parser

Signed-off-by: Roger Wang <hey@rogerw.io>

vllm/entrypoints/openai/chat_completion/serving.py

@@ -906,6 +906,12 @@ class OpenAIServingChat(OpenAIServing):
                         harmony_tools_streamed[i] |= tools_streamed_flag
                     # handle streaming deltas for tools with named tool_choice
                     elif tool_choice_function_name:
+                        if (
+                            reasoning_parser
+                            and not reasoning_end_arr[i]
+                            and prompt_is_reasoning_end_arr[i]
+                        ):
+                            reasoning_end_arr[i] = True
                         if (
                             reasoning_parser
                             and not reasoning_end_arr[i]
@@ -929,11 +935,8 @@ class OpenAIServingChat(OpenAIServing):
                             # i.e {"enable_thinking": False},
                             # set reasoning status to end.
                             # Only keep 'content', remove 'reasoning'.
-                            if (
-                                reasoning_parser.is_reasoning_end(
-                                    as_list(output.token_ids)
-                                )
-                                or prompt_is_reasoning_end_arr[i]
+                            if reasoning_parser.is_reasoning_end(
+                                as_list(output.token_ids)
                             ):
                                 reasoning_end_arr[i] = True
                                 if delta_message and delta_message.content:
@@ -1122,14 +1125,19 @@ class OpenAIServingChat(OpenAIServing):
 
                     # when only reasoning
                     elif reasoning_parser:
-                        delta_message = reasoning_parser.extract_reasoning_streaming(
-                            previous_text,
-                            current_text,
-                            delta_text,
-                            previous_token_ids,
-                            current_token_ids,
-                            output.token_ids,
-                        )
+                        if prompt_is_reasoning_end_arr[i]:
+                            delta_message = DeltaMessage(content=delta_text)
+                        else:
+                            delta_message = (
+                                reasoning_parser.extract_reasoning_streaming(
+                                    previous_text,
+                                    current_text,
+                                    delta_text,
+                                    previous_token_ids,
+                                    current_token_ids,
+                                    output.token_ids,
+                                )
+                            )
                     # handle streaming just a content delta
                     else:
                         delta_message = DeltaMessage(content=delta_text)

vllm/model_executor/layers/linear.py

@@ -685,8 +685,13 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         self,
         param: Parameter,
         loaded_weight: torch.Tensor,
-        loaded_shard_id: int | None = None,
+        loaded_shard_id: tuple[int, ...] | int | None = None,
     ):
+        if isinstance(loaded_shard_id, tuple):
+            raise NotImplementedError(
+                "Shard id with multiple indices is not supported in weight_loader, "
+                "please use weight_loader_v2 instead."
+            )
         # Special case for GGUF
         # initialize GGUF param after we know the quantize type
         is_gguf_weight = getattr(param, "is_gguf_weight", False)
@@ -770,6 +775,8 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         if output_dim is not None:
             shard_offset = sum(self.output_sizes[:loaded_shard_id])
             shard_size = self.output_sizes[loaded_shard_id]
+            shard_offset //= self.tp_size
+            shard_size //= self.tp_size
 
             if isinstance(param, BlockQuantScaleParameter):
                 weight_block_size = getattr(self, "weight_block_size", None)
@@ -777,9 +784,6 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
                     weight_block_size, shard_size, shard_offset
                 )
 
-            shard_offset //= self.tp_size
-            shard_size //= self.tp_size
-
             # Special case for quantization.
             # If quantized, we need to adjust the offset and size to account
             # for the packing.
@@ -825,7 +829,10 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         param_data.copy_(loaded_weight)
 
     def _load_fused_module_from_checkpoint(
-        self, param: BasevLLMParameter, loaded_weight: torch.Tensor
+        self,
+        param: BasevLLMParameter,
+        loaded_weight: torch.Tensor,
+        output_sizes: list[int] | None = None,
     ):
         """
         Handle special case for models where MLP layers are already
@@ -839,7 +846,8 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
 
         current_shard_offset = 0
         shard_offsets: list[tuple[int, int, int]] = []
-        for i, output_size in enumerate(self.output_sizes):
+        output_sizes = output_sizes or self.output_sizes
+        for i, output_size in enumerate(output_sizes):
             shard_offsets.append((i, current_shard_offset, output_size))
             current_shard_offset += output_size
 
