| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable |
| 5 | from typing import ClassVar, Literal |
| 6 | |
| 7 | import torch |
| 8 | from torch import nn |
| 9 | from transformers import Glm5NextTextConfig |
| 10 | |
| 11 | from vllm.config import ParallelConfig, VllmConfig |
| 12 | from vllm.distributed import ( |
| 13 | get_ep_group, |
| 14 | get_pp_group, |
| 15 | get_tensor_model_parallel_rank, |
| 16 | get_tensor_model_parallel_world_size, |
| 17 | tensor_model_parallel_all_gather, |
| 18 | ) |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp |
| 21 | from vllm.model_executor.layers.fused_moe import ( |
| 22 | FusedMoEFactory, |
| 23 | GateLinear, |
| 24 | fused_moe_make_expert_params_mapping, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.fusion.mm_input_norm import build_mm_input_norm |
| 27 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 28 | from vllm.model_executor.layers.linear import ( |
| 29 | MergedColumnParallelLinear, |
| 30 | RowParallelLinear, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 33 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 34 | MambaStateCopyFunc, |
| 35 | MambaStateCopyFuncCalculator, |
| 36 | MambaStateDtypeCalculator, |
| 37 | MambaStateShapeCalculator, |
| 38 | ) |
| 39 | from vllm.model_executor.layers.mhc import ( |
| 40 | MHCFusedPostPreOp, |
| 41 | MHCPostOp, |
| 42 | MHCPreOp, |
| 43 | hc_contract, |
| 44 | hc_expand, |
| 45 | ) |
| 46 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 47 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 48 | GroupShape, |
| 49 | scaled_dequantize, |
| 50 | ) |
| 51 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 52 | ParallelLMHead, |
| 53 | VocabParallelEmbedding, |
| 54 | ) |
| 55 | from vllm.model_executor.model_loader.weight_utils import ( |
| 56 | default_weight_loader, |
| 57 | maybe_remap_kv_scale_name, |
| 58 | ) |
| 59 | from vllm.model_executor.models.deepseek_v2 import _get_moe_router_dtype |
| 60 | from vllm.model_executor.models.glm4_1v import ( |
| 61 | Glm4vDummyInputsBuilder, |
| 62 | Glm4vForConditionalGeneration, |
| 63 | ) |
| 64 | from vllm.model_executor.models.interfaces import ( |
| 65 | HasInnerState, |
| 66 | IsHybrid, |
| 67 | MixtureOfExperts, |
| 68 | SupportsPP, |
| 69 | ) |
| 70 | from vllm.model_executor.models.utils import ( |
| 71 | AutoWeightsLoader, |
| 72 | PPMissingLayer, |
| 73 | init_vllm_registered_model, |
| 74 | is_pp_missing_parameter, |
| 75 | make_layers, |
| 76 | maybe_prefix, |
| 77 | sequence_parallel_chunk, |
| 78 | ) |
| 79 | from vllm.models.common.ops.sequence_parallel import ( |
| 80 | sp_all_gather, |
| 81 | sp_reduce_scatter, |
| 82 | sp_shard, |
| 83 | ) |
| 84 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 85 | from vllm.platforms import current_platform |
| 86 | from vllm.sequence import IntermediateTensors |
| 87 | |
| 88 | from .attention import Glm5NextMLAAttention |
| 89 | from .kda import Glm5NextLinearAttention |
| 90 | from .multimodal import ( |
| 91 | Glm5NextMultiModalProcessor, |
| 92 | Glm5NextProcessingInfo, |
| 93 | Glm5NextVisionTransformer, |
| 94 | ) |
| 95 | |
| 96 | logger = init_logger(__name__) |
| 97 | |
| 98 | _MHC_TAU = 0.05 |
| 99 | """mHC routing temperature. A GLM-5.3-Flash trained value that neither the |
| 100 | checkpoint nor `Glm5NextTextConfig` carries.""" |
| 101 | |
| 102 | _MHC_POST_MULT_VALUE = 2.0 |
| 103 | """mHC post-multiplier. A GLM-5.3-Flash trained value that neither the |
| 104 | checkpoint nor `Glm5NextTextConfig` carries.""" |
| 105 | |
| 106 | _LOGIT_SCALE = 1.0 |
| 107 | """Output logit scale. A GLM-5.3-Flash trained value that neither the |
| 108 | checkpoint nor `Glm5NextTextConfig` carries.""" |
| 109 | |
| 110 | _VISION_RMS_NORM_EPS = 1e-6 |
| 111 | """Vision tower RMSNorm epsilon. |
| 112 | |
| 113 | GLM-5.3-Flash checkpoints ship `vision_config.rms_norm_eps = 1e-5`, but the |
| 114 | vision tower was trained with 1e-6. Serving with 1e-5 drifts the RMSNorm and |
| 115 | produces repetitive/degraded image descriptions, so force the trained value |
| 116 | regardless of the checkpoint field. |
| 117 | """ |
| 118 | |
| 119 | |
| 120 | def _is_moe(config: Glm5NextTextConfig) -> bool: |
| 121 | return config.n_routed_experts is not None |
| 122 | |
| 123 | |
| 124 | def _is_kda_layer(config: Glm5NextTextConfig, layer_idx: int) -> bool: |
| 125 | layer_types = config.layer_types |
| 126 | return ( |
| 127 | layer_types is not None |
| 128 | and layer_idx < len(layer_types) |
| 129 | and layer_types[layer_idx] == "linear_attention" |
| 130 | ) |
| 131 | |
| 132 | |
| 133 | def _is_linear_attn(config: Glm5NextTextConfig) -> bool: |
| 134 | layer_types = config.layer_types |
| 135 | return layer_types is not None and "linear_attention" in layer_types |
| 136 | |
| 137 | |
| 138 | def _validate_supported_config(config: Glm5NextTextConfig) -> None: |
| 139 | """Reject checkpoints using config options this implementation lacks. |
| 140 | |
| 141 | The kpool indexer kernels always keep the incomplete trailing pool, so a |
| 142 | checkpoint asking otherwise would be served silently wrong. |
| 143 | """ |
| 144 | if config.index_topk is not None and not config.index_kpool_always_select_tail: |
| 145 | raise NotImplementedError( |
| 146 | "GLM-5.3 sparse indexer requires index_kpool_always_select_tail=True" |
| 147 | ) |
| 148 | |
| 149 | |
| 150 | class Glm5NextMLP(nn.Module): |
| 151 | def __init__( |
| 152 | self, |
| 153 | hidden_size: int, |
| 154 | intermediate_size: int, |
| 155 | hidden_act: str, |
| 156 | quant_config: QuantizationConfig | None = None, |
| 157 | reduce_results: bool = True, |
| 158 | is_sequence_parallel=False, |
| 159 | prefix: str = "", |
| 160 | swiglu_limit: float | None = None, |
| 161 | ) -> None: |
| 162 | super().__init__() |
| 163 | |
| 164 | # If is_sequence_parallel, the input and output tensors are sharded |
| 165 | # across the ranks within the tp_group. In this case the weights are |
| 166 | # replicated and no collective ops are needed. |
| 167 | # Otherwise we use standard TP with an allreduce at the end. |
| 168 | self.gate_up_proj = MergedColumnParallelLinear( |
| 169 | hidden_size, |
| 170 | [intermediate_size] * 2, |
| 171 | bias=False, |
| 172 | quant_config=quant_config, |
| 173 | disable_tp=is_sequence_parallel, |
| 174 | prefix=f"{prefix}.gate_up_proj", |
| 175 | ) |
| 176 | self.down_proj = RowParallelLinear( |
| 177 | intermediate_size, |
| 178 | hidden_size, |
| 179 | bias=False, |
| 180 | quant_config=quant_config, |
| 181 | reduce_results=reduce_results, |
| 182 | disable_tp=is_sequence_parallel, |
| 183 | prefix=f"{prefix}.down_proj", |
| 184 | ) |
| 185 | if hidden_act != "silu": |
| 186 | raise ValueError( |
| 187 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 188 | ) |
