[feat][Ascend][Mindspore]: support model-impl of mindspore (#9234)

docs/references/mindspore_models.md

@@ -0,0 +1,162 @@
+# MindSpore Models
+
+## Introduction
+
+SGLang support run MindSpore framework models, this doc guide users to run mindspore models with SGLang.
+
+## Requirements
+
+MindSpore with SGLang current only support Ascend Npu device, users need first install Ascend CANN software packages.
+The CANN software packages can download from the [Ascend Official Websites](https://www.hiascend.com). The version depends on the MindSpore version [MindSpore Installation](https://www.mindspore.cn/install)
+
+## Supported Models
+
+Currently, the following models are supported:
+
+- **Qwen3**: Dense models supported. MoE models coming soon.
+- *More models coming soon...*
+
+## Installation
+
+> **Note**: Currently, MindSpore models are provided by an independent package `sgl-mindspore`, which needs to be installed separately.
+
+```shell
+git clone https://github.com/chz34/sgl-mindspore.git
+cd sgl-mindspore
+pip install -e .
+```
+
+You will need to install the following packages, due to the support of tensor conversion through `dlpack` on 3rd devices, the minimum version of  `PyTorch` is 2.7.1
+
+```shell
+pip install mindspore
+pip install "torch>=2.7.1"
+pip install "torch_npu>=2.7.1"
+pip install triton_ascend
+```
+
+```shell
+pip install -e "python[all_npu]"
+```
+
+## Run Model
+
+Current SGLang-MindSpore support Qwen3 dense model, this doc uses Qwen3-8B as example.
+
+### Offline infer
+
+Use the following script for offline infer:
+
+```python
+import sglang as sgl
+
+# Initialize the engine with MindSpore backend
+llm = sgl.Engine(
+    model_path="/path/to/your/model",  # Local model path
+    device="npu",                      # Use NPU device
+    model_impl="mindspore",            # MindSpore implementation
+    attention_backend="ascend",        # Attention backend
+    tp_size=1,                         # Tensor parallelism size
+    dp_size=1                          # Data parallelism size
+)
+
+# Generate text
+prompts = [
+    "Hello, my name is",
+    "The capital of France is",
+    "The future of AI is"
+]
+
+sampling_params = {"temperature": 0.01, "top_p": 0.9}
+outputs = llm.generate(prompts, sampling_params)
+
+for prompt, output in zip(prompts, outputs):
+    print(f"Prompt: {prompt}")
+    print(f"Generated: {output['text']}")
+    print("---")
+```
+
+### Start server
+
+Launch a server with MindSpore backend:
+
+```bash
+# Basic server startup
+python3 -m sglang.launch_server \
+    --model-path /path/to/your/model \
+    --host 0.0.0.0 \
+    --device npu \
+    --model-impl mindspore \
+    --attention-backend ascend \
+    --tp-size 1 \
+    --dp-size 1
+```
+
+For distributed server with multiple nodes:
+
+```bash
+# Multi-node distributed server
+python3 -m sglang.launch_server \
+    --model-path /path/to/your/model \
+    --host 0.0.0.0 \
+    --device npu \
+    --model-impl mindspore \
+    --attention-backend ascend \
+    --dist-init-addr 127.0.0.1:29500 \
+    --nnodes 2 \
+    --node-rank 0 \
+    --tp-size 4 \
+    --dp-size 2
+```
+
+## Troubleshooting
+
+#### Debug Mode
+
+Enable sglang debug logging by log-level argument.
+
+```bash
+python3 -m sglang.launch_server \
+    --model-path /path/to/your/model \
+    --host 0.0.0.0 \
+    --device npu \
+    --model-impl mindspore \
+    --attention-backend ascend \
+    --log-level DEBUG
+```
+
+Enable mindspore info and debug logging by setting environments.
+
+```bash
+export GLOG_v=1  # INFO
+export GLOG_v=0  # DEBUG
+```
+
+#### Explicitly select devices
+
+Use the following environment variable to explicitly select the devices to use.
+
+```shell
+export ASCEND_RT_VISIBLE_DEVICES=4,5,6,7  # to set device
+```
+
+#### Some communication environment issues
+
+In case of some environment with special communication environment, users need set some environment variables.
+
+```shell
+export MS_ENABLE_LCCL=off # current not support LCCL communication mode in SGLang-MindSpore
+```
+
+#### Some dependencies of protobuf
+
+In case of some environment with special protobuf version, users need set some environment variables to avoid binary version mismatch.
+
+```shell
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python  # to avoid protobuf binary version mismatch
+```
+
+## Support
+For MindSpore-specific issues:
+
+- Refer to the [MindSpore documentation](https://www.mindspore.cn/)

