, non_spec_token_indx |
| 876 | ) |
| 877 | k_proj_states_non_spec = k_proj_states.index_select( |
| 878 | 0, non_spec_token_indx |
| 879 | ) |
| 880 | v_proj_states_non_spec = v_proj_states.index_select( |
| 881 | 0, non_spec_token_indx |
| 882 | ) |
| 883 | g1_non_spec = g1.index_select(1, non_spec_token_indx) |
| 884 | beta_non_spec = beta.index_select(1, non_spec_token_indx) |
| 885 | else: |
| 886 | q_proj_states_spec = None |
| 887 | k_proj_states_spec = None |
| 888 | v_proj_states_spec = None |
| 889 | g1_spec = None |
| 890 | beta_spec = None |
| 891 | q_proj_states_non_spec = q_proj_states |
| 892 | k_proj_states_non_spec = k_proj_states |
| 893 | v_proj_states_non_spec = v_proj_states |
| 894 | g1_non_spec = g1 |
| 895 | beta_non_spec = beta |
| 896 | |
| 897 | q_conv_weights = self.q_conv1d.weight.view( |
| 898 | self.q_conv1d.weight.size(0), self.q_conv1d.weight.size(2) |
| 899 | ) |
| 900 | k_conv_weights = self.k_conv1d.weight.view( |
| 901 | self.k_conv1d.weight.size(0), self.k_conv1d.weight.size(2) |
| 902 | ) |
| 903 | v_conv_weights = self.v_conv1d.weight.view( |
| 904 | self.v_conv1d.weight.size(0), self.v_conv1d.weight.size(2) |
| 905 | ) |
| 906 | |
| 907 | if spec_sequence_masks is not None: |
| 908 | assert q_proj_states_spec is not None |
| 909 | assert k_proj_states_spec is not None |
| 910 | assert v_proj_states_spec is not None |
| 911 | assert spec_state_indices is not None |
| 912 | q_spec = causal_conv1d_update( |
| 913 | q_proj_states_spec, |
| 914 | conv_state_q, |
| 915 | q_conv_weights, |
| 916 | self.q_conv1d.bias, |
| 917 | activation="silu", |
| 918 | conv_state_indices=spec_state_indices[:, 0][ |
| 919 | : attn_metadata.num_spec_decodes |
| 920 | ], |
| 921 | num_accepted_tokens=num_accepted_tokens, |
| 922 | query_start_loc=spec_query_start_loc, |
| 923 | max_query_len=spec_state_indices.size(-1), |
| 924 | validate_data=False, |
| 925 | ) |
| 926 | k_spec = causal_conv1d_update( |
| 927 | k_proj_states_spec, |
| 928 | conv_state_k, |
| 929 | k_conv_weights, |
| 930 | self.k_conv1d.bias, |
| 931 | activation="silu", |
| 932 | conv_state_indices=spec_state_indices[:, 0][ |
| 933 | : attn_metadata.num_spec_decodes |
| 934 | ], |
| 935 | num_accepted_tokens=num_accepted_tokens, |
| 936 | query_start_loc=spec_query_start_loc, |
| 937 | max_query_len=spec_state_indices.size(-1), |
| 938 | validate_data=False, |
| 939 | ) |
| 940 | v_spec = causal_conv1d_update( |
| 941 | v_proj_states_spec, |
| 942 | conv_state_v, |
| 943 | v_conv_weights, |
| 944 | self.v_conv1d.bias, |
| 945 | activation="silu", |
| 946 | conv_state_indices=spec_state_indices[:, 0][ |
| 947 | : attn_metadata.num_spec_decodes |
| 948 | ], |
| 949 | num_accepted_tokens=num_accepted_tokens, |
| 950 | query_start_loc=spec_query_start_loc, |
| 951 | max_query_len=spec_state_indices.size(-1), |
| 952 | validate_data=False, |
| 953 | ) |
| 954 | else: |
| 955 | q_spec = k_spec = v_spec = None |
| 956 | |
| 957 | if attn_metadata.num_prefills > 0: |
| 958 | assert q_proj_states_non_spec is not None |
| 959 | assert k_proj_states_non_spec is not None |
| 960 | assert v_proj_states_non_spec is not None |
