[pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

tests/metrics/test_metrics_standalone.py

@@ -114,6 +114,7 @@ try:
 except Exception as e:
     print(f"   ⚠️  Prometheus test skipped: {e}")
     import traceback
+
     traceback.print_exc()
 
 print("\n" + "=" * 60)

unsloth/metrics/stats.py

@@ -110,7 +110,7 @@ class InferenceStats:
 
     def record_first_token(self, request_id: str, timestamp: Optional[float] = None):
         """Record when the first token was generated.
-        
+
         Args:
             request_id: Unique identifier for the request
             timestamp: Optional timestamp. If None, uses current time.
@@ -119,7 +119,9 @@ class InferenceStats:
             if request_id in self._active_requests:
                 req = self._active_requests[request_id]
                 if req.first_token_time is None:
-                    req.first_token_time = timestamp if timestamp is not None else time.time()
+                    req.first_token_time = (
+                        timestamp if timestamp is not None else time.time()
+                    )
                     if req.scheduled_time is None:
                         req.scheduled_time = req.first_token_time
 

unsloth/models/_utils.py

@@ -1917,20 +1917,20 @@ def _patch_training_metrics(Trainer):
                     if hasattr(tensor, "shape") and len(tensor.shape) > 0:
                         batch_size = tensor.shape[0]
                         break
-        
+
         # Track step duration (includes both forward and backward passes)
         step_start = time.time()
-        
+
         # Call original training_step
         try:
             result = original_training_step(self, model, inputs, *args, **kwargs)
         except Exception as e:
             # Re-raise exception but don't track metrics on error
             raise
-        
+
         step_end = time.time()
         step_duration = step_end - step_start
-        
+
         # Extract loss and other info from result
         loss_value = None
         if isinstance(result, (int, float)):

unsloth/models/vision.py

@@ -365,22 +365,28 @@ def unsloth_base_fast_generate(
                     else:
                         total_tokens = sequences.shape[-1]
                     num_generation_tokens = max(0, total_tokens - num_prompt_tokens)
-            
+
             # Estimate timing (simplified)
             # Note: These are estimations. For more accurate metrics, consider hooking into
             # the generation process itself (e.g., via LogitsProcessor or StoppingCriteria)
             if num_generation_tokens > 0:
                 # Estimate first token time
-                estimated_first_token_time = start_time + (e2e_latency / (num_generation_tokens + 1))
-                collector.inference_stats.record_first_token(request_id, timestamp=estimated_first_token_time)
-                
+                estimated_first_token_time = start_time + (
+                    e2e_latency / (num_generation_tokens + 1)
+                )
+                collector.inference_stats.record_first_token(
+                    request_id, timestamp = estimated_first_token_time
+                )
+
                 # Record tokens (simplified - records all at once after generation)
                 for _ in range(num_generation_tokens):
                     collector.inference_stats.record_token(request_id)
 
             # Determine finish reason (simplified - could be improved)
             finish_reason = "stop"  # Default
-            if isinstance(output, (dict, type(output))) and hasattr(output, "finish_reason"):
+            if isinstance(output, (dict, type(output))) and hasattr(
+                output, "finish_reason"
+            ):
                 finish_reason = output.finish_reason
             elif isinstance(output, dict) and "finish_reason" in output:
                 finish_reason = output["finish_reason"]