[Feat] Logger redesign (#1015)

## Purpose
Current logger has self reference problem, causing npu to out of memory.
and the logger output is incomplete after the vllm has stopped.

## Modifications 
refactor logger, python use std logger lib, add handler to redirect to
c++ spd log.
use async logger for better performance.
do not create empty log file, all log files are in same directory for
better user experience.
capture vllm log by add provider to it, avoid hook vllm logger.

## Test
verified using jenkins pipeline, works as expect, no performance impact.

Author: dante159753

.github/workflows/pull-request.yml

@@ -101,7 +101,7 @@ jobs:
           cache-from: type=gha,scope=npu
           cache-to: type=gha,mode=max,scope=npu,ignore-error=true
 
-  test-build-vllm-ascend-v0-20-2:
+  test-build-vllm-ascend-deepseekv4:
     timeout-minutes: 25
     runs-on: ubuntu-24.04-arm
     steps:
@@ -116,14 +116,14 @@ jobs:
         uses: docker/build-push-action@v5
         with:
           context: .
-          file: ./docker/Dockerfile.ucm-vllm-ascend.a2-v0.20.2
+          file: ./docker/Dockerfile.ucm-vllm-ascend.a2-deepseekv4
           build-args: |
             PIP_INDEX_URL=https://pypi.org/simple
-          tags: ucm-npu-v0.20.2:latest
+          tags: ucm-npu-deepseekv4:latest
           push: false
           load: false
-          cache-from: type=gha,scope=npu-v0.20.2
-          cache-to: type=gha,mode=max,scope=npu-v0.20.2,ignore-error=true
+          cache-from: type=gha,scope=npu-deepseekv4
+          cache-to: type=gha,mode=max,scope=npu-deepseekv4,ignore-error=true
 
   test-build-vllm-cuda-v0-20-2:
     timeout-minutes: 25

…er/Dockerfile.ucm-vllm-ascend.a2-v0.20.2 …Dockerfile.ucm-vllm-ascend.a2-deepseekv4docker/Dockerfile.ucm-vllm-ascend.a2-v0.20.2 renamed to docker/Dockerfile.ucm-vllm-ascend.a2-deepseekv4 docker/Dockerfile.ucm-vllm-ascend.a2-v0.20.2 renamed to docker/Dockerfile.ucm-vllm-ascend.a2-deepseekv4

@@ -1,26 +1,25 @@
 # Set to other image if needed
 ARG IMAGE_SOURCE="quay.io/ascend"
-ARG IMAGE_NAME_VERSION="vllm-ascend:v0.11.0"
+ARG IMAGE_NAME_VERSION="vllm-ascend:v0.18.0"
 
 FROM ${IMAGE_SOURCE}/${IMAGE_NAME_VERSION}
 
 ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
 ARG INSTALL_MODE="source"
 
-# Apply the UCM monkey patch for vllm & vllm_ascend
-ENV ENABLE_UCM_PATCH=1
-
 WORKDIR /workspace
 
 # Install unified-cache-management
 COPY . /workspace/unified-cache-management
 
 RUN pip config set global.index-url ${PIP_INDEX_URL}
 
+RUN pip install --no-cache-dir "transformers==5.4"
+
 # Build or link package
 RUN if [ "${INSTALL_MODE}" != "package" ]; then \
         pip install --no-cache-dir build cmake && \
-        export WORKSPACE=/workspace SKIP_TAR=1 && \
+        export WORKSPACE=/workspace SKIP_TAR=1 ENABLE_SPARSE=false && \
         bash /workspace/unified-cache-management/scripts/build_ascend.sh; \
     else \
         ln -s /workspace/unified-cache-management /workspace/package; \
@@ -29,9 +28,4 @@ RUN if [ "${INSTALL_MODE}" != "package" ]; then \
 # Install UCM
 RUN pip install /workspace/package/uc_manager-*.whl
 
