# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project # Adapted from # https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py # Copyright 2023 The vLLM team. # # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX # and OPT implementations in this library. It has been modified from its # original forms to accommodate minor architectural differences compared # to GPT-NeoX and OPT used by the Meta AI team that trained the model. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Inference-only A.X K1 model.""" import typing from collections.abc import Callable, Iterable from itertools import islice import torch from torch import nn from vllm._aiter_ops import rocm_aiter_ops from vllm.compilation.decorators import support_torch_compile from vllm.config import CacheConfig, ParallelConfig, VllmConfig from vllm.distributed import ( get_ep_group, get_pp_group, get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size, tensor_model_parallel_all_gather, ) from vllm.logger import init_logger from vllm.model_executor.layers.attention import Attention from vllm.model_executor.layers.fused_moe import SharedFusedMoE from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import ( ColumnParallelLinear, MergedColumnParallelLinear, ReplicatedLinear, RowParallelLinear, ) from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding, ) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name, ) from vllm.model_executor.models.deepseek_v2 import ( DeepseekAttention, DeepseekV2MLP, yarn_get_mscale, ) from vllm.model_executor.models.utils import sequence_parallel_chunk from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs.AXK1 import AXK1Config from .interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP from .utils import ( PPMissingLayer, is_pp_missing_parameter, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix, ) logger = init_logger(__name__) class AXK1MLP(DeepseekV2MLP): pass class AXK1MoE(nn.Module): def __init__( self, config: AXK1Config, parallel_config: ParallelConfig, quant_config: QuantizationConfig | None = None, prefix: str = "", ): super().__init__() self.tp_size = get_tensor_model_parallel_world_size() self.tp_rank = get_tensor_model_parallel_rank() self.routed_scaling_factor = config.routed_scaling_factor self.ep_group = get_ep_group().device_group self.ep_rank = get_ep_group().rank_in_group self.ep_size = self.ep_group.size() self.n_routed_experts: int = config.n_routed_experts self.n_shared_experts: int = config.n_shared_experts self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe if config.hidden_act != "silu": raise ValueError( f"Unsupported activation: {config.hidden_act}. " "Only silu is supported for now." ) self.gate = ReplicatedLinear( config.hidden_size, config.n_routed_experts, bias=False, quant_config=None, prefix=f"{prefix}.gate", ) if config.topk_method == "noaux_tc": self.gate.e_score_correction_bias = nn.Parameter( torch.empty(config.n_routed_experts, dtype=torch.float32) ) else: self.gate.e_score_correction_bias = None # Load balancing settings. eplb_config = parallel_config.eplb_config self.enable_eplb = parallel_config.enable_eplb self.n_redundant_experts = eplb_config.num_redundant_experts self.n_logical_experts = self.n_routed_experts self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts self.n_local_physical_experts = self.n_physical_experts // self.ep_size self.physical_expert_start = self.ep_rank * self.n_local_physical_experts self.physical_expert_end = ( self.physical_expert_start + self.n_local_physical_experts ) self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled() self.is_fusion_moe_shared_experts_enabled = ( rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() ) if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled: self.shared_experts = None else: intermediate_size = config.moe_intermediate_size * config.n_shared_experts self.shared_experts = AXK1MLP( hidden_size=config.hidden_size, intermediate_size=intermediate_size, hidden_act=config.hidden_act, quant_config=quant_config, is_sequence_parallel=self.is_sequence_parallel, reduce_results=False, prefix=f"{prefix}.shared_experts", ) self.experts = SharedFusedMoE( shared_experts=self.shared_experts, gate=self.gate, num_experts=config.n_routed_experts, top_k=config.num_experts_per_tok, hidden_size=config.hidden_size, intermediate_size=config.moe_intermediate_size, reduce_results=False, renormalize=config.norm_topk_prob, quant_config=quant_config, use_grouped_topk=True, num_expert_group=config.n_group, topk_group=config.topk_group, prefix=f"{prefix}.experts", scoring_func=config.scoring_func, # we do scaling outside, set factor to 1.0 to avoid double mul # aiter applies routed_scaling_factor internally routed_scaling_factor=1.0 if not self.is_rocm_aiter_moe_enabled else self.routed_scaling_factor, e_score_correction_bias=self.gate.e_score_correction_bias, enable_eplb=self.enable_eplb, num_redundant_experts=self.n_redundant_experts, is_sequence_parallel=self.is_sequence_parallel, n_shared_experts=config.n_shared_experts if self.is_fusion_moe_shared_experts_enabled else None, ) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: num_tokens, hidden_dim = hidden_states.shape hidden_states = hidden_states.view(-1, hidden_dim) # Chunk the hidden states so they aren't replicated across TP ranks. # This avoids duplicate computation in self.experts. # TODO: We can replace the all_reduce at the end of attn with a # reduce_scatter instead of chunking here. if self.is_sequence_parallel: hidden_states = sequence_parallel_chunk(hidden_states) if self.experts.is_internal_router: # In this case, the gate/router runs inside the FusedMoE class fused_moe_out = self.experts( hidden_states=hidden_states, router_logits=hidden_states ) else: # router_logits: (num_tokens, n_experts) router_logits, _ = self.gate(hidden_states) fused_moe_out = self.experts( hidden_states=hidden_states, router_logits=router_logits ) shared_output, final_hidden_states = fused_moe_out if self.shared_experts is None: assert shared_output is None # Fix FP16 overflow # See AXK1DecoderLayer for more details. if hidden_states.dtype != torch.float16: if not self.is_rocm_aiter_moe_enabled: final_hidden_states *= self.routed_scaling_factor elif self.shared_experts is not None: assert shared_output is not None shared_output *= 1.0 / self.routed_scaling_factor if self.shared_experts is not None: assert shared_output is not None final_hidden_states += shared_output if self.is_sequence_parallel: final_hidden_states = tensor_model_parallel_all_gather( final_hidden_states, 0 ) final_hidden_states = final_hidden_states[:num_tokens] elif self.tp_size > 1: final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( final_hidden_states ) return final_hidden_states.view(num_tokens, hidden_dim) def _get_llama_4_scaling( original_max_position_embeddings: int, scaling_beta: float, positions: torch.Tensor ) -> torch.Tensor: scaling = 1 + scaling_beta * torch.log( 1 + torch.floor(positions / original_max_position_embeddings) ) # Broadcast over num_heads and head_dim return scaling[..., None, None] class AXK1Attention(nn.Module): def __init__( self, vllm_config: VllmConfig, config: AXK1Config, hidden_size: int, num_heads: int, qk_nope_head_dim: int, qk_rope_head_dim: int, v_head_dim: int, q_lora_rank: int, kv_lora_rank: int, max_position_embeddings: int = 8192, cache_config: CacheConfig | None = None, quant_config: QuantizationConfig | None = None, topk_indices_buffer: torch.Tensor | None = None, prefix: str = "", ) -> None: super().