model: serve bare Qwen3Model backbone natively as an embedding model (#32457)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs_new/docs/supported-models/embedding_models.mdx

@@ -151,6 +151,12 @@ print("Embedding:", response["data"][0]["embedding"])
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>N/A</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Latest Qwen3-based text embedding model for semantic representation</td>
     </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Qwen3 (bare backbone)</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>`microsoft/harrier-oss-v1-0.6b`</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>N/A</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Bare <code>Qwen3Model</code> backbone (no LM head); served natively on SGLang's fused Qwen3 kernels and auto-classified as an embedding model</td>
+    </tr>
     <tr>
       <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>BGE</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>`BAAI/bge-large-en-v1.5`</td>

python/sglang/srt/configs/model_config.py

@@ -1696,6 +1696,7 @@ def is_generation_model(model_architectures: List[str], is_embedding: bool = Fal
         or "Qwen3ForRewardModel" in model_architectures
         or "Qwen2ForSequenceClassification" in model_architectures
         or "Qwen3ForSequenceClassification" in model_architectures
+        or "Qwen3Model" in model_architectures
         or "CLIPModel" in model_architectures
         or "BertModel" in model_architectures
         or "Contriever" in model_architectures

python/sglang/srt/models/qwen3_embedding.py

@@ -0,0 +1,124 @@
+import logging
+from typing import Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+
+from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
+from sglang.srt.models.qwen3 import Qwen3Model as Qwen3TransformerModel
+from sglang.srt.utils import add_prefix
+
+logger = logging.getLogger(__name__)
+
+
+class Qwen3Model(nn.Module):
+    """Bare Qwen3 backbone (no LM head) served as an embedding model.
+
+    Checkpoints exported as architectures=["Qwen3Model"], e.g.
+    microsoft/harrier-oss-v1-0.6b, have no native implementation to resolve to
+    and fall back to the Transformers backend. Registering the arch here runs
+    them on the native fused Qwen3 kernels instead.
+    """
+
+    def __init__(
+        self,
+        config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Qwen3TransformerModel(
+            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
+        )
+        # Use LAST + normalize=True for qwen3 embedding based on official implementation
+        # Reference: https://github.com/QwenLM/Qwen3-Embedding/blob/main/examples/qwen3_embedding_transformers.py#L55
+        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.model.get_input_embeddings()
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        get_embedding: bool = True,
+    ) -> EmbeddingPoolerOutput:
+        assert get_embedding, f"{self.__class__.__name__} is only used for embedding"
+
+        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
+        return self.pooler(hidden_states, forward_batch)
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            # Bare-backbone checkpoints omit the "model." prefix of the backbone
+            if not name.startswith("model.") and (
+                name.startswith("layers.")
+                or name.startswith("embed_tokens.")
+                or name.startswith("norm.")
+            ):
+                name = add_prefix(name, "model")
+
+            # Skip rotary embeddings and other non-parameter tensors
+            if "rotary_emb.inv_freq" in name or "projector" in name:
+                continue
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+
+            # Skip lm_head weights a non-tied checkpoint may carry (no LM head here)
+            if name.startswith("lm_head"):
+                continue
+
+            # Normalize kv cache scale names of quantized checkpoints
+            if "scale" in name:
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                if name in params_dict:
+                    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 = Qwen3Model

test/registered/unit/models/test_qwen3_embedding_registration.py

@@ -0,0 +1,61 @@
+"""Unit tests for native registration of the bare ``Qwen3Model`` embedding arch.
+
+Checkpoints such as ``microsoft/harrier-oss-v1-0.6b`` declare
+``architectures=["Qwen3Model"]`` (a bare Qwen3 backbone). These must resolve to
+the native SGLang implementation (``sglang.srt.models.qwen3_embedding.Qwen3Model``)
+and be served as an embedding model, NOT fall back to the Transformers backend.
+"""
+
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=10, suite="base-a-test-cpu")
+
+import unittest
+
+from sglang.srt.configs.model_config import is_generation_model
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestQwen3ModelEmbeddingRegistration(CustomTestCase):
+    def test_entry_class_is_native_qwen3model(self):
+        """The bare arch string maps to a native EntryClass named 'Qwen3Model'."""
+        from sglang.srt.models import qwen3_embedding
+
+        entry = qwen3_embedding.EntryClass
+        self.assertEqual(entry.__name__, "Qwen3Model")
+        self.assertEqual(entry.__module__, "sglang.srt.models.qwen3_embedding")
+        # It carries a LAST-token / normalized pooler, i.e. an embedding head.
+        self.assertTrue(hasattr(entry, "forward"))
+        self.assertTrue(hasattr(entry, "load_weights"))
+
+    def test_registry_resolves_native_not_transformers_fallback(self):
+        """ModelRegistry resolves 'Qwen3Model' to the native class, not the
+        TransformersForCausalLM fallback."""
+        from sglang.srt.models.registry import ModelRegistry
+
+        model_cls, resolved_arch = ModelRegistry.resolve_model_cls("Qwen3Model")
+        self.assertEqual(resolved_arch, "Qwen3Model")
+        self.assertEqual(model_cls.__name__, "Qwen3Model")
+        self.assertEqual(model_cls.__module__, "sglang.srt.models.qwen3_embedding")
+        self.assertNotIn("Transformers", model_cls.__name__)
+
+    def test_bare_qwen3model_classified_as_embedding(self):
+        """'Qwen3Model' is non-generative regardless of the --is-embedding flag."""
+        self.assertFalse(is_generation_model(["Qwen3Model"]))
+        self.assertFalse(is_generation_model(["Qwen3Model"], is_embedding=False))
+        self.assertFalse(is_generation_model(["Qwen3Model"], is_embedding=True))
+
+    def test_existing_qwen3_archs_unaffected(self):
+        """The generative / classification archs keep their prior behavior."""
+        # Qwen3ForCausalLM is generative by default, embedding only with the flag
+        # (this is how Qwen3-Embedding-0.6B is served).
+        self.assertTrue(is_generation_model(["Qwen3ForCausalLM"]))
+        self.assertTrue(is_generation_model(["Qwen3ForCausalLM"], is_embedding=False))
+        self.assertFalse(is_generation_model(["Qwen3ForCausalLM"], is_embedding=True))
+        # Sequence-classification / reward archs stay non-generative.
+        self.assertFalse(is_generation_model(["Qwen3ForSequenceClassification"]))
+        self.assertFalse(is_generation_model(["Qwen3ForRewardModel"]))
+
+
+if __name__ == "__main__":
+    unittest.main()