dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

feat(MoE): wire full ix_fused_moe_forward as Tier 0 dispatch

Browse Source 
ix_bridge.py: expose all 6 ixformer::infer functions + fused_moe_forward()
qwen3_5.py: 4-tier MoE dispatch (fused C++ → CUB topk → ix topk → PyTorch)
patch_ops.sh: deploy ix_moe_bridge.cpp to 4 search paths for JIT
This commit is contained in:
EX Engine
2026-08-10 03:36:39 +00:00
parent 1be9449883
commit 388f6b2d1a
3 changed files with 192 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
qwen3_6_scripts/patch_ops.sh
View File

@@ -185,9 +185,11 @@ if [ -d "$EX_ENGINE_SRC/python" ]; then
EX_DST="$VLLM/model_executor/models/ex_engine"
mkdir -p "$EX_DST/python" "$EX_DST/csrc"
cp "$EX_ENGINE_SRC/python/"*.py "$EX_DST/python/" 2>/dev/null || true
# ix_moe_bridge.cpp for JIT compile
# ix_moe_bridge.cpp for JIT compile — deploy to ALL search paths
cp "$EX_ENGINE_SRC/csrc/ix_moe_bridge.cpp" "$EX_DST/csrc/" 2>/dev/null || true
cp "$EX_ENGINE_SRC/csrc/ix_moe_bridge.cpp" "$EX_DST/python/" 2>/dev/null || true
cp "$EX_ENGINE_SRC/csrc/ix_moe_bridge.cpp" "/workspace/ex_engine/csrc/" 2>/dev/null || true
cp "$EX_ENGINE_SRC/csrc/ix_moe_bridge.cpp" "/workspace/qwen3_6_scripts/" 2>/dev/null || true
touch "$EX_DST/__init__.py"
touch "$EX_DST/python/__init__.py"
# Copy built .so files