@@ -864,23 +872,38 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
         self,
         param: BasevLLMParameter,
         loaded_weight: torch.Tensor,
-        loaded_shard_id: int | None = None,
+        loaded_shard_id: tuple[int, ...] | int | None = None,
     ):
-        if loaded_shard_id is None:
+        if loaded_shard_id is None or isinstance(loaded_shard_id, tuple):
             if isinstance(param, PerTensorScaleParameter):
                 param.load_merged_column_weight(loaded_weight=loaded_weight, shard_id=0)
                 return
             elif type(param) in (RowvLLMParameter, BasevLLMParameter):
                 param.load_merged_column_weight(loaded_weight=loaded_weight)
                 return
+            output_sizes = (
+                [self.output_sizes[idx] for idx in loaded_shard_id]
+                if loaded_shard_id
+                else None
+            )
+            if isinstance(param, BlockQuantScaleParameter):
+                weight_block_size = getattr(self, "weight_block_size", None)
+                output_sizes = [
+                    adjust_block_scale_shard(weight_block_size, size, 0)[0]
+                    for size in (output_sizes or self.output_sizes)
+                ]
             # TODO: @dsikka - move to parameter.py
-            self._load_fused_module_from_checkpoint(param, loaded_weight)
+            self._load_fused_module_from_checkpoint(
+                param, loaded_weight, output_sizes=output_sizes
+            )
             return
 
         assert loaded_shard_id < len(self.output_sizes)
 
         shard_offset = sum(self.output_sizes[:loaded_shard_id])
         shard_size = self.output_sizes[loaded_shard_id]
+        shard_offset //= self.tp_size
+        shard_size //= self.tp_size
 
         if isinstance(param, BlockQuantScaleParameter):
             weight_block_size = getattr(self, "weight_block_size", None)
@@ -888,9 +911,6 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
                 weight_block_size, shard_size, shard_offset
             )
 
-        shard_offset //= self.tp_size
-        shard_size //= self.tp_size
-
         param.load_merged_column_weight(
             loaded_weight=loaded_weight,
             shard_id=loaded_shard_id,

vllm/model_executor/models/qwen3_5.py

@@ -30,36 +30,20 @@ from collections.abc import Callable, Iterable
 import torch
 from einops import rearrange
 from torch import nn
-from transformers.activations import ACT2FN
 
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import (
-    CacheConfig,
-    ModelConfig,
-    SpeculativeConfig,
     VllmConfig,
-    get_current_vllm_config,
 )
 from vllm.distributed import (
-    divide,
     get_pp_group,
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
 )
 from vllm.logger import init_logger
 from vllm.model_executor.layers.layernorm import (
     GemmaRMSNorm as Qwen3_5RMSNorm,
 )
-from vllm.model_executor.layers.layernorm import RMSNormGated
-from vllm.model_executor.layers.linear import (
-    ColumnParallelLinear,
-    MergedColumnParallelLinear,
-    RowParallelLinear,
-)
+from vllm.model_executor.layers.linear import MergedColumnParallelLinear
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.mamba.mamba_mixer2 import (
-    mamba_v2_sharded_weight_loader,
-)
 from vllm.model_executor.layers.mamba.mamba_utils import (
     MambaStateCopyFunc,
     MambaStateCopyFuncCalculator,
@@ -73,11 +57,8 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 )
 from vllm.model_executor.model_loader.weight_utils import (
     default_weight_loader,
-    sharded_weight_loader,
 )
-from vllm.model_executor.utils import set_weight_attrs
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs.qwen3_5 import (
     Qwen3_5Config,
@@ -99,7 +80,6 @@ from .interfaces import (
 )
 from .qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP
 from .qwen3_next import (
-    ChunkGatedDeltaRule,
     Qwen3NextAttention,
     Qwen3NextDecoderLayer,
     Qwen3NextGatedDeltaNet,
@@ -139,154 +119,31 @@ class Qwen3_5MoeProcessingInfo(Qwen3VLProcessingInfo):
 