| 189 | |
| 190 | self.swiglu_limit = swiglu_limit |
| 191 | if self.swiglu_limit is not None: |
| 192 | self.act_fn = SiluAndMulWithClamp(swiglu_limit=self.swiglu_limit) |
| 193 | else: |
| 194 | self.act_fn = SiluAndMul() |
| 195 | |
| 196 | def forward(self, x): |
| 197 | gate_up, _ = self.gate_up_proj(x) |
| 198 | x = self.act_fn(gate_up) |
| 199 | x, _ = self.down_proj(x) |
| 200 | return x |
| 201 | |
| 202 | |
| 203 | class Glm5NextMoE(nn.Module): |
| 204 | def __init__( |
| 205 | self, |
| 206 | config: Glm5NextTextConfig, |
| 207 | parallel_config: ParallelConfig, |
| 208 | quant_config: QuantizationConfig | None = None, |
| 209 | prefix: str = "", |
| 210 | apply_routed_scale_to_output: bool = False, |
| 211 | ): |
| 212 | super().__init__() |
| 213 | self.tp_size = get_tensor_model_parallel_world_size() |
| 214 | self.tp_rank = get_tensor_model_parallel_rank() |
| 215 | |
| 216 | self.routed_scaling_factor = config.routed_scaling_factor |
| 217 | |
| 218 | self.ep_group = get_ep_group().device_group |
| 219 | self.ep_rank = get_ep_group().rank_in_group |
| 220 | self.ep_size = self.ep_group.size() |
| 221 | self.n_routed_experts: int = config.n_routed_experts |
| 222 | self.n_shared_experts: int = config.n_shared_experts |
| 223 | |
| 224 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 225 | |
| 226 | if config.hidden_act != "silu": |
| 227 | raise ValueError( |
| 228 | f"Unsupported activation: {config.hidden_act}. " |
| 229 | "Only silu is supported for now." |
| 230 | ) |
| 231 | |
| 232 | self.router_dtype = _get_moe_router_dtype(config) |
| 233 | self.gate = GateLinear( |
| 234 | config.hidden_size, |
| 235 | config.n_routed_experts, |
| 236 | out_dtype=self.router_dtype, |
| 237 | prefix=f"{prefix}.gate", |
| 238 | ) |
| 239 | if config.topk_method == "noaux_tc": |
| 240 | self.gate.e_score_correction_bias = nn.Parameter( |
| 241 | torch.empty(config.n_routed_experts, dtype=torch.float32) |
| 242 | ) |
| 243 | else: |
| 244 | self.gate.e_score_correction_bias = None |
| 245 | |
| 246 | # Load balancing settings. |
| 247 | eplb_config = parallel_config.eplb_config |
| 248 | self.enable_eplb = parallel_config.enable_eplb |
| 249 | |
| 250 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 251 | self.n_logical_experts = self.n_routed_experts |
| 252 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 253 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 254 | |
| 255 | self.physical_expert_start = self.ep_rank * self.n_local_physical_experts |
| 256 | self.physical_expert_end = ( |
| 257 | self.physical_expert_start + self.n_local_physical_experts |
| 258 | ) |
| 259 | |
| 260 | swiglu_limit = config.swiglu_limit |
| 261 | if config.n_shared_experts is None: |
| 262 | self.shared_experts = None |
| 263 | else: |
| 264 | intermediate_size = config.moe_intermediate_size * config.n_shared_experts |
| 265 | |
| 266 | self.shared_experts = Glm5NextMLP( |
| 267 | hidden_size=config.hidden_size, |
| 268 | intermediate_size=intermediate_size, |
| 269 | hidden_act=config.hidden_act, |
| 270 | quant_config=quant_config, |
| 271 | is_sequence_parallel=self.is_sequence_parallel, |
| 272 | reduce_results=False, |
| 273 | prefix=f"{prefix}.shared_experts", |
| 274 | swiglu_limit=swiglu_limit, |
| 275 | ) |
| 276 | |
| 277 | self.experts = FusedMoEFactory( |
| 278 | shared_experts=self.shared_experts, |
| 279 | gate=self.gate, |
| 280 | num_experts=config.n_routed_experts, |
| 281 | top_k=config.num_experts_per_tok, |
| 282 | hidden_size=config.hidden_size, |
| 283 | intermediate_size=config.moe_intermediate_size, |
| 284 | renormalize=config.norm_topk_prob, |
| 285 | quant_config=quant_config, |
| 286 | use_grouped_topk=True, |
| 287 | num_expert_group=config.n_group, |
| 288 | topk_group=config.topk_group, |
| 289 | prefix=f"{prefix}.experts", |
| 290 | scoring_func=config.scoring_func, |
| 291 | routed_scaling_factor=self.routed_scaling_factor, |
| 292 | apply_routed_scale_to_output=apply_routed_scale_to_output, |
| 293 | e_score_correction_bias=self.gate.e_score_correction_bias, |
| 294 | enable_eplb=self.enable_eplb, |
| 295 | num_redundant_experts=self.n_redundant_experts, |
| 296 | is_sequence_parallel=self.is_sequence_parallel, |
| 297 | n_shared_experts=None, |
| 298 | router_logits_dtype=self.gate.out_dtype, |
| 299 | swiglu_limit=swiglu_limit, |
| 300 | ) |
| 301 | |
| 302 | def forward( |
| 303 | self, |
| 304 | hidden_states: torch.Tensor, |
| 305 | already_sequence_parallel: bool = False, |
| 306 | ) -> torch.Tensor: |
| 307 | num_tokens, hidden_dim = hidden_states.shape |
| 308 | |
| 309 | # Chunk the hidden states so they aren't replicated across TP ranks. |
| 310 | # This avoids duplicate computation in self.experts. |
| 311 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 312 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 313 | |
| 314 | # MoERunner holds the gate (passed to FusedMoEFactory) and computes |
| 315 | # the router logits itself, so nothing is precomputed here (matches |
| 316 | # DeepseekV2MoE; `router_logits` is a placeholder). |
| 317 | final_hidden_states = self.experts( |
| 318 | hidden_states=hidden_states, router_logits=hidden_states |
| 319 | ) |
| 320 | |
| 321 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 322 | final_hidden_states = tensor_model_parallel_all_gather( |
| 323 | final_hidden_states, 0 |
| 324 | ) |
| 325 | final_hidden_states = final_hidden_states[:num_tokens] |
| 326 | |
| 327 | return final_hidden_states.view(num_tokens, hidden_dim) |
| 328 | |
| 329 | |
| 330 | class Glm5NextDecoderLayer(nn.Module): |
| 331 | def __init__( |
| 332 | self, |
| 333 | vllm_config: VllmConfig, |
| 334 | config: Glm5NextTextConfig, |
| 335 | layer_idx: int, |
| 336 | prefix: str = "", |
| 337 | topk_indices_buffer: torch.Tensor | None = None, |
| 338 | is_mtp_layer: bool = False, |
| 339 | **kwargs, |
| 340 | ) -> None: |
| 341 | super().__init__() |
| 342 | |
| 343 | cache_config = vllm_config.cache_config |
| 344 | quant_config = vllm_config.quant_config |
| 345 | parallel_config = vllm_config.parallel_config |
| 346 | |
| 347 | self.hidden_size = config.hidden_size |
| 348 | self.layer_idx = layer_idx |
| 349 | self.is_moe = _is_moe(config) |
| 350 | self.num_hidden_layers = config.num_hidden_layers |
| 351 | self.rms_norm_eps = config.rms_norm_eps |
| 352 | self.num_experts = config.n_routed_experts |
| 353 | self.is_mtp_layer = is_mtp_layer |
| 354 | self.mhc = config.mhc |
| 355 | is_kda_layer = not is_mtp_layer and _is_kda_layer(config, layer_idx) |
| 356 | self.layer_kind = "kda" if is_kda_layer else "mla" |
| 357 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 358 | |
| 359 | if is_kda_layer: |
| 360 | self.self_attn = Glm5NextLinearAttention( |