python/sglang/srt/configs/model_config.py

@@ -47,6 +47,7 @@ class ModelImpl(str, Enum):
     AUTO = "auto"
     SGLANG = "sglang"
     TRANSFORMERS = "transformers"
+    MINDSPORE = "mindspore"
 
 
 def is_deepseek_nsa(config: PretrainedConfig) -> bool:

python/sglang/srt/model_executor/mindspore_runner.py

@@ -0,0 +1,118 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the SGLang project
+"""ms_runner launch MindSpore distributed modules."""
+
+import multiprocessing as mp
+import os
+import sys
+from pathlib import Path
+
+import mindspore as ms
+import torch
+from mindspore._c_expression import GroupOptions
+from mindspore.communication import create_group
+
+from sglang.srt.distributed.parallel_state import _groups
+
+
+class _Tmp:
+    def __init__(self):
+        self.sched_p = None
+
+    def set_sched_process(self, p):
+        self.sched_p = p
+
+    def __del__(self):
+        if self.sched_p:
+            self.sched_p.kill()
+
+
+_tmp = _Tmp()
+
+
+def _get_host_and_ip(distributed_init_method):
+    try:
+        _, ip_str, port_str = distributed_init_method.split(":")
+        ip = ip_str.split("/")[-1]
+        port = int(port_str)
+    except Exception as e:
+        raise RuntimeError(
+            "Cannot get host and port information from %s, error: %s!"
+            % (distributed_init_method, str(e))
+        )
+
+    return ip, port
+
+
+def run_scheduler_init(rank, local_rank, world_size, master_addr, master_port):
+    with open(str(Path() / "schedule.log"), "w") as scheduler_f:
+        # For Python outputs.
+        sys.stdout = scheduler_f
+        sys.stderr = scheduler_f
+        # For C++ outputs.
+        os.dup2(scheduler_f.fileno(), 1)
+        os.dup2(scheduler_f.fileno(), 2)
+        os.environ["DEVICE_ID"] = str(local_rank)
+        os.environ["MS_WORKER_NUM"] = str(world_size)
+        os.environ["MS_ROLE"] = "MS_SCHED"
+        os.environ["MS_NODE_ID"] = str(rank)
+        os.environ["MS_SCHED_HOST"] = str(master_addr)
+        os.environ["MS_SCHED_PORT"] = str(master_port)
+        # This function is blocked until the whole cluster exits.
+        ms.communication.init()
+
+
+def set_ms_parallel_env(rank, local_rank, world_size, init_method):
+    master_addr, master_port = _get_host_and_ip(init_method)
+    # change port avoiding port conflicts with torch
+    master_port = master_port + 35 if master_port < 65500 else master_port - 35
+    if not os.getenv("MS_ROLE"):
+        if rank == 0:
+            # Create a subprocess for scheduler of MindSpore, just for internal collaboration, not for collective communication
+            sched_p = mp.Process(
+                target=run_scheduler_init,
+                args=(rank, local_rank, world_size, master_addr, master_port),
+            )
+            sched_p.start()
+            global _tmp
+            _tmp.set_sched_process(sched_p)
+
+        os.environ["DEVICE_ID"] = str(local_rank)
+        os.environ["MS_WORKER_NUM"] = str(world_size)
+        os.environ["MS_ROLE"] = "MS_WORKER"
+        os.environ["MS_NODE_ID"] = str(rank)
+        os.environ["MS_SCHED_HOST"] = str(master_addr)
+        os.environ["MS_SCHED_PORT"] = str(master_port)
+
+
+def reuse_hccl_comm():
+    for group_name, group in _groups.items():
+        # Torch ProcessGroupHccl
+        device_group = group().device_group
+        hccl_comm_handle = device_group._get_backend(torch.device("npu")).get_hccl_comm(
+            group().local_rank
+        )
+        print(
+            f"MindSpore reuse torch group: {device_group}, group_name: {group_name}, local rank: {group().local_rank},"
+            f"hccl communicator handle: {hex(hccl_comm_handle)}",
+            flush=True,
+        )
+        # Create MS communication group by hccl comm handle to reuse Torch group.
+        group_options = GroupOptions()
+        group_options.hccl_config = {"hccl_comm": hccl_comm_handle}
+        create_group(group_name, group().ranks, group_options)
+
+
+def init_ms_distributed(world_size, rank, local_rank, server_args, port):
+    if server_args.dist_init_addr:
+        dist_init_method = f"tcp://{server_args.dist_init_addr}"
+    else:
+        dist_init_method = f"tcp://{server_args.host}:{port}"
+    set_ms_parallel_env(rank, local_rank, world_size, dist_init_method)
+
+    ms.set_context(infer_boost="on", jit_level="O0")
+    ms.set_context(mode=ms.context.PYNATIVE_MODE)
+    ms.set_device("Ascend", local_rank)
+    ms.communication.init("hccl")
+    # After distributed job is initialized, reuse hccl comms for MindSpore.
+    reuse_hccl_comm()

