[Models] Step-3.5-Flash (#33523)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: i-zhangmingming <i-zhangmingming@stepfun.com>
Co-authored-by: xiewuxun <xiewuxun@stepfun.com>
Co-authored-by: zetaohong <i-hongzetao@stepfun.com>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>

docs/models/supported_models.md

@@ -456,6 +456,7 @@ th {
 | `StableLmForCausalLM` | StableLM | `stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc. | | |
 | `Starcoder2ForCausalLM` | Starcoder2 | `bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc. | | ✅︎ |
 | `Step1ForCausalLM` | Step-Audio | `stepfun-ai/Step-Audio-EditX`, etc. | ✅︎ | ✅︎ |
+| `Step3p5ForCausalLM` | Step-3.5-flash | `stepfun-ai/step-3.5-flash`, etc. |  | ✅︎ |
 | `TeleChat2ForCausalLM` | TeleChat2 | `Tele-AI/TeleChat2-3B`, `Tele-AI/TeleChat2-7B`, `Tele-AI/TeleChat2-35B`, etc. | ✅︎ | ✅︎ |
 | `TeleFLMForCausalLM` | TeleFLM | `CofeAI/FLM-2-52B-Instruct-2407`, `CofeAI/Tele-FLM`, etc. | ✅︎ | ✅︎ |
 | `XverseForCausalLM` | XVERSE | `xverse/XVERSE-7B-Chat`, `xverse/XVERSE-13B-Chat`, `xverse/XVERSE-65B-Chat`, etc. | ✅︎ | ✅︎ |

tests/kernels/core/test_activation.py

@@ -17,6 +17,8 @@ from vllm.model_executor.layers.activation import (
     QuickGELU,
     SiluAndMul,
     SwigluOAIAndMul,
+    SwigluStepAndMul,
+    swiglustep_and_mul_triton,
 )
 from vllm.utils.torch_utils import set_random_seed
 
@@ -36,6 +38,7 @@ CUDA_DEVICES = [f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 e
         "gelu_tanh",
         "fatrelu",
         "swigluoai_and_mul",
+        "swiglustep_and_mul",
     ],
 )
 @pytest.mark.parametrize("num_tokens", NUM_TOKENS)
@@ -75,9 +78,12 @@ def test_act_and_mul(
     elif activation == "swigluoai_and_mul":
         layer = SwigluOAIAndMul()
         fn = torch.ops._C.swigluoai_and_mul
+    elif activation == "swiglustep_and_mul":
+        layer = SwigluStepAndMul()
+        fn = swiglustep_and_mul_triton
     out = layer(x)
     ref_out = layer.forward_native(x)
-    if activation == "swigluoai_and_mul":
+    if activation in ["swigluoai_and_mul", "swiglustep_and_mul"]:
         rtol = {
             # For fp16, change the relative tolerance from 1e-3 to 2e-3
             torch.float16: 2e-3,
@@ -104,7 +110,7 @@ def test_act_and_mul(
         opcheck(fn, (out, x, threshold))
     elif activation == "swigluoai_and_mul":
         opcheck(fn, (out, x, layer.alpha, layer.limit))
-    else:
+    elif activation != "swiglustep_and_mul":
         opcheck(fn, (out, x))
 
 

tests/models/registry.py

@@ -484,6 +484,9 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "Step1ForCausalLM": _HfExamplesInfo(
         "stepfun-ai/Step-Audio-EditX", trust_remote_code=True
     ),
+    "Step3p5ForCausalLM": _HfExamplesInfo(
+        "stepfun-ai/step-3.5-flash", is_available_online=False
+    ),
     "SmolLM3ForCausalLM": _HfExamplesInfo("HuggingFaceTB/SmolLM3-3B"),
     "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"),
     "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
@@ -1113,6 +1116,12 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
     "Qwen3NextMTP": _HfExamplesInfo(
         "Qwen/Qwen3-Next-80B-A3B-Instruct", min_transformers_version="4.56.3"
     ),
+    "Step3p5MTP": _HfExamplesInfo(
+        "stepfun-ai/Step-3.5-Flash",
+        trust_remote_code=True,
+        speculative_model="stepfun-ai/Step-3.5-Flash",
+        is_available_online=False,
+    ),
 }
 
