[NPU] initial support of asym_int4_rtn (#12484)

* initiail support of q4_1

* fix

* fix

* update

* update min to Z1

* update

* fix

* update

* fix style

* fix

* support qwen2 optimize_model=True mp version

* temp save

* fix

* fix style

* replace min with zero

* support split linear for q4_1

* fix lm_head with mixed_precision=True

* fix style

* revert test code

* add down proj back for q4_0

* remove print

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

@@ -52,6 +52,7 @@
                      "fp6_k": 30,
                      "sym_int4_rtn": 31,
                      "sym_int8_rtn": 32,
+                     "asym_int4_rtn": 33,
                      }
 
 # mixed precison from llama.cpp

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

@@ -84,8 +84,10 @@
 FP6_K = ggml_tensor_qtype["fp6_k"]
 SYM_INT4_RTN = ggml_tensor_qtype["sym_int4_rtn"]
 SYM_INT8_RTN = ggml_tensor_qtype["sym_int8_rtn"]
+ASYM_INT4_RTN = ggml_tensor_qtype["asym_int4_rtn"]
 RTN_DTYPE = {
     SYM_INT4_RTN: torch.uint8,
+    ASYM_INT4_RTN: torch.uint8,
     SYM_INT8_RTN: torch.int8,
 }
 
@@ -223,12 +225,16 @@ 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
-    if qtype in [SYM_INT8_RTN, SYM_INT4_RTN]:
+    if qtype in [SYM_INT8_RTN, SYM_INT4_RTN, ASYM_INT4_RTN]:
         dst_tensor = torch.empty(dst_size, dtype=RTN_DTYPE[qtype],
                                  device=device)
         dst_tensor = dst_tensor.reshape(tensor.shape[0], tensor.shape[-1] // QK)
-        scale = torch.empty(n // k, dtype=torch.float32,
-                            device=device)
+        if qtype == ASYM_INT4_RTN:
+            scale = torch.empty((n // k) * 2, dtype=torch.float32,
+                                device=device)
+        else:
+            scale = torch.empty(n // k, dtype=torch.float32,
+                                device=device)
     elif qtype == NF4:
         # Deepspeed zero3 requires unified dtype,
         # thus here uses bfloat16 consistent to other layers
@@ -244,7 +250,7 @@ def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
         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]:
-            if qtype in [SYM_INT8_RTN, SYM_INT4_RTN]:
+            if qtype in [SYM_INT8_RTN, SYM_INT4_RTN, ASYM_INT4_RTN]:
                 scale_ptr = ctypes.cast(scale.data.data_ptr(), ctypes.POINTER(ctypes.c_float))
                 if imatrix is None:
                     ggml.ggml_quantize_tensor_rtn(src, dst, scale_ptr, qtype, n,
@@ -269,7 +275,7 @@ def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
             ggml.ggml_quantize_tensor_with_weights(src, dst, qtype,
                                                    n // in_features, in_features,
                                                    hist, imatrix)
-    if qtype in [SYM_INT8_RTN, SYM_INT4_RTN]:
+    if qtype in [SYM_INT8_RTN, SYM_INT4_RTN, ASYM_INT4_RTN]:
         return dst_tensor, scale.type(torch.float16)
     else:
         return dst_tensor

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

@@ -103,6 +103,7 @@ def from_pretrained(cls, *args, **kwargs):
         qtype_map = {
             "sym_int4": "sym_int4_rtn",
             "sym_int8": "sym_int8_rtn",
+            "asym_int4": "asym_int4_rtn",
         }
 
         invalidInputError(
@@ -154,7 +155,7 @@ def from_pretrained(cls, *args, **kwargs):
                 f"but got {quantization_group_size}"
             )
         )
-        _args = copy.deepcopy(args)
+
         _kwargs = copy.deepcopy(kwargs)
 
         try:
@@ -270,6 +271,7 @@ def optimize_npu_model(cls, *args, **kwargs):
         with torch.no_grad():
             model.config.update({"mixed_precision": mixed_precision})
             model.config.update({"group_size": quantization_group_size})
+            model.config.update({"asym": qtype == "asym_int4_rtn"})
             optimize_llm_pre(model, qtype, mixed_precision,
                              quantization_group_size=quantization_group_size)
             cls.load_convert(qtype, model, "cpu", modules_to_not_convert,
@@ -416,9 +418,9 @@ def load_low_bit(cls, pretrained_model_name_or_path: str, *model_args, **kwargs)
         )
 
         invalidInputError(
-            qtype in ["sym_int8_rtn", "sym_int4_rtn"],
+            qtype in ["sym_int8_rtn", "sym_int4_rtn", "asym_int4_rtn"],
             f"Unknown bigdl_transformers_low_bit value: {qtype},"
-            f" expected: sym_int8_rtn, sym_int4_rtn. "
+            f" expected: sym_int8_rtn, sym_int4_rtn, asym_int4_rtn. "
         )
 
         if enable_cpp_backend:

