From be4198e754d43a86115468b64faf0a8833141209 Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Sat, 11 Apr 2026 23:19:48 +0000
Subject: [PATCH] Add CMM_BUILD_DATE cache-bust arg to Dockerfile

Docker was reusing cached layers for sitecustomize.py and
vllm_managed_mem.py COPY steps, causing stale files in the image.
CMM_BUILD_DATE arg invalidates the cache for all CMM file copies.
---
 vllm/Dockerfile | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/vllm/Dockerfile b/vllm/Dockerfile
index 7f60fd3..f1746a9 100644
--- a/vllm/Dockerfile
+++ b/vllm/Dockerfile
@@ -257,14 +257,15 @@ RUN pip uninstall -y pynvml && pip install nvidia-ml-py
 #   - cudaMemPrefetchAsync(GPU): actively migrates pages to GPU immediately,
 #     so model weight writes go to HBM/EGM, not system RAM
 #
-# vllm_managed_mem.py is the launcher that swaps the allocator before any
-# CUDA operations and patches vLLM's memory validation to understand the
-# larger managed memory space.
+# vllm_managed_mem.py is the launcher that patches vLLM's memory validation
+# to understand the larger managed memory space. No global allocator swap.
 #
 # sitecustomize.py is auto-loaded by Python in ALL subprocesses (including
-# vLLM's EngineCore). It applies the allocator swap and torch.cuda patches
-# before any CUDA operations in spawned processes.
+# vLLM's EngineCore). It sets VLLM_KV_CACHE_USE_MANAGED_MEMORY=1 and patches
+# MemorySnapshot/request_memory — does NOT swap the global allocator.
 # ==============================================================================
+ARG CMM_BUILD_DATE=default
+RUN echo "CMM build: ${CMM_BUILD_DATE}" > /dev/null
 COPY managed_alloc.cu /tmp/managed_alloc.cu
 RUN nvcc -shared -o /usr/local/lib/libmanaged_alloc.so \
     /tmp/managed_alloc.cu -Xcompiler -fPIC && rm /tmp/managed_alloc.cu