Fix WOQ int8 unpack weight (#1393)

Author: changwangss

intel_extension_for_transformers/transformers/llm/quantization/nn/modules.py

@@ -301,6 +301,8 @@ def recover_qparms(self):
  qzeros = torch.ops.bestlaop.acquire_woq_packw_info(self.weight, 10)
  if bits == 4:
  qzeros = qzeros // 16 + 8
+ else:
+ qzeros = (qzeros.to(torch.int32) + 128).to(torch.uint8)
  else:
  qzeros = None
 

intel_extension_for_transformers/transformers/llm/quantization/utils.py

@@ -54,24 +54,45 @@
 
 
 def unpack_weight(qweight, scales, qzeros, q_config):
+ sym = q_config.sym
  bits = q_config.bits
  wf = torch.tensor(list(range(0, 32, bits)), dtype=torch.int32).unsqueeze(0)
+
  zeros = torch.bitwise_right_shift(
  torch.unsqueeze(qzeros, 2).expand(-1, -1, 32 // bits), wf.unsqueeze(0)
  ).to(torch.int16 if bits == 8 else torch.int8)
  torch.bitwise_and(zeros, (2**bits) - 1, out=zeros)
  if bits == 8:
- zeros = zeros.to(torch.int8)
+ zeros = zeros.to(torch.int8 if sym else torch.uint8)
+ # due to INC minus one
  zeros = zeros + 1
- zeros = zeros.reshape(scales.shape)
+ try:
+ zeros = zeros.reshape(scales.shape)
+ except:
+ # zeros and scales have different iteam numbers.
+ # remove 1 (due to 0 + 1 in line 68)
+ zeros = zeros[zeros !=1]
+ zeros = zeros.reshape(scales.shape)
+
+ # due to INC asym return torch.uint8 but backend request int8,
+ # change it to int8 with offset 128
+ if not sym and bits == 8:
+ zeros = (zeros.to(torch.int32) - 128).to(torch.int8)
 
  weight = torch.bitwise_right_shift(
  torch.unsqueeze(qweight, 1).expand(-1, 32 // bits, -1), wf.unsqueeze(-1)
  ).to(torch.int16 if bits == 8 else torch.int8)
  torch.bitwise_and(weight, (2**bits) - 1, out=weight)
-
  if bits == 8:
- weight = weight.to(torch.int8)
+ # due to INC add shift bias for sym
+ if sym:
+ shift_bias = 2 ** (bits - 1)
+ weight -= shift_bias
+ weight = weight.to(torch.int8 if sym else torch.uint8)
+ # due to INC asym return torch.uint8 but backend request int8,
+ # change it to int8 with offset 128
+ if not sym:
+ weight = (weight.to(torch.int32) - 128). to(torch.int8)
  return weight, scales, zeros
 
 
@@ -238,7 +259,7 @@ def _replace_linear(
  model._modules[name].requires_grad_(False)
  if device == "cpu" or device == torch.device("cpu") or device == "auto":
  if quantization_config.weight_dtype in \
- ["fp8_e5m2", "fp8_e4m3", "nf4", "fp4", "int8", "int4_fullrange"]:
+ ["fp8_e5m2", "fp8_e4m3", "nf4", "fp4", "int4_fullrange"]:
  model._modules[name].set_fp_weights_bias(
  module.weight.data,
  None if module.bias is None else module.bias.data,
@@ -506,7 +527,7 @@ def default_calib_func(model):
 
  q_model = replace_linear(model, None, None, config, device=device)
  else:
- if config.weight_dtype not in ["nf4", "fp4", "int8", "int4_fullrange"]:
+ if config.weight_dtype not in ["nf4", "fp4", "int4_fullrange"]:
  inc_model = inc_model.export_compressed_model(use_optimum_format=True)
  inc_model.eval()
  q_model = replace_linear(inc_model, None, None, config, device=device)

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -195,7 +195,7 @@ def save_low_bit(
  return
 
  if self.quantization_config.weight_dtype not in \
- ["fp8_e5m2", "fp8_e4m3", "nf4", "fp4", "int8", "int4_fullrange"]:
+ ["fp8_e5m2", "fp8_e4m3", "nf4", "fp4", "int4_fullrange"]:
  convert_model_to_public(self)
  os.makedirs(save_directory, exist_ok=True)
  # use transformers original `save_pretrained` function
@@ -1171,7 +1171,7 @@ def load_low_bit(cls, pretrained_model_name_or_path, *model_args, **kwargs):
  else:
  model = model_class(config, *model_args, **kwargs)
  if config.quantization_config["weight_dtype"] not in \
- ["fp8_e5m2", "fp8_e4m3", "fp4", "nf4", "int8", "int4_fullrange"]:
+ ["fp8_e5m2", "fp8_e4m3", "fp4", "nf4", "int4_fullrange"]:
  model = build_woq_model(model, quantization_config)
  else:
  model = replace_linear(
@@ -1222,7 +1222,7 @@ def load_low_bit(cls, pretrained_model_name_or_path, *model_args, **kwargs):
  # Set model in evaluation mode to deactivate DropOut modules by default
  model.eval()
  if config.quantization_config["weight_dtype"] not in \
- ["fp8_e5m2", "fp8_e4m3", "int8", "nf4", "fp4" "int4_fullrange"]:
+ ["fp8_e5m2", "fp8_e4m3", "nf4", "fp4" "int4_fullrange"]:
  model = replace_linear(
  model,
  quantization_config=quantization_config,

tests/CI/test_quantization.py

@@ -433,7 +433,7 @@ def test_quantization_for_llm(self):
  )
  bit8_model.eval()
  output = bit8_model(dummy_input)
- self.assertTrue(isclose(float(output[0][0][0][0]), 0.1675747185945511, rel_tol=1e-04))
+ self.assertTrue(isclose(float(output[0][0][0][0]), 0.16759155690670013, rel_tol=1e-04))
 
  # GPTQ
  woq_config = GPTQConfig(bits=4,

tests/CI/test_weight_only.py

@@ -29,8 +29,14 @@
  Trainer
 )
 from intel_extension_for_transformers.transformers.modeling import AutoModelForCausalLM
-from intel_extension_for_transformers.transformers.llm.quantization.nn.modules import QuantizedLinearQBits, QuantizedLoraLinearQBits
-from intel_extension_for_transformers.transformers.llm.quantization.utils import convert_to_quantized_model, replace_linear
+from intel_extension_for_transformers.transformers.llm.quantization.nn.modules import (
+ QuantizedLinearQBits,
+ QuantizedLoraLinearQBits
+)
+from intel_extension_for_transformers.transformers.llm.quantization.utils import (
+ convert_to_quantized_model,
+ replace_linear
+)
 from intel_extension_for_transformers.transformers.llm.utils.generation import _beam_search, _greedy_search
 from intel_extension_for_transformers.transformers import RtnConfig