Add pipeline parallelism for Qwen2 and Qwen3 Model (#6250)

python/sglang/srt/models/qwen3_moe.py

@@ -17,6 +17,7 @@
 
 """Inference-only Qwen3MoE model compatible with HuggingFace weights."""
 
+import logging
 from dataclasses import dataclass
 from enum import Enum, auto
 from functools import partial
@@ -28,6 +29,7 @@ from torch import nn
 from transformers.configuration_utils import PretrainedConfig
 
 from sglang.srt.distributed import (
+    get_pp_group,
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
     split_tensor_along_last_dim,
@@ -57,12 +59,13 @@ from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.utils import get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2_moe import Qwen2MoeMLP as Qwen3MoeMLP
 from sglang.srt.models.qwen2_moe import Qwen2MoeModel
@@ -70,6 +73,8 @@ from sglang.srt.utils import add_prefix
 
 Qwen3MoeConfig = None
 
+logger = logging.getLogger(__name__)
+
 
 class Qwen3MoeSparseMoeBlock(nn.Module):
     def __init__(
@@ -516,6 +521,7 @@ class Qwen3MoeForCausalLM(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
         self.model = Qwen3MoeModel(
@@ -536,12 +542,31 @@ class Qwen3MoeForCausalLM(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
-    ) -> LogitsProcessorOutput:
-        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head, forward_batch
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        if self.pp_group.is_last_rank:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            return hidden_states
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
@@ -563,6 +588,17 @@ class Qwen3MoeForCausalLM(nn.Module):
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+
             if "rotary_emb.inv_freq" in name:
                 continue
             for param_name, weight_name, shard_id in stacked_params_mapping:
@@ -611,11 +647,14 @@ class Qwen3MoeForCausalLM(nn.Module):
                     if name not in params_dict:
                         continue
 
-                    param = params_dict[name]
-                    weight_loader = getattr(
-                        param, "weight_loader", default_weight_loader
-                    )
-                    weight_loader(param, loaded_weight)
+                    if name in params_dict.keys():
+                        param = params_dict[name]
+                        weight_loader = getattr(
+                            param, "weight_loader", default_weight_loader
+                        )
+                        weight_loader(param, loaded_weight)
+                    else:
+                        logger.warning(f"Parameter {name} not found in params_dict")
 
 
 EntryClass = Qwen3MoeForCausalLM