Add debug logging to benchmark parsing

app.py

@@ -503,31 +503,38 @@ def _run_benchmark():
             wall_time = time.time() - t0
             output_text = proc.stdout.strip()
 
+            # Log raw output for debugging
+            _log(f"Benchmark: {model_info['name']} stdout ({len(output_text)} chars): {output_text[:200]}")
+            _log(f"Benchmark: {model_info['name']} stderr ({len(proc.stderr)} chars): {proc.stderr[-500:]}")
+
             # Parse llama.cpp stderr for timing info
-            # llama.cpp outputs: "llama_perf_sampler_print: ... prompt N tokens ... M tokens/s"
-            # and "llama_perf_sampler_print: ... eval N tokens ... M tokens/s"
+            # Various formats depending on version:
+            # "prompt eval time = 1234.56 ms / 12 tokens (102.88 ms per token, 9.72 tokens per second)"
+            # "eval time = 5678.90 ms / 64 tokens (88.73 ms per token, 11.27 tokens per second)"
             prompt_tps = 0.0
             gen_tps = 0.0
             prompt_tok = 0
             gen_tok = 0
+            import re
             for line in proc.stderr.splitlines():
-                if "prompt eval time" in line:
-                    # e.g. "prompt eval time = 1234.56 ms / 12 tokens (102.88 ms per token, 9.72 tokens per second)"
-                    import re
-                    m = re.search(r"(\d+\.?\d*) tokens per second", line)
-                    if m:
-                        prompt_tps = float(m.group(1))
-                    m2 = re.search(r"/ (\d+) tokens", line)
-                    if m2:
-                        prompt_tok = int(m2.group(1))
-                elif "eval time" in line and "prompt" not in line:
-                    import re
-                    m = re.search(r"(\d+\.?\d*) tokens per second", line)
-                    if m:
-                        gen_tps = float(m.group(1))
-                    m2 = re.search(r"/ (\d+) tokens", line)
-                    if m2:
-                        gen_tok = int(m2.group(1))
+                # Match: "N.NN tokens per second"
+                tps_match = re.search(r"(\d+\.?\d*) tokens per second", line)
+                tok_match = re.search(r"/ (\d+) tokens", line)
+                if "prompt eval" in line.lower() or "prompt_eval" in line.lower():
+                    if tps_match:
+                        prompt_tps = float(tps_match.group(1))
+                    if tok_match:
+                        prompt_tok = int(tok_match.group(1))
+                elif ("eval time" in line.lower() or "eval_time" in line.lower()) and "prompt" not in line.lower():
+                    if tps_match:
+                        gen_tps = float(tps_match.group(1))
+                    if tok_match:
+                        gen_tok = int(tok_match.group(1))
+
+            # If parsing failed, try counting newlines as rough token estimate
+            if gen_tok == 0 and output_text:
+                gen_tok = len(output_text.split())
+                _log(f"Benchmark: parsing failed, estimated {gen_tok} tokens from word count")
 
             result.update({
                 "status": "success",