| 361 | config=config, |
| 362 | vllm_config=vllm_config, |
| 363 | prefix=f"{prefix}.self_attn", |
| 364 | ) |
| 365 | else: |
| 366 | # MLA layers require the latent head dims, which are guaranteed set |
| 367 | # on MLA configs; narrow away the `int | None`. |
| 368 | assert config.v_head_dim is not None |
| 369 | assert config.kv_lora_rank is not None |
| 370 | self.self_attn = Glm5NextMLAAttention( |
| 371 | vllm_config=vllm_config, |
| 372 | config=config, |
| 373 | hidden_size=self.hidden_size, |
| 374 | num_heads=config.num_attention_heads, |
| 375 | qk_nope_head_dim=config.qk_nope_head_dim, |
| 376 | qk_rope_head_dim=config.qk_rope_head_dim, |
| 377 | v_head_dim=config.v_head_dim, |
| 378 | q_lora_rank=config.q_lora_rank, |
| 379 | kv_lora_rank=config.kv_lora_rank, |
| 380 | max_position_embeddings=config.max_position_embeddings, |
| 381 | cache_config=cache_config, |
| 382 | quant_config=None, # MLA projections are BF16 in checkpoint |
| 383 | prefix=f"{prefix}.self_attn", |
| 384 | topk_indices_buffer=topk_indices_buffer, |
| 385 | skip_rope=config.mla_use_nope, |
| 386 | ) |
| 387 | |
| 388 | # MTP layers sit past the base model's hidden layers (layer_idx >= |
| 389 | # num_hidden_layers), so they're outside mlp_layer_types; default them |
| 390 | # to the last base layer's MLP type (sparse/MoE for these checkpoints). |
| 391 | mlp_layer_types = config.mlp_layer_types |
| 392 | mlp_type = ( |
| 393 | mlp_layer_types[layer_idx] |
| 394 | if layer_idx < len(mlp_layer_types) |
| 395 | else (mlp_layer_types[-1] if mlp_layer_types else "sparse") |
| 396 | ) |
| 397 | if self.is_moe and self.num_experts is not None and mlp_type == "sparse": |
| 398 | self.mlp = Glm5NextMoE( |
| 399 | config=config, |
| 400 | parallel_config=parallel_config, |
| 401 | quant_config=quant_config, |
| 402 | prefix=f"{prefix}.mlp", |
| 403 | ) |
| 404 | else: |
| 405 | self.mlp = Glm5NextMLP( |
| 406 | hidden_size=self.hidden_size, |
| 407 | intermediate_size=config.intermediate_size, |
| 408 | hidden_act=config.hidden_act, |
| 409 | quant_config=quant_config, |
| 410 | is_sequence_parallel=self.is_sequence_parallel, |
| 411 | prefix=f"{prefix}.mlp", |
| 412 | swiglu_limit=config.swiglu_limit, |
| 413 | ) |
| 414 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 415 | # Cached for the hot forward path (isinstance per layer per step). |
| 416 | self._mlp_is_moe = isinstance(self.mlp, Glm5NextMoE) |
| 417 | # In SP, the attention output projection leaves a partial sum; the |
| 418 | # decoder-layer reduce_scatter after attention completes it (DSv4 pattern). |
| 419 | # MTP layers use the non-mHC path which has no sp_reduce_scatter, so |
| 420 | # their o_proj must still reduce normally. |
| 421 | if self.is_sequence_parallel and not is_mtp_layer: |
| 422 | self.self_attn.o_proj.reduce_results = False |
| 423 | self.post_attention_layernorm = RMSNorm( |
| 424 | config.hidden_size, eps=config.rms_norm_eps |
| 425 | ) |
| 426 | |
| 427 | if self.mhc and not is_mtp_layer: |
| 428 | # mhc config |
| 429 | self.mhc_num_residual_streams = config.hc_mult |
| 430 | self.mhc_tau = _MHC_TAU |
| 431 | self.hc_eps = config.hc_eps |
| 432 | self.mhc_sinkhorn_iterations = config.hc_sinkhorn_iters |
| 433 | self.mhc_post_mult_value = _MHC_POST_MULT_VALUE |
| 434 | |
| 435 | n = config.hc_mult |
| 436 | d_model = n * self.hidden_size |
| 437 | mix_hc = (2 + n) * n |
| 438 | |
| 439 | self.n = n |
| 440 | |
| 441 | # attn hc |
| 442 | self.hc_attn_fn = nn.Parameter( |
| 443 | torch.empty(mix_hc, d_model, dtype=torch.float32) |
| 444 | ) |
| 445 | self.hc_attn_base = nn.Parameter(torch.empty(mix_hc, dtype=torch.float32)) |
| 446 | self.hc_attn_scale = nn.Parameter(torch.empty(3, dtype=torch.float32)) |
| 447 | |
| 448 | # ffn hc |
| 449 | self.hc_ffn_fn = nn.Parameter( |
| 450 | torch.empty(mix_hc, d_model, dtype=torch.float32) |
| 451 | ) |
| 452 | self.hc_ffn_base = nn.Parameter(torch.empty(mix_hc, dtype=torch.float32)) |
| 453 | self.hc_ffn_scale = nn.Parameter(torch.empty(3, dtype=torch.float32)) |
| 454 | |
| 455 | self.mhc_pre_op = MHCPreOp() |
| 456 | self.mhc_post_op = MHCPostOp() |
| 457 | self.mhc_fused_post_pre_op = MHCFusedPostPreOp() |
| 458 | |
| 459 | if vllm_config.kernel_config.enable_jit_warmup: |
| 460 | from vllm.model_executor.kernels.mhc.tilelang_kernels import ( |
| 461 | _HC_PRENORM_GEMM_TILELANG_KERNEL, |
| 462 | _MHC_FUSED_TILELANG_KERNEL, |
| 463 | _MHC_POST_TILELANG_KERNEL, |
| 464 | _MHC_PRE_BIG_FUSE_TILELANG_KERNEL, |
| 465 | ) |
| 466 | from vllm.utils.deep_gemm import is_deep_gemm_supported |
| 467 | |
| 468 | include_pre_gemm_splits = is_deep_gemm_supported() |
| 469 | _MHC_PRE_BIG_FUSE_TILELANG_KERNEL.register_warmup( |
| 470 | vllm_config, |
| 471 | hidden_size=self.hidden_size, |
| 472 | hc_mult=self.n, |
| 473 | use_norm_weight=True, |
| 474 | include_pre_gemm_splits=include_pre_gemm_splits, |
| 475 | include_broadcast_splits=False, |
| 476 | rms_eps=self.rms_norm_eps, |
| 477 | hc_pre_eps=self.hc_eps, |
| 478 | hc_sinkhorn_eps=self.hc_eps, |
| 479 | hc_post_mult_value=self.mhc_post_mult_value, |
| 480 | sinkhorn_repeat=self.mhc_sinkhorn_iterations, |
| 481 | norm_eps=( |
| 482 | self.input_layernorm.variance_epsilon, |
| 483 | self.post_attention_layernorm.variance_epsilon, |
| 484 | ), |
| 485 | ) |
| 486 | if not include_pre_gemm_splits: |
| 487 | _HC_PRENORM_GEMM_TILELANG_KERNEL.register_warmup( |
| 488 | vllm_config, |
| 489 | hidden_size=self.hidden_size, |
| 490 | hc_mult=self.n, |
| 491 | n_out=self.n * (2 + self.n), |
| 492 | ) |
| 493 | _MHC_POST_TILELANG_KERNEL.register_warmup( |
| 494 | hidden_size=self.hidden_size, |
| 495 | hc_mult=self.n, |
| 496 | ) |
| 497 | _MHC_FUSED_TILELANG_KERNEL.register_warmup( |
| 498 | vllm_config, |
| 499 | hidden_size=self.hidden_size, |
| 500 | hc_mult=self.n, |
| 501 | ) |
| 502 | |
| 503 | def forward( |
| 504 | self, |
| 505 | positions: torch.Tensor, |
| 506 | hidden_states: torch.Tensor, |
| 507 | residual: torch.Tensor | None = None, |
| 508 | post: torch.Tensor | None = None, |
| 509 | comb: torch.Tensor | None = None, |
| 510 | ) -> tuple[ |
| 511 | torch.Tensor, |
| 512 | torch.Tensor | None, |
| 513 | torch.Tensor | None, |
| 514 | torch.Tensor | None, |
| 515 | ]: |
| 516 | # 70B or MTP layers: KDA + MoE without HC. |
| 517 | if not self.mhc or self.is_mtp_layer: |
| 518 | residual = hidden_states |
| 519 | hidden_states = self.input_layernorm(hidden_states) |
| 520 | |
| 521 | attn_output = self.self_attn( |
| 522 | hidden_states=hidden_states, |
| 523 | positions=positions, |
| 524 | ) |
| 525 | hidden_states, residual = self.post_attention_layernorm( |
| 526 | attn_output, residual=residual |
| 527 | ) |
| 528 | hidden_states = self.mlp(hidden_states) |
| 529 | if self.is_mtp_layer: |