__init__() self.hidden_size = hidden_size self.qk_nope_head_dim = qk_nope_head_dim self.qk_rope_head_dim = qk_rope_head_dim self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim self.v_head_dim = v_head_dim self.q_lora_rank = q_lora_rank self.kv_lora_rank = kv_lora_rank self.num_heads = num_heads tp_size = get_tensor_model_parallel_world_size() assert num_heads % tp_size == 0 self.num_local_heads = num_heads // tp_size self.scaling = self.qk_head_dim**-0.5 self.max_position_embeddings = max_position_embeddings assert topk_indices_buffer is None, ( "topk_indices_buffer is not \ supported for AXK1Attention" ) if self.q_lora_rank is not None: self.q_a_proj = ReplicatedLinear( self.hidden_size, self.q_lora_rank, bias=False, quant_config=quant_config, prefix=f"{prefix}.q_a_proj", ) self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) self.q_b_proj = ColumnParallelLinear( q_lora_rank, self.num_heads * self.qk_head_dim, bias=False, quant_config=quant_config, prefix=f"{prefix}.q_b_proj", ) else: self.q_proj = ColumnParallelLinear( self.hidden_size, self.num_heads * self.qk_head_dim, bias=False, quant_config=quant_config, prefix=f"{prefix}.q_proj", ) self.kv_a_proj_with_mqa = ReplicatedLinear( self.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False, quant_config=quant_config, prefix=f"{prefix}.kv_a_proj_with_mqa", ) self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) self.kv_b_proj = ColumnParallelLinear( self.kv_lora_rank, self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), bias=False, quant_config=quant_config, prefix=f"{prefix}.kv_b_proj", ) # O projection. self.o_proj = RowParallelLinear( self.num_heads * self.v_head_dim, self.hidden_size, bias=False, quant_config=quant_config, prefix=f"{prefix}.o_proj", ) if config.rope_parameters["rope_type"] != "default": config.rope_parameters["rope_type"] = ( "deepseek_yarn" if config.rope_parameters.get("apply_yarn_scaling", True) else "deepseek_llama_scaling" ) self.rotary_emb = get_rope( qk_rope_head_dim, max_position=max_position_embeddings, rope_parameters=config.rope_parameters, is_neox_style=False, ) if config.rope_parameters["rope_type"] == "deepseek_yarn": mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) scaling_factor = config.rope_parameters["factor"] mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) self.scaling = self.scaling * mscale * mscale self.attn = Attention( self.num_local_heads, self.qk_head_dim, self.scaling, num_kv_heads=self.num_local_heads, cache_config=cache_config, quant_config=quant_config, prefix=f"{prefix}.attn", ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, llama_4_scaling: torch.Tensor | None, ) -> torch.Tensor: if self.q_lora_rank is not None: q = self.q_a_proj(hidden_states)[0] q = self.q_a_layernorm(q) q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim) else: q = self.q_proj(hidden_states)[0].view( -1, self.num_local_heads, self.qk_head_dim ) q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0] kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) latent_cache = latent_cache.unsqueeze(1) kv_a = self.kv_a_layernorm(kv_a) kv = self.kv_b_proj(kv_a)[0] kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim) k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) k_pe = latent_cache[:, :, self.kv_lora_rank :] q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe) q[..., self.qk_nope_head_dim :] = q_pe k = torch.empty_like(q) k[..., : self.qk_nope_head_dim] = k_nope k[..., self.qk_nope_head_dim :] = k_pe # Apply llama 4 scaling if provided if llama_4_scaling is not None: q *= llama_4_scaling # padding value to qk_head_dim for alignment v = torch.nn.functional.pad( v, [0, self.qk_head_dim - self.v_head_dim], value=0 ).view(-1, self.num_local_heads * self.qk_head_dim) attn_output = self.attn(q, k, v) attn_output = attn_output.view(-1, self.num_local_heads, self.qk_head_dim)[ ..., : self.v_head_dim ].reshape(-1, self.num_local_heads * self.v_head_dim) output, _ = self.o_proj(attn_output) return output class AXK1MLAAttention(nn.Module): """ Main reference: DeepseekV2 paper, and FlashInfer Implementation (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551). For more info see MLACommonImpl in: vllm/v1/attention/backends/mla/utils.py """ def __init__( self, vllm_config: VllmConfig, config: AXK1Config, hidden_size: int, num_heads: int, qk_nope_head_dim: int, qk_rope_head_dim: int, v_head_dim: int, q_lora_rank: int | None, kv_lora_rank: int, max_position_embeddings: int = 8192, cache_config: CacheConfig | None = None, quant_config: QuantizationConfig | None = None, prefix: str = "", topk_indices_buffer: torch.Tensor | None = None, ) -> None: super().