python/sglang/srt/model_executor/model_runner.py

@@ -42,6 +42,7 @@ from sglang.srt.configs.load_config import LoadConfig, LoadFormat
 from sglang.srt.configs.model_config import (
     AttentionArch,
     ModelConfig,
+    ModelImpl,
     get_nsa_index_head_dim,
     is_deepseek_nsa,
 )
@@ -317,6 +318,8 @@ class ModelRunner:
 
         if get_bool_env_var("SGLANG_DETECT_SLOW_RANK"):
             slow_rank_detector.execute()
+        # Init mindspore running environment when model impl is "mindspore"
+        self.init_mindspore_runner()
 
         # Update deep gemm configure
         if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
@@ -364,6 +367,20 @@ class ModelRunner:
         else:
             self.piecewise_cuda_graph_runner = None
 
+    def init_mindspore_runner(self):
+        # Init the mindspore runner
+        # for now, there is only some communication initialization work
+        if self.server_args.model_impl.lower() == ModelImpl.MINDSPORE and _is_npu:
+            from sglang.srt.model_executor.mindspore_runner import init_ms_distributed
+
+            init_ms_distributed(
+                world_size=self.tp_size * self.pp_size,
+                rank=self.tp_size * self.pp_rank + self.tp_rank,
+                local_rank=self.gpu_id,
+                server_args=self.server_args,
+                port=self.dist_port,
+            )
+
     def initialize(self, min_per_gpu_memory: float):
         server_args = self.server_args
 
@@ -2018,6 +2035,9 @@ class ModelRunner:
             # TODO: Currently, cuda graph only captures decode steps, which only exists for generation models
             return
 
+        if self.server_args.model_impl.lower() == ModelImpl.MINDSPORE:
+            return
+
         if self.device != "cpu" and self.server_args.disable_cuda_graph:
             return
 

python/sglang/srt/model_loader/utils.py

@@ -99,7 +99,9 @@ def get_model_architecture(model_config: ModelConfig) -> Tuple[Type[nn.Module],
     supported_archs = ModelRegistry.get_supported_archs()
     is_native_supported = any(arch in supported_archs for arch in architectures)
 
-    if not is_native_supported or model_config.model_impl == ModelImpl.TRANSFORMERS:
+    if model_config.model_impl == ModelImpl.MINDSPORE:
+        architectures = ["MindSporeForCausalLM"]
+    elif not is_native_supported or model_config.model_impl == ModelImpl.TRANSFORMERS:
         architectures = resolve_transformers_arch(model_config, architectures)
     return ModelRegistry.resolve_model_cls(architectures)
 