| 530 | # Return the unsummed pair: the MTP caller feeds it straight |
| 531 | # into shared_head's fused_add_rms_norm (one kernel instead of |
| 532 | # a separate residual-add + norm). The sum itself is unchanged |
| 533 | # (fp32-accumulated inside the fused kernel). |
| 534 | return hidden_states, residual, None, None |
| 535 | hidden_states = residual + hidden_states |
| 536 | return hidden_states, residual, None, None |
| 537 | |
| 538 | # mHC start. `post`/`comb` carry the previous layer's deferred |
| 539 | # hc_post inputs (its ffn-pre outputs); when present, fuse that |
| 540 | # hc_post with this layer's attn hc_pre into one kernel (inter-layer |
| 541 | # fusion). Layer 0 has no incoming state -> standalone hc_pre. |
| 542 | x = hidden_states |
| 543 | if post is None: |
| 544 | if self.layer_idx == 0: |
| 545 | x = hc_expand(x, self.n) |
| 546 | residual = x |
| 547 | post, comb, x = self.hc_pre( |
| 548 | x, |
| 549 | self.hc_attn_fn, |
| 550 | self.hc_attn_scale, |
| 551 | self.hc_attn_base, |
| 552 | norm_weight=self.input_layernorm.weight.data, |
| 553 | norm_eps=self.input_layernorm.variance_epsilon, |
| 554 | ) |
| 555 | else: |
| 556 | residual, post, comb, x = self.hc_fused_post_pre( |
| 557 | x, |
| 558 | residual, |
| 559 | post, |
| 560 | comb, |
| 561 | self.hc_attn_fn, |
| 562 | self.hc_attn_scale, |
| 563 | self.hc_attn_base, |
| 564 | norm_weight=self.input_layernorm.weight.data, |
| 565 | norm_eps=self.input_layernorm.variance_epsilon, |
| 566 | ) |
| 567 | |
| 568 | # Attention needs the full token sequence; mHC above ran on the SP |
| 569 | # shard. Gather for attention, scatter back afterward (DSv4 pattern). |
| 570 | if self.is_sequence_parallel: |
| 571 | x = sp_all_gather(x)[: positions.shape[0]] |
| 572 | |
| 573 | x = self.self_attn( |
| 574 | hidden_states=x, |
| 575 | positions=positions, |
| 576 | ) |
| 577 | |
| 578 | if self.is_sequence_parallel: |
| 579 | x = sp_reduce_scatter(x) |
| 580 | |
| 581 | # Fuse post-attn hc_post + pre-FFN hc_pre (+ RMSNorm) into one kernel. |
| 582 | residual, post, comb, x = self.hc_fused_post_pre( |
| 583 | x, |
| 584 | residual, |
| 585 | post, |
| 586 | comb, |
| 587 | self.hc_ffn_fn, |
| 588 | self.hc_ffn_scale, |
| 589 | self.hc_ffn_base, |
| 590 | norm_weight=self.post_attention_layernorm.weight.data, |
| 591 | norm_eps=self.post_attention_layernorm.variance_epsilon, |
| 592 | ) |
| 593 | |
| 594 | # Fully Connected |
| 595 | if self._mlp_is_moe: |
| 596 | x = self.mlp(x, already_sequence_parallel=self.is_sequence_parallel) |
| 597 | else: |
| 598 | x = self.mlp(x) |
| 599 | |
| 600 | # mHC end. The last mHC layer materializes its final hc_post (nothing |
| 601 | # to fuse with) then contracts; every other layer defers its hc_post to |
| 602 | # the next layer's fused pre, returning the state. |
| 603 | if self.layer_idx == self.num_hidden_layers - 1: |
| 604 | x = self.hc_post(x, residual, post, comb) |
| 605 | x = hc_contract(x, self.n) |
| 606 | return x, None, None, None |
| 607 | |
| 608 | return x, residual, post, comb |
| 609 | |
| 610 | def hc_pre( |
| 611 | self, |
| 612 | x: torch.Tensor, |
| 613 | hc_fn: torch.Tensor, |
| 614 | hc_scale: torch.Tensor, |
| 615 | hc_base: torch.Tensor, |
| 616 | norm_weight: torch.Tensor | None = None, |
| 617 | norm_eps: float = 0.0, |
| 618 | ): |
| 619 | post_mix, res_mix, layer_input = self.mhc_pre_op( |
| 620 | residual=x, |
| 621 | fn=hc_fn, |
| 622 | hc_scale=hc_scale, |
| 623 | hc_base=hc_base, |
| 624 | rms_eps=self.rms_norm_eps, |
| 625 | hc_pre_eps=self.hc_eps, |
| 626 | hc_sinkhorn_eps=self.hc_eps, |
| 627 | hc_post_mult_value=self.mhc_post_mult_value, |
| 628 | sinkhorn_repeat=self.mhc_sinkhorn_iterations, |
| 629 | norm_weight=norm_weight, |
| 630 | norm_eps=norm_eps, |
| 631 | ) |
| 632 | return post_mix, res_mix, layer_input |
| 633 | |
| 634 | def hc_post( |
| 635 | self, |
| 636 | x: torch.Tensor, |
| 637 | residual: torch.Tensor, |
| 638 | post: torch.Tensor, |
| 639 | comb: torch.Tensor, |
| 640 | ): |
| 641 | return self.mhc_post_op(x, residual, post, comb) |
| 642 | |
| 643 | def hc_fused_post_pre( |
| 644 | self, |
| 645 | x: torch.Tensor, |
| 646 | residual: torch.Tensor, |
| 647 | post: torch.Tensor, |
| 648 | comb: torch.Tensor, |
| 649 | hc_fn: torch.Tensor, |
| 650 | hc_scale: torch.Tensor, |
| 651 | hc_base: torch.Tensor, |
| 652 | norm_weight: torch.Tensor | None = None, |
| 653 | norm_eps: float = 0.0, |
| 654 | ): |
| 655 | return self.mhc_fused_post_pre_op( |
| 656 | x=x, |
| 657 | residual=residual, |
| 658 | post_layer_mix=post, |
| 659 | comb_res_mix=comb, |
| 660 | fn=hc_fn, |
| 661 | hc_scale=hc_scale, |
| 662 | hc_base=hc_base, |
| 663 | rms_eps=self.rms_norm_eps, |
| 664 | hc_pre_eps=self.hc_eps, |
| 665 | hc_sinkhorn_eps=self.hc_eps, |
| 666 | hc_post_mult_value=self.mhc_post_mult_value, |
| 667 | sinkhorn_repeat=self.mhc_sinkhorn_iterations, |
| 668 | n_splits=1, |
| 669 | tile_n=1, |
| 670 | norm_weight=norm_weight, |
| 671 | norm_eps=norm_eps, |
| 672 | ) |
| 673 | |
| 674 | |
| 675 | class Glm5NextModel(nn.Module): |
| 676 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 677 | super().__init__() |
| 678 | |
| 679 | config = vllm_config.model_config.hf_text_config |
| 680 | _validate_supported_config(config) |
| 681 | self.config = config |
| 682 | |
| 683 | self.vocab_size = config.vocab_size |
| 684 | self.device = current_platform.device_type |
| 685 | |
| 686 | self.is_v32 = config.index_topk is not None |
| 687 | if self.is_v32: |
| 688 | topk_tokens = config.index_topk |
| 689 | assert topk_tokens is not None |
| 690 | # Reserve room for the incomplete pool tail. |
| 691 | kpool = config.index_kpool |
| 692 | assert kpool is not None |
| 693 | buffer_width = topk_tokens + (kpool - 1 if kpool > 1 else 0) |
| 694 | # Sparse MLA tiles top-k in 128 columns; padded slots remain masked. |
| 695 | sparse_topk_block_n = 128 |
| 696 | buffer_width = ( |
| 697 | (buffer_width + sparse_topk_block_n - 1) // sparse_topk_block_n |
| 698 | ) * sparse_topk_block_n |
| 699 | topk_indices_buffer = torch.empty( |
| 700 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 701 | buffer_width, |
| 702 | dtype=torch.int32, |
| 703 | device=self.device, |
| 704 | ) |
| 705 | else: |
| 706 | # Full-MLA config (no kpool sparse indexer): no topk buffer. |
| 707 | topk_indices_buffer = None |
| 708 | |
| 709 | if get_pp_group().is_first_rank: |
| 710 | self.embed_tokens = VocabParallelEmbedding( |
| 711 | config.vocab_size, |
| 712 | config.hidden_size, |
| 713 | prefix=f"{prefix}.embed_tokens", |
| 714 | ) |
| 715 | else: |
| 716 | self.embed_tokens = PPMissingLayer() |
| 717 | |
| 718 | def get_layer(prefix: str): |
| 719 | layer_idx = int(prefix.rsplit(".", 1)[1]) |
| 720 | return Glm5NextDecoderLayer( |
