fix(bridge): rewrite ix_full_bridge.cpp for real ixformer_torch_ext namespace

nm -D confirmed real symbols are in ixformer_torch_ext:: not ixformer::infer::
Bridges: silu_and_mul, rms_norm, fused_add_rms_norm, rotary_embedding, reshape_and_cache
This commit is contained in:
Claude
2026-08-14 07:00:30 +00:00
parent 9e3157b444
commit d9ffc5159d
2 changed files with 123 additions and 489 deletions

View File

@@ -1,95 +1,95 @@
#!/bin/bash
# Build ix_full_bridge.so — bridges ixformer::infer C++ symbols to Python
# Compiled via torch.utils.cpp_extension at docker build time
# Runtime: dlopen links against libixformer.so in base image
# Build ix_full_bridge.so — bridges ixformer_torch_ext C++ symbols to Python
set -eo pipefail
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
VLLM_ROOT="${1:?usage: build_ix_bridge.sh VLLM_ROOT}"
BRIDGE_SRC="${SCRIPT_DIR}/ix_full_bridge.cpp"
MOE_SRC="${SCRIPT_DIR}/ix_moe_bridge.cpp"
if [ ! -f "$BRIDGE_SRC" ]; then
echo "[bridge] SKIP: $BRIDGE_SRC not found"
exit 0
fi
# Find ixformer .so directory
IX_LIB=""
for d in /usr/local/corex/lib/python3/dist-packages/ixformer \
/usr/local/corex/lib64 \
/usr/lib/python3/dist-packages/ixformer; do
if [ -d "$d" ]; then
IX_LIB="$d"
# Find ixformer .so with the real symbols
IX_TORCH_SO=""
for f in /usr/local/corex/lib/python3/dist-packages/ixformer/_ixformer_torch*.so; do
if [ -f "$f" ]; then
IX_TORCH_SO="$f"
break
fi
done
# Find torch library path
TORCH_LIB=$(python3 -c "import torch; print(torch.__path__[0] + '/lib')" 2>/dev/null)
IX_LIB_SO=""
for f in /usr/local/corex/lib/python3/dist-packages/ixformer/libixformer.so; do
if [ -f "$f" ]; then
IX_LIB_SO="$f"
break
fi
done
echo "[bridge] Building ix_full_bridge via torch.utils.cpp_extension..."
echo "[bridge] ixformer lib: ${IX_LIB:-not found}"
echo "[bridge] torch lib: ${TORCH_LIB:-not found}"
TORCH_LIB=$(python3 -c "import torch; print(torch.__path__[0] + '/lib')" 2>/dev/null)
IX_DIR=$(python3 -c "import ixformer; import os; print(os.path.dirname(ixformer.__file__))" 2>/dev/null)
echo "[bridge] IX_TORCH_SO=${IX_TORCH_SO}"
echo "[bridge] IX_LIB_SO=${IX_LIB_SO}"
echo "[bridge] TORCH_LIB=${TORCH_LIB}"
# Clear cached build (namespace changed)
rm -rf /root/.cache/torch_extensions/py310_cu102/ix_full_bridge
python3 << PYEOF
import torch
from torch.utils.cpp_extension import load
import os, shutil
import shutil, os
# Extra link flags: find libixformer.so and link
extra_ldflags = []
ix_lib = "${IX_LIB}"
# Link against _ixformer_torch .so (has ixformer_torch_ext:: symbols)
ix_torch = "${IX_TORCH_SO}"
if ix_torch and os.path.exists(ix_torch):
extra_ldflags.append(ix_torch)
extra_ldflags.append(f"-Wl,-rpath,{os.path.dirname(ix_torch)}")
# Also link libixformer.so (has launcher symbols)
ix_lib = "${IX_LIB_SO}"
if ix_lib and os.path.exists(ix_lib):
extra_ldflags.append(ix_lib)
# torch lib rpath
torch_lib = "${TORCH_LIB}"
# Search for libixformer.so
for d in [ix_lib, "/usr/local/corex/lib64", "/usr/local/corex/lib"]:
if d and os.path.exists(os.path.join(d, "libixformer.so")):
extra_ldflags.extend([f"-L{d}", "-lixformer"])
break
else:
# No explicit libixformer.so — symbols may be in already-loaded .so
# (ixformer is imported as Python module which loads the .so)
try:
import ixformer
print("[bridge] ixformer Python module available — symbols in process")
except:
print("[bridge] WARNING: no libixformer.so found, link may fail at runtime")
# Add torch lib to rpath
if torch_lib and os.path.isdir(torch_lib):
extra_ldflags.append(f"-Wl,-rpath,{torch_lib}")
try:
mod = load(
name="ix_full_bridge",
sources=["${BRIDGE_SRC}"],
extra_cflags=["-O2", "-std=c++17"],
extra_ldflags=extra_ldflags,
verbose=True,
)
# Copy compiled .so to VLLM_ROOT for import
import importlib
spec = importlib.util.find_spec("ix_full_bridge")
if spec and spec.origin:
print(f"[bridge] ldflags: {extra_ldflags}")
mod = load(
name="ix_full_bridge",
sources=["${BRIDGE_SRC}"],
extra_cflags=["-O2", "-std=c++17"],
extra_ldflags=extra_ldflags,
verbose=True,
)
# Find the compiled .so and copy to VLLM_ROOT
import importlib
spec = importlib.util.find_spec("ix_full_bridge")
if spec and spec.origin:
dest = os.path.join("${VLLM_ROOT}", "ix_full_bridge.so")
shutil.copy2(spec.origin, dest)
print(f"[bridge] SUCCESS: {dest}")
fns = [x for x in dir(mod) if not x.startswith("_")]
print(f"[bridge] functions: {fns}")
else:
# Search in cache
import glob
for so in glob.glob(os.path.expanduser("~/.cache/torch_extensions/**/ix_full_bridge*.so"), recursive=True):
dest = os.path.join("${VLLM_ROOT}", "ix_full_bridge.so")
shutil.copy2(spec.origin, dest)
print(f"[bridge] SUCCESS: {dest}")
shutil.copy2(so, dest)
print(f"[bridge] SUCCESS: {so} -> {dest}")
break
else:
# Try finding it in torch extension build dir
build_dir = os.path.expanduser("~/.cache/torch_extensions")
for root, dirs, files in os.walk(build_dir):
for f in files:
if f.startswith("ix_full_bridge") and f.endswith(".so"):
src = os.path.join(root, f)
dest = os.path.join("${VLLM_ROOT}", "ix_full_bridge.so")
shutil.copy2(src, dest)
print(f"[bridge] SUCCESS: {src} -> {dest}")
break
except Exception as e:
print(f"[bridge] FAILED: {e}")
raise SystemExit(1)
print("[bridge] WARNING: could not find compiled .so")
PYEOF
echo "[bridge] Build complete"