adapt glm4.5 moe

Signed-off-by: whx-sjtu <2952154980@qq.com>

Author: whx-sjtu

vllm_ascend/ascend_forward_context.py

@@ -23,6 +23,7 @@ class FusedMoEState(Enum):
 
 # TODO(zzzzwwjj): add soc_version to choose branch
 def get_fused_moe_state(ep_size: int, with_prefill: bool):
+    return FusedMoEState.AllGather
     enable_chunk_mc2 = envs.VLLM_ASCEND_ENABLE_CHUNK_MC2
     if ep_size == 1:
         return FusedMoEState.AllGather

vllm_ascend/models/__init__.py

@@ -13,6 +13,7 @@ def register_model():
         AscendQwen2_5_VLForConditionalGeneration  # noqa: F401
     from .qwen2_vl import AscendQwen2VLForConditionalGeneration  # noqa: F401
     from .qwen3 import CustomQwen3ForCausalLM  # noqa: F401
+    from .glm4_moe import CustomGlm4MoeForCausalLM
 
     ModelRegistry.register_model(
         "DeepSeekMTPModel",
@@ -64,3 +65,6 @@ def register_model():
 
     ModelRegistry.register_model(
         "Qwen3ForCausalLM", "vllm_ascend.models.qwen3:CustomQwen3ForCausalLM")
+
+    ModelRegistry.register_model(
+        "Glm4MoeForCausalLM", "vllm_ascend.models.glm4_moe:CustomGlm4MoeForCausalLM")

vllm_ascend/models/glm4_moe.py

@@ -0,0 +1,351 @@
+import typing
+from collections.abc import Callable, Iterable
+from typing import Any, Optional, Union
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
+from vllm.distributed import get_pp_group
+from vllm.logger import init_logger
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead, VocabParallelEmbedding)
+from vllm.sequence import IntermediateTensors
+from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP
+
+from vllm.model_executor.models.utils import (PPMissingLayer, make_empty_intermediate_tensors_factory,
+                    make_layers, maybe_prefix)
+from vllm.distributed import (get_dp_group, get_pp_group,
+                              get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size,
+                              get_tp_group, split_tensor_along_last_dim,
+                              tensor_model_parallel_reduce_scatter)
+from vllm.model_executor.models.glm4_moe import Glm4MoeForCausalLM, Glm4MoeDecoderLayer, Glm4MoeModel, Glm4MoeAttention, Glm4MoeMLP
+from vllm_ascend.ops.fused_moe import AscendFusedMoE
+from vllm.forward_context import get_forward_context
+logger = init_logger(__name__)
+
+
+class CustomGlm4MoE(nn.Module):
+
+    top_k: int
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        enable_eplb: bool = False,
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_shared_experts = config.n_shared_experts
+        if self.tp_size > config.n_routed_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {config.n_routed_experts}.")
+
+        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")
+        
+        self.gate.e_score_correction_bias = (nn.Parameter(
+            torch.empty(config.n_routed_experts,)))
+
+        self.experts = AscendFusedMoE(
+            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,
+            e_score_correction_bias=self.gate.e_score_correction_bias)
+
+        if config.n_shared_experts is not None:
+            intermediate_size = (config.moe_intermediate_size *
+                                 config.n_shared_experts)
+            self.shared_experts = Glm4MoeMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=True,
+                prefix=f"{prefix}.shared_experts",
+            )
+        else:
+            self.shared_experts = None  # type: ignore
+        CustomGlm4MoE.top_k = config.num_experts_per_tok
+
+        self.dp_size = get_dp_group().world_size
+
+        self.tp_group = get_tp_group().device_group
+        self.tp_rank = get_tp_group().rank_in_group
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor) -> torch.Tensor:
+        forward_context = get_forward_context()
+        # when profile runs, force experts to load balanced tokens
+        # to avoid high memory consumption on a single rank.
+        enable_force_load_balance = forward_context.in_profile_run
+
+        is_prefill = forward_context.with_prefill
+
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+
+        experts_hidden_states = self.experts(
+            hidden_states=hidden_states,
+            router_logits=router_logits,
+            is_prefill=is_prefill,
+            top_k=CustomGlm4MoE.top_k,
+            enable_force_load_balance=enable_force_load_balance,
+            shared_experts=self.shared_experts,
+            gate=None,
+        )
+
+        hidden_states = (
+            experts_hidden_states[0] * self.routed_scaling_factor +
+            experts_hidden_states[1])
+
+        return hidden_states
+
+class CustomGlm4MoeDecoderLayer(Glm4MoeDecoderLayer):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        enable_eplb: bool = False,
+    ) -> None:
+        nn.Module.__init__(self)
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          131072)
+        # 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
+
+        self.self_attn = Glm4MoeAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            head_dim=config.head_dim,
+            rms_norm_eps=config.rms_norm_eps,
+            qkv_bias=config.attention_bias,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+            use_qk_norm=config.use_qk_norm,
+        )
+
+        if (config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace):
+            self.mlp = AscendSparseMoeBlock(
+                config=config,
+                quant_config=quant_config,
+                prefix=f"{prefix}.mlp",
+            )
+        else:
+            self.mlp = Glm4MoeMLP(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.routed_scaling_factor = config.routed_scaling_factor
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        residual: Optional[torch.Tensor],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+        hidden_states = self.self_attn(positions=positions,
+                                       hidden_states=hidden_states)
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile
+class CustomGlm4MoeModel(Glm4MoeModel):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        nn.Module.__init__(self)
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        # enable_eplb = vllm_config.parallel_config.enable_eplb
+        self.config = config
+
+        self.vocab_size = config.vocab_size
+
+        if get_pp_group().is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                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: CustomGlm4MoeDecoderLayer(
+                config=config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+                # enable_eplb=enable_eplb,
+            ),
+            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 get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(positions, hidden_states, residual)
+
+        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 CustomGlm4MoeForCausalLM(Glm4MoeForCausalLM):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        nn.Module.__init__(self)
+        SupportsPP.__init__(self)
+        SupportsLoRA.__init__(self)
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        self.config = config
+        self.quant_config = quant_config
+        self.model = CustomGlm4MoeModel(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)
+        else:
+            self.lm_head = PPMissingLayer()
+        if self.config.tie_word_embeddings:
+            self.lm_head.weight = self.model.embed_tokens.weight
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+        self.expert_weights = []
+
+        # Set MoE hyperparameters
+        self.num_moe_layers = (config.num_hidden_layers -
+                               config.first_k_dense_replace)
+        self.num_expert_groups = config.n_group
+
+        self.moe_layers: list[FusedMoE] = []
+        example_moe = None
+        for layer in self.model.layers:
+            if isinstance(layer, PPMissingLayer):
+                continue
+
+            assert isinstance(layer, CustomGlm4MoeDecoderLayer)
+            if isinstance(layer.mlp, AscendSparseMoeBlock):
+                # Pick last one layer since the first ones may be dense layers.
+                example_moe = layer.mlp
+                self.moe_layers.append(layer.mlp.experts)
+
+        if example_moe is None:
+            raise RuntimeError("No Glm4MoE layer found in model.layers.")
+
+        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 forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states