arch(probe): add build-time CoreX API discovery — stop guessing interfaces

probe_corex_api.py runs during docker build BEFORE qwen3_5.py deployment:
1. Lists ALL .py files in base image's vllm/model_executor/models/
2. For each corex_gdn/corex_moe/corex_fa2: import → inspect signatures
3. If import fails: AST parse the .py file directly for class/method defs
4. Checks native qwen3_5.py for corex references before we overwrite it
5. Checks .so files exist (libcorex_gdn.so etc)
6. Dumps everything to /workspace/corex_probe_result.json

Next deploy's build log will show EXACTLY what the corex API looks like.
Then we write real dispatch code against real signatures, not guesses.
This commit is contained in:
Claude
2026-08-08 11:16:58 +00:00
parent ee09550263
commit dbfe20fd1c
2 changed files with 214 additions and 0 deletions

View File

@@ -66,6 +66,14 @@ else
echo "[patch_ops] WARNING: transformers/models not found"
fi
# 1b. CoreX API probe — MUST run BEFORE deploying our qwen3_5.py
# Discovers corex_gdn.py, corex_moe.py, corex_fa2.py interfaces from base image.
# Also inspects native qwen3_5.py before we overwrite it.
# Results go to /workspace/corex_probe_result.json + build log.
echo "[patch_ops] Running CoreX API probe..."
python3 ./probe_corex_api.py 2>&1
echo "[patch_ops] CoreX probe done (see above for results)"
# 2. Model module — qwen3_5.py with CoreX dispatch (CCCL env_dispatch pattern).
# Our version tries to import corex_gdn/corex_moe from the base image.
# If they exist → uses fused CUDA kernels (10x faster).

