From 9c001e9769116cba36261eb06e50a579cafc785a Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Sun, 7 Dec 2025 09:17:59 +0000
Subject: [PATCH 1/8] upd

---
 python/pyproject.toml                          | 2 +-
 python/pyproject_cpu.toml                      | 2 +-
 python/pyproject_other.toml                    | 2 +-
 python/pyproject_xpu.toml                      | 2 +-
 python/sglang/srt/configs/qwen3_omni.py        | 4 ----
 scripts/ci/ci_install_dependency.sh            | 8 ++++++++
 sgl-kernel/CMakeLists.txt                      | 2 +-
 sgl-kernel/python/sgl_kernel/_fa4_interface.py | 6 +++---
 8 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/python/pyproject.toml b/python/pyproject.toml
index 85428cba02d8..2bf00e5dbb33 100755
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -69,7 +69,7 @@ dependencies = [
   "torchvision",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.0.0rc0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.27",
diff --git a/python/pyproject_cpu.toml b/python/pyproject_cpu.toml
index a537bc3607ef..5421f2b7cfbf 100644
--- a/python/pyproject_cpu.toml
+++ b/python/pyproject_cpu.toml
@@ -59,7 +59,7 @@ dependencies = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.0.0rc0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.27",
diff --git a/python/pyproject_other.toml b/python/pyproject_other.toml
index 3e854caa726a..ae1622363c41 100755
--- a/python/pyproject_other.toml
+++ b/python/pyproject_other.toml
@@ -59,7 +59,7 @@ runtime_common = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.0.0rc0",
   "uvicorn",
   "uvloop",
   "xgrammar==0.1.27",
diff --git a/python/pyproject_xpu.toml b/python/pyproject_xpu.toml
index a058e6f48a82..3d7dc969a4b5 100644
--- a/python/pyproject_xpu.toml
+++ b/python/pyproject_xpu.toml
@@ -63,7 +63,7 @@ dependencies = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==4.57.1",
+  "transformers==5.0.0rc0",
   "uvicorn",
   "uvloop",
   # "xgrammar==0.1.24", , xgrammar depends on CUDA PyTorch and Triton only
diff --git a/python/sglang/srt/configs/qwen3_omni.py b/python/sglang/srt/configs/qwen3_omni.py
index d42e98a9a07b..8baea892335d 100644
--- a/python/sglang/srt/configs/qwen3_omni.py
+++ b/python/sglang/srt/configs/qwen3_omni.py
@@ -1,6 +1,5 @@
 from transformers import PretrainedConfig
 from transformers.configuration_utils import layer_type_validation
-from transformers.modeling_rope_utils import rope_config_validation
 
 from sglang.utils import logger
 
@@ -168,7 +167,6 @@ def __init__(
         # BC: if there is a 'type' field, move it to 'rope_type'.
         if self.rope_scaling is not None and "type" in self.rope_scaling:
             self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-        rope_config_validation(self)
 
         # MoE arguments
         self.decoder_sparse_step = decoder_sparse_step
@@ -311,7 +309,6 @@ def __init__(
         # BC: if there is a 'type' field, move it to 'rope_type'.
         if self.rope_scaling is not None and "type" in self.rope_scaling:
             self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-        rope_config_validation(self)
 
         self.layer_types = layer_types
         if self.layer_types is None:
@@ -405,7 +402,6 @@ def __init__(
         # BC: if there is a 'type' field, move it to 'rope_type'.
         if self.rope_scaling is not None and "type" in self.rope_scaling:
             self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-        rope_config_validation(self)
 
         # MoE arguments
         self.decoder_sparse_step = decoder_sparse_step
diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh
index 7d0735602452..96dc3a2410f2 100755
--- a/scripts/ci/ci_install_dependency.sh
+++ b/scripts/ci/ci_install_dependency.sh
@@ -147,3 +147,11 @@ python3 -c "import torch; print(torch.version.cuda)"
 
