sglang 0.4.7.post1__py3-none-any.whl → 0.4.8__py3-none-any.whl

sglang/srt/layers/moe/topk.py

@@ -28,10 +28,18 @@ from sglang.srt.managers.expert_location_dispatch import (
     topk_ids_logical_to_physical,
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.utils import get_compiler_backend, is_cuda, is_hip
+from sglang.srt.utils import (
+    cpu_has_amx_support,
+    get_compiler_backend,
+    is_cpu,
+    is_cuda,
+    is_hip,
+)
 
 _is_cuda = is_cuda()
 _is_hip = is_hip()
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
 
 if _is_cuda:
     from sgl_kernel import moe_fused_gate
@@ -40,7 +48,7 @@ if _is_cuda or _is_hip:
     from sgl_kernel import topk_softmax
 
 
-def fused_topk_native(
+def fused_topk_torch_native(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
     topk: int,
@@ -61,6 +69,20 @@ def fused_topk_native(
     return topk_weights, topk_ids
 
 
+def fused_topk_cpu(
+    hidden_states: torch.Tensor,
+    gating_output: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+):
+    return torch.ops.sgl_kernel.topk_softmax_cpu(
+        hidden_states=hidden_states,
+        gating_output=gating_output,
+        topk=topk,
+        renormalize=renormalize,
+    )
+
+
 def fused_topk(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
@@ -115,7 +137,7 @@ def _fused_topk_postprocess(
 
 # This is used by the Deepseek V2/V3/R1 series models
 @torch.compile(dynamic=True, backend=get_compiler_backend())
-def grouped_topk(
+def grouped_topk_gpu(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
     topk: int,
@@ -171,6 +193,32 @@ def grouped_topk(
     return topk_weights, topk_ids
 
 
+def grouped_topk_cpu(
+    hidden_states: torch.Tensor,
+    gating_output: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+    num_expert_group: int = 0,
+    topk_group: int = 0,
+    num_fused_shared_experts: int = 0,
+    routed_scaling_factor: Optional[float] = None,
+    num_token_non_padded: Optional[torch.Tensor] = None,
+    expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
+):
+    assert expert_location_dispatch_info is None
+    return torch.ops.sgl_kernel.grouped_topk_cpu(
+        hidden_states,
+        gating_output,
+        topk,
+        renormalize,
+        num_expert_group,
+        topk_group,
+        num_fused_shared_experts,
+        routed_scaling_factor,
+        num_token_non_padded,
+    )
+
+
 def biased_grouped_topk_impl(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
@@ -258,7 +306,7 @@ def _biased_grouped_topk_postprocess(
     return topk_ids
 
 
-def biased_grouped_topk(
+def biased_grouped_topk_gpu(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
     correction_bias: torch.Tensor,
@@ -322,6 +370,45 @@ def biased_grouped_topk(
         )
 
 
+def biased_grouped_topk_cpu(
+    hidden_states: torch.Tensor,
+    gating_output: torch.Tensor,
+    correction_bias: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+    num_expert_group: int = 0,
+    topk_group: int = 0,
+    compiled: bool = True,
+    num_fused_shared_experts: int = 0,
+    routed_scaling_factor: Optional[float] = None,
+    num_token_non_padded: Optional[torch.Tensor] = None,
+    expert_location_dispatch_info: Optional[ExpertLocationDispatchInfo] = None,
+):
+    assert expert_location_dispatch_info is None
+    return torch.ops.sgl_kernel.biased_grouped_topk_cpu(
+        hidden_states,
+        gating_output,
+        correction_bias,
+        topk,
+        renormalize,
+        num_expert_group,
+        topk_group,
+        num_fused_shared_experts,
+        routed_scaling_factor,
+        num_token_non_padded,
+    )
+
+
+if _is_cpu and _is_cpu_amx_available:
+    biased_grouped_topk = biased_grouped_topk_cpu
+    grouped_topk = grouped_topk_cpu
+    fused_topk_native = fused_topk_cpu
+else:
+    biased_grouped_topk = biased_grouped_topk_gpu
+    grouped_topk = grouped_topk_gpu
+    fused_topk_native = fused_topk_torch_native
+
+
 def select_experts(
     hidden_states: torch.Tensor,
     router_logits: torch.Tensor,

