Init NPU quantize method and support q8_0_rtn (#11452)

* q8_0_rtn

* fix float point

python/llm/src/ipex_llm/ggml/model/llama/llama_cpp.py

@@ -991,6 +991,33 @@ _lib.ggml_quantize_tensor_with_weights.argtypes = [
 _lib.ggml_quantize_tensor_with_weights.restype = ctypes.c_size_t
 
 
+# GGML API
+def ggml_quantize_tensor_rtn(
+    src,  # type: ctypes.Array[ctypes.c_float] # type: ignore
+    dst: ctypes.c_void_p,
+    scale_ptr,  # type: ctypes.Array[ctypes.c_float] # type: ignore
+    qtype: ctypes.c_int,
+    n: ctypes.c_size_t,
+    k: ctypes.c_int,
+    hist,  # type: ctypes.Array[ctypes.c_int64] # type: ignore
+    scale_search: ctypes.c_bool,
+) -> int:
+    return _lib.ggml_quantize_tensor_rtn(src, dst, scale_ptr, qtype, n, k, hist, scale_search)
+
+
+_lib.ggml_quantize_tensor_rtn.argtypes = [
+    ctypes.POINTER(ctypes.c_float),
+    ctypes.c_void_p,
+    ctypes.POINTER(ctypes.c_float),
+    ctypes.c_int,
+    ctypes.c_size_t,
+    ctypes.c_int,
+    ctypes.POINTER(ctypes.c_int64),
+    ctypes.c_bool,
+]
+_lib.ggml_quantize_tensor_rtn.restype = ctypes.c_size_t
+
+
 def ggml_type_size(qtype: ctypes.c_int) -> int:
     return _lib.ggml_type_size(qtype)
 

python/llm/src/ipex_llm/ggml/quantize.py

@@ -50,6 +50,8 @@ ggml_tensor_qtype = {"sym_int4": 2,   # q4_0 in ggml
                      "q5_k": 28,
                      "fp6": 29,
                      "fp6_k": 30,
+                     "sym_int4_rtn": 31,
+                     "sym_int8_rtn": 32,
                      }
 
 # mixed precison from llama.cpp

python/llm/src/ipex_llm/transformers/low_bit_linear.py

@@ -81,6 +81,7 @@ Q4_K = ggml_tensor_qtype["q4_k"]
 Q6_K = ggml_tensor_qtype["q6_k"]
 Q5_K = ggml_tensor_qtype["q5_k"]
 FP6_K = ggml_tensor_qtype["fp6_k"]
+SYM_INT8_RTN = ggml_tensor_qtype["sym_int8_rtn"]
 
 
 # For sym_int4
@@ -216,14 +217,27 @@ def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
                       f"Last dim of input tensor must be multiple of {QK}")
 
     dst_size = (n // QK) * block_size_in_bytes
-    dst_tensor = torch.empty(dst_size, dtype=torch.uint8,
-                             device=device)
+    if qtype in [SYM_INT8_RTN]:
+        dst_tensor = torch.empty(dst_size, dtype=torch.int8,
+                                 device=device)
+        scale = torch.empty(n // k, dtype=torch.float32,
+                            device=device)
+    else:
+        dst_tensor = torch.empty(dst_size, dtype=torch.uint8,
+                                 device=device)
 
     if not convert_shape_only and device != 'meta':
         dst = ctypes.c_void_p(dst_tensor.data.data_ptr())
         hist = (ctypes.c_int64 * 16)()
         if qtype not in [IQ2_XXS, IQ2_XS, Q2_K, IQ1_S, Q4_K, Q6_K, Q5_K, FP6_K]:
-            ggml.ggml_quantize_tensor(src, dst, qtype, n, k, hist, enable_scale_search)
+            if qtype in [SYM_INT8_RTN]:
+                scale_ptr = ctypes.cast(scale.data.data_ptr(), ctypes.POINTER(ctypes.c_float))
+                ggml.ggml_quantize_tensor_rtn(src, dst, scale_ptr, qtype, n,
+                                              k, hist, enable_scale_search)
+                dst_tensor = dst_tensor.reshape_as(tensor)
+                return dst_tensor, scale.type(torch.float16)
+            else:
+                ggml.ggml_quantize_tensor(src, dst, qtype, n, k, hist, enable_scale_search)
         else:
             if imatrix is not None:
                 # quantize with importance matrix

python/llm/src/ipex_llm/transformers/npu_model.py

@@ -77,7 +77,7 @@ class _BaseAutoModelClass:
             from intel_npu_acceleration_library.dtypes import int8, int4
             qtype_map = {
                 'sym_int4': int4,
-                'sym_int8': int8,
+                'sym_int8': "sym_int8_rtn",
                 'fp16': torch.half,
                 'fp32': torch.float,
             }
@@ -119,9 +119,12 @@ class _BaseAutoModelClass:
             from intel_npu_acceleration_library.compiler import create_npu_kernels
             with torch.no_grad():
                 optimize_llm(model)
-                if not qtype.is_floating_point:
-                    model = quantize_model(model, qtype)
-                create_npu_kernels(model)
+                if qtype == "sym_int8_rtn":
+                    cls.load_convert(qtype, model, *args, **kwargs)
+                else:
+                    if not qtype.is_floating_point:
+                        model = quantize_model(model, qtype)
+                    create_npu_kernels(model)
             model = model.eval()
         except ImportError as _e:
             # for intel_npu_acceleration_library < 1.1.0
@@ -133,6 +136,11 @@ class _BaseAutoModelClass:
 
         return model
 
+    @classmethod
+    def load_convert(cls, q_k, optimize_model, *arg, **kwarg):
+        from ipex_llm.transformers.npu_models.convert import replace_with_QuantizedLinear
+        replace_with_QuantizedLinear(optimize_model, q_k)
+
     @staticmethod
     def save_low_bit(self, model_dir: str, *args, **kwargs):
         os.makedirs(model_dir, exist_ok=True)

python/llm/src/ipex_llm/transformers/npu_models/convert.py

@@ -15,6 +15,49 @@
 
 
 import torch
+from intel_npu_acceleration_library.nn import QuantizedLinear
+
+
+def module_optimization(func) -> torch.nn.Module:
+    """Optimize recursively a torch.nn.Module with a specific function.
+
+    The function `func` get called recursively to every module in the network.
+
+    Args:
+        func (Callable): optimization function
+
+    Returns:
+        torch.nn.Module: optimized module
+    """
+
+    def wrapper(model: torch.nn.Module, qtype, *args, **kwargs):
+        """Recursively apply the optimization function.
+
+        Args:
+            model (torch.nn.Module): original module
+            args (Any): positional arguments
+            kwargs (Any): keyword arguments
+
+        """
+        for name, layer in model.named_children():
+            new_layer = func(layer, qtype, *args, **kwargs)
+            if new_layer:
+                model.add_module(name, new_layer)
+                wrapper(new_layer, qtype, *args, **kwargs)
+            else:
+                wrapper(layer, qtype, *args, **kwargs)
+
+    return wrapper
+
+
+@module_optimization
+def replace_with_QuantizedLinear(layer, qtype):
+    from ipex_llm.transformers.low_bit_linear import ggml_convert_qtype
+    from ipex_llm.ggml.quantize import ggml_tensor_qtype
+    iqtype = ggml_tensor_qtype[qtype]
+    if isinstance(layer, torch.nn.Linear):
+        qweights, scale = ggml_convert_qtype(layer.weight.data, iqtype, 'cpu')
+        return QuantizedLinear(qweights, scale, layer.bias)
 
 
 def convert_forward(m, target_m, new_forward):