__init__() self.hidden_size = hidden_size self.qk_nope_head_dim = qk_nope_head_dim self.qk_rope_head_dim = qk_rope_head_dim self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim self.v_head_dim = v_head_dim self.q_lora_rank = q_lora_rank self.kv_lora_rank = kv_lora_rank self.num_heads = num_heads tp_size = get_tensor_model_parallel_world_size() assert num_heads % tp_size == 0 self.num_local_heads = num_heads // tp_size self.scaling = self.qk_head_dim**-0.5 self.max_position_embeddings = max_position_embeddings if self.q_lora_rank is not None: self.fused_qkv_a_proj = MergedColumnParallelLinear( self.hidden_size, [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], bias=False, quant_config=quant_config, prefix=f"{prefix}.fused_qkv_a_proj", disable_tp=True, ) else: self.kv_a_proj_with_mqa = ReplicatedLinear( self.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False, quant_config=quant_config, prefix=f"{prefix}.kv_a_proj_with_mqa", ) if self.q_lora_rank is not None: self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) self.q_b_proj = ColumnParallelLinear( self.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False, quant_config=quant_config, prefix=f"{prefix}.q_b_proj", ) else: self.q_proj = ColumnParallelLinear( self.hidden_size, self.num_heads * self.qk_head_dim, bias=False, quant_config=quant_config, prefix=f"{prefix}.q_proj", ) self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) self.kv_b_proj = ColumnParallelLinear( self.kv_lora_rank, self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), bias=False, quant_config=quant_config, prefix=f"{prefix}.kv_b_proj", ) self.o_proj = RowParallelLinear( self.num_heads * self.v_head_dim, self.hidden_size, bias=False, quant_config=quant_config, prefix=f"{prefix}.o_proj", ) if config.rope_parameters["rope_type"] != "default": config.rope_parameters["rope_type"] = ( "deepseek_yarn" if config.rope_parameters.get("apply_yarn_scaling", True) else "deepseek_llama_scaling" ) self.rotary_emb = get_rope( qk_rope_head_dim, max_position=max_position_embeddings, rope_parameters=config.rope_parameters, is_neox_style=False, ) if config.rope_parameters["rope_type"] == "deepseek_yarn": mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) scaling_factor = config.rope_parameters["factor"] mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) self.scaling = self.scaling * mscale * mscale mla_modules = MLAModules( kv_a_layernorm=self.kv_a_layernorm, kv_b_proj=self.kv_b_proj, rotary_emb=self.rotary_emb, o_proj=self.o_proj, fused_qkv_a_proj=self.fused_qkv_a_proj if self.q_lora_rank is not None else None, kv_a_proj_with_mqa=self.kv_a_proj_with_mqa if self.q_lora_rank is None else None, q_a_layernorm=self.q_a_layernorm if self.q_lora_rank is not None else None, q_b_proj=self.q_b_proj if self.q_lora_rank is not None else None, q_proj=self.q_proj if self.q_lora_rank is None else None, indexer=None, indexer_rotary_emb=None, is_sparse=False, topk_indices_buffer=topk_indices_buffer, ) self.mla_attn = MultiHeadLatentAttentionWrapper( self.hidden_size, self.num_local_heads, self.scaling, self.qk_nope_head_dim, self.qk_rope_head_dim, self.v_head_dim, self.q_lora_rank, self.kv_lora_rank, mla_modules, cache_config, quant_config, prefix, ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, llama_4_scaling: torch.Tensor | None, ) -> torch.Tensor: return self.mla_attn(positions, hidden_states, llama_4_scaling) class AXK1DecoderLayer(nn.Module): def __init__( self, vllm_config: VllmConfig, prefix: str, config: AXK1Config | None = None, ) -> None: super().__init__() if config is None: config = vllm_config.model_config.hf_config model_config = vllm_config.model_config cache_config = vllm_config.cache_config quant_config = vllm_config.quant_config parallel_config = vllm_config.parallel_config self.config = config self.hidden_size = config.hidden_size max_position_embeddings = config.max_position_embeddings # DecoderLayers are created with `make_layers` which passes the prefix # with the layer's index. layer_idx = int(prefix.split(sep=".")