docker: Upgrade TileLang to 0.1.14 (#3309)

docker/Dockerfile

@@ -86,10 +86,18 @@ RUN pip install git+https://github.com/ISEEKYAN/mbridge.git@89eb10887887bc74853f
 
 RUN pip install flash-linear-attention==0.5.2
 # required for DeepSeek V4
-RUN pip install tilelang==0.1.8 -f https://tile-ai.github.io/whl/nightly/cu128/
-RUN pip install --no-deps tile_kernels==1.0.0
+RUN pip install tilelang==0.1.14 apache-tvm-ffi==0.1.11
+# TileKernels target-import fix: https://github.com/deepseek-ai/TileKernels/pull/25.
+COPY docker/patch/tile_kernels_target.patch /tmp/tile_kernels_target.patch
+RUN pip install --no-deps tile_kernels==1.0.0 && \
+    TILE_KERNELS_DIR=$(python3 -c 'import importlib.util; print(importlib.util.find_spec("tile_kernels").submodule_search_locations[0])') && \
+    patch --fuzz=0 -p1 -d "$TILE_KERNELS_DIR" < /tmp/tile_kernels_target.patch && \
+    rm /tmp/tile_kernels_target.patch
 # FlashQLA backend for Qwen GDN linear-attention layers (requires SM90+, CUDA 12.8+, PyTorch 2.8+; built on tilelang above)
-RUN pip install -v --no-build-isolation "git+https://github.com/QwenLM/FlashQLA.git"
+# Its dependency pins would downgrade TileLang and TVM-FFI; use the versions above.
+# FlashQLA FP16 on Blackwell retains a pre-existing compile failure.
+# Fixed after 0.1.14 by https://github.com/tile-ai/tilelang/pull/3158.
+RUN pip install -v --no-deps --no-build-isolation "git+https://github.com/QwenLM/FlashQLA.git@7c7dfe16416ad21b1d03258189fc8d3b8460ae06"
 # Prefer the prebuilt wheel; fall back to a source build while the current
 # arch's wheels release doesn't carry it yet (QEMU-emulated arm64 source
 # builds are the expensive path this avoids).

docker/patch/tile_kernels_target.patch

@@ -0,0 +1,6 @@
+diff --git a/quant/common.py b/quant/common.py
+--- a/quant/common.py
++++ b/quant/common.py
+@@ -7 +7 @@
+-from tilelang.utils.target import determine_target
++from tilelang.backend.target import determine_target