| 961 | q = causal_conv1d_fn( |
| 962 | q_proj_states_non_spec.transpose(0, 1), |
| 963 | q_conv_weights, |
| 964 | self.q_conv1d.bias, |
| 965 | activation="silu", |
| 966 | conv_states=conv_state_q, |
| 967 | has_initial_state=has_initial_state, |
| 968 | cache_indices=state_indices, |
| 969 | query_start_loc=query_start_loc, |
| 970 | metadata=attn_metadata, |
| 971 | ).transpose(0, 1) |
| 972 | k = causal_conv1d_fn( |
| 973 | k_proj_states_non_spec.transpose(0, 1), |
| 974 | k_conv_weights, |
| 975 | self.k_conv1d.bias, |
| 976 | activation="silu", |
| 977 | conv_states=conv_state_k, |
| 978 | has_initial_state=has_initial_state, |
| 979 | cache_indices=state_indices, |
| 980 | query_start_loc=query_start_loc, |
| 981 | metadata=attn_metadata, |
| 982 | ).transpose(0, 1) |
| 983 | v = causal_conv1d_fn( |
| 984 | v_proj_states_non_spec.transpose(0, 1), |
| 985 | v_conv_weights, |
| 986 | self.v_conv1d.bias, |
| 987 | activation="silu", |
| 988 | conv_states=conv_state_v, |
| 989 | has_initial_state=has_initial_state, |
| 990 | cache_indices=state_indices, |
| 991 | query_start_loc=query_start_loc, |
| 992 | metadata=attn_metadata, |
| 993 | ).transpose(0, 1) |
| 994 | elif attn_metadata.num_decodes > 0: |
| 995 | assert q_proj_states_non_spec is not None |
| 996 | assert k_proj_states_non_spec is not None |
| 997 | assert v_proj_states_non_spec is not None |
| 998 | assert state_indices is not None |
| 999 | decode_indices = state_indices[:num_actual_tokens] |
| 1000 | q = causal_conv1d_update( |
| 1001 | q_proj_states_non_spec, |
| 1002 | conv_state_q, |
| 1003 | q_conv_weights, |
| 1004 | self.q_conv1d.bias, |
| 1005 | activation="silu", |
| 1006 | conv_state_indices=decode_indices, |
| 1007 | validate_data=True, |
| 1008 | ) |
| 1009 | k = causal_conv1d_update( |
| 1010 | k_proj_states_non_spec, |
| 1011 | conv_state_k, |
| 1012 | k_conv_weights, |
| 1013 | self.k_conv1d.bias, |
| 1014 | activation="silu", |
| 1015 | conv_state_indices=decode_indices, |
| 1016 | validate_data=True, |
| 1017 | ) |
| 1018 | v = causal_conv1d_update( |
| 1019 | v_proj_states_non_spec, |
| 1020 | conv_state_v, |
| 1021 | v_conv_weights, |
| 1022 | self.v_conv1d.bias, |
| 1023 | activation="silu", |
| 1024 | conv_state_indices=decode_indices, |
| 1025 | validate_data=True, |
| 1026 | ) |
| 1027 | else: |
| 1028 | q = k = v = None |
| 1029 | |
| 1030 | if q_spec is not None: |
| 1031 | q_spec, k_spec, v_spec = map( |
| 1032 | lambda x: rearrange(x, "n (h d) -> 1 n h d", d=self.head_dim), |
| 1033 | (q_spec, k_spec, v_spec), |
| 1034 | ) |
| 1035 | if q is not None: |
| 1036 | q, k, v = map( |
| 1037 | lambda x: rearrange(x, "n (h d) -> 1 n h d", d=self.head_dim), |
| 1038 | (q, k, v), |
| 1039 | ) |
| 1040 | if spec_sequence_masks is not None: |
| 1041 | assert g1_spec is not None |
| 1042 | assert beta_spec is not None |
| 1043 | assert spec_query_start_loc is not None |
| 1044 | assert spec_state_indices is not None |
| 1045 | out_spec, _ = fused_recurrent_kda_fwd( |
| 1046 | q=q_spec.contiguous(), |
| 1047 | k=k_spec.contiguous(), |
| 1048 | v=v_spec.contiguous(), |
| 1049 | g=g1_spec.contiguous(), |