[-1]) self.layer_idx = layer_idx # verify MLA attention specific fields qk_nope_head_dim = config.qk_nope_head_dim qk_rope_head_dim = config.qk_rope_head_dim v_head_dim = config.v_head_dim kv_lora_rank = config.kv_lora_rank use_mha = all(dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)) self.use_mha = use_mha if use_mha: attn_cls = DeepseekAttention elif model_config.use_mla: attn_cls = AXK1MLAAttention else: attn_cls = AXK1Attention self.self_attn = attn_cls( vllm_config=vllm_config, config=config, hidden_size=self.hidden_size, num_heads=config.num_attention_heads, qk_nope_head_dim=qk_nope_head_dim, qk_rope_head_dim=qk_rope_head_dim, v_head_dim=v_head_dim, q_lora_rank=config.q_lora_rank, kv_lora_rank=kv_lora_rank, max_position_embeddings=max_position_embeddings, cache_config=cache_config, quant_config=quant_config, prefix=f"{prefix}.self_attn", topk_indices_buffer=None, ) self.is_layer_sparse = self._is_layer_sparse() if self.is_layer_sparse: self.mlp = AXK1MoE( config=config, parallel_config=parallel_config, quant_config=quant_config, prefix=f"{prefix}.mlp", ) else: self.mlp = AXK1MLP( hidden_size=config.hidden_size, intermediate_size=config.intermediate_size, hidden_act=config.hidden_act, quant_config=quant_config, prefix=f"{prefix}.mlp", ) self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_attention_layernorm = RMSNorm( config.hidden_size, eps=config.rms_norm_eps ) self.post_mlp_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.routed_scaling_factor = config.routed_scaling_factor def _is_layer_sparse(self) -> bool: return ( self.config.n_routed_experts is not None and self.layer_idx >= self.config.first_k_dense_replace and self.layer_idx % self.config.moe_layer_freq == 0 ) def forward( self, positions: torch.Tensor, hidden_states: torch.Tensor, residual: torch.Tensor | None, llama_4_scaling: torch.Tensor | None = None, ) -> tuple[torch.Tensor, torch.Tensor]: # Self Attention if residual is None: residual = hidden_states.clone() hidden_states = self.input_layernorm(hidden_states) else: hidden_states, residual = self.input_layernorm(hidden_states, residual) attn_kwargs = { "positions": positions, "hidden_states": hidden_states, } if not self.use_mha: attn_kwargs["llama_4_scaling"] = llama_4_scaling hidden_states = self.self_attn(**attn_kwargs) if ( not isinstance(self.self_attn, DeepseekAttention) and hidden_states.dtype == torch.float16 ): # Fix FP16 overflow # We scale both hidden_states and residual before # rmsnorm, and rmsnorm result would not affect by scale. hidden_states *= 1.0 / self.routed_scaling_factor if self.layer_idx == 0: # The residual is shared by all layers, we only scale it on # first layer. residual *= 1.0 / self.routed_scaling_factor # Fully Connected hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) hidden_states = self.mlp(hidden_states) if self.is_layer_sparse: hidden_states = self.post_mlp_layernorm(hidden_states) if isinstance(self.mlp, AXK1MLP) and hidden_states.dtype == torch.float16: # Fix FP16 overflow # Scaling the AXK1MLP output, it is the input of # input_layernorm of next decoder layer. # The scaling of AXK1MOE output would be done in the forward # of AXK1MOE hidden_states *= 1.0 / self.routed_scaling_factor return hidden_states, residual @support_torch_compile class AXK1Model(nn.Module): fall_back_to_pt_during_load = False def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config: AXK1Config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config self.config = config self.device = current_platform.device_type self.vocab_size = config.vocab_size if get_pp_group().is_first_rank: self.embed_tokens = VocabParallelEmbedding( config.vocab_size, config.hidden_size, quant_config=quant_config, prefix=f"{prefix}.embed_tokens", ) else: self.embed_tokens = PPMissingLayer() self.start_layer, self.end_layer, self.layers = make_layers( config.num_hidden_layers, lambda prefix: AXK1DecoderLayer(vllm_config, prefix), prefix=f"{prefix}.layers", ) if get_pp_group().is_last_rank: self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) else: