[Model] Add support for SenseNova-Loop

docs/CN/source/cookbook/sensenova_loop_deployment.rst

@@ -0,0 +1,60 @@
+.. _sensenova_loop_deployment:
+
+SenseNova Loop 部署指南
+===========================
+
+启动服务
+--------
+
+下面以 4 卡部署为例，在仓库根目录执行：
+
+.. code-block:: bash
+
+    CUDA_VISIBLE_DEVICES=0,1,2,3 LOADWORKER=4 LIGHTLLM_TRITON_AUTOTUNE_LEVEL=1 \
+    python -m lightllm.server.api_server \
+        --model_dir /path/to/SenseNova-Loop \
+        --model_name sensenova_loop \
+        --tp 4 \
+        --port 8000
+
+启用 MTP
+--------
+
+checkpoint 已包含 MTP 权重，增加以下参数即可启用：
+
+.. code-block:: bash
+
+    CUDA_VISIBLE_DEVICES=0,1,2,3 LOADWORKER=4 LIGHTLLM_TRITON_AUTOTUNE_LEVEL=1 \
+    python -m lightllm.server.api_server \
+        --model_dir /path/to/SenseNova-Loop \
+        --model_name sensenova_loop \
+        --mtp_mode eagle_with_att \
+        --mtp_step 3 \
+        --tp 4 \
+        --port 8000
+
+2P1D 部署
+---------
+
+下面的脚本默认使用 8 张 GPU，启动两个 TP2 Prefill 服务和一个 TP4 Decode 服务，并启用 MTP：
+
+.. code-block:: bash
+
+    MODEL_DIR=/path/to/SenseNova-Loop PORT=8000 \
+    bash test/start_scripts/sensenova_loop/sensenova_loop_pd_2p1d.sh
+
+可通过 ``GPU_IDS`` 指定显卡列表，添加 ``DRY_RUN=1`` 可先查看启动命令。
+按 ``Ctrl+C`` 会关闭所有服务。
+
+请求示例
+--------
+
+.. code-block:: bash
+
+    curl http://localhost:8000/v1/chat/completions \
+        -H 'Content-Type: application/json' \
+        -d '{
+          "model": "sensenova_loop",
+          "messages": [{"role": "user", "content": "你好，请用一句话介绍自己。"}],
+          "max_tokens": 128
+        }'

docs/CN/source/index.rst

@@ -67,6 +67,7 @@ Lightllm 整合了众多的开源方案的优点，包括但不限于 FasterTran
 
    GLM-4.7-Flash 部署 <cookbook/glm4_deployment>
    Qwen3.5 部署 <cookbook/qwen35_deployment>
+   SenseNova Loop 部署 <cookbook/sensenova_loop_deployment>
 
 .. toctree::
    :maxdepth: 1

docs/EN/source/cookbook/sensenova_loop_deployment.rst

@@ -0,0 +1,60 @@
+.. _sensenova_loop_deployment:
+
+SenseNova Loop Deployment Guide
+==============================
+
+Start the Server
+----------------
+
+This example starts a BF16 service on four GPUs. Run it from the repository root:
+
+.. code-block:: bash
+
+    CUDA_VISIBLE_DEVICES=0,1,2,3 LOADWORKER=4 LIGHTLLM_TRITON_AUTOTUNE_LEVEL=1 \
+    python -m lightllm.server.api_server \
+        --model_dir /path/to/SenseNova-Loop \
+        --model_name sensenova_loop \
+        --tp 4 \
+        --port 8000
+
+Enable MTP
+----------
+
+The checkpoint includes MTP weights. Add the following options to enable MTP:
+
+.. code-block:: bash
+
+    CUDA_VISIBLE_DEVICES=0,1,2,3 LOADWORKER=4 LIGHTLLM_TRITON_AUTOTUNE_LEVEL=1 \
+    python -m lightllm.server.api_server \
+        --model_dir /path/to/SenseNova-Loop \
+        --model_name sensenova_loop \
+        --mtp_mode eagle_with_att \
+        --mtp_step 3 \
+        --tp 4 \
+        --port 8000
+
+2P1D Deployment
+---------------
+
+This script uses eight GPUs to start two TP2 prefill services and one TP4 decode service, with MTP enabled:
+
+.. code-block:: bash
+
+    MODEL_DIR=/path/to/SenseNova-Loop PORT=8000 \
+    bash test/start_scripts/sensenova_loop/sensenova_loop_pd_2p1d.sh
+
+Set ``GPU_IDS`` to select the GPUs, or add ``DRY_RUN=1`` to preview the commands.
+Press ``Ctrl+C`` to stop all services.
+
+Example Request
+---------------
+
+.. code-block:: bash
+
+    curl http://localhost:8000/v1/chat/completions \
+        -H 'Content-Type: application/json' \
+        -d '{
+          "model": "sensenova_loop",
+          "messages": [{"role": "user", "content": "Hello, introduce yourself in one sentence."}],
+          "max_tokens": 128
+        }'

