[CPU] [Quantization] Add GPTQ/AWQ 4bits quantization support for CPU (#22685)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

test/srt/cpu/utils.py

@@ -402,3 +402,39 @@ def unpack_and_dequant_awq(
     fp16_weight = qdq_weight_T.T
 
     return fp16_weight, zeros
+
+
+def unpack_4bit_to_32bit_signed(qweight, qzeros):
+    # Unpack 4-bit values and interpret them as signed integers
+    unpacked_weights = torch.zeros(
+        (qweight.shape[0] * 8, qweight.shape[1]),
+        dtype=torch.int8,
+        device=qweight.device,
+        requires_grad=False,
+    )
+    unpacked_zeros = torch.zeros(
+        (qzeros.shape[0], qzeros.shape[1] * 8),
+        dtype=torch.int8,
+        device=qzeros.device,
+        requires_grad=False,
+    )
+
+    for row in range(unpacked_weights.shape[0]):
+        i = row % 8
+        unpacked_weights[row, :] = (qweight[row // 8, :] >> (4 * i)) & 0xF
+
+    for col in range(unpacked_zeros.shape[1]):
+        i = col % 8
+        unpacked_zeros[:, col] = (qzeros[:, col // 8] >> (4 * i)) & 0xF
+
+    return unpacked_weights, unpacked_zeros + 1
+
+
+def unpack_and_dequant_gptq(qweight, qzeros, scales):
+    unpacked_qweight, unpacked_qzeros = unpack_4bit_to_32bit_signed(qweight, qzeros)
+    group_size = unpacked_qweight.shape[0] // scales.shape[0]
+    scales = scales.repeat_interleave(group_size, dim=0)
+    unpacked_qzeros = unpacked_qzeros.repeat_interleave(group_size, dim=0)
+    unpacked_qweight = (unpacked_qweight - unpacked_qzeros) * scales
+
+    return unpacked_qweight.T