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

@@ -88,19 +88,26 @@ def replace_with_QuantizedLinear(layer, qtype, device, modules_to_not_convert,
     from ipex_llm.ggml.quantize import ggml_tensor_qtype
     iqtype = ggml_tensor_qtype[qtype]
     if isinstance(layer, torch.nn.Linear) and not hasattr(layer, "qtype"):
-        if qtype == "sym_int4_rtn":
+        if qtype in ["sym_int4_rtn", "asym_int4_rtn"]:
             # workaround for qwen2-7B & int4
-            if (layer.in_features == 3584 and layer.out_features == 152064) or \
-               (layer.in_features == 18944 and layer.out_features == 3584):
+            if (layer.in_features == 3584 and layer.out_features == 152064):
+                qtype = "sym_int8_rtn"
+                iqtype = ggml_tensor_qtype[qtype]
+        if qtype == "sym_int4_rtn":
+            if (layer.in_features == 18944 and layer.out_features == 3584):
                 qtype = "sym_int8_rtn"
                 iqtype = ggml_tensor_qtype[qtype]
         enable_scale_search = os.environ.get("IPEX_LLM_NPU_QUANTIZATION_OPT", "0") != "0"
         qweights, scale = ggml_convert_qtype(layer.weight.data.to(torch.float32),
                                              iqtype, device=device,
                                              enable_scale_search=enable_scale_search,
                                              imatrix=imatrix)
-        return QuantizedLinear(qweights, scale, layer.bias,
-                               group_size=group_size)
+        zero = None
+        # split scale to scale & zero
+        if qtype == "asym_int4_rtn":
+            scale, zero = torch.split(scale, scale.shape[0] // 2)
+        return QuantizedLinear(qweights, scale, zero, layer.bias,
+                               group_size=group_size, qtype=qtype)
 
 
 @module_optimization
@@ -111,12 +118,21 @@ def replace_with_DequantizedLinear(layer, qtype, device, modules_to_not_convert,
     from ipex_llm.ggml.quantize import ggml_tensor_qtype
     iqtype = ggml_tensor_qtype[qtype]
     if isinstance(layer, torch.nn.Linear) and not hasattr(layer, "qtype"):
+        if qtype in ["sym_int4_rtn", "asym_int4_rtn"]:
+            # workaround for qwen2-7B & int4
+            if (layer.in_features == 3584 and layer.out_features == 152064):
+                qtype = "sym_int8_rtn"
+                iqtype = ggml_tensor_qtype[qtype]
         enable_scale_search = os.environ.get("IPEX_LLM_NPU_QUANTIZATION_OPT", "0") != "0"
         qweights, scale = ggml_convert_qtype(layer.weight.data.to(torch.float32),
                                              iqtype, device=device,
                                              enable_scale_search=enable_scale_search,
                                              imatrix=imatrix)
-        return DequantizedLinear(qweights, scale, layer.bias)
+        zero = None
+        # split scale to scale & zero
+        if qtype == "asym_int4_rtn":
+            scale, zero = torch.split(scale, scale.shape[0] // 2)
+        return DequantizedLinear(qweights, scale, zero, layer.bias, qtype)
 
 
 @module_optimization

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

@@ -128,7 +128,7 @@ def optimize_llm_pre(model: torch.nn.Module, qtype, mixed_precision,
         from ipex_llm.transformers.npu_models.common import split_linears
         if quantization_group_size == 0:
             n_splits_linear = 1
-            if qtype == "sym_int8_rtn":
+            if qtype in ["sym_int8_rtn", "asym_int4_rtn"]:
                 # do not split mlp down_proj for Qwen2-7B & sym_int8
                 n_splits_down_proj = 1
             else:
@@ -154,18 +154,21 @@ def optimize_llm_pre(model: torch.nn.Module, qtype, mixed_precision,
                 # workaround for MiniCPM-2B
                 new_lm_head_0 = SlicedLMHead(model.lm_head_0.weight, split_num=split_num,
                                              bias=model.lm_head_0.bias, use_split=True,
-                                             group_size=quantization_group_size)
+                                             group_size=quantization_group_size,
+                                             asym=(qtype == "asym_int4_rtn"))
                 del model.lm_head_0
                 model.lm_head_0 = new_lm_head_0
                 new_lm_head_1 = SlicedLMHead(model.lm_head_1.weight, split_num=split_num,
                                              bias=model.lm_head_1.bias, use_split=True,
-                                             group_size=quantization_group_size)
+                                             group_size=quantization_group_size,
+                                             asym=(qtype == "asym_int4_rtn"))
                 del model.lm_head_1
                 model.lm_head_1 = new_lm_head_1
             else:
                 new_lm_head = SlicedLMHead(model.lm_head.weight, split_num=split_num,
                                            bias=model.lm_head.bias, use_split=True,
-                                           group_size=quantization_group_size)
+                                           group_size=quantization_group_size,
+                                           asym=(qtype == "asym_int4_rtn"))
                 del model.lm_head
                 model.lm_head = new_lm_head
 
