fix(benchmarks): separate quality measurement from efficiency proxy (issue #63)

- Add --quality flag to run_benchmarks.py that delegates to llama-perplexity
- Clarify token/sec is an efficiency metric, not perplexity
- Ollama cannot provide true logprob-based PPL (no logprob API)
- Quality gate now runs llama-perplexity binary directly when requested

Closes #63

benchmarks/run_benchmarks.py

@@ -1,17 +1,26 @@
 #!/usr/bin/env python3
 """
-TurboQuant Benchmarking Suite — Multi-Backend (Issue #29)
+TurboQuant Benchmarking Suite — Multi-Backend (Issue #29, #63)
 
 Supports Ollama and llama-server backends with KV cache type configuration.
 Measures: TTFT, tokens/sec, latency, peak memory.
 
+Perplexity (quality) is NOT measured here tokens/sec is a throughput proxy.
+For actual quality (logprob-based PPL), use --quality flag which delegates to
+llama-perplexity binary, since Ollama lacks logprob support (issue #63).
+
 Usage:
-    # Ollama (default)
+    # Ollama (efficiency only)
     python3 benchmarks/run_benchmarks.py --backend ollama --model llama3
 
-    # llama-server with turbo4 KV
+    # llama-server with turbo4 KV + quality gate in one shot
     python3 benchmarks/run_benchmarks.py --backend llama-server \
-        --url http://localhost:11434 --model qwen3.5 --kv-type turbo4
+        --url http://localhost:11434 --model qwen3.5 --kv-type turbo4 --quality
+
+    # Quality gate only (separate tool)
+    python3 benchmarks/run_perplexity.py --model ~/models/qwen3.5-27b.gguf \
+        --llama-cpp ~/turboquant/llama.cpp-fork/build/bin/llama-perplexity \
+        --corpus corpora/wiki.test.raw --context 2048
 """
 
 import argparse
@@ -108,9 +117,7 @@ def run_llama_server(prompt: str, model: str, url: str, kv_type: str = "f16",
         completion_tokens = usage.get("completion_tokens", 0)
         prompt_tokens = usage.get("prompt_tokens", 0)
 
-        # llama-server includes timing in x_* headers or we estimate
         if elapsed > 0 and completion_tokens > 0:
-            # Subtract estimated prompt eval time (rough)
             tokens_per_sec = completion_tokens / max(elapsed - 0.1, 0.01)
 
         return {
@@ -128,8 +135,10 @@ def run_llama_server(prompt: str, model: str, url: str, kv_type: str = "f16",
 
 
 def run_benchmark_suite(backend: str, model: str, url: str, kv_type: str,
-                        prompts_file: str, output_file: str, timeout: int = 120):
-    """Run the full benchmark suite."""
+                        prompts_file: str, output_file: str, timeout: int = 120,
+                        measure_quality: bool = False, quality_corpus: str = None,
+                        llama_cpp_bin: str = None, context: int = 2048, threads: int = 4):
+    """Run the full benchmark suite, optionally measuring perplexity in parallel."""
     if not os.path.exists(prompts_file):
         print(f"ERROR: {prompts_file} not found")
         sys.exit(1)
@@ -191,15 +200,76 @@ def run_benchmark_suite(backend: str, model: str, url: str, kv_type: str,
         }
     }
 