| 1050 | beta=beta_spec.contiguous(), |
| 1051 | scale=self.head_dim**-0.5, |
| 1052 | initial_state=recurrent_state_active, |
| 1053 | inplace_final_state=True, |
| 1054 | cu_seqlens=spec_query_start_loc[: attn_metadata.num_spec_decodes + 1], |
| 1055 | ssm_state_indices=spec_state_indices, |
| 1056 | num_accepted_tokens=num_accepted_tokens, |
| 1057 | use_qk_l2norm_in_kernel=True, |
| 1058 | ) |
| 1059 | else: |
| 1060 | out_spec = None |
| 1061 | |
| 1062 | if attn_metadata.num_prefills > 0: |
| 1063 | assert q is not None and k is not None and v is not None |
| 1064 | assert g1_non_spec is not None and beta_non_spec is not None |
| 1065 | assert state_indices is not None |
| 1066 | assert has_initial_state is not None |
| 1067 | zero_idx = state_indices[~has_initial_state] |
| 1068 | recurrent_state[zero_idx] = 0 |
| 1069 | initial_state = recurrent_state_active[state_indices].contiguous() |
| 1070 | out, last_state = chunk_kda( |
| 1071 | q=q, |
| 1072 | k=k, |
| 1073 | v=v, |
| 1074 | g=g1_non_spec, |
| 1075 | beta=beta_non_spec, |
| 1076 | initial_state=initial_state, |
| 1077 | output_final_state=True, |
| 1078 | use_qk_l2norm_in_kernel=True, |
| 1079 | safe_gate=self.safe_gate, |
| 1080 | lower_bound=self.lower_bound, |
| 1081 | cu_seqlens=query_start_loc, |
| 1082 | ) |
| 1083 | recurrent_state_active[state_indices] = last_state |
| 1084 | elif attn_metadata.num_decodes > 0: |
| 1085 | assert q is not None and k is not None and v is not None |
| 1086 | assert g1_non_spec is not None and beta_non_spec is not None |
| 1087 | assert state_indices is not None |
| 1088 | assert query_start_loc is not None |
| 1089 | out, _ = fused_recurrent_kda( |
| 1090 | q=q, |
| 1091 | k=k, |
| 1092 | v=v, |
| 1093 | g=g1_non_spec, |
| 1094 | beta=beta_non_spec, |
| 1095 | initial_state=recurrent_state_active, |
| 1096 | use_qk_l2norm_in_kernel=True, |
| 1097 | safe_gate=self.safe_gate, |
| 1098 | lower_bound=self.lower_bound, |
| 1099 | cu_seqlens=query_start_loc[: attn_metadata.num_decodes + 1], |
| 1100 | ssm_state_indices=state_indices, |
| 1101 | ) |
| 1102 | else: |
| 1103 | out = None |
| 1104 | |
| 1105 | if spec_sequence_masks is not None and out is not None: |
| 1106 | assert spec_token_indx is not None and non_spec_token_indx is not None |
| 1107 | merged_out = torch.empty( |
| 1108 | (1, num_actual_tokens, *out_spec.shape[2:]), |
| 1109 | dtype=out_spec.dtype, |
| 1110 | device=out_spec.device, |
| 1111 | ) |
| 1112 | merged_out.index_copy_(1, spec_token_indx, out_spec) |
| 1113 | merged_out.index_copy_(1, non_spec_token_indx, out) |
| 1114 | core_attn_out[0, :num_actual_tokens] = merged_out[0, :num_actual_tokens] |
| 1115 | elif spec_sequence_masks is not None: |
| 1116 | core_attn_out[0, :num_actual_tokens] = out_spec[0, :num_actual_tokens] |
| 1117 | else: |
| 1118 | core_attn_out[0, :num_actual_tokens] = out[0, :num_actual_tokens] |
| 1119 | |
| 1120 | |
| 1121 | class BailingMoeV3MoE(nn.Module): |
| 1122 | def __init__( |
| 1123 | self, |
| 1124 | config: PreTrainedConfig, |
| 1125 | quant_config: QuantizationConfig | None = None, |
| 1126 | layer_id: int = 0, |
| 1127 | prefix: str = "", |
