GLM-4 Update

benchmarks/kernels/benchmark_moe.py

@@ -576,7 +576,11 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif config.architectures[0] in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM"):
+    elif config.architectures[0] in (
+        "DeepseekV3ForCausalLM",
+        "DeepseekV2ForCausalLM",
+        "Glm4MoeForCausalLM",
+    ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -321,10 +321,12 @@ def main(args: argparse.Namespace):
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
-    elif config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]:
+    elif (
+        config.architectures[0] in ["Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"]
+        or config.architectures[0] == "Glm4MoeForCausalLM"
+    ):
         E = config.num_experts
         topk = config.num_experts_per_tok
-
     else:
         # Support for llama4
         config = config.get_text_config()

docs/models/supported_models.md

@@ -574,6 +574,7 @@ Specified using `--task generate`.
 | `Gemma3ForConditionalGeneration` | Gemma 3 | T + I<sup>+</sup> | `google/gemma-3-4b-it`, `google/gemma-3-27b-it`, etc. | ✅︎ | ✅︎ | ⚠️ |
 | `GLM4VForCausalLM`<sup>^</sup> | GLM-4V | T + I | `THUDM/glm-4v-9b`, `THUDM/cogagent-9b-20241220`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Glm4vForConditionalGeneration` | GLM-4.1V-Thinking | T + I<sup>E+</sup> + V<sup>E+</sup> | `THUDM/GLM-4.1V-9B-Thinkg`, etc. | ✅︎ | ✅︎ | ✅︎ |
+| `Glm4MoeForCausalLM` | GLM-4-MoE | T + I<sup>E+</sup> + V<sup>E+</sup> | `THUDM/GLM-4-MoE-100B-A10B`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `GraniteSpeechForConditionalGeneration` | Granite Speech | T + A | `ibm-granite/granite-speech-3.3-8b` | ✅︎ | ✅︎ | ✅︎ |
 | `H2OVLChatModel` | H2OVL | T + I<sup>E+</sup> | `h2oai/h2ovl-mississippi-800m`, `h2oai/h2ovl-mississippi-2b`, etc. | | ✅︎ | ✅︎ |
 | `Idefics3ForConditionalGeneration` | Idefics3 | T + I | `HuggingFaceM4/Idefics3-8B-Llama3`, etc. | ✅︎ | | ✅︎ |

tests/models/registry.py

@@ -357,6 +357,7 @@ def check_available_online(
                                         trust_remote_code=True,
                                         hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
     "Glm4vForConditionalGeneration": _HfExamplesInfo("THUDM/GLM-4.1V-9B-Thinking", min_transformers_version="4.53"),  # noqa: E501
+    "Glm4MoeForCausalLM": _HfExamplesInfo("THUDM/GLM-4-MoE-100B-A10B", min_transformers_version="4.54"),   # noqa: E501
     "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
                                       extras={"2b": "h2oai/h2ovl-mississippi-2b"},  # noqa: E501
                                       max_transformers_version="4.48",  # noqa: E501
@@ -469,6 +470,8 @@ def check_available_online(
                                             is_available_online=False,
                                             speculative_model="openbmb/MiniCPM-2B-sft-bf16",
                                             tokenizer="openbmb/MiniCPM-2B-sft-bf16"),
+    "Glm4MoeMTPModel": _HfExamplesInfo("THUDM/GLM-4-MoE",
+                                        speculative_model="THUDM/GLM-4-MoE"),
     "MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                     trust_remote_code=True,
                                     speculative_model="XiaomiMiMo/MiMo-7B-RL")

vllm/config.py

@@ -2545,7 +2545,8 @@ def __post_init__(self):
 
 
 SpeculativeMethod = Literal["ngram", "eagle", "eagle3", "medusa",
-                            "mlp_speculator", "draft_model", "deepseek_mtp"]
+                            "mlp_speculator", "draft_model", "deepseek_mtp",
+                            "glm4_moe_mtp"]
 SpeculativeAcceptanceMethod = Literal["rejection_sampler",
                                       "typical_acceptance_sampler"]
 
@@ -2805,8 +2806,10 @@ def __post_init__(self):
                 elif (self.draft_model_config.hf_config.model_type ==
                       "mlp_speculator"):
                     self.method = "mlp_speculator"
-                elif (self.draft_model_config.hf_config.model_type ==
-                      "deepseek_mtp"):
+                elif (self.draft_model_config.hf_config.model_type
+                      == "deepseek_mtp"
+                      or self.draft_model_config.hf_config.model_type
+                      == "glm4_moe_mtp"):
                     self.method = "deepseek_mtp"
                     if self.num_speculative_tokens > 1:
                         logger.warning(

vllm/engine/arg_utils.py

@@ -1507,7 +1507,8 @@ def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
                     is_ngram_enabled = True
                 elif speculative_method == "medusa":
                     is_medusa_enabled = True
-                elif speculative_method in ("eagle", "eagle3", "deepseek_mtp"):
+                elif speculative_method in ("eagle", "eagle3", "deepseek_mtp",
+                                            "glm4_moe_mtp"):
                     is_eagle_enabled = True
             else:
                 speculative_model = self.speculative_config.get("model")