python/sglang/srt/models/mindspore.py

@@ -0,0 +1,305 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the SGLang project
+import logging
+from typing import Any, Iterable, Optional, Tuple
+
+import torch
+
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from sglang.srt.layers.logits_processor import LogitsProcessorOutput
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.models.registry import import_model_classes
+from sglang.srt.utils import is_npu
+
+_is_npu = is_npu()
+
+if _is_npu:
+    import mindspore as ms
+    import numpy as np
+    import torch_npu
+    from mindspore import Tensor, mint, mutable
+
+logger = logging.getLogger(__name__)
+
+
+def tensor_torch2ms(x: torch.Tensor):
+    if x is None or not isinstance(x, torch.Tensor):
+        return x
+
+    # torch tensor -> dlpack -> mindspore tensor
+    pt_dlpack = torch.utils.dlpack.to_dlpack(x)
+    ms_tensor = ms.utils.dlpack.from_dlpack(pt_dlpack)
+    return ms_tensor
+
+
+def tensor_ms2torch(x: ms.Tensor):
+    if x is None or not isinstance(x, ms.Tensor):
+        return x
+
+    # ms tensor -> dlpack -> torch tensor
+    ms_dlpack = ms.utils.dlpack.to_dlpack(x)
+    torch_tensor = torch.utils.dlpack.from_dlpack(ms_dlpack)
+    torch_npu.npu.synchronize()
+    return torch_tensor
+
+
+# Adapt from: https://gitee.com/mindspore/vllm-mindspore/blob/master/vllm_mindspore/model_executor/models/attention_mask.py
+class LowerTriangularMask:
+    r"""
+    Provide Infer model attention mask.
+    Args:
+        dtype (ms dtype): The compute type of Infer model.
+        max_model_len (int): The max model length of Infer model.
+    """
+
+    def __init__(self, dtype, max_model_len, decode_mask_coeff=-10000.0):
+        self.dtype = dtype
+        self.max_model_len = max_model_len
+        self.cached_mask_len = 8 * 1024
+        self.decode_mask_coeff = decode_mask_coeff
+
+        prefill_mask_coeff = 1.0 if self.dtype == ms.bfloat16 else -10000.0
+        self.prefill_mask = Tensor(
+            np.triu(np.ones(shape=(128, 128), dtype=np.float16), k=1)
+            * prefill_mask_coeff,
+            dtype=self.dtype,
+        )
+
+        self.hard_mask = mint.zeros((1, 1), dtype=dtype)
+        self.decode_mask = (
+            Tensor(
+                np.triu(
+                    np.ones(
+                        shape=(self.cached_mask_len, self.cached_mask_len),
+                        dtype=np.int8,
+                    ),
+                    k=1,
+                ),
+                dtype=self.dtype,
+            )
+            * self.decode_mask_coeff
+        )
+
+    def create_mask(self, query_lens_np, seq_lens_np):
+        """
+        when query_lens_np = [3], seq_lens_np = [6], decode_mask_coeff = 1
+        init attention mask
+        0 0 0 0 0 0
+        0 0 0 0 0 0
+        0 0 0 0 0 0
+        """
+        max_seq_len = seq_lens_np.max().item()
+        total_q_len = query_lens_np.sum().item()
+        attention_mask = mint.zeros((total_q_len, max_seq_len), dtype=self.dtype)
+
+        req_num = query_lens_np.shape[0]
+        current_row = 0
+        for i in range(req_num):
+            q_len = query_lens_np[i].item()
+            current_row += q_len
+            # skip row when q_len <= 1, to decrease execute time
+            if q_len <= 1:
+                continue
+            seq_len = seq_lens_np[i].item()
+            context_len = seq_len - q_len
+            """
+            set the right half to 1
+            0 0 0 1 1 1
+            0 0 0 1 1 1
+            0 0 0 1 1 1
+            """
+            attention_mask[current_row - q_len : current_row, context_len:] = (
+                self.decode_mask_coeff
+            )
+            """
+            set the lower triangle of the right half to 0
+            0 0 0 0 1 1
+            0 0 0 0 0 1
+            0 0 0 0 0 0
+            """
+            right_tensor = attention_mask[
+                current_row - q_len : current_row, context_len:seq_len
+            ]
+
+            # use masked_fill_ to inplace modify attention_mask
+            right_tensor.masked_fill_(right_tensor.tril() == self.decode_mask_coeff, 0)
+
+        return attention_mask
+
+    def gen_attention_mask(
+        self,
+        is_prefill: bool,
+        position_ids: Tensor,
+        query_lens_np: np.ndarray,
+        seq_lens_np: np.ndarray,
+    ):
+        max_query_len = query_lens_np.max()
+        max_seq_len = seq_lens_np.max()
+        if is_prefill:
+            attention_mask = self.prefill_mask
+        elif max_query_len > 1:
+            if max_seq_len <= self.cached_mask_len:
+                attention_mask = mint.index_select(self.decode_mask, 0, position_ids)
+            else:
+                attention_mask = self.create_mask(query_lens_np, seq_lens_np)
+        else:
+            attention_mask = self.hard_mask
+        return attention_mask
+
+
+class MindSporeForCausalLM(torch.nn.Module):
+    def __init__(
+        self,
+        config: Any,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+
+        ms.set_context(graph_kernel_flags="--disable_pass=gather_pre_rms_norm_fusion")
+        ms.set_kernel_launch_capture(False)
+
+        logger.info(
+            "MindSporeForCausalLM tp size %d tp rank %d",
+            get_tensor_model_parallel_world_size(),
+            get_tensor_model_parallel_rank(),
+        )
+        if get_tensor_model_parallel_world_size() not in (1, 2, 4, 8):