| 1128 | ) -> None: |
| 1129 | super().__init__() |
| 1130 | self.num_experts = config.num_experts |
| 1131 | self.top_k = config.num_experts_per_tok |
| 1132 | self.hidden_size = config.hidden_size |
| 1133 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 1134 | self.gate = GateLinear( |
| 1135 | config.hidden_size, |
| 1136 | config.num_experts, |
| 1137 | out_dtype=torch.float32, |
| 1138 | params_dtype=torch.float32, |
| 1139 | prefix=f"{prefix}.gate", |
| 1140 | ) |
| 1141 | self.gate.expert_bias = nn.Parameter( |
| 1142 | torch.empty((config.num_experts,), dtype=torch.float32) |
| 1143 | ) |
| 1144 | self.expert_swiglu_limit = _get_layer_swiglu_limit( |
| 1145 | getattr(config, "expert_swiglu_limit_list", None), layer_id |
| 1146 | ) |
| 1147 | self.share_expert_swiglu_limit = _get_layer_swiglu_limit( |
| 1148 | getattr(config, "share_expert_swiglu_limit_list", None), layer_id |
| 1149 | ) |
| 1150 | shared_intermediate = ( |
| 1151 | config.moe_shared_expert_intermediate_size * config.num_shared_experts |
| 1152 | ) |
| 1153 | self.shared_experts = BailingMoeV3MLP( |
| 1154 | intermediate_size=_get_block_fp8_mlp_padded_intermediate_size( |
| 1155 | quant_config, shared_intermediate, f"{prefix}.shared_experts" |
| 1156 | ), |
| 1157 | config=config, |
| 1158 | quant_config=quant_config, |
| 1159 | reduce_results=False, |
| 1160 | prefix=f"{prefix}.shared_experts", |
| 1161 | swiglu_limit=self.share_expert_swiglu_limit, |
| 1162 | ) |
| 1163 | self.experts = FusedMoEFactory( |
| 1164 | shared_experts=self.shared_experts, |
| 1165 | num_experts=self.num_experts, |
| 1166 | top_k=self.top_k, |
| 1167 | hidden_size=self.hidden_size, |
| 1168 | intermediate_size=config.moe_intermediate_size, |
| 1169 | renormalize=True, |
| 1170 | quant_config=quant_config, |
| 1171 | activation="silu", |
| 1172 | swiglu_limit=self.expert_swiglu_limit, |
| 1173 | prefix=f"{prefix}.experts", |
| 1174 | scoring_func="sigmoid", |
| 1175 | routed_scaling_factor=self.routed_scaling_factor, |
| 1176 | e_score_correction_bias=self.gate.expert_bias, |
| 1177 | num_expert_group=config.n_group, |
| 1178 | topk_group=config.topk_group, |
| 1179 | use_grouped_topk=True, |
| 1180 | router_logits_dtype=torch.float32, |
| 1181 | ) |
| 1182 | |
| 1183 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 1184 | num_tokens, hidden_size = hidden_states.shape |
| 1185 | hidden_states = hidden_states.contiguous().view(-1, hidden_size) |
| 1186 | router_logits, _ = self.gate(hidden_states) |
| 1187 | hidden_states = self.experts( |
| 1188 | hidden_states=hidden_states, router_logits=router_logits |
| 1189 | ) |
| 1190 | return hidden_states.view(num_tokens, hidden_size) |
| 1191 | |
| 1192 | |
| 1193 | class BailingMoeV3DecoderLayer(nn.Module): |
| 1194 | def __init__( |
| 1195 | self, |
| 1196 | config: PreTrainedConfig, |
| 1197 | quant_config: QuantizationConfig | None = None, |
| 1198 | layer_id: int = 0, |
| 1199 | prefix: str = "layer", |
| 1200 | model_config: ModelConfig | None = None, |
| 1201 | cache_config: CacheConfig | None = None, |
| 1202 | num_speculative_tokens: int = 0, |
| 1203 | ) -> None: |