 # Prepare the CI runner (cleanup HuggingFace cache, etc.)
 bash "${SCRIPT_DIR}/prepare_runner.sh"
+
+if [ -d "/usr/local/lib/python3.10/dist-packages/flash_attn" ]; then
+    echo "Directory /usr/local/lib/python3.10/dist-packages/flash_attn exists."
+else
+    echo "Directory /usr/local/lib/python3.10/dist-packages/flash_attn does not exist."
+fi
+
+rm -rf /usr/local/lib/python3.10/dist-packages/flash_attn
diff --git a/sgl-kernel/CMakeLists.txt b/sgl-kernel/CMakeLists.txt
index 079eb2f45a3e..c4b6172690ea 100644
--- a/sgl-kernel/CMakeLists.txt
+++ b/sgl-kernel/CMakeLists.txt
@@ -595,7 +595,7 @@ install(DIRECTORY "${repo-triton_SOURCE_DIR}/python/triton_kernels/triton_kernel
 if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_SM100A)
     # flash_attn/cute
     install(DIRECTORY "${repo-flash-attention_SOURCE_DIR}/flash_attn/cute/"
-            DESTINATION "flash_attn/cute"
+            DESTINATION "flash_attn_origin/cute"
             PATTERN ".git*" EXCLUDE
             PATTERN "__pycache__" EXCLUDE)
     endif()
diff --git a/sgl-kernel/python/sgl_kernel/_fa4_interface.py b/sgl-kernel/python/sgl_kernel/_fa4_interface.py
index 5411cfe16bb0..ffd0563f56dc 100644
--- a/sgl-kernel/python/sgl_kernel/_fa4_interface.py
+++ b/sgl-kernel/python/sgl_kernel/_fa4_interface.py
@@ -19,9 +19,9 @@
 import cutlass.cute as cute
 import torch
 from cutlass.cute.runtime import from_dlpack
-from flash_attn.cute import utils
-from flash_attn.cute.flash_fwd import FlashAttentionForwardSm90
-from flash_attn.cute.flash_fwd_sm100 import FlashAttentionForwardSm100
+from flash_attn_origin.cute import utils
+from flash_attn_origin.cute.flash_fwd import FlashAttentionForwardSm90
+from flash_attn_origin.cute.flash_fwd_sm100 import FlashAttentionForwardSm100
 
 
 def maybe_contiguous(x):

From fd57e5b247127920f2c55a0160bb4a5dd16435c4 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Mon, 8 Dec 2025 04:11:20 +0000
Subject: [PATCH 2/8] upd

---
 sgl-kernel/CMakeLists.txt                     | 37 +++++++++++++++++--
 .../python/sgl_kernel/_fa4_interface.py       |  6 +--
 2 files changed, 37 insertions(+), 6 deletions(-)

