fix: pad token perf in perplexity scoring, SGLang SSRF validation

- Use -100 ignore_index for pad tokens in perplexity computation (avoids
  redundant softmax on padding positions)
- Block URL-based model paths in SGLang create_sglang_runtime (SSRF protection)

soup_cli/utils/quality.py

@@ -63,7 +63,10 @@ def compute_perplexity_scores(
             shift_labels = input_ids[:, 1:].contiguous()
             shift_mask = attention_mask[:, 1:].contiguous()
 
-            loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+            # Set pad positions to -100 so CrossEntropyLoss skips them
+            shift_labels = shift_labels.masked_fill(shift_mask == 0, -100)
+
+            loss_fct = torch.nn.CrossEntropyLoss(reduction="none", ignore_index=-100)
             per_token_loss = loss_fct(
                 shift_logits.view(-1, shift_logits.size(-1)),
                 shift_labels.view(-1),

soup_cli/utils/sglang.py

@@ -47,8 +47,18 @@ def create_sglang_runtime(
     Returns:
         (runtime, runtime_model_name) tuple.
     """
+    import re
+
     import sglang as sgl
 
+    # SSRF protection: block URL-based model paths
+    for path_val in (model_path, base_model):
+        if path_val and re.match(r'^https?://', path_val):
+            raise ValueError(
+                "model_path/base_model must be a local path or HuggingFace model ID, "
+                "not a URL"
+            )
+
     # For LoRA adapters, load the base model
     if is_adapter and base_model:
         runtime = sgl.Runtime(