docs/EN/source/index.rst

@@ -66,6 +66,7 @@ Documentation List
 
    GLM-4.7-Flash Deployment <cookbook/glm4_deployment>
    Qwen3.5 Deployment <cookbook/qwen35_deployment>
+   SenseNova Loop Deployment <cookbook/sensenova_loop_deployment>
 
 .. toctree::
    :maxdepth: 1

lightllm/common/basemodel/triton_kernel/att/decode_att/gqa/mtp_diverse/stage1_single_token.py

@@ -308,7 +308,7 @@ def mtp_diverse_stage1_single_token(
 
     Lq, Lk = int(q.shape[2]), int(k.shape[2])
     assert Lq == Lk
-    assert Lk in {16, 32, 64, 128}
+    assert Lk in {16, 32, 64, 128, 256}
     batch = int(B_req_idx.shape[0])
     kv_head_num = int(k.shape[1])
     q_head_num = int(q.shape[1])

lightllm/common/basemodel/triton_kernel/att/decode_att/gqa/mtp_diverse/stage2_single_token.py

@@ -164,7 +164,7 @@ def mtp_diverse_stage2_single_token(
     num_stages = run_config["num_stages"]
 
     Lk = mid_out.shape[-1]
-    assert Lk in {16, 32, 64, 128}
+    assert Lk in {16, 32, 64, 128, 256}
     batch, head_num = mid_out.shape[0], mid_out.shape[1]
     grid = (batch, head_num)
 

lightllm/common/req_manager/linear_att.py

@@ -90,10 +90,7 @@ def save_state(self, req: "InferReq", buffer_idx: int, state_cache_manager: Line
         dst_ssm_state.copy_(gpu_ssm_state, non_blocking=True)
 
     def get_mamba_cache(self, layer_idx_in_all: int):
-        assert (
-            0 <= layer_idx_in_all < self.linear_config.all_layer_num
-        ), f"invalid transformer layer index {layer_idx_in_all}"
-        layer_idx_in_linear = layer_idx_in_all - (layer_idx_in_all // self.linear_config.full_attention_interval)
+        layer_idx_in_linear = self.linear_config.get_linear_state_layer_index(layer_idx_in_all)
         conv_states = self.req_to_conv_state.buffer[layer_idx_in_linear]
         ssm_states = self.req_to_ssm_state.buffer[layer_idx_in_linear]
         return conv_states, ssm_states

lightllm/common/state_cache_manager/__init__.py

@@ -17,6 +17,6 @@ def get_hybrid_cache_config():
         return DeepseekV4CacheConfig.from_model_config(model_cfg, args)
     if model_type in ("glm5_next", "glm5_next_text"):
         return Glm5NextCacheConfig.from_model_config(model_cfg, args)
-    if model_type in ("qwen3_5", "qwen3_5_moe", "qwen3_5_text", "qwen3_5_moe_text"):
+    if model_type in ("qwen3_5", "sensenova_loop", "qwen3_5_moe", "qwen3_5_text", "qwen3_5_moe_text"):
         return LinearAttCacheConfig.load_from_args()
     raise ValueError("No hybrid state-cache layout registered for this model")