| 721 | vllm_config=vllm_config, |
| 722 | config=config, |
| 723 | layer_idx=layer_idx, |
| 724 | prefix=prefix, |
| 725 | topk_indices_buffer=topk_indices_buffer, |
| 726 | ) |
| 727 | |
| 728 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 729 | config.num_hidden_layers, |
| 730 | get_layer, |
| 731 | prefix=f"{prefix}.layers", |
| 732 | ) |
| 733 | # The active slice is fixed after construction; cache it so forward |
| 734 | # doesn't rebuild the slice (a fresh list) every step. |
| 735 | self._active_layers = self.layers[self.start_layer : self.end_layer] |
| 736 | |
| 737 | if get_pp_group().is_last_rank: |
| 738 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 739 | else: |
| 740 | self.norm = PPMissingLayer() |
| 741 | |
| 742 | self.is_sequence_parallel = ( |
| 743 | vllm_config.parallel_config.use_sequence_parallel_moe |
| 744 | ) |
| 745 | |
| 746 | world_size = get_tensor_model_parallel_world_size() |
| 747 | assert config.num_attention_heads % world_size == 0, ( |
| 748 | "num_attention_heads must be divisible by world_size" |
| 749 | ) |
| 750 | |
| 751 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 752 | return self.embed_tokens(input_ids) |
| 753 | |
| 754 | def forward( |
| 755 | self, |
| 756 | input_ids: torch.Tensor | None, |
| 757 | positions: torch.Tensor, |
| 758 | intermediate_tensors: IntermediateTensors | None, |
| 759 | inputs_embeds: torch.Tensor | None = None, |
| 760 | **kwargs, |
| 761 | ) -> torch.Tensor: |
| 762 | if get_pp_group().is_first_rank: |
| 763 | if inputs_embeds is not None: |
| 764 | hidden_states = inputs_embeds |
| 765 | else: |
| 766 | hidden_states = self.embed_input_ids(input_ids) |
| 767 | residual = None |
| 768 | post = None |
| 769 | comb = None |
| 770 | else: |
| 771 | assert intermediate_tensors is not None |
| 772 | hidden_states = intermediate_tensors["hidden_states"] |
| 773 | residual = intermediate_tensors["residual"] |
| 774 | # post/comb (deferred mHC hc_post state) are not propagated across |
| 775 | # PP ranks; the receiving rank's first mHC layer uses standalone pre. |
| 776 | post = None |
| 777 | comb = None |
| 778 | |
| 779 | full_num_tokens = positions.shape[0] |
| 780 | if self.is_sequence_parallel: |
| 781 | hidden_states = sp_shard(hidden_states) |
| 782 | |
| 783 | for layer in self._active_layers: |
| 784 | hidden_states, residual, post, comb = layer( |
| 785 | positions, hidden_states, residual, post, comb |
| 786 | ) |
| 787 | |
| 788 | if not get_pp_group().is_last_rank: |
| 789 | # PP is gated off for GLM-5.3-Flash (no make_empty_intermediate_tensors), |
| 790 | # so this branch is not exercised. post/comb are the deferred |
| 791 | # hc_post state of this rank's last mHC layer; a future PP path |
| 792 | # would need to propagate them, but for now they are dropped (the |
| 793 | # receiving rank's first layer would fall back to standalone pre). |
| 794 | return IntermediateTensors( |
| 795 | {"hidden_states": hidden_states, "residual": residual} |
| 796 | ) |
| 797 | |
| 798 | if self.is_sequence_parallel: |
| 799 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 800 | |
| 801 | hidden_states = self.norm(hidden_states) |
| 802 | return hidden_states |
| 803 | |
| 804 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 805 | stacked_params_mapping = [ |
| 806 | # (param_name, shard_name, shard_id) |
| 807 | (".gate_up_proj", ".gate_proj", 0), |
| 808 | (".gate_up_proj", ".up_proj", 1), |
| 809 | # MLA: fuse q_a_proj and kv_a_proj_with_mqa |
| 810 | (".fused_qkv_a_proj", ".q_a_proj", 0), |
| 811 | (".fused_qkv_a_proj", ".kv_a_proj_with_mqa", 1), |
| 812 | # Indexer: fuse wk and weights_proj |
| 813 | (".wk_weights_proj", ".wk", 0), |
| 814 | (".wk_weights_proj", ".weights_proj", 1), |
| 815 | # KDA: merge q, k, v, b, f_a, g_a projections into one GEMM |
| 816 | (".in_proj_qkvbfg_a", ".q_proj", 0), |
| 817 | (".in_proj_qkvbfg_a", ".k_proj", 1), |
| 818 | (".in_proj_qkvbfg_a", ".v_proj", 2), |
| 819 | (".in_proj_qkvbfg_a", ".b_proj", 3), |
| 820 | (".in_proj_qkvbfg_a", ".f_a_proj", 4), |
| 821 | (".in_proj_qkvbfg_a", ".g_a_proj", 5), |
| 822 | ] |
| 823 | if _is_moe(self.config): |
| 824 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 825 | # (param_name, weight_name, expert_id, shard_id) |
| 826 | # EPLB: the mapping enumerates physical experts, so it must cover |
| 827 | # the redundant replicas or their slots are never loaded. |
| 828 | num_redundant_experts = next( |
| 829 | ( |
| 830 | layer.mlp.n_redundant_experts |
| 831 | for layer in self.layers |
| 832 | if isinstance(layer, Glm5NextDecoderLayer) |
| 833 | and isinstance(layer.mlp, Glm5NextMoE) |
| 834 | ), |
| 835 | 0, |
| 836 | ) |
| 837 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 838 | self, |
| 839 | ckpt_gate_proj_name="gate_proj", |
| 840 | ckpt_down_proj_name="down_proj", |
| 841 | ckpt_up_proj_name="up_proj", |
| 842 | num_experts=self.config.n_routed_experts, |
| 843 | num_redundant_experts=num_redundant_experts, |
| 844 | ) |
| 845 | else: |
| 846 | expert_params_mapping = [] |
| 847 | params_dict = dict(self.named_parameters()) |
| 848 | loaded_params: set[str] = set() |
| 849 | |
| 850 | # GLM-5.3-Flash NoPE checkpoints omit the RoPE rows from |
| 851 | # ``kv_a_proj_with_mqa``; pad them with zeros for the model shape. |
| 852 | kv_a_pad_size = 0 |
| 853 | if self.config.mla_use_nope and self.config.qk_rope_head_dim > 0: |
| 854 | kv_a_pad_size = self.config.qk_rope_head_dim |
| 855 | |
| 856 | _pending_wk_fp8: dict = {} |
| 857 | |
| 858 | for args in weights: |
| 859 | name, loaded_weight = args[:2] |
| 860 | kwargs: dict = args[2] if len(args) > 2 else {} |
| 861 | if "rotary_emb.inv_freq" in name: |
| 862 | continue |
| 863 | |
| 864 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) |
| 865 | if spec_layer is not None: |
| 866 | continue # skip spec decode layers for main model |
| 867 | if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name: |
| 868 | # Models trained using ColossalAI may include these tensors in |
| 869 | # the checkpoint. Skip them. |
| 870 | continue |
| 871 | |
| 872 | # Handle FP8 indexer WK: dequantize to BF16 for fusion with |
| 873 | # weights_proj into wk_weights_proj. |
| 874 | if _try_load_fp8_indexer_wk( |
| 875 | name, |
| 876 | loaded_weight, |
| 877 | _pending_wk_fp8, |
| 878 | params_dict, |
| 879 | loaded_params, |
| 880 | ): |
| 881 | continue |
| 882 | |
| 883 | # FP8 checkpoint: dequantize BF16-kept MLA projections |
| 884 | # (q_a_proj / kv_a_proj_with_mqa / o_proj) to BF16. |
| 885 | if _try_load_fp8_attn_proj( |
| 886 | name, |
| 887 | loaded_weight, |
| 888 | _pending_wk_fp8, |
| 889 | params_dict, |
| 890 | loaded_params, |
| 891 | kv_a_pad_size, |
| 892 | ): |
| 893 | continue |
| 894 | |