diff --git a/sgl-kernel/CMakeLists.txt b/sgl-kernel/CMakeLists.txt
index c4b6172690ea..9a438591a7aa 100644
--- a/sgl-kernel/CMakeLists.txt
+++ b/sgl-kernel/CMakeLists.txt
@@ -593,9 +593,40 @@ install(DIRECTORY "${repo-triton_SOURCE_DIR}/python/triton_kernels/triton_kernel
 # ============================ Extra Install: FA4 ============================= #
 # TODO: find a better install condition.
 if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_SM100A)
-    # flash_attn/cute
-    install(DIRECTORY "${repo-flash-attention_SOURCE_DIR}/flash_attn/cute/"
+
+    set(FLASH_ATTN_CUTE_SRC "${repo-flash-attention_SOURCE_DIR}/flash_attn/cute")
+    set(FLASH_ATTN_CUTE_DST "${CMAKE_CURRENT_BINARY_DIR}/flash_attn_origin/cute")
+
+    file(MAKE_DIRECTORY "${FLASH_ATTN_CUTE_DST}")
+
+    file(COPY "${FLASH_ATTN_CUTE_SRC}/"
+         DESTINATION "${FLASH_ATTN_CUTE_DST}"
+         PATTERN ".git*" EXCLUDE
+         PATTERN "__pycache__" EXCLUDE)
+
+    file(GLOB_RECURSE FLASH_ATTN_CUTE_DST_PY
+         "${FLASH_ATTN_CUTE_DST}/*.py")
+
+    foreach(FILE_PATH IN LISTS FLASH_ATTN_CUTE_DST_PY)
+        file(READ "${FILE_PATH}" FILE_CONTENT)
+
+        set(MODIFIED_CONTENT "${FILE_CONTENT}")
+
+        # The main goal is to avoid using "flash_attn" so that other libraries (such as transformers) do not mistakenly assume that "flash_attn" is already installed.
+
+        string(REPLACE "flash_attn.cute"
+                       "flash_attn_origin.cute"
+                       MODIFIED_CONTENT "${MODIFIED_CONTENT}")
+
+        if (NOT FILE_CONTENT STREQUAL MODIFIED_CONTENT)
+            file(WRITE "${FILE_PATH}" "${MODIFIED_CONTENT}")
+            message(STATUS "  - [FA4 Patch] Patched: ${FILE_PATH}")
+        endif()
+    endforeach()
+
+    install(DIRECTORY "${FLASH_ATTN_CUTE_DST}/"
             DESTINATION "flash_attn_origin/cute"
             PATTERN ".git*" EXCLUDE
             PATTERN "__pycache__" EXCLUDE)
-    endif()
+
+endif()
diff --git a/sgl-kernel/python/sgl_kernel/_fa4_interface.py b/sgl-kernel/python/sgl_kernel/_fa4_interface.py
index ffd0563f56dc..5411cfe16bb0 100644
--- a/sgl-kernel/python/sgl_kernel/_fa4_interface.py
+++ b/sgl-kernel/python/sgl_kernel/_fa4_interface.py
@@ -19,9 +19,9 @@
 import cutlass.cute as cute
 import torch
 from cutlass.cute.runtime import from_dlpack
-from flash_attn_origin.cute import utils
-from flash_attn_origin.cute.flash_fwd import FlashAttentionForwardSm90
-from flash_attn_origin.cute.flash_fwd_sm100 import FlashAttentionForwardSm100
+from flash_attn.cute import utils
+from flash_attn.cute.flash_fwd import FlashAttentionForwardSm90
+from flash_attn.cute.flash_fwd_sm100 import FlashAttentionForwardSm100
 
 
 def maybe_contiguous(x):

From 48351837c7765a4b6a06222a9add7a5c46fe4327 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Mon, 8 Dec 2025 05:14:28 +0000
Subject: [PATCH 3/8] update

---
 scripts/ci/ci_install_dependency.sh | 45 +++++++++++++++++++++++++----
 1 file changed, 40 insertions(+), 5 deletions(-)

diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh
index 96dc3a2410f2..371761d3e53c 100755
--- a/scripts/ci/ci_install_dependency.sh
+++ b/scripts/ci/ci_install_dependency.sh
@@ -1,5 +1,18 @@
 #!/bin/bash
 # Install the dependency in CI.
+
+PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
+FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
+
+if [ -d "$FLASH_ATTN_PATH" ]; then
+    echo "8 Directory $FLASH_ATTN_PATH exists. Removing..."
+    rm -rf "$FLASH_ATTN_PATH"
+else
+    echo "11 Directory $FLASH_ATTN_PATH does not exist."
+fi
+
+rm -rf $FLASH_ATTN_PATH
+
 set -euxo pipefail
 
 IS_BLACKWELL=${IS_BLACKWELL:-0}
@@ -80,6 +93,7 @@ if [ "$IS_BLACKWELL" = "1" ]; then
     # Clean up existing installations
     $PIP_CMD uninstall -y sgl-kernel sglang $PIP_INSTALL_SUFFIX || true
     $PIP_CMD uninstall -y flashinfer-python flashinfer-cubin flashinfer-jit-cache $PIP_INSTALL_SUFFIX || true
+    $PIP_CMD install "importlib-metadata>=4.7" $PIP_INSTALL_SUFFIX
 else
     # In normal cases, we use uv, which is much faster than pip.
     pip install --upgrade pip
@@ -106,6 +120,16 @@ $PIP_CMD install -e "python[${EXTRAS}]" --extra-index-url https://download.pytor
 # Install router for pd-disagg test
 $PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX
 
+PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
+FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
+
+if [ -d "$FLASH_ATTN_PATH" ]; then
+    echo "126 Directory $FLASH_ATTN_PATH exists. Removing..."
+    rm -rf "$FLASH_ATTN_PATH"
+else
+    echo "129 Directory $FLASH_ATTN_PATH does not exist."
+fi
+
 # Install sgl-kernel
 SGL_KERNEL_VERSION_FROM_KERNEL=$(grep -Po '(?<=^version = ")[^"]*' sgl-kernel/pyproject.toml)
 SGL_KERNEL_VERSION_FROM_SRT=$(grep -Po -m1 '(?<=sgl-kernel==)[0-9A-Za-z\.\-]+' python/pyproject.toml)
@@ -119,6 +143,14 @@ if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ]; then
     else
         WHEEL_ARCH="x86_64"
     fi
+
+    WHEEL_PATH="sgl-kernel/dist/sgl_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl"
+    if [ -f "$WHEEL_PATH" ]; then
+        echo "147 Wheel file $WHEEL_PATH exists."
+    else
+        echo "149 Error: Wheel file $WHEEL_PATH does not exist."
+    fi
+
     $PIP_CMD install sgl-kernel/dist/sgl_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl --force-reinstall $PIP_INSTALL_SUFFIX
 else
     $PIP_CMD install sgl-kernel==${SGL_KERNEL_VERSION_FROM_SRT} --force-reinstall $PIP_INSTALL_SUFFIX
@@ -148,10 +180,13 @@ python3 -c "import torch; print(torch.version.cuda)"
 # Prepare the CI runner (cleanup HuggingFace cache, etc.)
 bash "${SCRIPT_DIR}/prepare_runner.sh"
 
-if [ -d "/usr/local/lib/python3.10/dist-packages/flash_attn" ]; then
-    echo "Directory /usr/local/lib/python3.10/dist-packages/flash_attn exists."
+PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
+FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
+
+if [ -d "$FLASH_ATTN_PATH" ]; then
+    echo "178 Directory $FLASH_ATTN_PATH exists. Removing..."
+    rm -rf "$FLASH_ATTN_PATH"
+    echo "Bug this should not happen"
 else
-    echo "Directory /usr/local/lib/python3.10/dist-packages/flash_attn does not exist."
+    echo "181 Directory $FLASH_ATTN_PATH does not exist."
 fi
-
-rm -rf /usr/local/lib/python3.10/dist-packages/flash_attn

From d39f49382c4850091138684da7a145fb7b53f305 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Mon, 8 Dec 2025 05:16:22 +0000
Subject: [PATCH 4/8] update

---
 scripts/ci/ci_install_dependency.sh | 8 --------
 1 file changed, 8 deletions(-)

diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh
index 371761d3e53c..92b8d570194b 100755
--- a/scripts/ci/ci_install_dependency.sh
+++ b/scripts/ci/ci_install_dependency.sh
@@ -143,14 +143,6 @@ if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ]; then
     else
         WHEEL_ARCH="x86_64"
     fi
-
-    WHEEL_PATH="sgl-kernel/dist/sgl_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl"
-    if [ -f "$WHEEL_PATH" ]; then
-        echo "147 Wheel file $WHEEL_PATH exists."
-    else
-        echo "149 Error: Wheel file $WHEEL_PATH does not exist."
-    fi
-
     $PIP_CMD install sgl-kernel/dist/sgl_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl --force-reinstall $PIP_INSTALL_SUFFIX
 else
     $PIP_CMD install sgl-kernel==${SGL_KERNEL_VERSION_FROM_SRT} --force-reinstall $PIP_INSTALL_SUFFIX

From bdddf7b7bc566e4ce7bab8a66f3883d40966a888 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Mon, 8 Dec 2025 05:18:26 +0000
Subject: [PATCH 5/8] updatee

---
 scripts/ci/ci_install_dependency.sh | 1 -
 1 file changed, 1 deletion(-)

diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh
index 92b8d570194b..0140d98b1405 100755
--- a/scripts/ci/ci_install_dependency.sh
+++ b/scripts/ci/ci_install_dependency.sh
@@ -93,7 +93,6 @@ if [ "$IS_BLACKWELL" = "1" ]; then
     # Clean up existing installations
     $PIP_CMD uninstall -y sgl-kernel sglang $PIP_INSTALL_SUFFIX || true
     $PIP_CMD uninstall -y flashinfer-python flashinfer-cubin flashinfer-jit-cache $PIP_INSTALL_SUFFIX || true
-    $PIP_CMD install "importlib-metadata>=4.7" $PIP_INSTALL_SUFFIX
 else
     # In normal cases, we use uv, which is much faster than pip.
     pip install --upgrade pip

