support qwen2 running on ascend npu device (#7022)

Co-authored-by: 刁莹煜 <diaoyingyu1@hisilicon.com>

docs/backend/native_api.ipynb

@@ -51,6 +51,10 @@
     "server_process, port = launch_server_cmd(\n",
     "    \"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0\"\n",
     ")\n",
+    "## To run qwen2.5-0.5b-instruct model on the Ascend-Npu, you can execute the following command:\n",
+    "# server_process, port = launch_server_cmd(\n",
+    "#     \"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0 --device npu --tp 2 --attention-backend torch_native\"\n",
+    "# )\n",
     "\n",
     "wait_for_server(f\"http://localhost:{port}\")"
    ]

python/sglang/srt/_custom_ops.py

@@ -4,7 +4,7 @@ from typing import List, Tuple
 
 import torch
 
-from sglang.srt.utils import get_bool_env_var, is_hip, is_hpu
+from sglang.srt.utils import get_bool_env_var, is_hip, is_hpu, is_npu
 
 logger = logging.getLogger(__name__)
 use_vllm_custom_allreduce = get_bool_env_var(
@@ -25,7 +25,7 @@ if not is_hpu():
             logger.warning("Failed to import from custom_ar with %r", e)
 
 
-if not is_hip():
+if not is_hip() and not is_npu():
     if use_vllm_custom_allreduce:
         custom_op = torch.ops._C_custom_ar
     else:

python/sglang/srt/layers/activation.py

@@ -29,10 +29,11 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.utils import is_cuda, set_weight_attrs
+from sglang.srt.utils import is_cuda, is_npu, set_weight_attrs
 from sglang.utils import resolve_obj_by_qualname
 
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
 if _is_cuda:
     from sgl_kernel import gelu_and_mul, gelu_tanh_and_mul, silu_and_mul
@@ -184,7 +185,7 @@ def get_cross_encoder_activation_function(config: PretrainedConfig):
         return nn.Identity()
 
 
-if not _is_cuda:
+if not _is_cuda and not _is_npu:
     logger.info(
         "sgl-kernel is not available on Non-NV platforms. Fallback to other kernel libraries."
     )

python/sglang/srt/layers/layernorm.py

@@ -20,10 +20,11 @@ import torch
 import torch.nn as nn
 
 from sglang.srt.custom_op import CustomOp
-from sglang.srt.utils import get_bool_env_var, is_cuda, is_hip
+from sglang.srt.utils import get_bool_env_var, is_cuda, is_hip, is_npu
 
 _is_cuda = is_cuda()
 _is_hip = is_hip()
+_is_npu = is_npu()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _is_cuda:
@@ -187,7 +188,7 @@ class Gemma3RMSNorm(nn.Module):
         return f"{tuple(self.weight.shape)}, eps={self.eps}"
 
 
-if not (_is_cuda or _is_hip):
+if not (_is_cuda or _is_hip or _is_npu):
     logger.info(
         "sgl-kernel layernorm implementation is not available on current platform. Fallback to other kernel libraries."
     )

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -17,11 +17,12 @@ from sglang.srt.layers.quantization.utils import (
     per_tensor_dequantize,
     replace_parameter,
 )
-from sglang.srt.utils import is_cuda, set_weight_attrs
+from sglang.srt.utils import is_cuda, is_npu, set_weight_attrs
 
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
-if not _is_cuda:
+if not _is_cuda and not _is_npu:
     from vllm import _custom_ops as vllm_ops
     from vllm._custom_ops import scaled_fp8_quant
 

python/sglang/srt/layers/quantization/fp8.py

@@ -67,6 +67,7 @@ from sglang.srt.utils import (
     get_bool_env_var,
     is_cuda,
     is_hip,
+    is_npu,
     log_info_on_rank0,
     print_warning_once,
     set_weight_attrs,
@@ -74,6 +75,7 @@ from sglang.srt.utils import (
 
 _is_hip = is_hip()
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
 _is_fp8_fnuz = is_fp8_fnuz()
 
@@ -86,7 +88,7 @@ if _is_hip:
     from aiter.fused_moe_bf16_asm import asm_moe, ck_moe_2stages
     from aiter.ops.shuffle import shuffle_weight
 
-if not _is_cuda:
+if not _is_cuda and not _is_npu:
     from vllm._custom_ops import scaled_fp8_quant
 