+            # MatMulAllReduce only support tp size in (1, 2, 4, 8)
+            ms.set_context(graph_kernel_flags="--disable_pass=MatMulAllReduce")
+
+        arch = self.get_arch(self.config)
+        self.model = arch(config=config, quant_config=quant_config)
+
+        self.casual_mask = LowerTriangularMask(
+            self.config.param_dtype, self.config.max_position_embeddings
+        )
+        self.key_cache = []
+        self.value_cache = []
+
+    def get_arch(self, config):
+        # Get all implemented models
+        mindspore_models = import_model_classes("sgl_mindspore.models")
+
+        # Get arch from config
+        architectures = config.architectures
+        if isinstance(architectures, str):
+            architectures = [architectures]
+        if not architectures:
+            logger.warning("No model architectures are specified")
+
+        for arch in architectures:
+            if arch in mindspore_models:
+                return mindspore_models[arch]
+        if arch is None:
+            raise ValueError(f"Unsupported arch {architectures}")
+
+    @property
+    def use_mla(self):
+        return self.config.architectures[0] in ("DeepseekV3ForCausalLM")
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        self.model.load_weights(weights)
+        for _, cell in self.model.cells_and_names():
+            quant_method = getattr(cell, "quant_method", None)
+            if quant_method is not None:
+                quant_method.process_weights_after_loading(cell)
+
+    def get_kvcache(self, forward_batch: ForwardBatch):
+        def prepare_cache(cache_list, is_key_cache):
+            for i in range(self.config.num_hidden_layers):
+                if is_key_cache:
+                    cache = forward_batch.token_to_kv_pool.get_key_buffer(i)
+                else:
+                    cache = forward_batch.token_to_kv_pool.get_value_buffer(i)
+                cache_ms = tensor_torch2ms(cache)
+                if cache_ms.ndim == 3:
+                    cache_ms = mint.unsqueeze(cache_ms, 2)
+                cache_list.append(cache_ms)
+
+        if self.use_mla:
+            if not self.key_cache:
+                prepare_cache(self.key_cache, is_key_cache=True)
+            return mutable(self.key_cache)
+
+        if self.key_cache and self.value_cache:
+            return mutable(self.key_cache), mutable(self.value_cache)
+
+        prepare_cache(self.key_cache, is_key_cache=True)
+        prepare_cache(self.value_cache, is_key_cache=False)
+
+        return mutable(self.key_cache), mutable(self.value_cache)
+
+    def prepare_inputs(self, input_ids, positions, forward_batch):
+        if self.use_mla:
+            key_cache = self.get_kvcache(forward_batch)
+        else:
+            key_cache, value_cache = self.get_kvcache(forward_batch)
+
+        # Different processing for the mindspore attention operator
+        # Without any prefix cache => Use FlashAttentionScore
+        # With cache => Use PagedAttention, no matter the query length is 1 or not
+        is_prefill = forward_batch.forward_mode.is_extend()
+        is_prefill = is_prefill and forward_batch.extend_prefix_lens.sum().item() == 0
+
+        batch_valid_length = forward_batch.seq_lens.cpu().numpy()
+
+        if forward_batch.extend_seq_lens is not None:
+            q_seq_lens = forward_batch.extend_seq_lens.cpu().numpy()
+        else:
+            q_seq_lens = np.ones([forward_batch.batch_size], dtype=np.int32)
+
+        page_size = forward_batch.token_to_kv_pool.page_size
+        block_tables = tensor_torch2ms(
+            (
+                forward_batch.req_to_token_pool.req_to_token[
+                    forward_batch.req_pool_indices, : forward_batch.seq_lens.max()
+                ][:, ::page_size]
+                // page_size
+            )
+        ).to(ms.int32)
+
+        model_inputs = {}
+        model_inputs["input_ids"] = tensor_torch2ms(input_ids).to(ms.int32)
+        model_inputs["batch_valid_length"] = ms.Tensor(
+            batch_valid_length, dtype=ms.int32
+        )
+        model_inputs["position_ids"] = tensor_torch2ms(positions)
+        model_inputs["q_seq_lens"] = ms.Tensor(q_seq_lens, dtype=ms.int32)
+        model_inputs["attention_mask"] = self.casual_mask.gen_attention_mask(
+            is_prefill, model_inputs["position_ids"], q_seq_lens, batch_valid_length
+        ).contiguous()
+        model_inputs["out_cache_loc"] = tensor_torch2ms(forward_batch.out_cache_loc).to(
+            ms.int32
+        )
+        model_inputs["is_prefill"] = is_prefill
+        model_inputs["key_cache"] = key_cache
+        if not self.use_mla:
+            model_inputs["value_cache"] = value_cache
+        model_inputs["block_tables"] = block_tables
+        return model_inputs
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> Tensor:
+        # prepare base inputs
+        model_inputs = self.prepare_inputs(input_ids, positions, forward_batch)
+        # prepare model inputs
+        model_inputs = self.model.prepare_inputs(forward_batch, model_inputs)
+
+        logits = self.model(**model_inputs)
+
+        # TODO: npu tensor ms2torch error to be fix, remain issues of torch_npu to get tensor from dlpack
+        logits_result = LogitsProcessorOutput(next_token_logits=tensor_ms2torch(logits))
+        return logits_result
+
+
+EntryClass = [MindSporeForCausalLM]

