Add the min, max to embedding 4bit

#1506
Please refer it for more context information.

Case:
Failed to export the embedding 4bit model with the following command:
python torchchat.py export stories110m --quantize torchchat/quant_config/mobile.json --output-pte-path stories110me4.pte

Solution:
Add the min, max to embedding 4bit.

Test:
The 4bit,8bit of embedding are exported successfully by the commands:
Change torchchat/quant_config/mobile.json to:
{
    "embedding": {"bitwidth": 8, "groupsize" : 32},
    "linear:a8w4dq": {"groupsize" : 256}
}
python torchchat.py export stories110m --quantize torchchat/quant_config/mobile.json --output-pte-path stories110me8.pte

{
"embedding": {"bitwidth": 4, "groupsize" : 32},
"linear:a8w4dq": {"groupsize" : 256}
}
python torchchat.py export stories110m --quantize torchchat/quant_config/mobile.json --output-pte-path stories110me4.pte

Signed-off-by: jijie <[email protected]>
Co-authored-by: jijie <[email protected]>

Author: TheBetterSolution

torchchat/utils/quantize.py

@@ -747,7 +747,7 @@ def et_forward(self, indices: torch.Tensor) -> torch.Tensor:
             )
         else:
             return torch.ops.quantized_decomposed.embedding_4bit.dtype(
-                self.weight, self.scales, None, 0, 0, indices, dtype=self.dtype
+                self.weight, self.scales, None, -8, 7, indices, dtype=self.dtype
             )
 
     @torch.no_grad()