lightllm/common/state_cache_manager/glm5_next.py

@@ -21,10 +21,6 @@ def from_model_config(cls, config, args):
         linear = config["linear_attn_config"]
         tp = args.tp // args.dp
         assert linear["num_heads"] % tp == 0
-        layers = config["num_hidden_layers"]
-        assert config["layer_types"] == [
-            "deepseek_sparse_attention" if i % 4 == 3 else "linear_attention" for i in range(layers)
-        ]
         dtype = get_torch_dtype(args.data_type)
         # Raw index keys and compression scores only live in request tails.
         packed_dim = config["kv_lora_rank"] + cls.INDEX_PADDING_BYTES // dtype.itemsize
@@ -41,11 +37,9 @@ def from_model_config(cls, config, args):
             head_linear_k_dim=linear["head_dim"],
             head_linear_v_dim=linear["head_dim"],
             conv_kernel_size=linear["short_conv_kernel_size"],
-            linear_layer_num=len(linear["kda_layers"]),
             conv_state_dtype=dtype,
             ssm_state_dtype=get_torch_dtype(args.linear_att_ssm_data_type),
-            full_attention_interval=4,
-            all_layer_num=layers,
+            layer_types=config["layer_types"],
             draft_full_att_kv_layer_num=get_added_mtp_kv_layer_num() if args.mtp_mode is not None else 0,
             index_kpool=config["index_kpool"],
             index_head_dim=config["index_head_dim"],

lightllm/common/state_cache_manager/linear_att.py

@@ -30,12 +30,32 @@ class LinearAttCacheConfig:
     head_linear_k_dim: int
     head_linear_v_dim: int
     conv_kernel_size: int
-    linear_layer_num: int
     conv_state_dtype: torch.dtype
     ssm_state_dtype: torch.dtype
-    full_attention_interval: int
-    all_layer_num: int  # 包括 linear att 和 full att 的层加起来的层数
+    layer_types: tuple[str, ...]
     draft_full_att_kv_layer_num: int = 0
+    linear_layer_num: int = dataclasses.field(init=False)
+    all_layer_num: int = dataclasses.field(init=False)
+
+    def __post_init__(self):
+        self.layer_types = tuple(self.layer_types)
+        indices = []
+        full_count = linear_count = 0
+        for kind in self.layer_types:
+            if kind == "linear_attention":
+                indices.append(linear_count)
+                linear_count += 1
+            else:
+                indices.append(full_count)
+                full_count += 1
+        # MTP draft layers only use full-attention KV slots after the main model.
+        indices.extend(range(full_count, full_count + self.draft_full_att_kv_layer_num))
+        self._layer_indices = tuple(indices)
+        self.linear_layer_num = linear_count
+        self.all_layer_num = len(self.layer_types)
+
+    def get_linear_state_layer_index(self, layer_index: int) -> int:
+        return self._layer_indices[layer_index]
 
     def get_conv_dim(self):
         # 第一项对应q的参数，第二项对应k的参数，第三项对应v的参数
@@ -49,7 +69,6 @@ def get_conv_dim(self):
 
     def get_main_model_full_att_layer_num(self):
         full_att_layer_num = self.all_layer_num - self.linear_layer_num
-        assert full_att_layer_num == self.all_layer_num // self.full_attention_interval
         return full_att_layer_num
 
     def get_full_att_kv_layer_num_with_draft_model(self):
@@ -58,16 +77,7 @@ def get_full_att_kv_layer_num_with_draft_model(self):
     def get_full_att_kv_layer_index(self, layer_index: int) -> int:
         """Map a global target/draft layer index to the packed full-attention cache."""
 