 
 class Qwen3_5GatedDeltaNet(Qwen3NextGatedDeltaNet):
-    def __init__(
-        self,
-        config: Qwen3_5TextConfig | Qwen3_5MoeTextConfig,
-        model_config: ModelConfig | None = None,
-        cache_config: CacheConfig | None = None,
-        quant_config: QuantizationConfig | None = None,
-        speculative_config: SpeculativeConfig | None = None,
-        prefix: str = "",
-    ) -> None:
-        super(Qwen3NextGatedDeltaNet, self).__init__()
-        self.tp_size = get_tensor_model_parallel_world_size()
-        self.tp_rank = get_tensor_model_parallel_rank()
-        self.hidden_size = config.hidden_size
-        self.num_v_heads = config.linear_num_value_heads
-        self.num_k_heads = config.linear_num_key_heads
-        self.head_k_dim = config.linear_key_head_dim
-        self.head_v_dim = config.linear_value_head_dim
-        self.key_dim = self.head_k_dim * self.num_k_heads
-        self.value_dim = self.head_v_dim * self.num_v_heads
-
-        self.conv_kernel_size = config.linear_conv_kernel_dim
-        self.layer_idx = extract_layer_index(prefix)
-        self.activation = config.hidden_act
-        self.act = ACT2FN[config.hidden_act]
-        self.layer_norm_epsilon = config.rms_norm_eps
-        self.prefix = prefix
-
-        self.config = config
-        self.model_config = model_config
-        self.cache_config = cache_config
-        self.quant_config = quant_config
-        self.speculative_config = speculative_config
-        self.num_spec = (
-            self.speculative_config.num_speculative_tokens
-            if self.speculative_config
-            else 0
-        )
-
-        # QKV
-        self.conv_dim = self.key_dim * 2 + self.value_dim
-        self.conv1d = ColumnParallelLinear(
-            input_size=self.conv_kernel_size,
-            output_size=self.conv_dim,
-            bias=False,
-            prefix=f"{prefix}.conv1d",
-        )
-        self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)
-
-        self.in_proj_qkv = MergedColumnParallelLinear(
-            input_size=self.hidden_size,
-            output_sizes=[self.key_dim, self.key_dim, self.value_dim],
-            bias=False,
-            quant_config=quant_config,
-            prefix=f"{prefix}.in_proj_qkv",
-        )
-        self.in_proj_z = ColumnParallelLinear(
-            input_size=self.hidden_size,
-            output_size=self.value_dim,
-            bias=False,
-            quant_config=quant_config,
-            prefix=f"{prefix}.in_proj_z",
-        )
-        self.in_proj_b = ColumnParallelLinear(
-            input_size=self.hidden_size,
-            output_size=self.num_v_heads,
-            bias=False,
-            quant_config=quant_config,
-            prefix=f"{prefix}.in_proj_b",
-        )
-        self.in_proj_a = ColumnParallelLinear(
-            input_size=self.hidden_size,
-            output_size=self.num_v_heads,
-            bias=False,
-            quant_config=quant_config,
-            prefix=f"{prefix}.in_proj_a",
-        )
-
-        query_key_settings = (self.key_dim, 0, False)
-        value_settings = (self.value_dim, 0, False)
-
-        delattr(self.conv1d.weight, "weight_loader")
-        set_weight_attrs(
-            self.conv1d.weight,
-            {
-                "weight_loader": mamba_v2_sharded_weight_loader(
-                    [
-                        query_key_settings,
-                        query_key_settings,
-                        value_settings,
-                    ],
-                    self.tp_size,
-                    self.tp_rank,
-                )
-            },
-        )
-
-        # selective projection used to make dt, B and C input dependant
-
-        # time step projection (discretization)
-        # instantiate once and copy inv_dt in init_weights of PretrainedModel
-        self.dt_bias = nn.Parameter(
-            torch.ones(self.num_v_heads // self.tp_size),
-        )
-        self.A_log = nn.Parameter(
-            torch.empty(
-                divide(self.num_v_heads, self.tp_size),
-            )
-        )
-
-        set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
-        set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
-
-        self.norm = RMSNormGated(
-            self.head_v_dim,
-            eps=self.layer_norm_epsilon,
-            group_size=None,
-            norm_before_gate=True,
-            device=current_platform.current_device(),
-            dtype=config.dtype,
-        )
-
-        self.out_proj = RowParallelLinear(
-            self.value_dim,
-            self.hidden_size,
-            bias=False,
-            input_is_parallel=True,
-            quant_config=quant_config,
-            prefix=f"{prefix}.out_proj",
-        )
-
-        self.chunk_gated_delta_rule = ChunkGatedDeltaRule()
-
-        compilation_config = get_current_vllm_config().compilation_config
-        if prefix in compilation_config.static_forward_context:
-            raise ValueError(f"Duplicate layer name: {prefix}")
-        compilation_config.static_forward_context[prefix] = self
-
     def fix_query_key_value_ordering(
         self,
-        mixed_qkv,
-        z,
-        b,
-        a,
+        mixed_qkvz: torch.Tensor,
+        mixed_ba: torch.Tensor,
     ):
         raise NotImplementedError(
             "Qwen3.5 Series dont need to fix query key value ordering"
         )
 