View File

@@ -0,0 +1,206 @@
"""
CoreX API probe — runs at Docker build time (in patch_ops.sh).
Discovers the real interfaces of corex_gdn.py, corex_moe.py, corex_fa2.py
from the base Docker image. Outputs:
1. /workspace/corex_probe_result.json — machine-readable API map
2. stdout — human-readable summary for build log
This is NOT runtime code. It runs once during `docker build`.
"""
import importlib
import inspect
import json
import os
import sys
PROBE_RESULT = {}
def probe_module(module_path, name):
"""Try to import a module and extract its public API."""
result = {"available": False, "classes": {}, "functions": {}, "error": None}
try:
# Try direct import first
mod = importlib.import_module(module_path)
result["available"] = True
result["file"] = getattr(mod, "__file__", "unknown")
for attr_name in dir(mod):
if attr_name.startswith("_"):
continue
obj = getattr(mod, attr_name)
if inspect.isclass(obj):
cls_info = {
"bases": [b.__name__ for b in obj.__bases__],
"methods": {},
}
for method_name in dir(obj):
if method_name.startswith("_") and method_name != "__init__":
continue
method = getattr(obj, method_name, None)
if callable(method):
try:
sig = str(inspect.signature(method))
cls_info["methods"][method_name] = sig
except (ValueError, TypeError):
cls_info["methods"][method_name] = "(unknown)"
result["classes"][attr_name] = cls_info
elif callable(obj):
try:
sig = str(inspect.signature(obj))
result["functions"][attr_name] = sig
except (ValueError, TypeError):
result["functions"][attr_name] = "(unknown)"
except ImportError as e:
result["error"] = f"ImportError: {e}"
except Exception as e:
result["error"] = f"{type(e).__name__}: {e}"
return result
def probe_file_directly(filepath, name):
"""If import fails, try to read the file and extract class/function defs."""
result = {"available": False, "classes": {}, "functions": {}, "error": None}
if not os.path.exists(filepath):
result["error"] = f"File not found: {filepath}"
return result
result["available"] = True
result["file"] = filepath
result["size"] = os.path.getsize(filepath)
try:
import ast
with open(filepath) as f:
tree = ast.parse(f.read())
for node in ast.walk(tree):
if isinstance(node, ast.ClassDef):
methods = {}
for item in node.body:
if isinstance(item, (ast.FunctionDef, ast.AsyncFunctionDef)):
args = []
for arg in item.args.args:
args.append(arg.arg)
methods[item.name] = f"({', '.join(args)})"
result["classes"][node.name] = {
"bases": [ast.dump(b) for b in node.bases],
"methods": methods,
"lineno": node.lineno,
}
elif isinstance(node, ast.FunctionDef) and node.col_offset == 0:
args = [arg.arg for arg in node.args.args]
result["functions"][node.name] = {
"signature": f"({', '.join(args)})",
"lineno": node.lineno,
}
except Exception as e:
result["error"] = f"AST parse error: {e}"
return result
# Probe paths
VLLM_MODELS = None
for p in [
"/usr/local/corex/lib/python3/dist-packages/vllm/model_executor/models",
"/usr/local/corex/lib64/python3/dist-packages/vllm/model_executor/models",
]:
if os.path.isdir(p):
VLLM_MODELS = p
break
print("=" * 70)
print("[corex_probe] CoreX API Discovery — Build Time")
print("=" * 70)
if VLLM_MODELS:
print(f"[corex_probe] vllm models dir: {VLLM_MODELS}")
# List ALL .py files in models dir to find corex modules
all_files = sorted(os.listdir(VLLM_MODELS))
corex_files = [f for f in all_files if "corex" in f.lower()]
print(f"[corex_probe] CoreX files found: {corex_files}")
print(f"[corex_probe] All model files: {[f for f in all_files if f.endswith('.py')]}")
# Probe each corex module
for target in ["corex_gdn", "corex_moe", "corex_fa2"]:
filepath = os.path.join(VLLM_MODELS, f"{target}.py")
# Try import first
result = probe_module(f"vllm.model_executor.models.{target}", target)
# If import failed, try AST parse
if not result["available"]:
print(f"[corex_probe] {target}: import failed ({result['error']}), trying AST...")
result = probe_file_directly(filepath, target)
PROBE_RESULT[target] = result
if result["available"]:
print(f"[corex_probe] {target}: FOUND at {result.get('file', filepath)}")
if result.get("size"):
print(f"[corex_probe] size: {result['size']} bytes")
for cls_name, cls_info in result.get("classes", {}).items():
print(f"[corex_probe] class {cls_name}:")
for method_name, sig in cls_info.get("methods", {}).items():
print(f"[corex_probe] {method_name}{sig}")
for func_name, func_info in result.get("functions", {}).items():
if isinstance(func_info, dict):
print(f"[corex_probe] def {func_name}{func_info['signature']} (line {func_info['lineno']})")
else:
print(f"[corex_probe] def {func_name}{func_info}")
else:
print(f"[corex_probe] {target}: NOT AVAILABLE — {result.get('error', 'unknown')}")
# Also probe the native qwen3_5.py BEFORE we overwrite it
native_qw = os.path.join(VLLM_MODELS, "qwen3_5.py")
if os.path.exists(native_qw):
sz = os.path.getsize(native_qw)
print(f"[corex_probe] Native qwen3_5.py: {sz} bytes")
# Check if it imports corex
with open(native_qw) as f:
content = f.read()
for keyword in ["corex_gdn", "corex_moe", "corex_fa2", "CoreXGDN", "CoreXMoE"]:
if keyword in content:
print(f"[corex_probe] → references '{keyword}'")
PROBE_RESULT["native_qwen3_5"] = {
"size": sz,
"line_count": content.count("\n"),
"has_corex_gdn": "corex_gdn" in content,
"has_corex_moe": "corex_moe" in content,
"has_corex_fa2": "corex_fa2" in content,
}
else:
print(f"[corex_probe] Native qwen3_5.py: NOT FOUND (will deploy ours)")
PROBE_RESULT["native_qwen3_5"] = {"size": 0, "exists": False}
else:
print("[corex_probe] ERROR: vllm models directory not found")
PROBE_RESULT["error"] = "vllm models dir not found"
# Also check .so files
for so_name, env_var in [
("libcorex_gdn.so", "VLLM_COREX_GDN_LIBRARY"),
("libcorex_moe.so", "VLLM_COREX_MOE_LIBRARY"),
("libcorex_fa2.so", "VLLM_COREX_FA2_LIBRARY"),
]:
path = os.environ.get(env_var, f"/usr/local/corex/lib64/{so_name}")
exists = os.path.exists(path)
size = os.path.getsize(path) if exists else 0
print(f"[corex_probe] {so_name}: {'EXISTS' if exists else 'MISSING'} ({size} bytes) at {path}")
PROBE_RESULT[so_name] = {"exists": exists, "size": size, "path": path}
# Write results
output_path = "/workspace/corex_probe_result.json"
with open(output_path, "w") as f:
json.dump(PROBE_RESULT, f, indent=2, default=str)
print(f"\n[corex_probe] Results written to {output_path}")
print("=" * 70)