sglang 0.4.4.post4__py3-none-any.whl → 0.4.5.post1__py3-none-any.whl

sglang/srt/layers/moe/fused_moe_triton/configs/E=288,N=64,device_name=NVIDIA_A800-SXM4-80GB.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -342,6 +342,7 @@ def fused_moe_kernel(
     use_fp8_w8a8: tl.constexpr,
     use_int8_w8a8: tl.constexpr,
     use_int8_w8a16: tl.constexpr,
+    per_channel_quant: tl.constexpr,
     even_Ks: tl.constexpr,
 ):
     """
@@ -416,20 +417,7 @@ def fused_moe_kernel(
         )
         b_scale = tl.load(b_scale_ptrs)
 
-    if use_fp8_w8a8:
-        # block-wise
-        if group_k > 0 and group_n > 0:
-            a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
-            offs_bsn = offs_bn // group_n
-            b_scale_ptrs = (
-                b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
-            )
-        # tensor-wise
-        else:
-            a_scale = tl.load(a_scale_ptr)
-            b_scale = tl.load(b_scale_ptr + off_experts)
-
-    if use_int8_w8a8:
+    if use_fp8_w8a8 or use_int8_w8a8:
         # block-wise
         if group_k > 0 and group_n > 0:
             a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
@@ -438,8 +426,7 @@ def fused_moe_kernel(
                 b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
             )
         # channel-wise
-        else:
-            # Load per-column scale for weights
+        elif per_channel_quant:
             b_scale_ptrs = (
                 b_scale_ptr + off_experts * stride_bse + offs_bn[None, :] * stride_bsn
             )
@@ -447,6 +434,10 @@ def fused_moe_kernel(
             # Load per-token scale for activations
             a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
             a_scale = tl.load(a_scale_ptrs, mask=token_mask, other=0.0)[:, None]
+        # tensor-wise
+        else:
+            a_scale = tl.load(a_scale_ptr)
+            b_scale = tl.load(b_scale_ptr + off_experts)
 
     # -----------------------------------------------------------
     # Iterate to compute a block of the C matrix.
@@ -711,12 +702,12 @@ def moe_align_block_size(
             num_tokens_post_pad,
         )
     else:
-        token_cnts_buffer = torch.zeros(
+        token_cnts_buffer = torch.empty(
             (num_experts + 1) * num_experts,
             dtype=torch.int32,
             device=topk_ids.device,
         )
-        cumsum_buffer = torch.zeros(
+        cumsum_buffer = torch.empty(
             num_experts + 1, dtype=torch.int32, device=topk_ids.device
         )
 
@@ -753,6 +744,7 @@ def invoke_fused_moe_kernel(
     use_int8_w8a8: bool,
     use_int8_w8a16: bool,
     use_int4_w4a16: bool,
+    per_channel_quant: bool,
     block_shape: Optional[List[int]] = None,
     no_combine: bool = False,
 ) -> None:
@@ -765,6 +757,8 @@ def invoke_fused_moe_kernel(
         from sglang.srt.layers.quantization.fp8_kernel import (
             sglang_per_token_group_quant_fp8,
         )
+    else:
+        from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
 