+    def create_qkvz_proj(
+        self,
+        hidden_size: int,
+        key_dim: int,
+        value_dim: int,
+        quant_config: QuantizationConfig | None,
+        prefix: str,
+    ) -> MergedColumnParallelLinear:
+        return MergedColumnParallelLinear(
+            input_size=hidden_size,
+            output_sizes=[key_dim, key_dim, value_dim, value_dim],
+            bias=False,
+            quant_config=quant_config,
+            prefix=prefix,
+        )
+
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -303,11 +160,13 @@ class Qwen3_5GatedDeltaNet(Qwen3NextGatedDeltaNet):
         # ============================================================
         # Part 1: Input Projection
         # ============================================================
-        mixed_qkv, _ = self.in_proj_qkv(hidden_states)
-        z, _ = self.in_proj_z(hidden_states)
+        mixed_qkvz, _ = self.in_proj_qkvz(hidden_states)
+        qkv_size = (self.key_dim * 2 + self.value_dim) // self.tp_size
+        z_size = self.value_dim // self.tp_size
+        mixed_qkv, z = mixed_qkvz.split([qkv_size, z_size], dim=-1)
         z = z.reshape(z.size(0), -1, self.head_v_dim)
-        b, _ = self.in_proj_b(hidden_states)
-        a, _ = self.in_proj_a(hidden_states)
+        ba, _ = self.in_proj_ba(hidden_states)
+        b, a = ba.chunk(2, dim=-1)
 
         b = b.contiguous()
         a = a.contiguous()
@@ -506,11 +365,18 @@ class Qwen3_5Model(Qwen3NextModel):
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
+            # self attention
             ("qkv_proj", "q_proj", "q"),
             ("qkv_proj", "k_proj", "k"),
             ("qkv_proj", "v_proj", "v"),
+            # mlp
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
+            # GDN
+            ("in_proj_qkvz", "in_proj_qkv", (0, 1, 2)),
+            ("in_proj_qkvz", "in_proj_z", 3),
+            ("in_proj_ba", "in_proj_b", 0),
+            ("in_proj_ba", "in_proj_a", 1),
         ]
 
         params_dict = dict(self.named_parameters())
@@ -657,6 +523,9 @@ class Qwen3_5ForCausalLMBase(
             "v_proj",
         ],
         "gate_up_proj": ["gate_proj", "up_proj"],
+        # GDN fused projections.
+        "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"],
+        "in_proj_ba": ["in_proj_b", "in_proj_a"],
     }
 
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
@@ -754,7 +623,12 @@ class Qwen3_5MoeForCausalLM(Qwen3_5ForCausalLMBase, QwenNextMixtureOfExperts):
     dummy_inputs=Qwen3VLDummyInputsBuilder,
 )
 class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration, IsHybrid):
-    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+    packed_modules_mapping = Qwen3VLForConditionalGeneration.packed_modules_mapping | {
+        "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"],
+        "in_proj_ba": ["in_proj_b", "in_proj_a"],
+    }
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"):
         # protocols have not __init__ method, so we need to use nn.Module.__init__
         nn.Module.__init__(self)
         config: Qwen3_5Config = vllm_config.model_config.hf_config
@@ -962,7 +836,7 @@ class Qwen3_5_MoeMixtureOfExperts(MixtureOfExperts):
 class Qwen3_5MoeForConditionalGeneration(
     Qwen3_5ForConditionalGeneration, Qwen3_5_MoeMixtureOfExperts
 ):
-    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"):
         # protocols have not __init__ method, so we need to use nn.Module.__init__
         nn.Module.__init__(self)
         config: Qwen3_5MoeConfig = vllm_config.model_config.hf_config