 _TRANSFORMERS_BACKEND_MODELS = {

vllm/config/speculative.py

@@ -41,6 +41,7 @@ MTPModelTypes = Literal[
     "longcat_flash_mtp",
     "mtp",
     "pangu_ultra_moe_mtp",
+    "step3p5_mtp",
 ]
 EagleModelTypes = Literal["eagle", "eagle3", MTPModelTypes]
 SpeculativeMethod = Literal[
@@ -264,6 +265,11 @@ class SpeculativeConfig:
                 {"n_predict": n_predict, "architectures": ["LongCatFlashMTPModel"]}
             )
 
+        if hf_config.model_type == "step3p5":
+            hf_config.model_type = "step3p5_mtp"
+            n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
+            hf_config.update({"n_predict": n_predict, "architectures": ["Step3p5MTP"]})
+
         if initial_architecture == "MistralLarge3ForCausalLM":
             hf_config.update({"architectures": ["EagleMistralLarge3ForCausalLM"]})
 

vllm/model_executor/layers/activation.py

@@ -17,11 +17,63 @@ from vllm.logger import init_logger
 from vllm.model_executor.custom_op import CustomOp
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
 from vllm.utils.collection_utils import LazyDict
 
 logger = init_logger(__name__)
 
 
+@triton.jit
+def _swiglustep_and_mul_kernel(
+    o_ptr,
+    o_stride,
+    x_ptr,
+    x_stride,
+    limit: tl.constexpr,
+    d: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+) -> None:
+    i = tl.program_id(axis=0).to(tl.int64)
+    j = tl.program_id(axis=1)
+    o_row_ptr = o_ptr + o_stride * i
+    x_row_ptr = x_ptr + x_stride * i
+    offsets = j * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
+    mask = offsets < d
+
+    gate = tl.load(x_row_ptr + offsets, mask=mask).to(tl.float32)
+    up = tl.load(x_row_ptr + offsets + d, mask=mask).to(tl.float32)
+
+    gate_silu = tl.sigmoid(gate) * gate
+    gate_clamped = tl.minimum(gate_silu, limit)
+    up_clamped = tl.minimum(tl.maximum(up, -limit), limit)
+
+    result = gate_clamped * up_clamped
+    result = result.to(x_ptr.dtype.element_ty)
+    tl.store(o_row_ptr + offsets, result, mask=mask)
+
+
+def swiglustep_and_mul_triton(
+    output: torch.Tensor, input: torch.Tensor, limit: float = 7.0
+):
+    b, n = input.shape
+    assert input.ndim == 2
+    assert n % 2 == 0
+    d = n // 2
+
+    def grid(meta):
+        return (b, triton.cdiv(d, meta["BLOCK_SIZE"]))
+
+    _swiglustep_and_mul_kernel[grid](
+        output,
+        output.stride(0),
+        input,
+        input.stride(0),
+        limit=limit,
+        d=d,
+        BLOCK_SIZE=1024,
+    )
+
+
 # --8<-- [start:fatrelu_and_mul]
 @CustomOp.register("fatrelu_and_mul")
 class FatreluAndMul(CustomOp):
@@ -304,6 +356,44 @@ class SwigluOAIAndMul(CustomOp):
         return f"alpha={repr(self.alpha)}, limit={repr(self.limit)}"
 