     assert topk_weights.stride(1) == 1
     assert sorted_token_ids.stride(0) == 1
@@ -775,10 +769,15 @@ def invoke_fused_moe_kernel(
         if block_shape is None:
             # activation tensor-wise fp8 quantization, dynamic or static
             padded_size = padding_size
+            # activations apply per-token quantization when weights apply per-channel quantization by default
             if _is_cuda:
-                A, A_scale = sgl_scaled_fp8_quant(A, A_scale)
+                A, A_scale = sgl_scaled_fp8_quant(
+                    A, A_scale, use_per_token_if_dynamic=per_channel_quant
+                )
             else:
-                A, A_scale = vllm_ops.scaled_fp8_quant(A, A_scale)
+                A, A_scale = vllm_ops.scaled_fp8_quant(
+                    A, A_scale, use_per_token_if_dynamic=per_channel_quant
+                )
         else:
             # activation block-wise fp8 quantization
             assert len(block_shape) == 2
@@ -794,6 +793,9 @@ def invoke_fused_moe_kernel(
         assert B_scale is not None
         if block_shape is None:
             # activation channel-wise int8 quantization
+            assert (
+                per_channel_quant
+            ), "int8 quantization only supports channel-wise quantization except for block-wise quantization"
             A, A_scale = per_token_quant_int8(A)
         else:
             # activation block-wise int8 quantization
@@ -902,6 +904,7 @@ def invoke_fused_moe_kernel(
             use_fp8_w8a8=use_fp8_w8a8,
             use_int8_w8a8=use_int8_w8a8,
             use_int8_w8a16=use_int8_w8a16,
+            per_channel_quant=per_channel_quant,
             even_Ks=even_Ks,
             **config,
         )
@@ -953,7 +956,7 @@ def get_moe_configs(
     logger.warning(
         (
             "Using default MoE config. Performance might be sub-optimal! "
-            "Config file not found at %s"
+            "Config file not found at %s, you can tune the config with https://github.com/sgl-project/sglang/blob/main/benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py."
         ),
         config_file_path,
     )
@@ -1079,10 +1082,12 @@ def inplace_fused_experts(
     topk_weights: torch.Tensor,
     topk_ids: torch.Tensor,
     activation: str = "silu",
+    apply_router_weight_on_input: bool = False,
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1099,10 +1104,12 @@ def inplace_fused_experts(
         topk_ids,
         True,
         activation,
+        apply_router_weight_on_input,
         use_fp8_w8a8,
         use_int8_w8a8,
         use_int8_w8a16,
         use_int4_w4a16,
+        per_channel_quant,
         w1_scale,
         w2_scale,
         w1_zp,
@@ -1120,10 +1127,12 @@ def inplace_fused_experts_fake(
     topk_weights: torch.Tensor,
     topk_ids: torch.Tensor,
     activation: str = "silu",
+    apply_router_weight_on_input: bool = False,
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1150,10 +1159,12 @@ def outplace_fused_experts(
     topk_weights: torch.Tensor,
     topk_ids: torch.Tensor,
     activation: str = "silu",
+    apply_router_weight_on_input: bool = False,
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1171,10 +1182,12 @@ def outplace_fused_experts(
         topk_ids,
         False,
         activation,
+        apply_router_weight_on_input,
         use_fp8_w8a8,
         use_int8_w8a8,
         use_int8_w8a16,
         use_int4_w4a16,
+        per_channel_quant,
         w1_scale,
         w2_scale,
         w1_zp,
@@ -1193,10 +1206,12 @@ def outplace_fused_experts_fake(
     topk_weights: torch.Tensor,
     topk_ids: torch.Tensor,
     activation: str = "silu",
+    apply_router_weight_on_input: bool = False,
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1225,10 +1240,12 @@ def fused_experts(
     topk_ids: torch.Tensor,
     inplace: bool = False,
     activation: str = "silu",
+    apply_router_weight_on_input: bool = False,
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1247,10 +1264,12 @@ def fused_experts(
             topk_weights,
             topk_ids,
             activation,
+            apply_router_weight_on_input,
             use_fp8_w8a8,
             use_int8_w8a8,
             use_int8_w8a16,
             use_int4_w4a16,
+            per_channel_quant,
             w1_scale,
             w2_scale,
             w1_zp,
@@ -1268,10 +1287,12 @@ def fused_experts(
             topk_weights,
             topk_ids,
             activation,
+            apply_router_weight_on_input,
             use_fp8_w8a8,
             use_int8_w8a8,
             use_int8_w8a16,
             use_int4_w4a16,
+            per_channel_quant,
             w1_scale,
             w2_scale,
             w1_zp,
@@ -1291,10 +1312,12 @@ def fused_experts_impl(
     topk_ids: torch.Tensor,
     inplace: bool = False,
     activation: str = "silu",
+    apply_router_weight_on_input: bool = False,
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1423,7 +1446,7 @@ def fused_experts_impl(
             sorted_token_ids,
             expert_ids,
             num_tokens_post_padded,
-            False,
+            apply_router_weight_on_input,
             topk_ids.shape[1],
             config,
             compute_type=compute_type,
@@ -1431,6 +1454,7 @@ def fused_experts_impl(
             use_int8_w8a8=use_int8_w8a8,
             use_int8_w8a16=use_int8_w8a16,
             use_int4_w4a16=use_int4_w4a16,
+            per_channel_quant=per_channel_quant,
             block_shape=block_shape,
         )
         if activation == "silu":
@@ -1456,7 +1480,7 @@ def fused_experts_impl(
             (
                 intermediate_cache3
                 if not no_combine and topk_ids.shape[1] != 1
-                else out_hidden_states[begin_chunk_idx:end_chunk_idx]
+                else out_hidden_states[begin_chunk_idx:end_chunk_idx].unsqueeze(0)
             ),
             a2_scale,
             w2_scale,
@@ -1466,7 +1490,7 @@ def fused_experts_impl(
             sorted_token_ids,
             expert_ids,
             num_tokens_post_padded,
-            True,
+            not apply_router_weight_on_input,
             1,
             config,
             compute_type=compute_type,
@@ -1474,6 +1498,7 @@ def fused_experts_impl(
             use_int8_w8a8=use_int8_w8a8,
             use_int8_w8a16=use_int8_w8a16,
             use_int4_w4a16=use_int4_w4a16,
+            per_channel_quant=per_channel_quant,
             block_shape=block_shape,
         )
 
