[sync] vLLM 2026-08-03 — 5233368da..42ab184ea

.sync/vllm-sha

@@ -1 +1 @@
-82ae4164ee016d4daecd2033c26f5c0827984a80
+5233368daccd3d84293bab7b04bacae1433ea0e4

aphrodite/model_executor/layers/fused_moe/experts/ocp_mx_emulation_moe.py

@@ -68,9 +68,15 @@ def __init__(
         self.quantization_emulation = True
 
         if self.ocp_mx_scheme in {
+            OCP_MX_Scheme.w_mxfp4,
+            OCP_MX_Scheme.w_mxfp6_e3m2,
+            OCP_MX_Scheme.w_mxfp6_e2m3,
+        }:
+            # Weight-only schemes leave activations unquantized.
+            self._quant_dtype = None
+        elif self.ocp_mx_scheme in {
             OCP_MX_Scheme.w_mxfp4_a_mxfp4,
         }:
-            # Weight has to be dequantized for mxfp4 emulation.
             self._quant_dtype = "mxfp4"
         elif self.ocp_mx_scheme in [
             OCP_MX_Scheme.w_mxfp4_a_mxfp6_e3m2,

aphrodite/model_executor/layers/fused_moe/experts/triton_moe.py

@@ -129,6 +129,7 @@ def _supports_activation(activation: MoEActivation) -> bool:
             MoEActivation.SILU,
             MoEActivation.GELU,
             MoEActivation.GELU_TANH,
+            MoEActivation.SITU,
             MoEActivation.SWIGLUOAI,
             MoEActivation.SWIGLUOAI_UNINTERLEAVE,
             MoEActivation.SWIGLUSTEP,

aphrodite/model_executor/layers/fused_moe/oracle/mxfp4.py

@@ -333,7 +333,10 @@ def _get_priority_backends() -> list[Mxfp4MoeBackend]:
     backend-level ``is_supported_config`` check filters by device capability).
     """
     if current_platform.is_rocm():
-        return [Mxfp4MoeBackend.AITER_MXFP4_BF16]
+        return [
+            Mxfp4MoeBackend.AITER_MXFP4_BF16,
+            Mxfp4MoeBackend.EMULATION,
+        ]
     if current_platform.is_xpu():
         return [Mxfp4MoeBackend.XPU]
     _AVAILABLE_BACKENDS = [
@@ -1091,8 +1094,12 @@ def swap_every_two_rows(x, axis=-1):
             w13_bias,
             w2_bias,
         )
-    elif mxfp4_backend == Mxfp4MoeBackend.XPU:
-        # No additional transformation needed for XPU backend
+    elif mxfp4_backend in (
+        Mxfp4MoeBackend.XPU,
+        Mxfp4MoeBackend.EMULATION,
+    ):
+        # No additional transformation is needed: XPU consumes the checkpoint
+        # layout directly, while emulation dequantizes that layout at runtime.
         return (
             w13_weight,
             w2_weight,
@@ -1460,7 +1467,7 @@ def shuffle_weight(w: torch.Tensor) -> torch.Tensor:
     else:
         raise ValueError(
             f"Unsupported mxfp4_backend for Mxfp4MoEMethod: {mxfp4_backend}. "
-            f"Expected TRTLLM, Triton, AITER, or XPU backend."
+            f"Expected TRTLLM, Triton, AITER, XPU, or emulation backend."
         )
 
 

aphrodite/model_executor/layers/quantization/mxfp4.py

@@ -15,6 +15,9 @@
     SharedExperts,
 )
 from aphrodite.model_executor.layers.fused_moe import modular_kernel as mk
+from aphrodite.model_executor.layers.fused_moe.config import (
+    mxfp4_w4a16_moe_quant_config,
+)
 from aphrodite.model_executor.layers.fused_moe.oracle.mxfp4 import (
     TRITON_BACKENDS,
     Mxfp4MoeBackend,
@@ -832,6 +835,18 @@ def get_fused_moe_quant_config(
             w1_scale = layer.w13_weight_scale
             w2_scale = layer.w2_weight_scale
 
+        if self.mxfp4_backend == Mxfp4MoeBackend.EMULATION:
+            # Canonical ``mxfp4`` checkpoints are weight-only W4A16. The
+            # generic EMULATION config is W4A4, so preserve BF16 activations
+            # while the fallback dequantizes only the weights.
+            return mxfp4_w4a16_moe_quant_config(
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                w1_bias=w1_bias,
+                w2_bias=w2_bias,
+                gemm1_clamp_limit=swiglu_limit,
+            )
+
         return make_mxfp4_moe_quant_config(
             mxfp4_backend=self.mxfp4_backend,
             w1_scale=w1_scale,