From f0f147fa58b876f388fcb583334a2ba6ffed405f Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Mon, 8 Dec 2025 05:19:39 +0000
Subject: [PATCH 6/8] update

---
 scripts/ci/ci_install_dependency.sh | 13 -------------
 1 file changed, 13 deletions(-)

diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh
index 0140d98b1405..549659a15243 100755
--- a/scripts/ci/ci_install_dependency.sh
+++ b/scripts/ci/ci_install_dependency.sh
@@ -1,18 +1,5 @@
 #!/bin/bash
 # Install the dependency in CI.
-
-PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
-FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
-
-if [ -d "$FLASH_ATTN_PATH" ]; then
-    echo "8 Directory $FLASH_ATTN_PATH exists. Removing..."
-    rm -rf "$FLASH_ATTN_PATH"
-else
-    echo "11 Directory $FLASH_ATTN_PATH does not exist."
-fi
-
-rm -rf $FLASH_ATTN_PATH
-
 set -euxo pipefail
 
 IS_BLACKWELL=${IS_BLACKWELL:-0}

From bb72126490c9993658bef5e103b4dbcadbdd9b83 Mon Sep 17 00:00:00 2001
From: yhyang201 <yhyang201@gmail.com>
Date: Mon, 8 Dec 2025 06:12:36 +0000
Subject: [PATCH 7/8] upd

---
 sgl-kernel/python/sgl_kernel/_fa4_interface.py | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sgl-kernel/python/sgl_kernel/_fa4_interface.py b/sgl-kernel/python/sgl_kernel/_fa4_interface.py
index 5411cfe16bb0..ffd0563f56dc 100644
--- a/sgl-kernel/python/sgl_kernel/_fa4_interface.py
+++ b/sgl-kernel/python/sgl_kernel/_fa4_interface.py
@@ -19,9 +19,9 @@
 import cutlass.cute as cute
 import torch
 from cutlass.cute.runtime import from_dlpack
-from flash_attn.cute import utils
-from flash_attn.cute.flash_fwd import FlashAttentionForwardSm90
-from flash_attn.cute.flash_fwd_sm100 import FlashAttentionForwardSm100
+from flash_attn_origin.cute import utils
+from flash_attn_origin.cute.flash_fwd import FlashAttentionForwardSm90
+from flash_attn_origin.cute.flash_fwd_sm100 import FlashAttentionForwardSm100
 
 
 def maybe_contiguous(x):

From dfb66260b56e4bff62f7c3d6b8923216d6643df1 Mon Sep 17 00:00:00 2001
From: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Date: Mon, 8 Dec 2025 16:06:54 +0800
Subject: [PATCH 8/8] Improve echo messages in ci_install_dependency.sh

---
 scripts/ci/ci_install_dependency.sh | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh
index 549659a15243..09fdb4104a7e 100755
--- a/scripts/ci/ci_install_dependency.sh
+++ b/scripts/ci/ci_install_dependency.sh
@@ -110,10 +110,10 @@ PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
 FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
 
 if [ -d "$FLASH_ATTN_PATH" ]; then
-    echo "126 Directory $FLASH_ATTN_PATH exists. Removing..."
+    echo "Directory $FLASH_ATTN_PATH exists. Removing..."
     rm -rf "$FLASH_ATTN_PATH"
 else
-    echo "129 Directory $FLASH_ATTN_PATH does not exist."
+    echo "Directory $FLASH_ATTN_PATH does not exist."
 fi
 
 # Install sgl-kernel
@@ -162,9 +162,9 @@ PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
 FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
 
 if [ -d "$FLASH_ATTN_PATH" ]; then
-    echo "178 Directory $FLASH_ATTN_PATH exists. Removing..."
+    echo "Directory $FLASH_ATTN_PATH exists. Removing..."
     rm -rf "$FLASH_ATTN_PATH"
-    echo "Bug this should not happen"
+    echo "error: this should not happen"
 else
-    echo "181 Directory $FLASH_ATTN_PATH does not exist."
+    echo "Directory $FLASH_ATTN_PATH does not exist."
 fi