From e54f08c8602c3bca11e6c16e69c5815f9c277bc7 Mon Sep 17 00:00:00 2001 From: Polisetty V R K Jyothendra Varma Date: Fri, 9 Oct 2026 07:56:46 +0530 Subject: [PATCH] [XPU] Misc changes to support XPU (#39817) Signed-off-by: P V R K Jyothendra Varma --- python/sglang/srt/arg_groups/fields/exec_.py | 2 ++ python/sglang/srt/arg_groups/overrides.py | 5 ----- python/sglang/srt/layers/quantization/unquant.py | 2 ++ python/sglang/srt/models/hunyuan_v3_nextn.py | 4 ++-- test/registered/xpu/test_dsa_indexer_xpu.py | 3 +-- 5 files changed, 7 insertions(+), 9 deletions(-) diff --git a/python/sglang/srt/arg_groups/fields/exec_.py b/python/sglang/srt/arg_groups/fields/exec_.py index 6ff90b7ec4..7bfcaf0f2a 100644 --- a/python/sglang/srt/arg_groups/fields/exec_.py +++ b/python/sglang/srt/arg_groups/fields/exec_.py @@ -223,6 +223,7 @@ class ExecKernel(msgspec.Struct): "triton", "aiter", "trtllm", + "intel_xpu", ], resolvable=True, ), @@ -263,6 +264,7 @@ class ExecKernel(msgspec.Struct): "triton", "aiter", "trtllm", + "intel_xpu", ], resolvable=True, ), diff --git a/python/sglang/srt/arg_groups/overrides.py b/python/sglang/srt/arg_groups/overrides.py index bb76413502..eda212a2a5 100644 --- a/python/sglang/srt/arg_groups/overrides.py +++ b/python/sglang/srt/arg_groups/overrides.py @@ -586,11 +586,6 @@ def _dsa_split_backend_resolution(view: Any) -> dict: declared["dsa_prefill_backend"] = "intel_xpu" if view.dsa_decode_backend is None: declared["dsa_decode_backend"] = "intel_xpu" - # sgl-kernel topk ops (the default) are CUDA-only; fall back to the - # torch-native topk implementation on XPU, unless the user already - # picked a different backend explicitly (e.g. "flashinfer"). - if view.dsa_topk_backend == "sgl-kernel": - declared["dsa_topk_backend"] = "torch" logger.warning( "Set DSA backends for XPU: prefill=%s, decode=%s, topk=%s.", declared.get("dsa_prefill_backend", view.dsa_prefill_backend), diff --git a/python/sglang/srt/layers/quantization/unquant.py b/python/sglang/srt/layers/quantization/unquant.py index 60defa37c6..ac239656cd 100644 --- a/python/sglang/srt/layers/quantization/unquant.py +++ b/python/sglang/srt/layers/quantization/unquant.py @@ -1274,8 +1274,10 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, BaseFusedOp): b1=getattr(layer, "w13_weight_bias", None), b2=getattr(layer, "w2_weight_bias", None), activation=moe_runner_config.activation, + routed_scaling_factor=moe_runner_config.routed_scaling_factor, gemm1_alpha=moe_runner_config.gemm1_alpha, gemm1_limit=moe_runner_config.gemm1_clamp_limit, + swiglu_limit=moe_runner_config.swiglu_limit, ) return StandardCombineInput(hidden_states=output) else: diff --git a/python/sglang/srt/models/hunyuan_v3_nextn.py b/python/sglang/srt/models/hunyuan_v3_nextn.py index f831d54efa..c700fb6f3b 100644 --- a/python/sglang/srt/models/hunyuan_v3_nextn.py +++ b/python/sglang/srt/models/hunyuan_v3_nextn.py @@ -154,8 +154,8 @@ class HYV3ForCausalLMNextN(nn.Module): del self.lm_head.weight self.model.embed_tokens.weight = embed self.lm_head.weight = head - torch.cuda.empty_cache() - torch.cuda.synchronize() + torch.get_device_module().empty_cache() + torch.get_device_module().synchronize() def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): nextn_layer_id = self.config.num_hidden_layers diff --git a/test/registered/xpu/test_dsa_indexer_xpu.py b/test/registered/xpu/test_dsa_indexer_xpu.py index a05a326b54..c1420ec4e0 100644 --- a/test/registered/xpu/test_dsa_indexer_xpu.py +++ b/test/registered/xpu/test_dsa_indexer_xpu.py @@ -220,7 +220,7 @@ class MockModelRunner: "enable_deterministic_inference": False, "dsa_prefill_backend": "intel_xpu", "dsa_decode_backend": "intel_xpu", - "dsa_topk_backend": "torch", # XPU uses torch.topk fallback + "dsa_topk_backend": "sgl-kernel", "dsa_paged_mqa_logits_backend": "auto", "disaggregation_mode": "null", "enable_two_batch_overlap": False, @@ -238,7 +238,6 @@ class TestDSAIndexerXPU(CustomTestCase): server_args = ServerArgs(model_path="dummy") server_args.dsa_prefill_backend = "intel_xpu" server_args.dsa_decode_backend = "intel_xpu" - server_args.dsa_topk_backend = "torch" # Disable CUDA-only JIT topk-v2 (TileLang requires CUDA_HOME) envs.SGLANG_OPT_USE_TOPK_V2.set(False) set_global_server_args_for_scheduler(server_args)