 
+# --8<-- [start:swiglustep_and_mul]
+@CustomOp.register("swiglustep_and_mul")
+class SwigluStepAndMul(CustomOp):
+    """An activation function for SwiGLU with clamping.
+
+    Computes x -> silu(x[:d]).clamp(max=limit) * x[d:].clamp(-limit, limit)
+    where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    def __init__(self, limit: float = 7.0):
+        super().__init__()
+        if limit is None:
+            raise ValueError("SwigluStepAndMul requires limit to be set.")
+        self.limit = limit
+
+    def forward_native(self, x: torch.Tensor) -> torch.Tensor:
+        """PyTorch-native implementation equivalent to forward()."""
+        gate, up = x.chunk(2, dim=-1)
+        gate = F.silu(gate)
+        gate = gate.clamp(max=self.limit)
+        up = up.clamp(min=-self.limit, max=self.limit)
+        return gate * up
+
+    def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        swiglustep_and_mul_triton(out, x, self.limit)
+        return out
+
+    def extra_repr(self) -> str:
+        return f"limit={repr(self.limit)}"
+
+
 # --8<-- [start:gelu_new]
 @CustomOp.register("gelu_new")
 class NewGELU(CustomOp):

vllm/model_executor/layers/fused_moe/deep_gemm_moe.py

@@ -144,7 +144,7 @@ class DeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
 
     @staticmethod
     def _supports_activation(activation: str) -> bool:
-        return activation in ["silu"]
+        return activation in ["silu", "swiglustep"]
 
     @staticmethod
     def _supports_parallel_config(moe_parallel_config: FusedMoEParallelConfig) -> bool:

vllm/model_executor/layers/fused_moe/fused_moe.py

@@ -1939,7 +1939,7 @@ class TritonExperts(mk.FusedMoEPermuteExpertsUnpermute):
 
     @staticmethod
     def _supports_activation(activation: str) -> bool:
-        return activation in ["silu", "gelu", "swigluoai"]
+        return activation in ["silu", "gelu", "swigluoai", "swiglustep"]
 
     @staticmethod
     def _supports_parallel_config(moe_parallel_config: FusedMoEParallelConfig) -> bool:

vllm/model_executor/layers/fused_moe/utils.py

@@ -358,6 +358,11 @@ def apply_moe_activation(
         torch.ops._C.gelu_and_mul(output, input)
     elif activation == "swigluoai":
         torch.ops._C.swigluoai_and_mul(output, input)
+    elif activation == "swiglustep":
+        from vllm.model_executor.layers.activation import swiglustep_and_mul_triton
+
+        swiglustep_and_mul_triton(output, input)
+
     # Activations without gated multiplication
     elif activation == SILU_NO_MUL:
         output.copy_(F.silu(input))

vllm/model_executor/models/registry.py

@@ -189,6 +189,7 @@ _TEXT_GENERATION_MODELS = {
     "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
     "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
     "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
+    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
     "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
     "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
     "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
@@ -495,6 +496,7 @@ _SPECULATIVE_DECODING_MODELS = {
     "MedusaModel": ("medusa", "Medusa"),
     "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
     "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
+    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
     # Temporarily disabled.
     # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
     # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),