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

@@ -14,14 +14,18 @@ from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz, scaled_fp8_qu
 from sglang.srt.layers.quantization.fp8_utils import normalize_e4m3fn_to_e4m3fnuz
 from sglang.srt.layers.quantization.utils import (
     all_close_1d,
+    cpu_has_amx_support,
     per_tensor_dequantize,
     replace_parameter,
 )
-from sglang.srt.utils import is_cuda, set_weight_attrs
+from sglang.srt.utils import is_cpu, is_cuda, is_npu, set_weight_attrs
 
 _is_cuda = is_cuda()
+_is_npu = is_npu()
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
 
-if not _is_cuda:
+if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available)):
     from vllm import _custom_ops as vllm_ops
     from vllm._custom_ops import scaled_fp8_quant
 

sglang/srt/layers/quantization/fp8.py

@@ -64,9 +64,12 @@ from sglang.srt.layers.quantization.utils import (
 )
 from sglang.srt.layers.utils import is_sm100_supported
 from sglang.srt.utils import (
+    cpu_has_amx_support,
     get_bool_env_var,
+    is_cpu,
     is_cuda,
     is_hip,
+    is_npu,
     log_info_on_rank0,
     print_warning_once,
     set_weight_attrs,
@@ -74,6 +77,9 @@ from sglang.srt.utils import (
 
 _is_hip = is_hip()
 _is_cuda = is_cuda()
+_is_npu = is_npu()
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
 
 _is_fp8_fnuz = is_fp8_fnuz()
 
@@ -82,10 +88,11 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
 if _is_hip:
     from aiter import ActivationType, QuantType
+    from aiter.fused_moe import fused_moe
     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 or _is_npu or (_is_cpu and _is_cpu_amx_available)):
     from vllm._custom_ops import scaled_fp8_quant
 
 
@@ -1045,15 +1052,15 @@ class Fp8MoEMethod:
         if _use_hip_int4:
             # TODO: add triton kernel and add check _use_aiter
             assert not no_combine, f"{no_combine=} is not supported."
-            return ck_moe_2stages(
+            return fused_moe(
                 x,
                 layer.w13_weight,
                 layer.w2_weight,
                 topk_weights,
                 topk_ids,
-                QuantType.per_Token,
-                layer.w13_weight_scale1,
-                layer.w2_weight_scale1,
+                quant_type=QuantType.per_Token,
+                w1_scale=layer.w13_weight_scale1,
+                w2_scale=layer.w2_weight_scale1,
                 activation=(
                     ActivationType.Silu if activation == "silu" else ActivationType.Gelu
                 ),
@@ -1062,31 +1069,32 @@ class Fp8MoEMethod:
         if _use_aiter:
             assert not no_combine, f"{no_combine=} is not supported."
             if self.block_quant:
-                # TODO(_use_aiter): FP8 block_quant only supports 'silu' for the time-being.
-                assert (
-                    activation == "silu"
-                ), f"_use_aiter: FP8 bloack_quant {activation=} will be supported later, unset _use_aiter"
-                return asm_moe(
+                return fused_moe(
                     x,
                     layer.w13_weight,
                     layer.w2_weight,
                     topk_weights,
                     topk_ids,
-                    layer.w13_weight_scale_inv,
-                    layer.w2_weight_scale_inv,
-                    block_shape=tuple(self.quant_config.weight_block_size),
+                    w1_scale=layer.w13_weight_scale_inv,
+                    w2_scale=layer.w2_weight_scale_inv,
+                    quant_type=QuantType.per_128x128,
+                    activation=(
+                        ActivationType.Silu
+                        if activation == "silu"
+                        else ActivationType.Gelu
+                    ),
                     expert_mask=None,
                 )
             else:
-                return ck_moe_2stages(
+                return fused_moe(
                     x,
                     layer.w13_weight,
                     layer.w2_weight,
                     topk_weights,
                     topk_ids,
-                    QuantType.per_Token,
-                    layer.w13_weight_scale1,
-                    layer.w2_weight_scale1,
+                    quant_type=QuantType.per_Token,
+                    w1_scale=layer.w13_weight_scale1,
+                    w2_scale=layer.w2_weight_scale1,
                     activation=(
                         ActivationType.Silu
                         if activation == "silu"

sglang/srt/layers/quantization/modelopt_quant.py

@@ -29,11 +29,17 @@ from sglang.srt.layers.quantization.utils import (
     requantize_with_max_scale,
 )
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.utils import is_cuda
+from sglang.srt.utils import is_cuda, next_power_of_2
 
 if is_cuda():
     from sgl_kernel import cutlass_scaled_fp4_mm, scaled_fp4_quant
 