vllm/model_executor/models/qwen3_next.py

@@ -44,6 +44,7 @@ from vllm.model_executor.layers.layernorm import (
 from vllm.model_executor.layers.layernorm import RMSNormGated
 from vllm.model_executor.layers.linear import (
     ColumnParallelLinear,
+    MergedColumnParallelLinear,
     QKVParallelLinear,
     ReplicatedLinear,
     RowParallelLinear,
@@ -406,19 +407,19 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
         self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)
 
         # projection of the input hidden states
-        self.projection_size_qkvz = self.key_dim * 2 + self.value_dim * 2
-        self.projection_size_ba = self.num_v_heads * 2
-        self.in_proj_qkvz = ColumnParallelLinear(
-            input_size=self.hidden_size,
-            output_size=self.projection_size_qkvz,
-            bias=False,
+        # Qwen3-Next and Qwen3.5 has a different qkv_proj layout,
+        # we need to create qkvz_proj adaptively here.
+        self.in_proj_qkvz = self.create_qkvz_proj(
+            hidden_size=self.hidden_size,
+            key_dim=self.key_dim,
+            value_dim=self.value_dim,
             quant_config=quant_config,
             prefix=f"{prefix}.in_proj_qkvz",
         )
         # ba_proj doesn't support blockwise fp8 quantization.
-        self.in_proj_ba = ColumnParallelLinear(
+        self.in_proj_ba = MergedColumnParallelLinear(
             input_size=self.hidden_size,
-            output_size=self.projection_size_ba,
+            output_sizes=[self.num_v_heads] * 2,
             bias=False,
             quant_config=quant_config,
             prefix=f"{prefix}.in_proj_ba",
@@ -484,10 +485,26 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
             raise ValueError(f"Duplicate layer name: {prefix}")
         compilation_config.static_forward_context[prefix] = self
 
+    def create_qkvz_proj(
+        self,
+        hidden_size: int,
+        key_dim: int,
+        value_dim: int,
+        quant_config: QuantizationConfig | None,
+        prefix: str,
+    ) -> MergedColumnParallelLinear:
+        return MergedColumnParallelLinear(
+            input_size=hidden_size,
+            output_sizes=[sum((key_dim, key_dim, value_dim, value_dim))],
+            bias=False,
+            quant_config=quant_config,
+            prefix=prefix,
+        )
+
     def fix_query_key_value_ordering(
         self,
-        mixed_qkvz,
-        mixed_ba,
+        mixed_qkvz: torch.Tensor,
+        mixed_ba: torch.Tensor,
     ):
         """
         Derives `query`, `key` and `value` tensors from `mixed_qkvzba`.

vllm/reasoning/qwen3_reasoning_parser.py

@@ -1,9 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+from collections.abc import Sequence
+
 from vllm.entrypoints.openai.chat_completion.protocol import (
     ChatCompletionRequest,
 )
+from vllm.entrypoints.openai.engine.protocol import DeltaMessage
 from vllm.entrypoints.openai.responses.protocol import (
     ResponsesRequest,
 )
@@ -12,13 +15,22 @@ from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser
 
 class Qwen3ReasoningParser(BaseThinkingReasoningParser):
     """
-    Reasoning parser for the Qwen3 model.
+    Reasoning parser for the Qwen3/Qwen3.5 model family.
 