| 895 | # Pad kv_a_proj_with_mqa for NoPE models |
| 896 | if kv_a_pad_size > 0 and ".kv_a_proj_with_mqa." in name: |
| 897 | pad = torch.zeros( |
| 898 | kv_a_pad_size, |
| 899 | *loaded_weight.shape[1:], |
| 900 | dtype=loaded_weight.dtype, |
| 901 | device=loaded_weight.device, |
| 902 | ) |
| 903 | loaded_weight = torch.cat([loaded_weight, pad], dim=0) |
| 904 | |
| 905 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 906 | if weight_name not in name: |
| 907 | continue |
| 908 | # We have mlp.experts[0].gate_proj in the checkpoint. |
| 909 | # Since we handle the experts below in expert_params_mapping, |
| 910 | # we need to skip here BEFORE we update the name, otherwise |
| 911 | # name will be updated to mlp.experts[0].gate_up_proj, which |
| 912 | # will then be updated below in expert_params_mapping |
| 913 | # for mlp.experts[0].gate_gate_up_proj, which breaks load. |
| 914 | if ("mlp.experts." in name) and name not in params_dict: |
| 915 | continue |
| 916 | name_mapped = name.replace(weight_name, param_name) |
| 917 | # QKV fusion: skip if fused module doesn't exist in model |
| 918 | if param_name == ".fused_qkv_a_proj" and name_mapped not in params_dict: |
| 919 | continue |
| 920 | name = name_mapped |
| 921 | # Skip loading extra bias for GPTQ models. |
| 922 | if name.endswith(".bias") and name not in params_dict: |
| 923 | continue |
| 924 | if is_pp_missing_parameter(name, self): |
| 925 | continue |
| 926 | param = params_dict[name] |
| 927 | weight_loader = param.weight_loader |
| 928 | weight_loader(param, loaded_weight, shard_id) |
| 929 | break |
| 930 | else: |
| 931 | is_expert_weight = False |
| 932 | for ( |
| 933 | param_name, |
| 934 | weight_name, |
| 935 | expert_id, |
| 936 | expert_shard_id, |
| 937 | ) in expert_params_mapping: |
| 938 | if weight_name not in name: |
| 939 | continue |
| 940 | # A checkpoint expert may map to several physical replicas |
| 941 | # under EPLB; keep `name` intact and try the next entry |
| 942 | # when this physical expert is not local to the rank. |
| 943 | is_expert_weight = True |
| 944 | name_mapped = name.replace(weight_name, param_name) |
| 945 | if is_pp_missing_parameter(name_mapped, self): |
| 946 | continue |
| 947 | param = params_dict[name_mapped] |
| 948 | weight_loader = param.weight_loader |
| 949 | success = weight_loader( |
| 950 | param, |
| 951 | loaded_weight, |
| 952 | name_mapped, |
| 953 | expert_id=expert_id, |
| 954 | shard_id=expert_shard_id, |
| 955 | return_success=True, |
| 956 | ) |
| 957 | if success: |
| 958 | name = name_mapped |
| 959 | break |
| 960 | else: |
| 961 | if is_expert_weight: |
| 962 | continue |
| 963 | # Skip loading extra bias for GPTQ models. |
| 964 | if ( |
| 965 | name.endswith(".bias") |
| 966 | and name not in params_dict |
| 967 | and not _is_linear_attn(self.config) |
| 968 | ): # noqa: E501 |
| 969 | continue |
| 970 | # Remapping the name of FP8 kv-scale. |
| 971 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 972 | if remapped_name is None: |
| 973 | continue |
| 974 | name = remapped_name |
| 975 | if is_pp_missing_parameter(name, self): |
| 976 | continue |
| 977 | |
| 978 | param = params_dict[name] |
| 979 | weight_loader = getattr( |
| 980 | param, "weight_loader", default_weight_loader |
| 981 | ) |
| 982 | weight_loader(param, loaded_weight, **kwargs) |
| 983 | loaded_params.add(name) |
| 984 | return loaded_params |
| 985 | |
| 986 | |
| 987 | class Glm5NextForCausalLM( |
| 988 | nn.Module, HasInnerState, SupportsPP, MixtureOfExperts, IsHybrid |
| 989 | ): |
| 990 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 991 | super().__init__() |
| 992 | self.model_config = vllm_config.model_config |
| 993 | self.vllm_config = vllm_config |
| 994 | self.config = self.model_config.hf_text_config |
| 995 | quant_config = vllm_config.quant_config |
| 996 | self.quant_config = quant_config |
| 997 | self.model = Glm5NextModel( |
| 998 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 999 | ) |
| 1000 | if get_pp_group().is_last_rank: |
| 1001 | self.lm_head = ParallelLMHead( |
| 1002 | self.config.vocab_size, |
| 1003 | self.config.hidden_size, |
| 1004 | quant_config=quant_config, |
| 1005 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1006 | ) |
| 1007 | else: |
| 1008 | self.lm_head = PPMissingLayer() |
| 1009 | self.logits_processor = LogitsProcessor( |
| 1010 | self.config.vocab_size, scale=_LOGIT_SCALE |
| 1011 | ) |
| 1012 | |
| 1013 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1014 | return self.model.embed_input_ids(input_ids) |
| 1015 | |
| 1016 | def forward( |
| 1017 | self, |
| 1018 | input_ids: torch.Tensor | None, |
| 1019 | positions: torch.Tensor, |
| 1020 | intermediate_tensors: IntermediateTensors | None = None, |
| 1021 | inputs_embeds: torch.Tensor | None = None, |
| 1022 | **kwargs, |
| 1023 | ) -> torch.Tensor | IntermediateTensors: |
| 1024 | hidden_states = self.model( |
| 1025 | input_ids, positions, intermediate_tensors, inputs_embeds, **kwargs |
| 1026 | ) |
| 1027 | return hidden_states |
| 1028 | |
| 1029 | @classmethod |
| 1030 | def get_mamba_state_dtype_from_config( |
| 1031 | cls, |
| 1032 | vllm_config: "VllmConfig", |
| 1033 | ) -> tuple[torch.dtype, torch.dtype]: |
| 1034 | return MambaStateDtypeCalculator.kda_state_dtype( |
| 1035 | vllm_config.model_config.dtype, vllm_config.cache_config.mamba_cache_dtype |
| 1036 | ) |
| 1037 | |
| 1038 | @classmethod |
| 1039 | def get_mamba_state_shape_from_config( |
| 1040 | cls, vllm_config: "VllmConfig" |
| 1041 | ) -> tuple[tuple[int, int], tuple[int, int, int]]: |
| 1042 | parallel_config = vllm_config.parallel_config |
| 1043 | hf_config = vllm_config.model_config.hf_text_config |
| 1044 | tp_size = parallel_config.tensor_parallel_size |
| 1045 | num_spec = ( |
| 1046 | vllm_config.speculative_config.num_speculative_tokens |
| 1047 | if vllm_config.speculative_config |
| 1048 | else 0 |
| 1049 | ) |
| 1050 | return MambaStateShapeCalculator.kda_state_shape( |
| 1051 | tp_size, |
| 1052 | hf_config.linear_num_heads, |
| 1053 | hf_config.linear_head_dim, |
| 1054 | conv_kernel_size=hf_config.linear_conv_kernel_dim, |
| 1055 | num_spec=num_spec, |
| 1056 | ) |
| 1057 | |
| 1058 | @classmethod |
| 1059 | def get_mamba_state_copy_func( |
| 1060 | cls, |
| 1061 | ) -> tuple[ |
| 1062 | MambaStateCopyFunc, MambaStateCopyFunc, MambaStateCopyFunc, MambaStateCopyFunc |
| 1063 | ]: |
| 1064 | return MambaStateCopyFuncCalculator.kda_state_copy_func() |
| 1065 | |
| 1066 | def compute_logits( |
| 1067 | self, |
| 1068 | hidden_states: torch.Tensor, |
| 1069 | ) -> torch.Tensor | None: |
| 1070 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1071 | return logits |
| 1072 | |
| 1073 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1074 | loader = AutoWeightsLoader(self) |