+try:
+    from flashinfer import fp4_quantize as fp4_quantize
+    from flashinfer.fused_moe import cutlass_fused_moe as flashinfer_cutlass_fused_moe
+except ImportError:
+    flashinfer_cutlass_fused_moe = None
+
 # Initialize logger for the module
 logger = logging.getLogger(__name__)
 
@@ -429,6 +435,9 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         layer.alpha = Parameter(
             layer.input_scale * layer.weight_scale_2, requires_grad=False
         )
+        layer.input_scale_inv = Parameter(
+            (1 / input_scale_2).to(torch.float32), requires_grad=False
+        )
 
         # Pad and blockwise interleave weight_scale
         scales = layer.weight_scale
@@ -467,7 +476,7 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         output_shape = [x_m, w_n]
 
         # Quantize BF16 or FP16 to (FP4 and interleaved block scale)
-        x_fp4, x_scale_interleaved = scaled_fp4_quant(x, 1 / layer.input_scale)
+        x_fp4, x_scale_interleaved = scaled_fp4_quant(x, layer.input_scale_inv)
 
         assert x_fp4.dtype == torch.uint8
         assert x_scale_interleaved.dtype == torch.float8_e4m3fn
@@ -521,6 +530,7 @@ class ModelOptNvFp4FusedMoEMethod:
                 " quantization. Please use Blackwell and"
                 " above."
             )
+        self.enable_flashinfer_moe = False
 
     def create_weights(
         self,
@@ -674,7 +684,10 @@ class ModelOptNvFp4FusedMoEMethod:
         w13_weight_scale_2 = layer.w13_weight_scale_2[:, 0]
         layer.w13_weight_scale_2 = Parameter(w13_weight_scale_2, requires_grad=False)
 
-        w13_input_scale = layer.w13_input_scale.max(dim=1).values.to(torch.float32)
+        if self.enable_flashinfer_moe:
+            w13_input_scale = layer.w13_input_scale.max().to(torch.float32)
+        else:
+            w13_input_scale = layer.w13_input_scale.max(dim=1).values.to(torch.float32)
         layer.g1_alphas = Parameter(
             (w13_input_scale * w13_weight_scale_2).to(torch.float32),
             requires_grad=False,
@@ -700,14 +713,19 @@ class ModelOptNvFp4FusedMoEMethod:
         layer.w13_weight = Parameter(layer.w13_weight.data, requires_grad=False)
 
         # GEMM 2
+        if self.enable_flashinfer_moe:
+            w2_input_scale = layer.w2_input_scale.max().to(torch.float32)
+        else:
+            w2_input_scale = layer.w2_input_scale
+
         layer.g2_alphas = Parameter(
-            (layer.w2_input_scale * layer.w2_weight_scale_2).to(torch.float32),
+            (w2_input_scale * layer.w2_weight_scale_2).to(torch.float32),
             requires_grad=False,
         )
 
         # This is for quantization, so we need to invert it.
         layer.w2_input_scale_quant = Parameter(
-            (1 / layer.w2_input_scale).to(torch.float32), requires_grad=False
+            (1 / w2_input_scale).to(torch.float32), requires_grad=False
         )
 
         assert (
@@ -727,11 +745,16 @@ class ModelOptNvFp4FusedMoEMethod:
         layer.cutlass_moe_params = CutlassMoEParams(
             CutlassMoEType.BlockscaledFP4,
             device,
-            num_experts=layer.num_experts,
+            num_experts=layer.num_experts,  # global num experts
             intermediate_size_per_partition=layer.w2_weight.shape[2] * 2,  # n
             hidden_size=layer.w13_weight.shape[2] * 2,
         )  # k
 
+    @property
+    def load_up_proj_weight_first(self) -> bool:
+        # FlashInfer CUTLASS kernel assumes [Up, Gate] Proj as W13
+        return self.enable_flashinfer_moe
+
     def apply(
         self,
         layer: torch.nn.Module,
@@ -750,11 +773,13 @@ class ModelOptNvFp4FusedMoEMethod:
         inplace: bool = True,
         no_combine: bool = False,
         routed_scaling_factor: Optional[float] = None,
+        ep_rank: Optional[int] = None,
+        ep_size: Optional[int] = None,
+        tp_rank: Optional[int] = None,
+        tp_size: Optional[int] = None,
     ) -> torch.Tensor:
 
         assert activation == "silu", "Only SiLU activation is supported."
-
-        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
         from sglang.srt.layers.moe.topk import select_experts
 
         topk_weights, topk_ids = select_experts(
@@ -771,6 +796,35 @@ class ModelOptNvFp4FusedMoEMethod:
             routed_scaling_factor=routed_scaling_factor,
         )
 
