[feature]Ascend quantization support (#7791)

Co-authored-by: ichernob <ichernobnn@gmail.com>
Co-authored-by: liupeng <liupeng374@huawei.com>

python/sglang/srt/utils.py

@@ -197,7 +197,7 @@ def get_int_env_var(name: str, default: int = 0) -> int:
 
 
 def support_triton(backend: str) -> bool:
-    return backend not in ["torch_native", "intel_amx"]
+    return backend not in ["torch_native", "intel_amx", "ascend"]
 
 
 try:
@@ -2782,3 +2782,101 @@ def lru_cache_frozenset(maxsize=128):
         return wrapper
 
     return decorator
+
+
+def apply_module_patch(target_module, target_function, wrappers):
+    original_module, original_function = parse_module_path(
+        target_module, target_function, False
+    )
+
+    original_function_id = id(original_function)
+
+    candidate = original_function
+    for wrapper in wrappers:
+        candidate = wrapper(candidate)
+    if target_function is not None:
+        setattr(original_module, target_function, candidate)
+
+    for key, value in sys.modules.copy().items():
+        if (
+            target_function is not None
+            and hasattr(value, target_function)
+            and id(getattr(value, target_function)) == original_function_id
+        ):
+            setattr(value, target_function, candidate)
+
+
+def parse_module_path(module_path, function_name, create_dummy):
+    from importlib.machinery import ModuleSpec
+
+    def create_dummy_module(full_path, parent=None):
+        """Create and register a placeholder module"""
+        dummy = types.ModuleType(full_path)
+        dummy.__file__ = "vllm_ascend.dummy_module.py"
+        dummy.__spec__ = ModuleSpec(full_path, None)
+        sys.modules[full_path] = dummy
+        if parent:
+            setattr(parent, full_path.split(".")[-1], dummy)
+        return dummy
+
+    def create_placeholder_function(func_name):
+        """Create dummy function that raises when called"""
+
+        def placeholder(*args, **kwargs):
+            raise NotImplementedError(f"Function {func_name} is a placeholder")
+
+        placeholder.__name__ = func_name
+        return placeholder
+
+    modules = module_path.split(".")
+    current_module = None
+    processed_path = []
+
+    for idx, part in enumerate(modules):
+        current_path = ".".join(modules[: idx + 1])
+        parent_path = ".".join(modules[:idx]) if idx > 0 else None
+
+        try:
+            current_module = importlib.import_module(current_path)
+        except ModuleNotFoundError:
+            # Handle missing module
+            parent = importlib.import_module(parent_path) if parent_path else None
+            if parent and hasattr(parent, part):
+                # Use existing attribute from parent
+                current_module = getattr(parent, part)
+                # Check for early function resolution
+                if function_name and hasattr(current_module, function_name):
+                    return current_module, getattr(current_module, function_name)
+                if function_name and create_dummy:
+                    ph_func = create_placeholder_function(function_name)
+                    setattr(current_module, function_name, ph_func)
+                    return current_module, ph_func
+                if function_name:
+                    raise AttributeError(
+                        f"Function {function_name} missing in {current_path}"
+                    )
+            else:
+                if not create_dummy:
+                    raise
+                # Create and register dummy module
+                current_module = create_dummy_module(
+                    current_path,
+                    parent=(
+                        importlib.import_module(parent_path) if parent_path else None
+                    ),
+                )
+
+        processed_path.append(part)
+
+    # Final function handling
+    final_module = sys.modules[module_path]
+    if function_name is not None:
+        if not hasattr(final_module, function_name):
+            if create_dummy:
+                ph_func = create_placeholder_function(function_name)
+                setattr(final_module, function_name, ph_func)
+            else:
+                setattr(final_module, function_name, None)
+        return final_module, getattr(final_module, function_name)
+
+    return final_module, None