@@ -176,11 +179,13 @@ def optimize_llm_pre(model: torch.nn.Module, qtype, mixed_precision,
             # Do not split lm_head and use sym_int8 instead when mixed_precison is True
             if quantization_group_size == 0:
                 # Do not split lm_head and use sym_int8 instead when mixed_precison is True
-                is_split = (not mixed_precision) and qtype == "sym_int4_rtn"
+                is_split = (not mixed_precision) and qtype in ["sym_int4_rtn", "asym_int4_rtn"]
                 split_num = 14 if is_split else 1
                 new_lm_head = SlicedLMHead(model.lm_head.weight, split_num=split_num,
                                            bias=model.lm_head.bias, use_split=True,
-                                           group_size=quantization_group_size)
+                                           group_size=quantization_group_size,
+                                           asym=((qtype == "asym_int4_rtn") and
+                                                 (not mixed_precision)))
             del model.lm_head
             model.lm_head = new_lm_head
 

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

@@ -129,16 +129,20 @@ def __init__(
         self,
         weight: torch.Tensor,
         scale: torch.Tensor,
+        zero: Optional[torch.Tensor] = None,
         bias: Optional[torch.Tensor] = None,
+        qtype: Optional[str] = "sym_int4_rtn",
         group_size: int = 0,
     ):
         """Initialize the QuantizedLinear class.
 
         Args:
             weight (torch.Tensor): Linear operation weight
             scale (torch.Tensor): Quantization scale
+            zero (Optional[torch.Tensor], optional): Quantization zero for asym_int4_rtn
             bias (Optional[torch.Tensor], optional): Linear operation optional bias.
                                                      Defaults to None.
+            qtype (Optional[str], optional): qtype of this Linear
 
         Raises:
             RuntimeError: Quantized weight must be in torch.int8 format
@@ -155,14 +159,19 @@ def __init__(
                 )
             )
         self.outC, self.inC = self.weight.shape
+        self.zero = None
         if group_size != 0:
             self.scale = Parameter(scale, requires_grad=False)
+            self.zero = Parameter(zero, requires_grad=False)
         else:
             if self.weight.dtype == torch.uint8:
                 # Int4 we need to double the input channels because weights are compressed
                 self.inC *= 2
             self.scale = Parameter(scale * math.sqrt(self.inC), requires_grad=False)
+            if zero is not None:
+                self.zero = Parameter(zero * math.sqrt(self.inC), requires_grad=False)
         self.bias = bias
+        self.qtype = qtype
         self.op_id = str(uuid.uuid4())
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -195,7 +204,8 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
                 )
             )
 
-        out = run_matmul(x, self.weight.data, self.scale.data, self.op_id)
+        zero_data = self.zero.data if self.zero is not None else None
+        out = run_matmul(x, self.weight.data, self.scale.data, zero_data, self.op_id)
 
         if self.bias is None:
             return out
@@ -209,14 +219,18 @@ def __init__(
         self,
         weight: torch.Tensor,
         scale: torch.Tensor,
+        zero: Optional[torch.Tensor] = None,
         bias: Optional[torch.Tensor] = None,
+        qtype: Optional[str] = "sym_int4_rtn",
     ):
         """Initialize the DequantizedLinear class.
         Args:
             weight (torch.Tensor): Linear operation quantized weight
             scale (torch.Tensor): Quantization scale
+            zero (Optional[torch.Tensor], optional): Quantization zero for asym_int4_rtn
             bias (Optional[torch.Tensor], optional): Linear operation optional bias.
                                                      Defaults to None.
+            qtype (Optional[str], optional): qtype of this Linear
         Raises:
             RuntimeError: Quantized weight must be in torch.int8 format
         """
@@ -240,6 +254,9 @@ def __init__(
             decompressed_weight = combined_weight.view(combined_weight.size(0), -1)
             dequantized_weight = decompressed_weight.to(torch.float32) * \
                 torch.unsqueeze(scale.to(torch.float32), dim=1)
+            if qtype == "asym_int4_rtn" and zero is not None:
+                dequantized_weight = dequantized_weight + torch.unsqueeze(zero.to(torch.float32),
+                                                                          dim=1)
             self.weight = Parameter(dequantized_weight, requires_grad=False).contiguous()
         else:
             dequantized_weight = weight.to(torch.float32) * \