@@ -1520,6 +1545,7 @@ def fused_moe(
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
     use_int4_w4a16: bool = False,
+    per_channel_quant: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
     w1_zp: Optional[torch.Tensor] = None,
@@ -1596,6 +1622,7 @@ def fused_moe(
         use_int8_w8a8=use_int8_w8a8,
         use_int8_w8a16=use_int8_w8a16,
         use_int4_w4a16=use_int4_w4a16,
+        per_channel_quant=per_channel_quant,
         w1_scale=w1_scale,
         w2_scale=w2_scale,
         w1_zp=w1_zp,

sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -128,6 +128,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
     ) -> torch.Tensor:
@@ -143,6 +144,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
             activation=activation,
+            apply_router_weight_on_input=apply_router_weight_on_input,
             inplace=inplace,
             no_combine=no_combine,
         )
@@ -160,6 +162,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
     ) -> torch.Tensor:
@@ -200,6 +203,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 topk_ids=topk_ids,
                 inplace=inplace and not no_combine,
                 activation=activation,
+                apply_router_weight_on_input=apply_router_weight_on_input,
                 no_combine=no_combine,
             )
 
@@ -276,6 +280,7 @@ class FusedMoE(torch.nn.Module):
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         use_presharded_weights: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
@@ -302,6 +307,7 @@ class FusedMoE(torch.nn.Module):
         self.custom_routing_function = custom_routing_function
         self.correction_bias = correction_bias
         self.activation = activation
+        self.apply_router_weight_on_input = apply_router_weight_on_input
         self.use_presharded_weights = use_presharded_weights
         self.inplace = inplace
         self.no_combine = no_combine
@@ -630,6 +636,7 @@ class FusedMoE(torch.nn.Module):
             custom_routing_function=self.custom_routing_function,
             correction_bias=self.correction_bias,
             activation=self.activation,
+            apply_router_weight_on_input=self.apply_router_weight_on_input,
         )
 
         if self.reduce_results and self.tp_size > 1:

sglang/srt/layers/moe/router.py

@@ -5,6 +5,9 @@ import triton
 import triton.language as tl
 
 from sglang.srt.layers.moe.topk import fused_topk
+from sglang.srt.utils import is_hip
+
+_is_hip = is_hip()
 
 
 @triton.jit
@@ -116,10 +119,13 @@ def fused_moe_router_impl(
     topk_ids = torch.empty((bs, topk), dtype=torch.int32, device=x.device)
 
     grid = lambda meta: (bs,)
+
+    min_num_warps = 16 if _is_hip else 32
+
     config = {
         "BLOCK_SIZE": triton.next_power_of_2(hidden_dim),
         "num_warps": max(
-            min(triton.next_power_of_2(triton.cdiv(hidden_dim, 256)), 32), 4
+            min(triton.next_power_of_2(triton.cdiv(hidden_dim, 256)), min_num_warps), 4
         ),
     }
 

sglang/srt/layers/moe/topk.py

@@ -12,6 +12,7 @@
 # limitations under the License.
 # ==============================================================================
 
+import math
 import os
 from typing import Callable, Optional
 
@@ -25,6 +26,8 @@ from sglang.srt.utils import get_compiler_backend, is_cuda, is_hip
 _is_cuda = is_cuda()
 _is_hip = is_hip()
 
+if _is_cuda:
+    from sgl_kernel import moe_fused_gate
 
 expert_distribution_recorder = ExpertDistributionRecorder()
 
@@ -209,6 +212,10 @@ def biased_grouped_topk_impl(
     return topk_weights.to(torch.float32), topk_ids.to(torch.int32)
 
