update transformers package version to 5.0.0rc0

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",

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",

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",

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

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

scripts/ci/ci_install_dependency.sh

@@ -106,6 +106,16 @@ $PIP_CMD install -e "python[${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)
@@ -147,3 +157,14 @@ python3 -c "import torch; print(torch.version.cuda)"
 
 # Prepare the CI runner (cleanup HuggingFace cache, etc.)
 bash "${SCRIPT_DIR}/prepare_runner.sh"
+
+PYTHON_LIB_PATH=$(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 "181 Directory $FLASH_ATTN_PATH does not exist."
+fi

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 "${repo-flash-attention_SOURCE_DIR}/flash_attn/cute/"
-            DESTINATION "flash_attn/cute"
+
+    set(FLASH_ATTN_CUTE_SRC "${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()