| 1204 | super().__init__() |
| 1205 | self.layer_id = layer_id |
| 1206 | self.hidden_size = config.hidden_size |
| 1207 | self.is_kda_layer = _is_kda_layer( |
| 1208 | layer_id, config.layer_group_size, config.num_hidden_layers |
| 1209 | ) |
| 1210 | if self.is_kda_layer: |
| 1211 | self.self_attn = BailingMoeV3KimiDeltaAttention( |
| 1212 | config, |
| 1213 | quant_config=quant_config, |
| 1214 | layer_id=layer_id, |
| 1215 | prefix=f"{prefix}.self_attn", |
| 1216 | model_config=model_config, |
| 1217 | cache_config=cache_config, |
| 1218 | num_speculative_tokens=num_speculative_tokens, |
| 1219 | ) |
| 1220 | else: |
| 1221 | self.self_attn = BailingMoeV3MLAAttention( |
| 1222 | config, |
| 1223 | quant_config=quant_config, |
| 1224 | layer_id=layer_id, |
| 1225 | prefix=f"{prefix}.self_attn", |
| 1226 | cache_config=cache_config, |
| 1227 | ) |
| 1228 | |
| 1229 | if config.num_experts is not None and layer_id >= config.first_k_dense_replace: |
| 1230 | self.mlp = BailingMoeV3MoE( |
| 1231 | config, |
| 1232 | quant_config=quant_config, |
| 1233 | layer_id=layer_id, |
| 1234 | prefix=f"{prefix}.mlp", |
| 1235 | ) |
| 1236 | else: |
| 1237 | self.mlp = BailingMoeV3MLP( |
| 1238 | intermediate_size=config.intermediate_size, |
| 1239 | config=config, |
| 1240 | quant_config=quant_config, |
| 1241 | reduce_results=True, |
| 1242 | prefix=f"{prefix}.mlp", |
| 1243 | ) |
| 1244 | |
| 1245 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1246 | self.post_attention_layernorm = RMSNorm( |
| 1247 | config.hidden_size, eps=config.rms_norm_eps |
| 1248 | ) |
| 1249 | |
| 1250 | def forward( |
| 1251 | self, |
| 1252 | hidden_states: torch.Tensor, |
| 1253 | positions: torch.Tensor, |
| 1254 | residual: torch.Tensor | None = None, |
| 1255 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 1256 | if residual is None: |
| 1257 | residual = hidden_states |
| 1258 | hidden_states = self.input_layernorm(hidden_states) |
| 1259 | else: |
| 1260 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 1261 | |
| 1262 | if self.is_kda_layer: |
| 1263 | attn_output = torch.zeros_like(hidden_states) |
| 1264 | self.self_attn(hidden_states, positions, attn_output) |
| 1265 | else: |
| 1266 | attn_output = self.self_attn(hidden_states, positions) |
| 1267 | |
| 1268 | hidden_states, residual = self.post_attention_layernorm(attn_output, residual) |
| 1269 | hidden_states = self.mlp(hidden_states) |
| 1270 | return hidden_states, residual |
| 1271 | |
| 1272 | |
| 1273 | @support_torch_compile( |
| 1274 | dynamic_arg_dims={ |
| 1275 | "input_ids": 0, |
| 1276 | "positions": -1, |
| 1277 | "intermediate_tensors": 0, |
| 1278 | "inputs_embeds": 0, |
| 1279 | } |
| 1280 | ) |
| 1281 | class BailingMoeV3Model(nn.Module): |
| 1282 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 1283 | super().__init__() |
| 1284 | config = vllm_config.model_config.hf_config |
| 1285 | model_config = vllm_config.model_config |
| 1286 | quant_config = vllm_config.quant_config |
| 1287 | cache_config = vllm_config.cache_config |
| 1288 | num_speculative_tokens = ( |
| 1289 | vllm_config.speculative_config.num_speculative_tokens |