| 1075 | return loader.load_weights(weights) |
| 1076 | |
| 1077 | |
| 1078 | @MULTIMODAL_REGISTRY.register_processor( |
| 1079 | Glm5NextMultiModalProcessor, |
| 1080 | info=Glm5NextProcessingInfo, |
| 1081 | dummy_inputs=Glm4vDummyInputsBuilder, |
| 1082 | ) |
| 1083 | class Glm5NextForConditionalGeneration( |
| 1084 | Glm4vForConditionalGeneration, HasInnerState, IsHybrid, MixtureOfExperts |
| 1085 | ): |
| 1086 | # The text model (KDA + dense-MLA + MoE) is a hybrid mamba model. The |
| 1087 | # multimodal wrapper must declare the same interfaces so vLLM treats it as |
| 1088 | # hybrid (auto-aligns mamba/attention block sizes, sizes the mamba state |
| 1089 | # cache); the mamba-state classmethods delegate to the text model. |
| 1090 | has_inner_state: ClassVar[Literal[True]] = True |
| 1091 | is_hybrid: ClassVar[Literal[True]] = True |
| 1092 | |
| 1093 | # GLM-5.3-Flash stores the dense-MLP gate/up as separate tensors (like |
| 1094 | # ``Glm4vMoeForConditionalGeneration``, ``glm4_moe`` and ``deepseek_v2``), |
| 1095 | # so the fused ``gate_up_proj`` must expand to its real shard names for |
| 1096 | # per-layer quant-scheme resolution. The identity ``gate_up_proj`` entry |
| 1097 | # inherited from ``Glm4vForConditionalGeneration`` (pre-fused gate_up_proj) |
| 1098 | # would otherwise route the module to ``global_quant_config`` and mismatch |
| 1099 | # at load for mixed-precision Quark checkpoints. |
| 1100 | packed_modules_mapping = { |
| 1101 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 1102 | } |
| 1103 | |
| 1104 | # NOTE: weight-prefix mapping is inherited from Glm4vForConditionalGeneration |
| 1105 | # (``model.visual.`` -> ``visual.``, ``model.language_model.`` -> |
| 1106 | # ``language_model.model.``, ``lm_head.`` -> ``language_model.lm_head.``), |
| 1107 | # matching the GLM-OCR / GLM-4V serialization convention. If the real |
| 1108 | # checkpoint's safetensors keys differ (e.g. ``language_model.model.`` with |
| 1109 | # no outer ``model.``), override ``hf_to_vllm_mapper`` accordingly. |
| 1110 | |
| 1111 | @classmethod |
| 1112 | def get_mamba_state_dtype_from_config(cls, vllm_config: VllmConfig): |
| 1113 | from .model import Glm5NextForCausalLM |
| 1114 | |
| 1115 | return Glm5NextForCausalLM.get_mamba_state_dtype_from_config(vllm_config) |
| 1116 | |
| 1117 | @classmethod |
| 1118 | def get_mamba_state_shape_from_config(cls, vllm_config: VllmConfig): |
| 1119 | from .model import Glm5NextForCausalLM |
| 1120 | |
| 1121 | return Glm5NextForCausalLM.get_mamba_state_shape_from_config(vllm_config) |
| 1122 | |
| 1123 | @classmethod |
| 1124 | def get_mamba_state_copy_func(cls): |
| 1125 | from .model import Glm5NextForCausalLM |
| 1126 | |
| 1127 | return Glm5NextForCausalLM.get_mamba_state_copy_func() |
| 1128 | |
| 1129 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1130 | super(Glm4vForConditionalGeneration, self).__init__() |
| 1131 | config = vllm_config.model_config.hf_config |
| 1132 | multimodal_config = vllm_config.model_config.multimodal_config |
| 1133 | assert multimodal_config is not None |
| 1134 | |
| 1135 | self.config = config |
| 1136 | self.model_config = vllm_config.model_config |
| 1137 | self.multimodal_config = multimodal_config |
| 1138 | self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" |
| 1139 | self.is_multimodal_pruning_enabled = ( |
| 1140 | multimodal_config.is_multimodal_pruning_enabled() |
| 1141 | ) |
| 1142 | |
| 1143 | with self._mark_tower_model(vllm_config, {"image", "video"}): |
| 1144 | self.visual = Glm5NextVisionTransformer( |
| 1145 | config.text_config, |
| 1146 | config.vision_config, |
| 1147 | norm_eps=_VISION_RMS_NORM_EPS, |
| 1148 | # Vision tower ships BF16 weights in this fp8 checkpoint (no |
| 1149 | # weight_scale_inv for visual.*), so it must NOT inherit the |
| 1150 | # global fp8 quant_config -- doing so incorrectly quantizes |
| 1151 | # the tower |
| 1152 | # and yields NaN image features. Mirrors the MLA/KDA proj |
| 1153 | # pattern (quant_config=None for BF16 submodules). |
| 1154 | quant_config=None, |
| 1155 | input_norm=build_mm_input_norm(self.model_config), |
| 1156 | prefix=maybe_prefix(prefix, "visual"), |
| 1157 | ) |
| 1158 | |
| 1159 | with self._mark_language_model(vllm_config): |
| 1160 | self.language_model = init_vllm_registered_model( |
| 1161 | vllm_config=vllm_config, |
| 1162 | hf_config=config.text_config, |
| 1163 | prefix=maybe_prefix(prefix, "language_model"), |
| 1164 | architectures=["Glm5NextForCausalLM"], |
| 1165 | ) |
| 1166 | |
| 1167 | self.set_moe_parameters() |
| 1168 | |
| 1169 | # Glm5NextForCausalLM does not implement make_empty_intermediate_tensors, |
| 1170 | # so pipeline parallelism is gated off (consistent with the text-only |
| 1171 | # model) and we intentionally do not alias it here. |
| 1172 | |
| 1173 | def set_moe_parameters(self) -> None: |
| 1174 | self.moe_mlp_layers = [ |
| 1175 | layer.mlp |
| 1176 | for layer in self.language_model.model.layers |
| 1177 | if isinstance(layer, Glm5NextDecoderLayer) |
| 1178 | and isinstance(layer.mlp, Glm5NextMoE) |
| 1179 | ] |
| 1180 | self.moe_layers = [moe.experts for moe in self.moe_mlp_layers] |
| 1181 | self.num_moe_layers = len(self.moe_layers) |
| 1182 | if not self.num_moe_layers: |
| 1183 | return |
| 1184 | example_moe = self.moe_mlp_layers[0] |
| 1185 | self.num_expert_groups = self.config.text_config.n_group |
| 1186 | self.num_logical_experts = example_moe.n_logical_experts |
| 1187 | self.num_physical_experts = example_moe.n_physical_experts |
| 1188 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 1189 | self.num_routed_experts = example_moe.n_routed_experts |
| 1190 | self.num_shared_experts = example_moe.n_shared_experts |
| 1191 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 1192 | |
| 1193 | def update_physical_experts_metadata( |
| 1194 | self, |
| 1195 | num_physical_experts: int, |
| 1196 | num_local_physical_experts: int, |
| 1197 | ) -> None: |
| 1198 | if not self.num_moe_layers: |
| 1199 | return |
| 1200 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1201 | self.num_physical_experts = num_physical_experts |
| 1202 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1203 | for moe in self.moe_mlp_layers: |
| 1204 | moe.n_physical_experts = num_physical_experts |
| 1205 | moe.n_redundant_experts = self.num_redundant_experts |
| 1206 | moe.experts.update_expert_map() |
| 1207 | |
| 1208 | def get_encoder_cudagraph_config(self): |
| 1209 | # This vision tower does not produce the absolute position embedding |
| 1210 | # buffer used by GLM4V. |
| 1211 | config = super().get_encoder_cudagraph_config() |
| 1212 | config.buffer_keys = [k for k in config.buffer_keys if k != "pos_embeds"] |
