[Intel GPU] Enable DeepSeek R1 inference on XPU (#18461)

Signed-off-by: P V R K Jyothendra Varma <polisetty.v.r.k.jyothendra.varma@intel.com>

python/sglang/srt/layers/moe/token_dispatcher/standard.py

@@ -30,7 +30,12 @@ from sglang.srt.layers.moe.utils import (
     get_moe_runner_backend,
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
-from sglang.srt.utils.common import get_bool_env_var, is_hip, is_sm120_supported
+from sglang.srt.utils.common import (
+    get_bool_env_var,
+    get_device,
+    is_hip,
+    is_sm120_supported,
+)
 
 _is_hip = is_hip()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
@@ -149,15 +154,16 @@ class StandardDispatcher(BaseDispatcher):
             and TopKOutputChecker.format_is_standard(topk_output)
         ):
             if self.local_expert_mapping is None:
+                device = get_device()
                 self.local_expert_mapping = torch.full(
-                    (self.num_experts,), -1, dtype=torch.int32, device="cuda"
+                    (self.num_experts,), -1, dtype=torch.int32, device=device
                 )
                 self.local_expert_mapping[
                     self.moe_ep_rank
                     * self.num_local_routed_experts : (self.moe_ep_rank + 1)
                     * self.num_local_routed_experts
                 ] = torch.arange(
-                    0, self.num_local_routed_experts, dtype=torch.int32, device="cuda"
+                    0, self.num_local_routed_experts, dtype=torch.int32, device=device
                 )
 
                 if self.num_local_shared_experts > 0:

python/sglang/srt/models/deepseek_common/deepseek_weight_loader.py

@@ -50,6 +50,7 @@ from sglang.srt.models.deepseek_common.utils import (
     _is_fp8_fnuz,
     _is_hip,
     _is_npu,
+    _is_xpu,
     _use_aiter_gfx95,
     awq_dequantize_func,
     enable_nextn_moe_bf16_cast_to_fp8,
@@ -497,7 +498,7 @@ class DeepseekV2WeightLoaderMixin:
                         )
 
                     if (
-                        _is_cuda
+                        (_is_cuda or _is_xpu)
                         and weight_block_size[0] == 128
                         and weight_block_size[1] == 128
                     ):

python/sglang/srt/models/deepseek_common/utils.py

@@ -31,6 +31,7 @@ from sglang.srt.utils import (
     is_hip,
     is_npu,
     is_nvidia_cublas_version_ge_12_9,
+    is_xpu,
 )
 
 _is_hip = is_hip()
@@ -40,6 +41,7 @@ _is_fp8_fnuz = is_fp8_fnuz()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
+_is_xpu = is_xpu()
 _device_sm = get_device_sm()
 _is_gfx95_supported = is_gfx95_supported()
 _use_aiter_gfx95 = _use_aiter and _is_gfx95_supported

python/sglang/srt/models/deepseek_v2.py

@@ -137,6 +137,7 @@ from sglang.srt.models.deepseek_common.utils import (
     _is_gfx95_supported,
     _is_hip,
     _is_npu,
+    _is_xpu,
     _use_aiter,
     _use_aiter_gfx95,
 )
@@ -677,6 +678,7 @@ class DeepseekV2MoE(nn.Module):
         )
         if (
             not _is_cuda
+            and not _is_xpu
             and not _use_aiter
             or isinstance(self.experts.quant_method, KTEPWrapperMethod)
         ):