fix: add missing whitespace around arithmetic operators (#459)

Fixed PEP 8 compliance issues by adding proper spacing around arithmetic
operators in f-strings and expressions. This resolves all flake8 E225/E226
formatting errors that were preventing pre-commit checks from passing.

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: XHPlus

llmc/compression/quantization/quarot.py

@@ -96,7 +96,7 @@ def get_orthogonal_matrix(self):
             raise ValueError(f'Unsupported mode {self.mode}')
 
     def block_transform(self, block):
-        logger.info(f'Start transform the {self.block_idx+1}-th block')
+        logger.info(f'Start transform the {self.block_idx + 1}-th block')
 
         if self.online_rotate:
             self.replace_rotate_linears(block)
@@ -108,7 +108,7 @@ def block_transform(self, block):
         gc.collect()
 
         logger.info(f'block:{block}')
-        logger.info(f'End transform the {self.block_idx+1}-th block')
+        logger.info(f'End transform the {self.block_idx + 1}-th block')
 
     @torch.no_grad()
     def subset_transform(self, block, subset):

llmc/compression/quantization/spqr.py

@@ -90,7 +90,7 @@ def block_transform_true_sequential(self, block, input_feat):
 
     @torch.no_grad()
     def block_transform(self, block, input_feat, *block_kwargs):
-        logger.info(f'Start transform the {self.block_idx+1}-th block')
+        logger.info(f'Start transform the {self.block_idx + 1}-th block')
 
         if self.true_sequential:
             self.block_transform_true_sequential(block, input_feat)
@@ -103,7 +103,7 @@ def block_transform(self, block, input_feat, *block_kwargs):
                 self.get_replacement_params(mode='fake_quant', w_only=True),
             )
 
-        logger.info(f'End transform the {self.block_idx+1}-th block')
+        logger.info(f'End transform the {self.block_idx + 1}-th block')
 
     @torch.no_grad()
     def subset_transform(self, layers_dict):

llmc/compression/quantization/tesseraq.py

@@ -176,7 +176,7 @@ def collect_block_qparams(self, block, input_feat):
 
     @torch.no_grad()
     def block_transform(self, block, input_feat, block_kwargs):
-        logger.info(f'Start transform the {self.block_idx+1}-th block')
+        logger.info(f'Start transform the {self.block_idx + 1}-th block')
 
         with torch.no_grad():
             block.float()
@@ -204,7 +204,7 @@ def block_transform(self, block, input_feat, block_kwargs):
         if self.reduce_memory:
             block.to(self.model_dtype)
 
-        logger.info(f'End transform the {self.block_idx+1}-th block')
+        logger.info(f'End transform the {self.block_idx + 1}-th block')
 
     def tesseraq_train(self, block):
         self.set_dynamic_tmp_quant(block, on=True)
@@ -273,8 +273,8 @@ def tesseraq_train(self, block):
                     norm = loss_scaler(loss, optimizer, parameters=params_r + params_s)
 
                 logger.info(
-                    f'block {self.block_idx} iter {i+1} loss:{loss.item():5f} \
-                    norm:{norm.item():4f} HR progress:{(1-thresholds[i])*100:1f}% '
+                    f'block {self.block_idx} iter {i + 1} loss:{loss.item():5f} \
+                    norm:{norm.item():4f} HR progress:{(1 - thresholds[i]) * 100:1f}% '
                 )
                 for p in params_r + params_s:
                     p.requires_grad = False

llmc/compression/sparsification/base_blockwise_sparsification.py

@@ -154,7 +154,7 @@ def block_opt(self, block):
             self.block_transform(block)
 
     def block_transform(self, block, input_feat, block_kwargs):
-        logger.info(f'Start transform the {self.block_idx+1}-th block')
+        logger.info(f'Start transform the {self.block_idx + 1}-th block')
         subsets = self.model.get_subsets_in_block(block)
         for index, subset in enumerate(subsets):
             if not self.filter_subset(subset):
@@ -174,7 +174,7 @@ def block_transform(self, block, input_feat, block_kwargs):
                 inspect_module,
                 subset_kwargs
             )
-        logger.info(f'End transform the {self.block_idx+1}-th block')
+        logger.info(f'End transform the {self.block_idx + 1}-th block')
 
     def filter_subset(self, subset):
         return True

llmc/compression/sparsification/dense.py

@@ -11,6 +11,6 @@ def __init__(self, model, sparsity_config, input, padding_mask, config):
         super().__init__(model, sparsity_config, input, padding_mask, config)
 
     def block_transform(self, block):
-        logger.info(f'Start transform the {self.block_idx+1}-th block')
+        logger.info(f'Start transform the {self.block_idx + 1}-th block')
         logger.info(block)
-        logger.info(f'End transform the {self.block_idx+1}-th block')
+        logger.info(f'End transform the {self.block_idx + 1}-th block')

llmc/compression/sparsification/shortgpt.py

@@ -30,7 +30,7 @@ def block_opt(self, block):
         self.input['data'] = output_feat
 
     def block_transform(self, input_feat, output_feat):
-        logger.info(f'Start transform the {self.block_idx+1}-th block')
+        logger.info(f'Start transform the {self.block_idx + 1}-th block')
         self.subset_transform(
             input_feat,
             output_feat

llmc/eval/eval_vqa.py

@@ -255,8 +255,8 @@ def _adjust_config(task_dict):
             gen_max_mem = torch.cuda.max_memory_allocated() / 1024 / 1024
 
             logger.info(f'peak memory: {gen_max_mem:.1f} MB.')
-            logger.info(f'prefill average time: {prefill *1000:.1f} ms.')
-            logger.info(f'decode average time: {decode *1000:.1f} ms.')
+            logger.info(f'prefill average time: {prefill * 1000:.1f} ms.')
+            logger.info(f'decode average time: {decode * 1000:.1f} ms.')
 
         if hasattr(lm, '_model'):
             del lm._model

llmc/models/mixtral.py

@@ -59,7 +59,8 @@ def get_subsets_in_block(self, block):
         return self._get_subsets_fused(block)
 
     def _get_subsets_legacy(self, block):
-        """transformers <5.0: block.block_sparse_moe with ModuleList experts."""
+        """Transformers <5.0: block.block_sparse_moe with ModuleList
+        experts."""
         moe = block.block_sparse_moe
         return [
             {
@@ -106,7 +107,7 @@ def _get_subsets_legacy(self, block):
         ]
 
     def _get_subsets_fused(self, block):
-        """transformers >=5.0: block.mlp with fused MixtralExperts."""
+        """Transformers >=5.0: block.mlp with fused MixtralExperts."""
         moe = block.mlp
         return [
             {