 

python/sglang/srt/layers/quantization/utils.py

@@ -6,11 +6,12 @@ from typing import List, Mapping, Tuple, Union
 import torch
 
 from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
-from sglang.srt.utils import is_cuda
+from sglang.srt.utils import is_cuda, is_npu
 
 _is_cuda = is_cuda()
+_is_npu = is_npu()
 
-if not _is_cuda:
+if not _is_cuda and not _is_npu:
     from vllm._custom_ops import scaled_fp8_quant
 
 

python/sglang/srt/layers/rotary_embedding.py

@@ -8,10 +8,11 @@ import torch
 import torch.nn as nn
 
 from sglang.srt.custom_op import CustomOp
-from sglang.srt.utils import is_cuda, is_hip
+from sglang.srt.utils import is_cuda, is_hip, is_npu
 
 _is_cuda = is_cuda()
 _is_hip = is_hip()
+_is_npu = is_npu()
 
 if _is_cuda:
     from sgl_kernel import apply_rope_with_cos_sin_cache_inplace
@@ -84,7 +85,7 @@ class RotaryEmbedding(CustomOp):
         if not _is_cuda:
             cache = cache.to(dtype)
 
-        if not _is_cuda or self.head_size not in [64, 128, 256, 512]:
+        if not (_is_cuda or _is_npu) or self.head_size not in [64, 128, 256, 512]:
             from vllm._custom_ops import rotary_embedding
 
             self.vllm_rotary_embedding = rotary_embedding

python/sglang/srt/utils.py

@@ -1291,6 +1291,15 @@ def get_hpu_memory_capacity():
         )
 
 
+def get_npu_memory_capacity():
+    try:
+        import torch_npu
+
+        return torch.npu.mem_get_info()[1] // 1024 // 1024  # unit: MB
+    except ImportError as e:
+        raise ImportError("torch_npu is required when run on npu device.")
+
+
 def get_device_memory_capacity(device: str = None):
     if is_cuda():
         gpu_mem = get_nvgpu_memory_capacity()
@@ -1298,6 +1307,8 @@ def get_device_memory_capacity(device: str = None):
         gpu_mem = get_amdgpu_memory_capacity()
     elif device == "hpu":
         gpu_mem = get_hpu_memory_capacity()
+    elif device == "npu":
+        gpu_mem = get_npu_memory_capacity()
     else:
         # GPU memory is not known yet or no GPU is available.
         gpu_mem = None
@@ -1423,6 +1434,11 @@ def get_device(device_id: Optional[int] = None) -> str:
             return "xpu"
         return "xpu:{}".format(device_id)
 
+    if hasattr(torch, "npu") and torch.npu.is_available():
+        if device_id == None:
+            return "npu"
+        return "npu:{}".format(device_id)
+
     if is_habana_available():
         try:
             import habana_frameworks.torch.hpu
@@ -1497,15 +1513,35 @@ def get_device_capability(device_id: int = 0) -> Tuple[int, int]:
     return major, minor
 
 
+def get_npu_compiler_config():
+    config = {
+        "frozen_parameter": True,
+        "tiling_schedule_optimize": True,
+        "topology_sorting_strategy": "StableRDFS",
+    }
+    return config
+
+
 def get_compiler_backend() -> str:
     if hasattr(torch, "hpu") and torch.hpu.is_available():
         return "hpu_backend"
 
     if hasattr(torch, "npu") and torch.npu.is_available():
-        import torchair
+        try:
+            import torchair
+            import torchair.ge_concrete_graph.ge_converter.experimental.patch_for_hcom_allreduce
+            from torchair.configs.compiler_config import CompilerConfig
+        except ImportError as e:
+            raise ImportError(
+                "NPU detected, but torchair package is not installed. "
+                "Please install torchair for torch.compile support on NPU."
+            )
+        compiler_config = CompilerConfig()
+        predefined_config = get_npu_compiler_config()
+        for k, v in predefined_config.items():
+            setattr(compiler_config.experimental_config, k, v)
 
-        config = torchair.CompilerConfig()
+        npu_backend = torchair.get_npu_backend(compiler_config=compiler_config)
-        npu_backend = torchair.get_npu_backend(compiler_config=config)
         return npu_backend
 
     return "inductor"