vllm/model_executor/models/step3p5_mtp.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from collections.abc import Iterable
+
+import torch
+import torch.nn as nn
+from transformers import PretrainedConfig
+
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.model_executor.layers.layernorm import GemmaRMSNorm
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.sequence import IntermediateTensors
+
+from .step3p5 import Step3p5DecoderLayer, get_spec_layer_idx_from_weight_name
+from .utils import maybe_prefix
+
+logger = init_logger(__name__)
+
+
+class SharedHead(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: QuantizationConfig | None = None,
+    ) -> None:
+        super().__init__()
+        self.norm = GemmaRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.head = ParallelLMHead(
+            config.vocab_size, config.hidden_size, quant_config=quant_config
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return self.norm(hidden_states)
+
+
+class Step3p5AMultiTokenPredictorLayer(nn.Module):
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        prefix: str,
+    ) -> None:
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        self.enorm = GemmaRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.hnorm = GemmaRMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
+        self.shared_head = SharedHead(config=config, quant_config=quant_config)
+        self.mtp_block = Step3p5DecoderLayer(
+            vllm_config,
+            prefix=f"{prefix}.mtp_block",
+        )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+        inputs_embeds: torch.Tensor | None = None,
+        spec_step_index: int = 0,
+    ) -> torch.Tensor:
+        assert inputs_embeds is not None
+        inputs_embeds = self.enorm(inputs_embeds)
+        previous_hidden_states = self.hnorm(previous_hidden_states)
+
+        hidden_states = self.eh_proj(
+            torch.cat([inputs_embeds, previous_hidden_states], dim=-1)
+        )
+
+        hidden_states = self.mtp_block(positions=positions, hidden_states=hidden_states)
+        return hidden_states
+
+
+class Step3p5AMultiTokenPredictor(nn.Module):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.mtp_start_layer_idx = config.num_hidden_layers
+        self.num_mtp_layers = config.num_nextn_predict_layers
+        # to map the exact layer index from weights
+        self.layers = torch.nn.ModuleDict(
+            {
+                str(idx): Step3p5AMultiTokenPredictorLayer(
+                    vllm_config,
+                    f"{prefix}.layers.{idx}",
+                )
+                for idx in range(
+                    self.mtp_start_layer_idx,
+                    self.mtp_start_layer_idx + self.num_mtp_layers,
+                )
+            }
+        )
+
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        previous_hidden_states: torch.Tensor,
+        inputs_embeds: torch.Tensor | None = None,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        current_step_idx = spec_step_idx % self.num_mtp_layers
+        return self.layers[str(self.mtp_start_layer_idx + current_step_idx)](
+            input_ids,
+            positions,
+            previous_hidden_states,
+            inputs_embeds,
+            current_step_idx,
+        )
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        current_step_idx = spec_step_idx % self.num_mtp_layers
+        mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)]
+        logits = self.logits_processor(
+            mtp_layer.shared_head.head, mtp_layer.shared_head(hidden_states)
+        )
+        return logits
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+
+class Step3p5MTP(nn.Module):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        self.config = vllm_config.model_config.hf_config
+        self.vllm_config = vllm_config
+        self.model = Step3p5AMultiTokenPredictor(
+            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.embed_input_ids(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor:
+        hidden_states = self.model(
+            input_ids, positions, hidden_states, inputs_embeds, spec_step_idx
+        )
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        spec_step_idx: int = 0,
+    ) -> torch.Tensor | None:
+        return self.model.compute_logits(hidden_states, spec_step_idx)
+
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        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),
+        ]
+
+        expert_params_mapping = [
+            (".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
+            (".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
+            (".moe.experts.w2_weight", ".moe.down_proj.weight", "w2"),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        loaded_params: set[str] = set()