+        if self.enable_flashinfer_moe:
+            assert (
+                not apply_router_weight_on_input
+            ), "apply_router_weight_on_input is not supported for Flashinfer"
+            # TRTLLM Cutlass moe takes in activations in BF16/Half/nvfp4 precision
+            # and fp4 quantized weights loaded from the checkpoint
+            output = flashinfer_cutlass_fused_moe(
+                x,
+                topk_ids.to(torch.int),
+                topk_weights,
+                layer.w13_weight.view(torch.long),
+                layer.w2_weight.view(torch.long),
+                x.dtype,
+                quant_scales=[
+                    layer.w13_input_scale_quant,
+                    layer.w13_blockscale_swizzled.view(torch.int32),
+                    layer.g1_alphas,
+                    layer.w2_input_scale_quant,
+                    layer.w2_blockscale_swizzled.view(torch.int32),
+                    layer.g2_alphas,
+                ],
+                ep_size=ep_size,
+                ep_rank=ep_rank,
+                tp_size=tp_size,
+                tp_rank=tp_rank,
+                tune_max_num_tokens=next_power_of_2(x.shape[0]),
+            )
+            return output[0]
+
         from sglang.srt.layers.moe.cutlass_moe import cutlass_moe_fp4
 
         return cutlass_moe_fp4(

sglang/srt/layers/quantization/utils.py

@@ -6,11 +6,14 @@ 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 cpu_has_amx_support, is_cpu, is_cuda, is_npu
 
 _is_cuda = is_cuda()
+_is_npu = is_npu()
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
 
-if not _is_cuda:
+if not (_is_cuda or _is_npu or (_is_cpu and _is_cpu_amx_available)):
     from vllm._custom_ops import scaled_fp8_quant
 
 

sglang/srt/layers/rotary_embedding.py

@@ -8,10 +8,13 @@ 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 cpu_has_amx_support, is_cpu, is_cuda, is_hip, is_npu
 
 _is_cuda = is_cuda()
 _is_hip = is_hip()
+_is_npu = is_npu()
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
 
 if _is_cuda:
     from sgl_kernel import apply_rope_with_cos_sin_cache_inplace
@@ -84,7 +87,9 @@ 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]
+        ) and not (_is_cpu and _is_cpu_amx_available):
             from vllm._custom_ops import rotary_embedding
 
             self.vllm_rotary_embedding = rotary_embedding
@@ -147,6 +152,26 @@ class RotaryEmbedding(CustomOp):
         key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
         return query, key
 
+    def forward_cpu(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        offsets: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        positions = torch.add(positions, offsets) if offsets is not None else positions
+        if _is_cpu_amx_available:
+            return torch.ops.sgl_kernel.rotary_embedding_cpu(
+                positions,
+                query,
+                key,
+                self.head_size,
+                self.cos_sin_cache,
+                self.is_neox_style,
+            )
+        else:
+            return self.forward_native(positions, query, key, offsets)
+
     def forward_cuda(
         self,
         positions: torch.Tensor,
@@ -696,6 +721,21 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbedding):
             key = key_rot
         return query.to(dtype), key.to(dtype)
 
+    def forward_cpu(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        offsets: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        positions = torch.add(positions, offsets) if offsets is not None else positions
+        if _is_cpu_amx_available:
+            return torch.ops.sgl_kernel.rotary_embedding_cpu(
+                positions, query, key, self.head_size, self.cos_sin_cache, False
+            )
+        else:
+            return self.forward_native(positions, query, key, offsets)
+
 
 class Llama3RotaryEmbedding(RotaryEmbedding):
 

sglang/srt/layers/sampler.py

@@ -91,7 +91,7 @@ class Sampler(nn.Module):
                         )
                     else:
                         batch_next_token_ids = top_k_top_p_sampling_from_probs(
-                            probs,
+                            probs.contiguous(),
                             sampling_info.top_ks,
                             sampling_info.top_ps,
                             filter_apply_order="joint",