self.norm = PPMissingLayer() self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( ["hidden_states", "residual"], config.hidden_size ) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: return self.embed_tokens(input_ids) def forward( self, input_ids: torch.Tensor | None, positions: torch.Tensor, intermediate_tensors: IntermediateTensors | None, inputs_embeds: torch.Tensor | None = None, ) -> torch.Tensor | IntermediateTensors: if get_pp_group().is_first_rank: if inputs_embeds is not None: hidden_states = inputs_embeds else: hidden_states = self.embed_input_ids(input_ids) residual = None else: assert intermediate_tensors is not None hidden_states = intermediate_tensors["hidden_states"] residual = intermediate_tensors["residual"] # Compute llama 4 scaling once per forward pass if enabled llama_4_scaling_config = getattr(self.config, "llama_4_scaling", None) llama_4_scaling: torch.Tensor | None if llama_4_scaling_config is not None: llama_4_scaling = _get_llama_4_scaling( original_max_position_embeddings=llama_4_scaling_config[ "original_max_position_embeddings" ], scaling_beta=llama_4_scaling_config["beta"], positions=positions, ) else: llama_4_scaling = None for layer in islice(self.layers, self.start_layer, self.end_layer): hidden_states, residual = layer( positions, hidden_states, residual, llama_4_scaling ) if not get_pp_group().is_last_rank: return IntermediateTensors( {"hidden_states": hidden_states, "residual": residual} ) hidden_states, _ = self.norm(hidden_states, residual) return hidden_states class AXK1MixtureOfExperts(MixtureOfExperts): moe_mlp_layers: list[AXK1MoE] """ List of MoE MLP layers in the model. """ def extract_moe_parameters(self, example_moe: AXK1MoE | None): if example_moe is None: self.num_moe_layers = 0 self.num_expert_groups = 0 self.num_logical_experts = 0 self.num_physical_experts = 0 self.num_local_physical_experts = 0 self.num_routed_experts = 0 self.num_shared_experts = 0 self.num_redundant_experts = 0 logger.warning("AXK1: No AXK1MoE layer found in model.layers.") else: self.num_logical_experts = example_moe.n_logical_experts self.num_physical_experts = example_moe.n_physical_experts self.num_local_physical_experts = example_moe.n_local_physical_experts self.num_routed_experts = example_moe.n_routed_experts self.num_shared_experts = example_moe.n_shared_experts self.num_redundant_experts = example_moe.n_redundant_experts def update_physical_experts_metadata( self, num_physical_experts: int, num_local_physical_experts: int, ) -> None: assert self.num_local_physical_experts == num_local_physical_experts self.num_physical_experts = num_physical_experts self.num_local_physical_experts = num_local_physical_experts self.num_redundant_experts = num_physical_experts - self.num_logical_experts for moe in self.moe_mlp_layers: moe.n_local_physical_experts = num_local_physical_experts moe.n_physical_experts = num_physical_experts moe.n_redundant_experts = self.num_redundant_experts moe.experts.update_expert_map() class AXK1ForCausalLM( nn.Module, SupportsPP, AXK1MixtureOfExperts, SupportsLoRA, SupportsEagle ): packed_modules_mapping = { "gate_up_proj": ["gate_proj", "up_proj"], } model_cls = AXK1Model def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config: AXK1Config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config self.config = config self.quant_config = quant_config qk_nope_head_dim = config.qk_nope_head_dim qk_rope_head_dim = config.qk_rope_head_dim self.use_mha = all(dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim)) if self.use_mha: self.packed_modules_mapping["qkv_proj"] = ["q_proj", "k_proj", "v_proj"] # `packed_modules_mapping` needs to be modified before # initializing AXK1Model, as it is passed inplace to # quantization config init and may be used to select the # quant_method for relevant layers during initialization. self.fuse_qkv_a_proj = config.q_lora_rank is not None if self.fuse_qkv_a_proj: self.packed_modules_mapping["fused_qkv_a_proj"] = [ "q_a_proj", "kv_a_proj_with_mqa", ] self.model = self.model_cls( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") ) if get_pp_group().is_last_rank: self.lm_head = ParallelLMHead( config.vocab_size, config.hidden_size, quant_config=quant_config, prefix=maybe_prefix(prefix, "lm_head"), ) else: self.lm_head = PPMissingLayer() self.logits_processor = LogitsProcessor(config.vocab_size) self.make_empty_intermediate_tensors = ( self.model.make_empty_intermediate_tensors ) # Set MoE hyperparameters self.num_moe_layers = ( self.config.num_hidden_layers - self.config.first_k_dense_replace ) self.set_moe_parameters() def set_moe_parameters(self): self.expert_weights = [] self.num_expert_groups = getattr(self.config, "n_group", 1) self.moe_layers = [] self.moe_mlp_layers = [] example_moe = None for layer in self.model.layers: if isinstance(layer, PPMissingLayer): continue assert isinstance(layer, AXK1DecoderLayer) if isinstance(layer.mlp, AXK1MoE): # Pick last one layer since the first ones may be dense layers. example_moe = layer.mlp self.moe_mlp_layers.append(layer.mlp) self.moe_layers.append(layer.mlp.experts) self.extract_moe_parameters(example_moe) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: return self.model.embed_input_ids(input_ids) def forward( self, input_ids: torch.Tensor | None, positions: torch.Tensor, intermediate_tensors: IntermediateTensors | None = None, inputs_embeds: torch.Tensor | None = None, ) -> torch.Tensor | IntermediateTensors: hidden_states = self.model( input_ids, positions, intermediate_tensors, inputs_embeds ) return hidden_states def compute_logits( self, hidden_states: torch.Tensor, ) -> torch.Tensor | None: logits = self.logits_processor(self.lm_head, hidden_states) return logits def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) return SharedFusedMoE.make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", ckpt_up_proj_name="up_proj", num_experts=self.config.n_routed_experts, num_redundant_experts=0, ) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: rocm_aiter_moe_shared_expert_enabled = ( rocm_aiter_ops.is_fusion_moe_shared_experts_enabled() ) stacked_params_mapping = [ # (param_name, shard_name, shard_id) ("gate_up_proj", "gate_proj", 0), ("gate_up_proj", "up_proj", 1), ] mla_params_mapping = [ ("fused_qkv_a_proj", "q_a_proj", 0), ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), ] mha_params_mapping = [ ("qkv_proj", "q_proj", "q"), ("qkv_proj", "k_proj", "k"), ("qkv_proj", "v_proj", "v"), ] if self.use_mha: stacked_params_mapping.extend(mha_params_mapping) else: stacked_params_mapping.extend(mla_params_mapping) # Params for weights, fp8 weight scales, fp8 activation scales # (param_name, weight_name, expert_id, shard_id) expert_params_mapping = SharedFusedMoE.make_expert_params_mapping( self, ckpt_gate_proj_name="gate_proj", ckpt_down_proj_name="down_proj", ckpt_up_proj_name="up_proj", num_experts=self.config.n_routed_experts + ( self.config.n_shared_experts if rocm_aiter_moe_shared_expert_enabled else 0 ), num_redundant_experts=self.num_redundant_experts, ) params_dict = dict(self.named_parameters()) loaded_params: set[str] = set() for name, loaded_weight in weights: if "rotary_emb.inv_freq" in name: continue spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) if spec_layer is not None: continue # skip spec decode layers for main model is_fusion_moe_shared_experts_layer = ( rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name) ) for param_name, weight_name, shard_id in stacked_params_mapping: # Skip non-stacked layers and experts (experts handled below). if weight_name not in name: continue # We have mlp.experts[0].gate_proj in the checkpoint. # Since we handle the experts below in expert_params_mapping, # we need to skip here BEFORE we update the name, otherwise # name will be updated to mlp.experts[0].gate_up_proj, which # will then be updated below in expert_params_mapping # for mlp.experts[0].gate_gate_up_proj, which breaks load. if ("mlp.experts." in name) and name not in params_dict: continue if is_fusion_moe_shared_experts_layer: continue name_mapped = name.replace(weight_name, param_name) # QKV fusion is optional, fall back to normal # weight loading if it's not enabled # if go with fusion option, then update name if ( param_name == "fused_qkv_a_proj" ) and name_mapped not in params_dict: continue else: name = name_mapped # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue if is_pp_missing_parameter(name, self): continue param = params_dict[name] weight_loader = param.weight_loader weight_loader(param, loaded_weight, shard_id) break else: is_expert_weight = False # Special handling: when AITER fusion_shared_experts is enabled, # checkpoints may provide a single widened shared_experts tensor # without explicit expert indices # (e.g. ...mlp.shared_experts.gate_proj.weight). # For models with multiple shared experts, split that tensor # evenly into per-shared-expert slices and load them into # appended expert slots mlp.experts.{n_routed_experts + j}.* # accordingly. num_chunks = 1 if is_fusion_moe_shared_experts_layer: num_chunks = getattr(self.config, "n_shared_experts", 1) or 1 # Determine split axis based on op type # gate/up: ColumnParallel → split along dim 0 # down: RowParallel → split along dim 1 split_dim = ( 1 if ("down_proj.weight" in name and loaded_weight.ndim > 1) else 0 ) total = loaded_weight.shape[split_dim] assert total % num_chunks == 0, ( f"Shared expert weight dim {total} " f"not divisible by num_chunks {num_chunks}" ) chunk_size = total // num_chunks for j in range(num_chunks): chunk_name = name weight_to_load = loaded_weight if is_fusion_moe_shared_experts_layer: chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size) if loaded_weight.ndim == 1: weight_to_load = loaded_weight[chunk_slice] elif split_dim == 0: weight_to_load = loaded_weight[chunk_slice, :] else: weight_to_load = loaded_weight[:, chunk_slice] # Synthesize an expert-style name so expert mapping # can route it chunk_name = name.replace( "mlp.shared_experts", f"mlp.experts.{self.config.n_routed_experts + j}", ) # Use expert_params_mapping to locate the destination # param and delegate to its expert-aware weight_loader # with expert_id. for mapping in expert_params_mapping: param_name, weight_name, expert_id, shard_id = mapping if weight_name not in chunk_name: continue # Anyway, this is an expert weight and should not be # attempted to load as other weights later is_expert_weight = True # Do not modify `name` since the loop may continue here # Instead, create a new variable name_mapped = chunk_name.replace(weight_name, param_name) if is_pp_missing_parameter(name_mapped, self): continue param = params_dict[name_mapped] # We should ask the weight loader to return success or # not here since otherwise we may skip experts with # other available replicas. weight_loader = typing.cast( Callable[..., bool], param.weight_loader ) success = weight_loader( param, weight_to_load, name_mapped, shard_id=shard_id, expert_id=expert_id, return_success=True, ) if success: if not is_fusion_moe_shared_experts_layer: name = name_mapped else: loaded_params.add(name_mapped) break else: if is_expert_weight: # We've checked that this is an expert weight # However it's not mapped locally to this rank # So we simply skip it continue # Skip loading extra bias for GPTQ models. if name.endswith(".bias") and name not in params_dict: continue # Remapping the name of FP8 kv-scale. name = maybe_remap_kv_scale_name(name, params_dict) if name is None: continue if is_pp_missing_parameter(name, self): continue param = params_dict[name] weight_loader = getattr( param, "weight_loader", default_weight_loader ) weight_loader(param, loaded_weight) if not is_fusion_moe_shared_experts_layer: loaded_params.add(name) return loaded_params def get_spec_layer_idx_from_weight_name( config: AXK1Config, weight_name: str ) -> int | None: if config.num_nextn_predict_layers and config.num_nextn_predict_layers > 0: layer_idx = config.num_hidden_layers for i in range(config.num_nextn_predict_layers): if weight_name.startswith(f"model.layers.{layer_idx + i}."): return layer_idx + i return None