 
+def is_power_of_two(n):
+    return n > 0 and math.log2(n).is_integer()
+
+
 def biased_grouped_topk(
     hidden_states: torch.Tensor,
     gating_output: torch.Tensor,
@@ -220,23 +227,37 @@ def biased_grouped_topk(
     compiled: bool = True,
     n_share_experts_fusion: int = 0,
 ):
-    biased_grouped_topk_fn = (
-        torch.compile(
-            biased_grouped_topk_impl, dynamic=True, backend=get_compiler_backend()
+    # TODO: moe_fused_gate kernel is not supported for n_share_experts_fusion > 0 now.
+    if (
+        _is_cuda
+        and n_share_experts_fusion == 0
+        and is_power_of_two(correction_bias.shape[0])
+    ):
+        return moe_fused_gate(
+            gating_output,
+            correction_bias,
+            num_expert_group,
+            topk_group,
+            topk,
+        )
+    else:
+        biased_grouped_topk_fn = (
+            torch.compile(
+                biased_grouped_topk_impl, dynamic=True, backend=get_compiler_backend()
+            )
+            if compiled
+            else biased_grouped_topk_impl
+        )
+        return biased_grouped_topk_fn(
+            hidden_states,
+            gating_output,
+            correction_bias,
+            topk,
+            renormalize,
+            num_expert_group,
+            topk_group,
+            n_share_experts_fusion=n_share_experts_fusion,
         )
-        if compiled
-        else biased_grouped_topk_impl
-    )
-    return biased_grouped_topk_fn(
-        hidden_states,
-        gating_output,
-        correction_bias,
-        topk,
-        renormalize,
-        num_expert_group,
-        topk_group,
-        n_share_experts_fusion=n_share_experts_fusion,
-    )
 
 
 def select_experts(

sglang/srt/layers/quantization/__init__.py

@@ -59,20 +59,20 @@ from sglang.srt.layers.quantization.compressed_tensors.compressed_tensors import
 )
 from sglang.srt.layers.quantization.fp8 import Fp8Config
 from sglang.srt.layers.quantization.gptq import GPTQConfig, GPTQMarlinConfig
-from sglang.srt.layers.quantization.modelopt_quant import ModelOptFp8Config
+from sglang.srt.layers.quantization.modelopt_quant import (
+    ModelOptFp4Config,
+    ModelOptFp8Config,
+)
 from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
 from sglang.srt.layers.quantization.w8a8_fp8 import W8A8Fp8Config
 from sglang.srt.layers.quantization.w8a8_int8 import W8A8Int8Config
-from sglang.srt.layers.vocab_parallel_embedding import (
-    ParallelLMHead,
-    UnquantizedEmbeddingMethod,
-)
 
 # Base quantization methods that don't depend on vllm
 BASE_QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
     "fp8": Fp8Config,
     "blockwise_int8": BlockInt8Config,
     "modelopt": ModelOptFp8Config,
+    "modelopt_fp4": ModelOptFp4Config,
     "w8a8_int8": W8A8Int8Config,
     "w8a8_fp8": W8A8Fp8Config,
     "moe_wna16": MoeWNA16Config,
@@ -176,6 +176,13 @@ def get_linear_quant_method(
     prefix: str,
     linear_method_cls: type,
 ):
+    # Move import here to avoid circular import. This is only used in monkey patching
+    # of vllm's QuantizationConfig.
+    from sglang.srt.layers.vocab_parallel_embedding import (
+        ParallelLMHead,
+        UnquantizedEmbeddingMethod,
+    )
+
     cloned_config = deepcopy(config)
     parallel_lm_head_quantized = (
         isinstance(layer, ParallelLMHead) and cloned_config.lm_head_quantized
@@ -280,6 +287,7 @@ def monkey_patch_moe_apply(class_obj: "FusedMoEMethodBase"):
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
     ):

sglang/srt/layers/quantization/blockwise_int8.py

@@ -370,6 +370,7 @@ class BlockInt8MoEMethod:
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
     ) -> torch.Tensor:
@@ -398,6 +399,7 @@ class BlockInt8MoEMethod:
             topk_ids=topk_ids,
             inplace=inplace,
             activation=activation,
+            apply_router_weight_on_input=apply_router_weight_on_input,
             use_int8_w8a8=True,
             w1_scale=(layer.w13_weight_scale_inv),
             w2_scale=(layer.w2_weight_scale_inv),