| 1290 | if vllm_config.speculative_config |
| 1291 | else 0 |
| 1292 | ) |
| 1293 | |
| 1294 | self.config = config |
| 1295 | self.vocab_size = config.vocab_size |
| 1296 | self.embed_dim = config.hidden_size |
| 1297 | self.num_layers = config.num_hidden_layers |
| 1298 | |
| 1299 | if get_pp_group().is_first_rank: |
| 1300 | self.word_embeddings = VocabParallelEmbedding( |
| 1301 | self.vocab_size, |
| 1302 | self.embed_dim, |
| 1303 | org_num_embeddings=self.vocab_size, |
| 1304 | ) |
| 1305 | else: |
| 1306 | self.word_embeddings = PPMissingLayer() |
| 1307 | |
| 1308 | def layer_fn(prefix: str): |
| 1309 | layer_idx = int(prefix.split(".")[-1]) |
| 1310 | return BailingMoeV3DecoderLayer( |
| 1311 | config=config, |
| 1312 | quant_config=quant_config, |
| 1313 | layer_id=layer_idx, |
| 1314 | prefix=prefix, |
| 1315 | model_config=model_config, |
| 1316 | cache_config=cache_config, |
| 1317 | num_speculative_tokens=num_speculative_tokens, |
| 1318 | ) |
| 1319 | |
| 1320 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 1321 | self.num_layers, layer_fn, prefix=f"{prefix}.layers" |
| 1322 | ) |
| 1323 | |
| 1324 | if get_pp_group().is_last_rank: |
| 1325 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1326 | else: |
| 1327 | self.norm = PPMissingLayer() |
| 1328 | |
| 1329 | @property |
| 1330 | def embed_tokens(self) -> nn.Module: |
| 1331 | return self.word_embeddings |
| 1332 | |
| 1333 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1334 | return self.word_embeddings(input_ids) |
| 1335 | |
| 1336 | def forward( |
| 1337 | self, |
| 1338 | input_ids: torch.Tensor | None, |
| 1339 | positions: torch.Tensor, |
| 1340 | intermediate_tensors: IntermediateTensors | None = None, |
| 1341 | inputs_embeds: torch.Tensor | None = None, |
| 1342 | ) -> torch.Tensor | IntermediateTensors: |
| 1343 | if get_pp_group().is_first_rank: |
| 1344 | hidden_states = ( |
| 1345 | self.word_embeddings(input_ids) |
| 1346 | if inputs_embeds is None |
| 1347 | else inputs_embeds |
| 1348 | ) |
| 1349 | residual = None |
| 1350 | else: |
| 1351 | assert intermediate_tensors is not None |
| 1352 | hidden_states = intermediate_tensors["hidden_states"] |
| 1353 | residual = intermediate_tensors["residual"] |
| 1354 | |
| 1355 | for layer in self.layers[self.start_layer : self.end_layer]: |
| 1356 | hidden_states, residual = layer(hidden_states, positions, residual) |
| 1357 | |
| 1358 | if not get_pp_group().is_last_rank: |
| 1359 | return IntermediateTensors( |
| 1360 | {"hidden_states": hidden_states, "residual": residual} |
| 1361 | ) |
| 1362 | if residual is not None: |
| 1363 | hidden_states, _ = self.norm(hidden_states, residual) |
| 1364 | else: |
| 1365 | hidden_states = self.norm(hidden_states) |
| 1366 | return hidden_states |
| 1367 | |
| 1368 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1369 | return fused_moe_make_expert_params_mapping( |
| 1370 | self, |
| 1371 | ckpt_gate_proj_name="gate_proj", |
| 1372 | ckpt_down_proj_name="down_proj", |
| 1373 | ckpt_up_proj_name="up_proj", |
| 1374 | num_experts=self.config.num_experts, |
| 1375 | ) |
| 1376 | |
| 1377 | |
