diff --git a/cmake/external_projects/deepgemm.cmake b/cmake/external_projects/deepgemm.cmake
index 72f7f5eb66..2ff69bd8b6 100644
--- a/cmake/external_projects/deepgemm.cmake
+++ b/cmake/external_projects/deepgemm.cmake
@@ -28,9 +28,11 @@ if(DEEPGEMM_SRC_DIR)
   message(STATUS "DeepGEMM using local DEEPGEMM_SRC_DIR: ${deepgemm_SOURCE_DIR}")
 else()
   # Keep in sync with tools/install_deepgemm.sh
-  set(_DEEPGEMM_UPSTREAM_REPO "https://github.com/deepseek-ai/DeepGEMM.git")
-  # Pinned to the tip of the nv_dev branch (SM120 support).
-  set(_DEEPGEMM_UPSTREAM_TAG "8b1392b978f5a03c828dd1711090d7fb50958b8a")
+  set(_DEEPGEMM_UPSTREAM_REPO "https://github.com/vllm-project/DeepGEMM.git")
+  # Pinned to the tip of the fork's dev branch: upstream 2.8.0 plus the SM120
+  # port, the SM90 paged-MQA kv_block=32/next_n=4 port, configurable SwiGLU
+  # alpha/beta, and SiTU for FP8/FP4 Mega MoE.
+  set(_DEEPGEMM_UPSTREAM_TAG "ad1f1726aa540a76c1d26d6a120effb8de21eaa4")
 
   set(_deepgemm_fc_root "${FETCHCONTENT_BASE_DIR}")
   if(NOT _deepgemm_fc_root)
@@ -51,7 +53,7 @@ else()
       BINARY_DIR "${_deepgemm_bin}"
       GIT_REPOSITORY "${_DEEPGEMM_UPSTREAM_REPO}"
       GIT_TAG "${_DEEPGEMM_UPSTREAM_TAG}"
-      GIT_SUBMODULES "third-party/cutlass" "third-party/fmt"
+      GIT_SUBMODULES "third-party/cutlass" "third-party/deep_jit"
       GIT_PROGRESS TRUE
     )
   endif()
@@ -119,13 +121,15 @@ if(DEEPGEMM_ARCHS)
   message(STATUS "DeepGEMM _C will be built for: ${_dg_pythons}")
 
   # add_custom_command does no implicit header scanning; glob explicitly so
-  # header-only edits in DeepGEMM/cutlass/fmt re-trigger the rebuild.
+  # header-only edits in DeepGEMM/cutlass/deep_jit re-trigger the rebuild.
   file(GLOB_RECURSE _dg_headers
     "${deepgemm_SOURCE_DIR}/csrc/*.h"
     "${deepgemm_SOURCE_DIR}/csrc/*.hpp"
     "${deepgemm_SOURCE_DIR}/deep_gemm/include/*.h"
     "${deepgemm_SOURCE_DIR}/deep_gemm/include/*.hpp"
-    "${deepgemm_SOURCE_DIR}/deep_gemm/include/*.cuh")
+    "${deepgemm_SOURCE_DIR}/deep_gemm/include/*.cuh"
+    "${deepgemm_SOURCE_DIR}/third-party/deep_jit/include/*.h"
+    "${deepgemm_SOURCE_DIR}/third-party/deep_jit/include/*.hpp")
 
   set(_dg_markers)
   set(_dg_seen_soabis)
diff --git a/docker/Dockerfile b/docker/Dockerfile
index 9b344552c5..b3e32777e7 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -131,6 +131,7 @@ RUN dnf install -y --setopt=install_weak_deps=False \
         git \
         curl \
         sudo \
+        elfutils-devel \
         rdma-core-devel \
     && dnf clean all \
     && rm -rf /var/cache/dnf
diff --git a/tools/build_deepgemm_C.py b/tools/build_deepgemm_C.py
index 67a527405e..f4438c6755 100644
--- a/tools/build_deepgemm_C.py
+++ b/tools/build_deepgemm_C.py
@@ -50,7 +50,7 @@ includes = [
     str(src / "deep_gemm/include"),
     str(src / "third-party/cutlass/include"),
     str(src / "third-party/cutlass/tools/util/include"),
-    str(src / "third-party/fmt/include"),
+    str(src / "third-party/deep_jit/include"),
     *cpp_extension.include_paths(device_type="cuda"),
 ]
 