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

@@ -36,6 +36,7 @@ def __init__(
         dtype: np.dtype = np.int8,
         use_split: bool = False,
         group_size: int = 0,
+        asym: bool = False,
     ):
         """Initialize the LMHeadLinear class.
 
@@ -54,11 +55,10 @@ def __init__(
         self.batch = batch
 
         self.split_num = split_num
-
         if use_split:
             input = self.parameter((1, self.batch, self.inC))
             res = self.dq_split_linear(input, self.split_num, self.outC, self.inC, wt_dtype=dtype,
-                                       scale_factor=(group_size == 0))
+                                       scale_factor=(group_size == 0), asym=asym)
         else:
             input = self.parameter((self.batch, self.inC))
             split_size = self.inC // split_num // 2 * 2
@@ -69,7 +69,7 @@ def __init__(
                 input_slice = self.slice(input, begin=[0, start_idx],
                                          end=[self.batch, end_idx])
                 linear_slice = self.linear(input_slice, outC, split_size, bias=False,
-                                           wt_dtype=dtype)
+                                           wt_dtype=dtype, asym=asym)
                 if i == 0:
                     res = linear_slice
                 else:
@@ -109,7 +109,7 @@ def run(
 
 
 class SlicedLMHead(nn.Module):
-    def __init__(self, weight, bias, split_num, use_split=False, group_size=0):
+    def __init__(self, weight, bias, split_num, use_split=False, group_size=0, asym=False):
         super().__init__()
         self.split_num = split_num
         self.outC, self.inC = weight.shape
@@ -128,6 +128,7 @@ def __init__(self, weight, bias, split_num, use_split=False, group_size=0):
             self.lm_heads.append(new_linear)
         self.bias = bias
         self.use_split = use_split
+        self.asym = asym
 
     def forward(self, hidden_states):
         if hidden_states.size(0) * hidden_states.size(1) == 1:
@@ -162,19 +163,33 @@ def get_fused_lm_head(self):
         np_dtype = np.uint8 if self.get_weight_dtype() == torch.uint8 else np.int8
         self.fused_lm_head = LMHeadLinear(self.inC, self.outC, 1, self.split_num,
                                           False, "NPU", dtype=np_dtype, use_split=self.use_split,
-                                          group_size=self.group_size)
+                                          group_size=self.group_size, asym=self.asym)
         if self.use_split:
             weights = []
             scales = []
+            zeros = []
             for i in range(self.split_num):
                 weights.append(self.lm_heads[i].weight)
                 scales.append(self.lm_heads[i].scale)
-            fused_lm_head_weights = (torch.stack(weights, axis=0).numpy(),
-                                     torch.stack(scales, axis=0).numpy())
+                if self.lm_heads[i].zero is not None:
+                    zeros.append(self.lm_heads[i].zero)
+            if len(zeros):
+                fused_lm_head_weights = [(torch.stack(weights, axis=0).numpy(),
+                                          torch.stack(scales, axis=0).numpy(),
+                                          torch.stack(zeros, axis=0).numpy())]
+            else:
+                fused_lm_head_weights = [(torch.stack(weights, axis=0).numpy(),
+                                          torch.stack(scales, axis=0).numpy())]
         else:
-            fused_lm_head_weights = [(self.lm_heads[i].weight.data.numpy(),
-                                      self.lm_heads[i].scale.data.numpy())
-                                     for i in range(self.split_num)]
+            if self.asym:
+                fused_lm_head_weights = [(self.lm_heads[i].weight.data.numpy(),
+                                          self.lm_heads[i].scale.data.numpy(),
+                                          self.lm_heads[i].zero.data.numpy())
+                                         for i in range(self.split_num)]
+            else:
+                fused_lm_head_weights = [(self.lm_heads[i].weight.data.numpy(),
+                                          self.lm_heads[i].scale.data.numpy())
+                                         for i in range(self.split_num)]
 
         self.fused_lm_head.set_weights(self.lm_heads[0].op_id,
                                        fused_lm_head_weights)