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name:
+                continue
+            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
+            if "embed_tokens" not in name and spec_layer is None:
+                continue
+            name = self._rewrite_spec_layer_name(spec_layer, name)
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                # Skip non-stacked layers and experts (experts handled below).
+                if weight_name not in name:
+                    continue
+                # We have mlp.experts[0].gate_proj in the checkpoint.
+                # Since we handle the experts below in expert_params_mapping,
+                # we need to skip here BEFORE we update the name, otherwise
+                # name will be updated to mlp.experts[0].gate_up_proj, which
+                # will then be updated below in expert_params_mapping
+                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
+                if ("mlp.experts." in name) and name not in params_dict:
+                    continue
+                if "experts" in name or "moe" in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                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:
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+                    # Skip loading extra bias for GPTQ models.
+                    if (
+                        name.endswith(".bias") or name.endswith("_bias")
+                    ) and name not in params_dict:
+                        continue
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    for expert_id in range(loaded_weight.shape[0]):
+                        loaded_weight_expert = loaded_weight[expert_id]
+                        weight_loader(
+                            param,
+                            loaded_weight_expert,
+                            name,
+                            shard_id=shard_id,
+                            expert_id=expert_id,
+                        )
+                    loaded_params.add(name)
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if (
+                        name.endswith(".bias")
+                        and name not in params_dict
+                        or "tok_embeddings" in name
+                    ):
+                        continue
+
+                    if spec_layer is not None and ".transformer." in name:
+                        name = name.replace(".transformer.", ".")
+                    if "shared_head" in name:
+                        name = name.replace("shared_head.output", "shared_head.head")
+                    if "embed_tokens" in name:
+                        assert (
+                            hasattr(self.config, "num_nextn_predict_layers")
+                            and self.config.num_nextn_predict_layers > 0
+                        )
+                        name = "model.embed_tokens.weight"
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        params_need_to_load = set(params_dict.keys())
+        # Some KV cache scales are optional: checkpoints may omit them and vLLM
+        # will fall back to default scales during initialization.
+        optional_params = {
+            name
+            for name, param in params_dict.items()
+            if name.endswith((".k_scale", ".v_scale", ".q_scale", ".prob_scale"))
+            and getattr(param, "numel", lambda: 0)() == 1
+            and getattr(param, "requires_grad", False) is False
+        }
+        params_need_to_load -= optional_params
+        if params_need_to_load != loaded_params:
+            missing_params = list(params_need_to_load - loaded_params)
+            param_name_example = missing_params[0]
+            raise RuntimeError(
+                "Some parameters like "
+                f"{param_name_example} are not in the checkpoint and will falsely "
+                "use random initialization"
+            )
+        return loaded_params
+
+    def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str:
+        """
+        Rewrite the weight name to match the format of the original model.
+        Add .mtp_block for modules in transformer layer block for spec layer
+        """
+        spec_layer_weight_names = [
+            "embed_tokens",
+            "enorm",
+            "hnorm",
+            "eh_proj",
+            "shared_head",
+        ]
+        spec_layer_weight = False
+        for weight_name in spec_layer_weight_names:
+            if weight_name in name:
+                spec_layer_weight = True
+                break
+        if not spec_layer_weight:
+            # treat rest weights as weights for transformer layer block
+            name = name.replace(
+                f"model.layers.{spec_layer}.", f"model.layers.{spec_layer}.mtp_block."
+            )
+        return name