| 1378 | class BailingMoeV3ForCausalLM(nn.Module, HasInnerState, IsHybrid, SupportsPP): |
| 1379 | hf_to_vllm_mapper = WeightsMapper(orig_to_new_substr={".attention.": ".self_attn."}) |
| 1380 | |
| 1381 | packed_modules_mapping = { |
| 1382 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 1383 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 1384 | "qkvb_proj": ["q_proj", "k_proj", "v_proj", "b_proj"], |
| 1385 | } |
| 1386 | |
| 1387 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 1388 | super().__init__() |
| 1389 | config = vllm_config.model_config.hf_config |
| 1390 | quant_config = vllm_config.quant_config |
| 1391 | _configure_ling_fp8_quant_config(quant_config, config) |
| 1392 | self.config = config |
| 1393 | self.quant_config = quant_config |
| 1394 | self.model = BailingMoeV3Model( |
| 1395 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1396 | ) |
| 1397 | if get_pp_group().is_last_rank: |
| 1398 | self.lm_head = ParallelLMHead( |
| 1399 | config.vocab_size, |
| 1400 | config.hidden_size, |
| 1401 | quant_config=quant_config, |
| 1402 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1403 | ) |
| 1404 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1405 | else: |
| 1406 | self.lm_head = PPMissingLayer() |
| 1407 | |
| 1408 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1409 | return self.model.embed_input_ids(input_ids) |
| 1410 | |
| 1411 | def forward( |
| 1412 | self, |
| 1413 | input_ids: torch.Tensor | None, |
| 1414 | positions: torch.Tensor, |
| 1415 | intermediate_tensors: IntermediateTensors | None = None, |
| 1416 | inputs_embeds: torch.Tensor | None = None, |
| 1417 | ) -> torch.Tensor | IntermediateTensors: |
| 1418 | return self.model(input_ids, positions, intermediate_tensors, inputs_embeds) |
| 1419 | |
| 1420 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 1421 | return self.logits_processor(self.lm_head, hidden_states) |
| 1422 | |
| 1423 | def make_empty_intermediate_tensors( |
| 1424 | self, batch_size: int, dtype: torch.dtype, device: torch.device |
| 1425 | ) -> IntermediateTensors: |
| 1426 | return IntermediateTensors( |
| 1427 | { |
| 1428 | "hidden_states": torch.zeros( |
| 1429 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 1430 | ), |
| 1431 | "residual": torch.zeros( |
| 1432 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 1433 | ), |
| 1434 | } |
| 1435 | ) |
| 1436 | |
| 1437 | @classmethod |
| 1438 | def get_mamba_state_shape_from_config( |
| 1439 | cls, vllm_config: VllmConfig |
| 1440 | ) -> tuple[tuple[int, int], tuple[int, int, int]]: |
| 1441 | return _get_kda_state_shape_for_config(vllm_config) |
| 1442 | |
| 1443 | @classmethod |
| 1444 | def get_mamba_state_dtype_from_config( |
| 1445 | cls, vllm_config: VllmConfig |
| 1446 | ) -> tuple[torch.dtype, torch.dtype]: |
| 1447 | return MambaStateDtypeCalculator.kda_state_dtype( |
| 1448 | vllm_config.model_config.dtype, |
| 1449 | vllm_config.cache_config.mamba_cache_dtype, |
| 1450 | ) |
| 1451 | |
| 1452 | @classmethod |
| 1453 | def get_mamba_state_copy_func(cls) -> tuple: |
| 1454 | return MambaStateCopyFuncCalculator.kda_state_copy_func() |
