From 589f62c95cf02849dbf1e9631b9b836213136130 Mon Sep 17 00:00:00 2001
From: lazypool <lazypool@proton.me>
Date: Sat, 18 Jul 2026 22:18:54 +0800
Subject: [PATCH] [Bugfix] Propagate VLLM_MARLIN_INPUT_DTYPE into INC/AutoRound
 quant path (#48904)

AutoRound (INC) checkpoints route through inc_wna16_linear.py and
inc_wna16_scheme.py, which construct Marlin linear/MoE methods but
never assign input_dtype.  The attribute defaults to None, so the
Marlin kernel never enters the int8-activation branch regardless of
the VLLM_MARLIN_INPUT_DTYPE env var.

Fix by setting method.input_dtype = get_marlin_input_dtype() after
construction, matching the pattern already used in auto_gptq.py:269
and auto_gptq.py:261 for the GPTQ/AWQ path.

Co-authored-by: noonghunna <10742901+noonghunna@users.noreply.github.com>
Co-authored-by: kevinb361 <91768207+kevinb361@users.noreply.github.com>
Signed-off-by: lazypool <lazypool@proton.me>
---
 .../layers/quantization/inc/schemes/inc_wna16_linear.py    | 5 ++++-
 .../layers/quantization/inc/schemes/inc_wna16_scheme.py    | 7 ++++++-
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_linear.py b/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_linear.py
index dd6c2fa2eaae..ebcf6dcaf862 100644
--- a/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_linear.py
+++ b/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_linear.py
@@ -10,6 +10,7 @@
 from vllm.model_executor.layers.quantization.auto_gptq import AutoGPTQConfig
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     check_marlin_supported,
+    get_marlin_input_dtype,
 )
 from vllm.model_executor.parameter import (
     GroupQuantScaleParameter,
@@ -62,7 +63,7 @@ def _build_gptq_method(self):
                 AutoGPTQLinearMethod,
             )
 
-            return AutoGPTQLinearMethod(
+            method = AutoGPTQLinearMethod(
                 AutoGPTQConfig(
                     weight_bits=self.layer_config.bits,
                     group_size=self.layer_config.group_size,
@@ -73,6 +74,8 @@ def _build_gptq_method(self):
                     full_config={},
                 )
             )
+            method.input_dtype = get_marlin_input_dtype()
+            return method
 
         raise NotImplementedError(
             f"INC quantization with bits={self.layer_config.bits}, "
diff --git a/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_scheme.py b/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_scheme.py
index 46ad24ea5b47..0354e096de55 100644
--- a/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_scheme.py
+++ b/vllm/model_executor/layers/quantization/inc/schemes/inc_wna16_scheme.py
@@ -6,6 +6,9 @@
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig
 from vllm.model_executor.layers.quantization.auto_gptq import AutoGPTQConfig
+from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+    get_marlin_input_dtype,
+)
 from vllm.platforms import current_platform
 from vllm.scalar_type import scalar_types
 
@@ -137,7 +140,7 @@ def _resolve_gptq_moe(layer: "torch.nn.Module", layer_config: "INCLayerConfig"):
         ) and check_moe_marlin_supports_layer(layer, layer_config.group_size)
 
     if use_marlin:
-        return AutoGPTQMoEMethod(
+        moe_method = AutoGPTQMoEMethod(
             AutoGPTQConfig(
                 weight_bits=layer_config.bits,
                 group_size=layer_config.group_size,
@@ -149,6 +152,8 @@ def _resolve_gptq_moe(layer: "torch.nn.Module", layer_config: "INCLayerConfig"):
             ),
             layer.moe_config,
         )
+        moe_method.input_dtype = get_marlin_input_dtype()
+        return moe_method
 
     moe_config = MoeWNA16Config.from_config(
         {