python/sglang/srt/server_args.py

@@ -2059,6 +2059,7 @@ class ServerArgs:
             "implementation is available.\n"
             '* "sglang" will use the SGLang model implementation.\n'
             '* "transformers" will use the Transformers model '
+            '* "mindspore" will use the MindSpore model '
             "implementation.\n",
         )
 
@@ -3957,6 +3958,9 @@ class ServerArgs:
             self.schedule_conservativeness >= 0
         ), "schedule_conservativeness must be non-negative"
 
+        if self.model_impl == "mindspore":
+            assert is_npu(), "MindSpore model impl is only supported on Ascend npu."
+
     def check_lora_server_args(self):
         assert self.max_loras_per_batch > 0, "max_loras_per_batch must be positive"
 

test/srt/ascend/test_mindspore_models.py

@@ -0,0 +1,64 @@
+"""
+Usage:
+python3 -m unittest test_mindspore_models.TestMindSporeQwen3.test_gsm8k
+"""
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+
+class TestMindSporeQwen3(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-8B"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--device",
+                "npu",
+                "--model-impl",
+                "mindspore",
+                "--attention-backend",
+                "ascend",
+                "--tp-size",
+                "1",
+                "--dp-size",
+                "1",
+                "--mem-fraction-static",
+                0.8,
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=200,
+            max_new_tokens=512,
+            parallel=128,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreater(metrics["accuracy"], 0.78)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/run_suite.py

@@ -536,6 +536,9 @@ suite_ascend = {
         TestFile("ascend/test_ascend_deepep.py", 400),
         TestFile("ascend/test_ascend_deepseek_mtp.py", 400),
     ],
+    "__not_in_ascend_ci__": [
+        TestFile("ascend/test_mindspore_models.py"),
+    ],
 }
 
 suites.update(suite_amd)