-        layer_index = int(layer_index)
-        if layer_index >= self.all_layer_num:
-            kv_layer_index = layer_index - self.linear_layer_num
-        else:
-            kv_layer_index = layer_index // self.full_attention_interval
-        assert 0 <= kv_layer_index < self.get_full_att_kv_layer_num_with_draft_model(), (
-            f"layer {layer_index} maps outside the packed full-attention KV cache: "
-            f"slot={kv_layer_index}, slots={self.get_full_att_kv_layer_num_with_draft_model()}"
-        )
-        return kv_layer_index
+        return self._layer_indices[layer_index]
 
     def get_conv_state_shape(self):
         # Base committed sliding-window state, without speculative MTP tail.
@@ -117,17 +127,13 @@ def load_from_args() -> "LinearAttCacheConfig":
         from transformers.configuration_utils import PretrainedConfig
 
         model_cfg, _ = PretrainedConfig.get_config_dict(model_path)
-        model_type = model_cfg["model_type"]
-        assert model_type in ["qwen3_5", "qwen3_5_moe", "qwen3_5_text", "qwen3_5_moe_text"]
-        llm_config = model_cfg
-        try:
-            llm_config = llm_config["text_config"]
-        except:
-            pass
-
-        n_layer = llm_config["num_hidden_layers"]
+        llm_config = model_cfg.get("text_config", model_cfg)
+        layer_types = llm_config["layer_types"]
+        if model_cfg["model_type"] == "sensenova_loop":
+            from lightllm.models.sensenova_loop.config import get_loop_layer_types
 
-        tp_world_size = get_env_start_args().tp // get_env_start_args().dp
+            layer_types = get_loop_layer_types(model_cfg["loop_settings"], llm_config)
+        tp_world_size = args.tp // args.dp
         full_att_dtype = (
             torch.uint8 if args.llm_kv_type in {"fp8kv_sph", "fp8kv_spt"} else get_torch_dtype(args.data_type)
         )
@@ -144,11 +150,9 @@ def load_from_args() -> "LinearAttCacheConfig":
             head_linear_k_dim=llm_config["linear_key_head_dim"],
             head_linear_v_dim=llm_config["linear_value_head_dim"],
             conv_kernel_size=llm_config["linear_conv_kernel_dim"],
-            linear_layer_num=n_layer - (n_layer // llm_config["full_attention_interval"]),
             conv_state_dtype=get_torch_dtype(args.data_type),
             ssm_state_dtype=get_torch_dtype(args.linear_att_ssm_data_type),
-            full_attention_interval=llm_config["full_attention_interval"],
-            all_layer_num=n_layer,
+            layer_types=layer_types,
             draft_full_att_kv_layer_num=get_added_mtp_kv_layer_num(),
         )
 

lightllm/models/builtin.py

@@ -114,6 +114,9 @@ def _tarsier_text_model_is(model_type):
 )
 ModelRegistry.register("qwen3", "lightllm.models.qwen3.model:Qwen3TpPartModel")
 ModelRegistry.register(["qwen3_5"], "lightllm.models.qwen3_5.model:Qwen3_5TpPartModel", is_multimodal=True)
+ModelRegistry.register(
+    "sensenova_loop", "lightllm.models.sensenova_loop.model:SenseNovaLoopTpPartModel", is_multimodal=True
+)
 ModelRegistry.register("qwen3_5_moe", "lightllm.models.qwen3_5_moe.model:Qwen3_5MOETpPartModel", is_multimodal=True)
 ModelRegistry.register("qwen3_moe", "lightllm.models.qwen3_moe.model:Qwen3MOEModel")
 ModelRegistry.register(
@@ -186,7 +189,7 @@ def _tarsier_text_model_is(model_type):
     "lightllm.models.qwen3_5_moe_mtp.model:Qwen3_5MoeMTPModel",
 )
 DraftModelRegistry.register(
-    ("qwen3_5", "qwen3_5_text"),
+    ("qwen3_5", "qwen3_5_text", "sensenova_loop"),
     ("vanilla_with_att", "eagle_with_att"),
     "lightllm.models.qwen3_5_mtp.model:Qwen3_5MTPModel",
 )

lightllm/models/qwen3next/model.py

@@ -11,7 +11,7 @@
 )
 from lightllm.models.qwen3next.infer_struct import Qwen3NextInferStateInfo
 from lightllm.utils.log_utils import init_logger
-from lightllm.utils.envs_utils import get_added_mtp_kv_layer_num, get_env_start_args
+from lightllm.utils.envs_utils import get_env_start_args
 from lightllm.common.kv_cache_mem_manager.qwen3next_mem_manager import (
     FP8StaticPerHeadQuantQwen3NextMemManager,
     FP8StaticPerTensorQuantQwen3NextMemManager,
@@ -25,7 +25,6 @@
 