-# Install Ascend custom ops if present
-RUN if [ -f /workspace/package/install_ascend_ops.sh ]; then \
-        cd /workspace/package && bash install_ascend_ops.sh; \
-    fi
-
 CMD ["/bin/bash"]

…er/Dockerfile.ucm-vllm-ascend.a2-v0.11.0 …kerfile.ucm-vllm-ascend.a2-v0.18.0glm5.1docker/Dockerfile.ucm-vllm-ascend.a2-v0.11.0 renamed to docker/Dockerfile.ucm-vllm-ascend.a2-v0.18.0glm5.1 docker/Dockerfile.ucm-vllm-ascend.a2-v0.11.0 renamed to docker/Dockerfile.ucm-vllm-ascend.a2-v0.18.0glm5.1

@@ -1,15 +1,12 @@
 # Set to other image if needed
-ARG IMAGE_SOURCE="vllm"
-ARG IMAGE_NAME_VERSION="vllm-openai:v0.11.0"
+ARG IMAGE_SOURCE="quay.io/ascend"
+ARG IMAGE_NAME_VERSION="vllm-ascend:v0.20.2rc1"
 
 FROM ${IMAGE_SOURCE}/${IMAGE_NAME_VERSION}
 
 ARG PIP_INDEX_URL="https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple"
 ARG INSTALL_MODE="source"
 
-# Apply the UCM monkey patch for vllm
-ENV ENABLE_UCM_PATCH=1
-
 WORKDIR /workspace
 
 # Install unified-cache-management
@@ -20,13 +17,13 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 # Build or link package
 RUN if [ "${INSTALL_MODE}" != "package" ]; then \
         pip install --no-cache-dir build cmake && \
-        export WORKSPACE=/workspace SKIP_TAR=1 && \
-        bash /workspace/unified-cache-management/scripts/build_cuda.sh; \
+        export WORKSPACE=/workspace SKIP_TAR=1 ENABLE_SPARSE=false && \
+        bash /workspace/unified-cache-management/scripts/build_ascend.sh; \
     else \
         ln -s /workspace/unified-cache-management /workspace/package; \
     fi
 
 # Install UCM
 RUN pip install /workspace/package/uc_manager-*.whl
 
-ENTRYPOINT ["/bin/bash"]
+CMD ["/bin/bash"]

docker/Dockerfile.ucm-vllm-cuda-v0.11.0 …Dockerfile.ucm-vllm-ascend.a2-v0.20.2rc1docker/Dockerfile.ucm-vllm-cuda-v0.11.0 renamed to docker/Dockerfile.ucm-vllm-ascend.a2-v0.20.2rc1 docker/Dockerfile.ucm-vllm-cuda-v0.11.0 renamed to docker/Dockerfile.ucm-vllm-ascend.a2-v0.20.2rc1

@@ -1,9 +1,24 @@
-from ucm.integration.vllm.patch.utils import patch_or_inject, when_imported
+import logging
+import os
+
+from ucm.integration.vllm.patch.utils import when_imported
+
+
+def _capture_enabled() -> bool:
+    value = os.getenv("UCM_CAPTURE_VLLM_LOG", "1").strip().lower()
+    return value in ("1", "true", "yes", "on")
 
 
 @when_imported("vllm.logger")
 def patch_logger(mod):
-    from ucm import logger
+    if not _capture_enabled():
+        return
+
+    from ucm.logger import UcmBridgeHandler, get_vllm_capture_handler
 
-    patch_or_inject(mod, "init_logger", logger.init_logger)
-    patch_or_inject(mod, "current_formatter_type", logger.current_formatter_type)
+    vllm_root = logging.getLogger("vllm")
+    if any(isinstance(h, UcmBridgeHandler) for h in vllm_root.handlers):
+        return
+    handler = get_vllm_capture_handler()
+    if handler is not None:
+        vllm_root.addHandler(handler)