vllm/reasoning/__init__.py

@@ -84,6 +84,10 @@ _REASONING_PARSERS_TO_REGISTER = {
         "step3_reasoning_parser",
         "Step3ReasoningParser",
     ),
+    "step3p5": (
+        "step3p5_reasoning_parser",
+        "Step3p5ReasoningParser",
+    ),
 }
 
 

vllm/reasoning/step3p5_reasoning_parser.py

+# 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,
+)
+from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser
+from vllm.tokenizers import TokenizerLike
+
+
+class Step3p5ReasoningParser(BaseThinkingReasoningParser):
+    """
+    Reasoning parser for Step3p5 model.
+
+    Step3p5 uses the <think>...</think> format, but it tends to emit an extra
+    newline immediately before and/or after the </think> token. This parser trims:
+      - the newline right before </think>
+      - the newline right after </think>
+    """
+
+    @property
+    def start_token(self) -> str:
+        return "<think>"
+
+    @property
+    def end_token(self) -> str:
+        return "</think>"
+
+    def __init__(self, tokenizer: TokenizerLike, *args, **kwargs):
+        super().__init__(tokenizer, *args, **kwargs)
+
+        # Used to hold a trailing "\n" from reasoning content so we can decide
+        # whether it is immediately before </think>.
+        self._pending_reasoning_newline = False
+
+        # Used to delay the reasoning end detection.
+        # This is necessary to remove the newline appears immediately after </think>,
+        # which may cause the end detection to be delayed by one round.
+        self.end_offset = 1
+
+    def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
+        if self.end_token_id in input_ids and self.end_offset > 0:
+            self.end_offset -= 1
+            return False
+        return self.end_offset < 1
+
+    def is_reasoning_end_streaming(
+        self, input_ids: Sequence[int], delta_ids: Sequence[int]
+    ) -> bool:
+        if self.end_token_id in input_ids and self.end_offset > 0:
+            self.end_offset -= 1
+            return False
+        return self.end_offset < 1
+
+    def extract_reasoning(
+        self,
+        model_output: str,
+        request: ChatCompletionRequest | ResponsesRequest,
+    ) -> tuple[str | None, str | None]:
+        reasoning, content = super().extract_reasoning(model_output, request)
+        if reasoning is not None:
+            reasoning = reasoning.removesuffix("\n")
+        if content is not None:
+            content = content.removeprefix("\n")
+        return reasoning or None, content or None
+
+    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:
+        # Drop the immediate newline that models often emit after </think>.
+        if previous_text.endswith(self.end_token) and delta_text:
+            if delta_text == "\n":
+                return None
+            elif delta_text.startswith("\n"):
+                remaining = delta_text.removeprefix("\n")
+                return DeltaMessage(content=remaining) if remaining else None
+
+        ret = super().extract_reasoning_streaming(
+            previous_text,
+            current_text,
+            delta_text,
+            previous_token_ids,
+            current_token_ids,
+            delta_token_ids,
+        )
+
+        if ret is None:
+            return None
+
+        # Compatibility path for models that don't generate the start token:
+        # treat everything before </think> as reasoning and everything after
+        # as content.
+        if (
+            self.start_token_id not in previous_token_ids
+            and self.start_token_id not in delta_token_ids
+        ):
+            if self.end_token_id in delta_token_ids:
+                end_index = delta_text.find(self.end_token)
+                reasoning = delta_text[:end_index]
+                content = delta_text[end_index + len(self.end_token) :]
+                ret = DeltaMessage(reasoning=reasoning, content=content or None)
+            elif self.end_token_id in previous_token_ids:
+                ret = DeltaMessage(content=delta_text)
+            else:
+                ret = DeltaMessage(reasoning=delta_text)
+
+        reasoning_to_output = ret.reasoning
+        content_to_output = ret.content
+
+        # Reasoning: handle the newline immediately before </think>.
+        if reasoning_to_output is not None:
+            if self._pending_reasoning_newline:
+                reasoning_to_output = "\n" + reasoning_to_output
+                self._pending_reasoning_newline = False
+
+            if reasoning_to_output.endswith("\n"):
+                reasoning_to_output = reasoning_to_output.removesuffix("\n")
+                if self.end_token in delta_text:
+                    # Trailing "\n" is right before </think>, drop it.
+                    self._pending_reasoning_newline = False
+                else:
+                    # Hold the trailing "\n" until we know whether </think> follows.
+                    self._pending_reasoning_newline = True
+
+        # Content: handle the newline immediately after </think>.
+        if content_to_output is not None:
+            # No need to get into parser again to remove newline after </think>.
+            self.end_offset -= 1
+
+            # If we have content, reasoning must have ended.
+            self._pending_reasoning_newline = False
+
+            if self.end_token in delta_text and content_to_output.startswith("\n"):
+                content_to_output = content_to_output.removeprefix("\n")
+
+        reasoning_to_output = reasoning_to_output or None
+        content_to_output = content_to_output or None
+        if reasoning_to_output is None and content_to_output is None:
+            return None
+
+        return DeltaMessage(reasoning=reasoning_to_output, content=content_to_output)

vllm/tool_parsers/__init__.py

@@ -134,6 +134,10 @@ _TOOL_PARSERS_TO_REGISTER = {
         "step3_tool_parser",
         "Step3ToolParser",
     ),
+    "step3p5": (
+        "step3p5_tool_parser",
+        "Step3p5ToolParser",
+    ),
     "xlam": (
         "xlam_tool_parser",
         "xLAMToolParser",

vllm/transformers_utils/config.py

@@ -97,6 +97,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
     ultravox="UltravoxConfig",
     step3_vl="Step3VLConfig",
     step3_text="Step3TextConfig",
+    step3p5="Step3p5Config",
     qwen3_asr="Qwen3ASRConfig",
     qwen3_next="Qwen3NextConfig",
     lfm2_moe="Lfm2MoeConfig",

vllm/transformers_utils/configs/__init__.py

@@ -52,6 +52,7 @@ _CLASS_TO_MODULE: dict[str, str] = {
     "Step3VLConfig": "vllm.transformers_utils.configs.step3_vl",
     "Step3VisionEncoderConfig": "vllm.transformers_utils.configs.step3_vl",
     "Step3TextConfig": "vllm.transformers_utils.configs.step3_vl",
+    "Step3p5Config": "vllm.transformers_utils.configs.step3p5",
     "Qwen3ASRConfig": "vllm.transformers_utils.configs.qwen3_asr",
     "Qwen3NextConfig": "vllm.transformers_utils.configs.qwen3_next",
     "Tarsier2Config": "vllm.transformers_utils.configs.tarsier2",
@@ -95,6 +96,7 @@ __all__ = [
     "Step3VLConfig",
     "Step3VisionEncoderConfig",
     "Step3TextConfig",
+    "Step3p5Config",
     "Qwen3ASRConfig",
     "Qwen3NextConfig",
     "Tarsier2Config",

