[Bug Fix] Align glm4_moe_nextn NPU MTP loading with qwen3 MTP (#25524)

python/sglang/srt/models/glm4_moe_nextn.py

@@ -12,9 +12,8 @@
 # limitations under the License.
 # ==============================================================================
 
-"""Inference-only GLM-4.5, GLM-4.6 Speculative Decoding."""
+"""Inference-only GLM-4.5, GLM-4.6 and GLM-4.7 Speculative Decoding."""
 
-import contextlib
 import logging
 from typing import Iterable, Optional, Tuple
 
@@ -23,7 +22,6 @@ from torch import nn
 from transformers import PretrainedConfig
 
 from sglang.srt.distributed import get_tensor_model_parallel_world_size
-from sglang.srt.environ import temp_set_env
 from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
 from sglang.srt.layers.dp_attention import is_dp_attention_enabled
 from sglang.srt.layers.layernorm import RMSNorm
@@ -36,7 +34,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.models.glm4_moe import Glm4MoeDecoderLayer, Glm4MoeForCausalLM
 from sglang.srt.server_args import get_global_server_args
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, is_npu
 
 logger = logging.getLogger(__name__)
 
@@ -51,7 +49,7 @@ class Glm4MoeModelNextN(nn.Module):
         super().__init__()
         if quant_config is not None and quant_config.get_name() == "modelopt_fp4":
             logger.warning(
-                "Overriding Glm4MoeForCausalLMNextN quant config for modelopt_fp4 GLM-4.5 / GLM-4.6 model."
+                "Overriding Glm4MoeForCausalLMNextN quant config for modelopt_fp4 GLM-4.5 / GLM-4.6 / GLM-4.7 model."
             )
             quant_config = None
 
@@ -128,11 +126,13 @@ class Glm4MoeForCausalLMNextN(Glm4MoeForCausalLM):
         nn.Module.__init__(self)
         self.config = config
         self.tp_size = get_tensor_model_parallel_world_size()
-        self.needs_quant_draft = (
-            get_global_server_args().speculative_draft_model_quantization is not None
-            or quant_config is not None
-        )
-        quant_config = quant_config if self.needs_quant_draft else None
+        if (
+            is_npu()
+            and get_global_server_args().speculative_draft_model_quantization is None
+        ):
+            quant_config = None
+        self.quant_config = quant_config
+
         self.model = Glm4MoeModelNextN(
             config, quant_config, prefix=add_prefix("model", prefix)
         )
@@ -156,21 +156,7 @@ class Glm4MoeForCausalLMNextN(Glm4MoeForCausalLM):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
     ) -> torch.Tensor:
-        # Support unquant speculative draft model
-        if self.needs_quant_draft:
-            cxt = contextlib.nullcontext()
-        else:
-            # SGLANG_DEEPEP_BF16_DISPATCH is deprecated, will need
-            # to be removed in the future and moved to a new
-            # --deepep-dispatcher-output-dtype server argument.
-            unquant_patch = {
-                "SGLANG_DEEPEP_BF16_DISPATCH": "1",
-            }
-            cxt = temp_set_env(allow_sglang=True, **unquant_patch)
-
-        with cxt:
-            hidden_states = self.model(input_ids, positions, forward_batch)
-
+        hidden_states = self.model(input_ids, positions, forward_batch)
         return self.logits_processor(
             input_ids, hidden_states, self.lm_head, forward_batch
         )