| 1455 | |
| 1456 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1457 | return self.model.get_expert_mapping() |
| 1458 | |
| 1459 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1460 | weights = self.hf_to_vllm_mapper.apply(weights) |
| 1461 | weights = _maybe_remap_ling_mxfp4_weight_names(weights, self.quant_config) |
| 1462 | params_dict = dict(self.named_parameters(remove_duplicate=False)) |
| 1463 | loaded_params: set[str] = set() |
| 1464 | stacked_mappings = [ |
| 1465 | (".gate_up_proj", ".gate_proj", 0), |
| 1466 | (".gate_up_proj", ".up_proj", 1), |
| 1467 | (".qkv_proj", ".q_proj", 0), |
| 1468 | (".qkv_proj", ".k_proj", 1), |
| 1469 | (".qkv_proj", ".v_proj", 2), |
| 1470 | (".qkvb_proj", ".q_proj", 0), |
| 1471 | (".qkvb_proj", ".k_proj", 1), |
| 1472 | (".qkvb_proj", ".v_proj", 2), |
| 1473 | (".qkvb_proj", ".b_proj", 3), |
| 1474 | (".fused_qkv_a_proj", ".q_a_proj", 0), |
| 1475 | (".fused_qkv_a_proj", ".kv_a_proj_with_mqa", 1), |
| 1476 | ] |
| 1477 | expert_mappings = list(self.get_expert_mapping()) |
| 1478 | |
| 1479 | def load_param(name: str, tensor: torch.Tensor, shard_id=None) -> bool: |
| 1480 | if name not in params_dict or is_pp_missing_parameter(name, self): |
| 1481 | return False |
| 1482 | param = params_dict[name] |
| 1483 | weight_loader: Callable[..., None] = getattr( |
| 1484 | param, "weight_loader", default_weight_loader |
| 1485 | ) |
| 1486 | if shard_id is None: |
| 1487 | weight_loader(param, tensor) |
| 1488 | elif isinstance(shard_id, int): |
| 1489 | weight_loader(param, tensor, shard_id) |
| 1490 | else: |
| 1491 | weight_loader( |
| 1492 | param, tensor, name, expert_id=shard_id[0], shard_id=shard_id[1] |
| 1493 | ) |
| 1494 | loaded_params.add(name) |
| 1495 | return True |
| 1496 | |
| 1497 | def normalize_name(name: str) -> str | None: |
| 1498 | if name.startswith("model.layers."): |
| 1499 | layer_idx = int(name.split("model.layers.")[1].split(".")[0]) |
| 1500 | if layer_idx >= self.config.num_hidden_layers: |
| 1501 | return None |
| 1502 | return maybe_remap_kv_scale_name(name, params_dict) |
| 1503 | |
| 1504 | for orig_name, weight in weights: |
| 1505 | name = normalize_name(orig_name) |
| 1506 | if name is None: |
| 1507 | continue |
| 1508 | weight = _maybe_pad_block_fp8_shared_expert_checkpoint_tensor( |
| 1509 | self.quant_config, |
| 1510 | self.config, |
| 1511 | name, |
| 1512 | weight, |
| 1513 | ) |
| 1514 | loaded = False |
| 1515 | for param_suf, weight_suf, stacked_shard_id in stacked_mappings: |
| 1516 | if weight_suf in name: |
| 1517 | mapped = name.replace(weight_suf, param_suf) |
| 1518 | if load_param(mapped, weight, stacked_shard_id): |
| 1519 | loaded = True |
| 1520 | break |
| 1521 | if loaded: |
| 1522 | continue |
| 1523 | |
| 1524 | if ".mlp.experts." in name: |
| 1525 | for param_name, weight_name, expert_id, shard_id in expert_mappings: |
| 1526 | if weight_name in name: |
| 1527 | mapped = name.replace(weight_name, param_name) |
| 1528 | load_param(mapped, weight, (expert_id, shard_id)) |
| 1529 | break |
| 1530 | continue |
| 1531 | |
| 1532 | load_param(name, weight) |
| 1533 | return loaded_params |