 
 class Qwen3NextTpPartModel(Qwen3MOEModel):
-
     # weight class
     pre_and_post_weight_class = Qwen3NextPreAndPostLayerWeight
     transformer_weight_class = Qwen3NextTransformerLayerWeight
@@ -60,31 +59,6 @@ def _init_config(self):
     def _init_mem_manager(self):
         assert self.config["num_attention_heads"] % self.tp_world_size_ == 0
         start_args: StartArgs = get_env_start_args()
-        ssm_dtype_dict = {"bfloat16": torch.bfloat16, "float32": torch.float32}
-        draft_full_att_kv_layer_num = get_added_mtp_kv_layer_num()
-        full_att_dtype = torch.uint8 if start_args.llm_kv_type in {"fp8kv_sph", "fp8kv_spt"} else self.data_type
-        self.linear_config = LinearAttCacheConfig(
-            tp_world_size=self.tp_world_size_,
-            full_att_all_num_kv_heads=self.config["num_key_value_heads"],
-            full_att_dtype=full_att_dtype,
-            full_att_num_kv_heads=self.num_kv_heads,
-            full_att_head_dim=self.config["head_dim"],
-            global_linear_k_heads=self.config["linear_num_key_heads"],
-            global_linear_v_heads=self.config["linear_num_value_heads"],
-            num_linear_k_heads=max(1, self.config["linear_num_key_heads"] // self.tp_world_size_),
-            num_linear_v_heads=max(1, self.config["linear_num_value_heads"] // self.tp_world_size_),
-            head_linear_k_dim=self.config["linear_key_head_dim"],
-            head_linear_v_dim=self.config["linear_value_head_dim"],
-            conv_kernel_size=self.config["linear_conv_kernel_dim"],
-            linear_layer_num=self.config["n_layer"]
-            - (self.config["n_layer"] // self.config["full_attention_interval"]),
-            conv_state_dtype=self.data_type,
-            ssm_state_dtype=ssm_dtype_dict[start_args.linear_att_ssm_data_type],
-            full_attention_interval=self.config["full_attention_interval"],
-            all_layer_num=self.config["n_layer"],
-            draft_full_att_kv_layer_num=draft_full_att_kv_layer_num,
-        )
-
         if start_args.llm_kv_type == "None":
             mem_manager_class = Qwen3NextMemManager
         elif start_args.llm_kv_type == "fp8kv_sph":
@@ -104,6 +78,7 @@ def _init_mem_manager(self):
         )
 
     def _init_req_manager(self):
+        self.linear_config = LinearAttCacheConfig.load_from_args()
         create_max_seq_len = 0
 
         if self.batch_max_tokens is not None:
@@ -112,7 +87,7 @@ def _init_req_manager(self):
             create_max_seq_len = max(create_max_seq_len, self.max_seq_length)
 
         self.req_manager = ReqManagerForMamba(
-            self.max_req_num, create_max_seq_len, None, linear_config=LinearAttCacheConfig.load_from_args()
+            self.max_req_num, create_max_seq_len, None, linear_config=self.linear_config
         )
         return
 

lightllm/models/sensenova_loop/config.py

@@ -0,0 +1,14 @@
+"""SenseNova Loop execution layout."""
+
+
+def get_loop_layer_indices(settings: dict, num_layers: int) -> tuple[int, ...]:
+    """Map execution depth to checkpoint layers; caches use execution depth."""
+    start, end, count = settings["start"], settings["end"], settings["eval_count"]
+    return tuple(range(start)) + tuple(range(start, end)) * count + tuple(range(end, num_layers))
+
+
+def get_loop_layer_types(settings: dict, text_config: dict) -> tuple[str, ...]:
+    num_layers = text_config["num_hidden_layers"]
+    indices = get_loop_layer_indices(settings, num_layers)
+    layer_types = text_config["layer_types"]
+    return tuple(layer_types[i] for i in indices)