| 1213 | return config |
| 1214 | |
| 1215 | |
| 1216 | def get_spec_layer_idx_from_weight_name( |
| 1217 | config: Glm5NextTextConfig, weight_name: str |
| 1218 | ) -> int | None: |
| 1219 | if hasattr(config, "num_nextn_predict_layers") and ( |
| 1220 | config.num_nextn_predict_layers > 0 |
| 1221 | ): |
| 1222 | layer_idx = config.num_hidden_layers |
| 1223 | for i in range(config.num_nextn_predict_layers): |
| 1224 | if weight_name.startswith( |
| 1225 | f"model.layers.{layer_idx + i}." |
| 1226 | ) or weight_name.startswith(f"layers.{layer_idx + i}."): |
| 1227 | return layer_idx + i |
| 1228 | return None |
| 1229 | |
| 1230 | |
| 1231 | def _try_load_fp8_indexer_wk(name, tensor, buf, params_dict, loaded_params): |
| 1232 | if "indexer.wk." not in name or "wk_weights" in name: |
| 1233 | return False |
| 1234 | is_weight = name.endswith(".weight") and tensor.dtype == torch.float8_e4m3fn |
| 1235 | is_scale = "weight_scale_inv" in name |
| 1236 | if not is_weight and not is_scale: |
| 1237 | return False |
| 1238 | layer_prefix = name.rsplit(".wk.", 1)[0] |
| 1239 | entry = buf.setdefault(layer_prefix, {}) |
| 1240 | entry["weight" if is_weight else "scale"] = tensor |
| 1241 | if "weight" not in entry or "scale" not in entry: |
| 1242 | return True |
| 1243 | |
| 1244 | weight_fp8, scale_inv = entry["weight"], entry["scale"] |
| 1245 | del buf[layer_prefix] |
| 1246 | block_size = weight_fp8.shape[1] // scale_inv.shape[1] |
| 1247 | weight_bf16 = scaled_dequantize( |
| 1248 | weight_fp8, |
| 1249 | scale_inv, |
| 1250 | group_shape=GroupShape(block_size, block_size), |
| 1251 | out_dtype=torch.bfloat16, |
| 1252 | ) |
| 1253 | |
| 1254 | fused_name = f"{layer_prefix}.wk_weights_proj.weight" |
| 1255 | param = params_dict[fused_name] |
| 1256 | param.weight_loader(param, weight_bf16, 0) |
| 1257 | loaded_params.add(fused_name) |
| 1258 | return True |
| 1259 | |
| 1260 | |
| 1261 | def _dequant_fp8_block( |
| 1262 | weight_fp8: torch.Tensor, |
| 1263 | scale_inv: torch.Tensor, |
| 1264 | block_size: int = 128, |
| 1265 | ) -> torch.Tensor: |
| 1266 | """Dequantize a block-FP8 (e4m3) weight with per-block scale to BF16. |
| 1267 | |
| 1268 | Unlike ``scaled_dequantize`` this tolerates a non-divisible (partial last |
| 1269 | block) shape by zero-padding to a multiple of ``block_size`` before the |
| 1270 | scale broadcast and trimming back afterwards (e.g. kv_a_proj_with_mqa is |
| 1271 | 576 rows = 4*128 + 64). |
| 1272 | """ |
| 1273 | out_dim, in_dim = weight_fp8.shape |
| 1274 | pad_out = (-out_dim) % block_size |
| 1275 | pad_in = (-in_dim) % block_size |
| 1276 | w = weight_fp8 |
| 1277 | if pad_out or pad_in: |
| 1278 | w = torch.nn.functional.pad(w, (0, pad_in, 0, pad_out)) |
| 1279 | # scale_inv is (ceil(out/block), ceil(in/block)); broadcast to (out, in). |
| 1280 | s = scale_inv.to(torch.float32) |
| 1281 | s_full = s.repeat_interleave(block_size, dim=0).repeat_interleave(block_size, dim=1) |
| 1282 | out = (w.to(torch.float32) * s_full).to(torch.bfloat16) |
| 1283 | return out[:out_dim, :in_dim].contiguous() |
| 1284 | |
| 1285 | |
| 1286 | # FP8 checkpoint projections that the MODEL keeps in BF16, so the block-FP8 |
| 1287 | # (weight + weight_scale_inv) must be dequantized to BF16 on load. |
| 1288 | # Maps checkpoint proj-suffix -> (buffer key, model target base, fused shard id |
| 1289 | # or None for a direct projection, whether NoPE rope-padding applies). |
| 1290 | _FP8_ATTN_PROJS = { |
| 1291 | ".q_a_proj.": ("q_a", "fused_qkv_a_proj", 0, False), |
| 1292 | ".kv_a_proj_with_mqa.": ("kv_a", "fused_qkv_a_proj", 1, True), |
| 1293 | ".q_b_proj.": ("q_b", "q_b_proj", None, False), |
| 1294 | ".o_proj.": ("o_proj", "o_proj", None, False), |
| 1295 | } |
| 1296 | |
| 1297 | |
| 1298 | def _try_load_fp8_attn_proj( |
| 1299 | name, |
| 1300 | tensor, |
| 1301 | buf, |
| 1302 | params_dict, |
| 1303 | loaded_params, |
| 1304 | kv_a_pad_size: int, |
| 1305 | ) -> bool: |
| 1306 | """Dequantize FP8 q_a_proj / kv_a_proj_with_mqa / o_proj to BF16 on load. |
| 1307 | |
| 1308 | The FP8 checkpoint stores these as block-FP8 (weight + weight_scale_inv), |
| 1309 | but the model holds them in BF16 (``fused_qkv_a_proj`` is always BF16 via |
| 1310 | DeepSeekV2FusedQkvAProjLinear; ``o_proj`` is excluded by |
| 1311 | modules_to_not_convert). When the model target is BF16 (no |
| 1312 | ``weight_scale_inv`` param) we dequantize; otherwise we return False so the |
| 1313 | normal stacked/direct path loads the FP8 tensor as-is. |
| 1314 | """ |
| 1315 | matched = None |
| 1316 | for suffix, info in _FP8_ATTN_PROJS.items(): |
| 1317 | if suffix in name: |
| 1318 | matched = (suffix, info) |
| 1319 | break |
| 1320 | if matched is None: |
| 1321 | return False |
| 1322 | suffix, (key, target_base, shard_id, is_kva) = matched |
| 1323 | is_weight = name.endswith(".weight") and tensor.dtype == torch.float8_e4m3fn |
| 1324 | # Need to accept both the DeepSeek-native ``weight_scale_inv`` and the Quark |
| 1325 | # ``weight_scale`` names before feeding the shared block dequant below. |
| 1326 | is_scale = "weight_scale_inv" in name or name.endswith(".weight_scale") |
| 1327 | if not is_weight and not is_scale: |
| 1328 | return False |
| 1329 | |
| 1330 | layer_prefix = name.rsplit(suffix, 1)[0] |
| 1331 | target_w = f"{layer_prefix}.{target_base}.weight" |
| 1332 | target_s = f"{layer_prefix}.{target_base}.weight_scale_inv" |
| 1333 | # If the model actually kept this projection in FP8, let the normal path |
| 1334 | # handle it (it has a weight_scale_inv param). |
| 1335 | if target_s in params_dict: |
| 1336 | return False |
| 1337 | |
| 1338 | entry = buf.setdefault(layer_prefix, {}).setdefault(key, {}) |
| 1339 | entry["weight" if is_weight else "scale"] = tensor |
| 1340 | if "weight" not in entry or "scale" not in entry: |
| 1341 | return True |
| 1342 | |
| 1343 | weight_fp8, scale_inv = entry["weight"], entry["scale"] |
| 1344 | buf[layer_prefix].pop(key, None) |
| 1345 | block_size = weight_fp8.shape[1] // scale_inv.shape[1] |
| 1346 | weight_bf16 = _dequant_fp8_block(weight_fp8, scale_inv, block_size) |
| 1347 | # NoPE: pad kv_a rope portion (kv_lora_rank -> kv_lora_rank + qk_rope_head_dim). |
| 1348 | if is_kva and kv_a_pad_size > 0: |
| 1349 | pad = torch.zeros( |
| 1350 | kv_a_pad_size, |
| 1351 | weight_bf16.shape[1], |
| 1352 | dtype=weight_bf16.dtype, |
| 1353 | device=weight_bf16.device, |
| 1354 | ) |
| 1355 | weight_bf16 = torch.cat([weight_bf16, pad], dim=0) |
| 1356 | |
| 1357 | param = params_dict[target_w] |
| 1358 | if shard_id is None: |
| 1359 | param.weight_loader(param, weight_bf16) |
| 1360 | else: |
| 1361 | param.weight_loader(param, weight_bf16, shard_id) |
| 1362 | loaded_params.add(target_w) |
| 1363 | return True |