vllm/transformers_utils/configs/step3p5.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class Step3p5Config(PretrainedConfig):
+    model_type = "step3p5"
+
+    def __init__(
+        self,
+        hidden_size: int = 5120,
+        intermediate_size: int = 13312,
+        num_attention_heads: int = 40,
+        num_attention_groups: int = 8,
+        num_hidden_layers: int = 48,
+        max_seq_len: int = 4096,
+        vocab_size: int = 65536,
+        rms_norm_eps: float = 1e-5,
+        moe_every_n_layer: int = 2,
+        use_moe: bool = False,
+        moe_intermediate_size: int = 10240,
+        moe_num_experts: int = 16,
+        moe_top_k: int = 4,
+        moe_layer_offset: int = 0,
+        rope_theta: float | list[float] | None = 500000,
+        rope_scaling: dict[str, Any] | None = None,
+        head_dim: int | None = None,
+        share_expert_dim: int | None = None,
+        norm_expert_weight: bool = True,
+        bos_token_id: list[int] | int | None = None,
+        eos_token_id: list[int] | int | None = None,
+        moe_router_activation: str = "softmax",
+        moe_router_scaling_factor: float = 1.0,
+        att_impl_type: str = "GQA",
+        use_head_wise_attn_gate: bool = False,
+        use_moe_router_bias: bool = True,
+        need_fp32_gate: bool = True,
+        layer_types: list[str] | None = None,
+        use_rope_layers: list[bool] | None = None,
+        yarn_only_types: list[str] | None = None,
+        attention_other_setting: dict[str, Any] | None = None,
+        num_nextn_predict_layers: int = 0,
+        swiglu_limits: list[float] | None = None,
+        swiglu_limits_shared: list[float] | None = None,
+        max_position_embeddings: int | None = None,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_groups = num_attention_groups
+        self.num_hidden_layers = num_hidden_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.rms_norm_eps = rms_norm_eps
+        self.use_moe = use_moe
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_every_n_layer = moe_every_n_layer
+        self.moe_num_experts = moe_num_experts
+        self.num_experts_per_tok = moe_top_k
+        self.moe_top_k = moe_top_k
+        self.moe_layer_offset = moe_layer_offset
+
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.head_dim = head_dim
+        if share_expert_dim is None:
+            self.share_expert_dim = self.moe_intermediate_size * self.moe_top_k
+        else:
+            self.share_expert_dim = share_expert_dim
+        self.norm_expert_weight = norm_expert_weight
+
+        self.max_position_embeddings = max_position_embeddings
+        self.moe_router_activation = moe_router_activation
+        self.moe_router_scaling_factor = moe_router_scaling_factor
+        self.use_moe_router_bias = use_moe_router_bias
+        self.need_fp32_gate = need_fp32_gate
+
+        self.att_impl_type = att_impl_type
+        self.use_head_wise_attn_gate = use_head_wise_attn_gate
+        self.layer_types = layer_types
+        self.use_rope_layers = use_rope_layers
+        self.yarn_only_types = yarn_only_types
+        self.attention_other_setting = attention_other_setting
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.swiglu_limits = swiglu_limits
+        self.swiglu_limits_shared = swiglu_limits_shared
+
+        resolved_bos_token_id = 1 if bos_token_id is None else bos_token_id
+        resolved_eos_token_id = [2, 3] if eos_token_id is None else eos_token_id
+        self.bos_token_id = resolved_bos_token_id
+        self.eos_token_id = resolved_eos_token_id
+
+        super().__init__(
+            bos_token_id=resolved_bos_token_id,
+            eos_token_id=resolved_eos_token_id,
+            **kwargs,
+        )