-    The Qwen3 model uses <think>...</think> tokens to denote reasoning text
-    within its output. The model provides a strict switch to disable reasoning
-    output via the 'enable_thinking=False' parameter. This parser extracts the
-    reasoning content enclosed by <think> and </think> tokens from the model's
-    output.
+    The Qwen3 model family uses <think>...</think> tokens to denote reasoning
+    text. Starting with Qwen3.5, the chat template places <think> in the
+    prompt so only </think> appears in the generated output. The model
+    provides a strict switch to disable reasoning output via the
+    'enable_thinking=False' parameter.
+
+    When thinking is disabled, the template places <think>\\n\\n</think>\\n\\n
+    in the prompt. The serving layer detects this via prompt_is_reasoning_end
+    and routes deltas as content without calling the streaming parser.
+
+    NOTE: Models up to the 2507 release (e.g., Qwen/Qwen3-235B-A22B-Instruct-2507)
+    use an older chat template where the model generates <think> itself.
+    This parser handles both styles: if <think> appears in the generated output
+    it is stripped before extraction (non-streaming) or skipped (streaming).
     """
 
     @property
@@ -37,31 +49,27 @@ class Qwen3ReasoningParser(BaseThinkingReasoningParser):
         """
         Extract reasoning content from the model output.
 
-        Qwen3 has stricter requirements - it needs both start and end tokens
-        to be present, unlike other models that work with just the end token.
+        The <think> token is placed in the prompt by the chat template,
+        so typically only </think> appears in the generated output.
+        If <think> is present (e.g. from a different template), it is
+        stripped before extraction.
 
-        For text <think>abc</think>xyz:
-        - 'abc' goes to reasoning
-        - 'xyz' goes to content
+        When thinking is disabled (no </think> in output), returns
+        (None, model_output) to indicate all output is content.
 
         Returns:
             tuple[Optional[str], Optional[str]]: reasoning content and content
         """
 
-        # Check if the model output contains both <think> and </think> tokens.
-        if self.start_token not in model_output or self.end_token not in model_output:
-            return None, model_output
-
-        # Check if the <think> is present in the model output, remove it
-        # if it is present.
+        # Strip <think> if present in the generated output.
         model_output_parts = model_output.partition(self.start_token)
         model_output = (
             model_output_parts[2] if model_output_parts[1] else model_output_parts[0]
         )
 
-        # Check if the model output contains the </think> tokens.
-        # If the end token is not found, return the model output as is.
         if self.end_token not in model_output:
+            # No end token means thinking is disabled or the model
+            # did not produce reasoning. Treat everything as content.
             return None, model_output
 
         # Extract reasoning content from the model output.
@@ -69,3 +77,57 @@ class Qwen3ReasoningParser(BaseThinkingReasoningParser):
 
         final_content = content or None
         return reasoning, final_content
+
+    def extract_reasoning_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+    ) -> DeltaMessage | None:
+        """
+        Extract reasoning content from a streaming delta.
+
+        Since <think> is placed in the prompt by the chat template, all
+        generated tokens before </think> are reasoning and tokens after
+        are content.
+
+        NOTE: When thinking is disabled, no think tokens appear in the
+        generated output. The serving layer detects this via
+        prompt_is_reasoning_end and routes deltas as content without
+        calling this method.
+        """
+        # Strip <think> from delta if present (old template / edge case
+        # where the model generates <think> itself).
+        if self.start_token_id in delta_token_ids:
+            start_idx = delta_text.find(self.start_token)
+            if start_idx >= 0:
+                delta_text = delta_text[start_idx + len(self.start_token) :]
+
+        if self.end_token_id in delta_token_ids:
+            # End token in this delta: split reasoning from content.
+            end_index = delta_text.find(self.end_token)
+            if end_index >= 0:
+                reasoning = delta_text[:end_index]
+                content = delta_text[end_index + len(self.end_token) :]
+                if not reasoning and not content:
+                    return None
+                return DeltaMessage(
+                    reasoning=reasoning if reasoning else None,
+                    content=content if content else None,
+                )
+            # end_token_id in IDs but not in text (already stripped)
+            return None
+
+        # No end token in this delta.
+        if not delta_text:
+            # Nothing left after stripping start token.
+            return None
+        elif self.end_token_id in previous_token_ids:
+            # End token already passed: everything is content now.
+            return DeltaMessage(content=delta_text)
+        else:
+            # No end token yet: still in reasoning phase.
+            return DeltaMessage(reasoning=delta_text)