diff --git a/tools/install_deepgemm.sh b/tools/install_deepgemm.sh
index f6a64d8796..db81bf4be3 100755
--- a/tools/install_deepgemm.sh
+++ b/tools/install_deepgemm.sh
@@ -6,9 +6,11 @@ set -e
 
 # Default values
 # Keep DEEPGEMM_GIT_REF in sync with cmake/external_projects/deepgemm.cmake
-DEEPGEMM_GIT_REPO="https://github.com/deepseek-ai/DeepGEMM.git"
-# NOTE: This is currently targeting the nv_dev branch tip due to sm120 support
-DEEPGEMM_GIT_REF="8b1392b978f5a03c828dd1711090d7fb50958b8a"
+DEEPGEMM_GIT_REPO="https://github.com/vllm-project/DeepGEMM.git"
+# NOTE: This targets the vLLM fork's dev branch tip, which carries the sm120
+# and sm90 paged-MQA ports plus the SwiGLU alpha/beta and SiTU Mega MoE
+# activations
+DEEPGEMM_GIT_REF="ad1f1726aa540a76c1d26d6a120effb8de21eaa4"
 WHEEL_DIR=""
 
 # Parse command line arguments
diff --git a/vllm/models/kimi_k3/nvidia/model.py b/vllm/models/kimi_k3/nvidia/model.py
index c4f0092f8d..6a26ce78b4 100644
--- a/vllm/models/kimi_k3/nvidia/model.py
+++ b/vllm/models/kimi_k3/nvidia/model.py
@@ -532,8 +532,10 @@ class KimiK3MegaMoEExperts(DeepseekV4MegaMoEExperts):
             symm_buffer,
             activation_clamp=activation_clamp,
             activation=self.activation,
-            situ_beta=self.activation_beta,
-            situ_linear_beta=self.activation_linear_beta,
+            # DeepGEMM names the SiTU gate tanh scale `activation_alpha` and the
+            # linear/up tanh scale `activation_beta`; beta=0 leaves up untouched.
+            activation_alpha=self.activation_beta or 1.0,
+            activation_beta=self.activation_linear_beta or 0.0,
             fast_math=fast_math,
         )
         return y
diff --git a/vllm/utils/deep_gemm.py b/vllm/utils/deep_gemm.py
index 3fc70c0f5d..d515d5131e 100644
--- a/vllm/utils/deep_gemm.py
+++ b/vllm/utils/deep_gemm.py
@@ -280,7 +280,11 @@ def _lazy_init() -> None:
     _fp8_gemm_nt_impl = getattr(_dg, "fp8_gemm_nt", None)
     _fp8_einsum_impl = getattr(_dg, "fp8_einsum", None)
     _grouped_impl = getattr(_dg, "m_grouped_fp8_gemm_nt_contiguous", None)
-    _grouped_masked_impl = getattr(_dg, "fp8_m_grouped_gemm_nt_masked", None)
+    # DeepGEMM 2.8.0 dropped the legacy `fp8_*` aliases; keep both spellings so
+    # an externally pinned older DeepGEMM still resolves.
+    _grouped_masked_impl = getattr(
+        _dg, "m_grouped_fp8_gemm_nt_masked", None
+    ) or getattr(_dg, "fp8_m_grouped_gemm_nt_masked", None)
     _grouped_fp4_impl = getattr(_dg, "m_grouped_fp8_fp4_gemm_nt_contiguous", None)
     # DeepGEMM exposes fp8_fp4_*_mqa_logits as the canonical symbols that
     # handle both the FP8 and FP4 Q/K paths via a tuple-typed `q`.