Support DeepEP for Kimi-k2-thinking through enabling gemm selection for compressed-tensor marlin wna16

vllm/model_executor/layers/fused_moe/fused_marlin_moe.py

@@ -501,7 +501,9 @@ def batched_fused_marlin_moe(
 class MarlinExpertsBase(mk.FusedMoEPermuteExpertsUnpermute):
     def __init__(self, quant_config: FusedMoEQuantConfig):
         # TODO (varun) : Enable activation quantization
-        assert quant_config.use_mxfp4_w4a16, "Supports only mxfp4_w4a16"
+        assert quant_config.use_mxfp4_w4a16 or quant_config.use_int4_w4a16, (
+            "Supports only mxfp4_w4a16 or int4_w4a16"
+        )
         super().__init__(quant_config)
 
     def moe_problem_size(
@@ -616,7 +618,11 @@ def apply(
             gating_output=None,
             topk_weights=topk_weights,
             topk_ids=topk_ids,
-            quant_type_id=scalar_types.float4_e2m1f.id,  # works only for w4a16
+            quant_type_id=(
+                scalar_types.uint4b8.id
+                if self.quant_config.use_int4_w4a16
+                else scalar_types.float4_e2m1f.id
+            ),  # works only for w4a16
             apply_router_weight_on_input=apply_router_weight_on_input,
             global_num_experts=global_num_experts,
             activation=activation,
@@ -720,8 +726,11 @@ def apply(
             w1_scale=self.w1_scale,
             w2_scale=self.w2_scale,
             gating_output=None,
-            quant_type_id=scalar_types.float4_e2m1f.id,  # works only for w4a16
-            apply_router_weight_on_input=apply_router_weight_on_input,
+            quant_type_id=(
+                scalar_types.uint4b8.id
+                if self.quant_config.use_int4_w4a16
+                else scalar_types.float4_e2m1f.id
+            ),  # works only for w4a16
             global_num_experts=global_num_experts,
             activation=activation,
             expert_map=expert_map,

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -35,7 +35,11 @@
 from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
     is_valid_flashinfer_cutlass_fused_moe,
 )
-from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe
+from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
+    BatchedMarlinExperts,
+    MarlinExperts,
+    fused_marlin_moe,
+)
 from vllm.model_executor.layers.quantization.compressed_tensors.schemes.compressed_tensors_wNa16 import (  # noqa
     WNA16_SUPPORTED_BITS,
     WNA16_SUPPORTED_TYPES_MAP,
@@ -1562,7 +1566,42 @@
     def get_fused_moe_quant_config(
         self, layer: torch.nn.Module
     ) -> FusedMoEQuantConfig | None:
-        return None
+        assert self.num_bits == 4 or self.num_bits == 8
+        config_builder = (
+            int4_w4a16_moe_quant_config
+            if self.num_bits == 4
+            else int8_w8a16_moe_quant_config
+        )
+
+        return config_builder(
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            w1_zp=None,
+            w2_zp=None,
+            block_shape=[0, self.group_size],
+        )
+
+    def select_gemm_impl(
+        self,
+        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
+        layer: torch.nn.Module,
+    ) -> mk.FusedMoEPermuteExpertsUnpermute:
+
+        layer.w13_weight = layer.w13_weight_packed
+        layer.w2_weight = layer.w2_weight_packed
+        assert all([w is not None for w in [layer.w13_weight, layer.w2_weight]])
+        assert self.moe_quant_config is not None
+        if (
+            prepare_finalize.activation_format
+            == mk.FusedMoEActivationFormat.BatchedExperts
+        ):
+            return BatchedMarlinExperts(
+                max_num_tokens=prepare_finalize.max_num_tokens_per_rank(),
+                num_dispatchers=prepare_finalize.num_dispatchers(),
+                quant_config=self.moe_quant_config,
+            )
+        else:
+            return MarlinExperts(self.moe_quant_config)
 
     def apply(
         self,