Fix ci bug

Author: gushiqiao

llmc/__main__.py

@@ -100,11 +100,12 @@ def main(config):
             org_model = MODEL_REGISTRY[config.model.type](
                 config.model.path, config.model.torch_dtype
             )
-            token_consist_eval = TokenConsistencyEval(tokenizer.get_tokenizer(), eval_config)
+            token_consist_eval = TokenConsistencyEval(tokenizer.get_tokenizer(),
+                                                      eval_config)
             consistency_ratio = token_consist_eval.eval(model, org_model)
             logger.info(f'Token consistency ratio: {consistency_ratio}')
             del org_model
-            
+
     if 'save' in config and config.save.get('save_fake', False):
         blockwise_opt.deploy('fake_quant')
         blockwise_opt.save_model(save_fake_path)

llmc/compression/quantization/gptq.py

@@ -239,13 +239,13 @@ def weight_transform(self, W, Hinv, Losses, tmp):
                 w, d = W1[:, i], Hinv1[i, i]
                 idx = i1 + i
 
-                if self.wquantizer.granularity == "per_group":
+                if self.wquantizer.granularity == 'per_group':
                     idx = i1 + i
                     if not self.static_groups:
                         if (i1 + i) % self.wquantizer.group_size == 0:
                             column_tensors = W[
                                 :,
-                                (i1 + i) : min(
+                                (i1 + i):min(
                                     (i1 + i + self.wquantizer.group_size),
                                     (self.columns - self.n_out),
                                 ),

llmc/eval/eval_base.py

@@ -114,4 +114,4 @@ def eval(self, model_llmc, model_org=None):
 
         gc.collect()
         torch.cuda.empty_cache()
-        return eval_res
+        return eval_res