+    # Issue #63: Optional quality measurement via llama-perplexity (Ollama lacks logprob)
+    if measure_quality:
+        print("\n" + "="*60)
+        print("Quality measurement requested — invoking llama-perplexity binary...")
+        llama_cpp_bin = llama_cpp_bin or "llama.cpp-fork/build/bin/llama-perplexity"
+        quality_corpus = quality_corpus or "corpora/wiki.test.raw"
+
+        if not os.path.exists(quality_corpus):
+            print(f"WARNING: quality corpus not found: {quality_corpus}")
+            suite["quality"] = {"perplexity": None, "passed": False, "error": f"Corpus missing: {quality_corpus}"}
+        elif not os.path.exists(llama_cpp_bin):
+            print(f"WARNING: llama-perplexity binary not found: {llama_cpp_bin}")
+            suite["quality"] = {"perplexity": None, "passed": False, "error": f"Binary missing: {llama_cpp_bin}"}
+        else:
+            cmd = [
+                llama_cpp_bin,
+                "-m", model,
+                "-f", quality_corpus,
+                "-c", str(context),
+                "-t", str(threads),
+                "--kv-type", kv_type,
+            ]
+            try:
+                start = time.time()
+                result = subprocess.run(cmd, capture_output=True, text=True, timeout=3600)
+                elapsed = time.time() - start
+                output = result.stdout + "\n" + result.stderr
+
+                ppl_match = re.search(r"perplexity[:\s]+(\d+\.?\d*)", output, re.IGNORECASE)
+                ppl = float(ppl_match.group(1)) if ppl_match else None
+                token_match = re.search(r"(\d+) tokens", output)
+                tokens = int(token_match.group(1)) if token_match else None
+
+                ppl_result = {
+                    "kv_type": kv_type,
+                    "perplexity": ppl,
+                    "tokens": tokens,
+                    "elapsed_seconds": round(elapsed, 1),
+                    "exit_code": result.returncode,
+                    "passed": result.returncode == 0,
+                    "output_tail": output.strip()[-500:] if output else "",
+                }
+                suite["quality"] = ppl_result
+                if ppl is not None:
+                    print(f"  Perplexity ({kv_type}): {ppl:.4f}")
+                else:
+                    print(f"  Perplexity: FAILED — could not parse output")
+            except subprocess.TimeoutExpired:
+                suite["quality"] = {"perplexity": None, "passed": False, "error": "Timeout after 3600s"}
+                print("  Perplexity: FAILED — timeout after 3600s")
+            except Exception as e:
+                suite["quality"] = {"perplexity": None, "passed": False, "error": str(e)}
+                print(f"  Perplexity: FAILED — {e}")
+        print("="*60)
+
     os.makedirs(os.path.dirname(output_file) or ".", exist_ok=True)
-    with open(output_file, "w") as f:
-        json.dump(suite, f, indent=2)
+    with open(output_file, "w") as fh:
+        json.dump(suite, fh, indent=2)
 
     s = suite["summary"]
     print(f"\n{'='*60}")
     print(f"RESULTS: {s['success']}/{s['total']} success | "
           f"Avg {s['avg_tok_per_sec']:.1f} tok/s | "
           f"Avg {s['avg_latency_s']:.2f}s latency")
+    if "quality" in suite:
+        q = suite["quality"]
+        if q.get("perplexity") is not None:
+            print(f"Quality: PPL = {q['perplexity']:.4f}")
+        else:
+            print(f"Quality: not available — {q.get('error','unknown')}")
     print(f"{'='*60}")
     print(f"Saved to {output_file}")
 
@@ -207,20 +277,45 @@ def run_benchmark_suite(backend: str, model: str, url: str, kv_type: str,
 def main():
     parser = argparse.ArgumentParser(description="TurboQuant Benchmark Suite")
     parser.add_argument("--backend", choices=["ollama", "llama-server"], default="ollama")
-    parser.add_argument("--model", required=True, help="Model name")
+    parser.add_argument("--model", required=True, help="Model name or path")
     parser.add_argument("--url", default="http://localhost:11434", help="Backend URL")
     parser.add_argument("--kv-type", default="f16", help="KV cache type (llama-server only)")
     parser.add_argument("--prompts", default="benchmarks/prompts.json", help="Prompts file")
     parser.add_argument("--output", default=None, help="Output file (auto-generated if omitted)")
     parser.add_argument("--timeout", type=int, default=120, help="Per-prompt timeout (s)")
+
+    # Issue #63: Quality measurement (Ollama lacks logprob → use llama-perplexity binary)
+    parser.add_argument("--quality", action="store_true", default=False,
+                        help="Also run quality measurement via llama-perplexity binary")
+    parser.add_argument("--llama-cpp", default="llama.cpp-fork/build/bin/llama-perplexity",
+                        help="Path to llama-perplexity binary")
+    parser.add_argument("--quality-corpus", default="corpora/wiki.test.raw",
+                        help="Test corpus for perplexity measurement")
+    parser.add_argument("--context", type=int, default=2048,
+                        help="Context length for quality measurement")
+    parser.add_argument("--threads", type=int, default=4,
+                        help="Thread count for quality measurement")
+
     args = parser.parse_args()
 
     if args.output is None:
         ts = int(time.time())
         args.output = f"benchmarks/results_{args.backend}_{args.kv_type}_{ts}.json"
 
-    run_benchmark_suite(args.backend, args.model, args.url, args.kv_type,
-                        args.prompts, args.output, args.timeout)
+    run_benchmark_suite(
+        backend=args.backend,
+        model=args.model,
+        url=args.url,
+        kv_type=args.kv_type,
+        prompts_file=args.prompts,
+        output_file=args.output,
+        timeout=args.timeout,
+        measure_quality=args.quality,
+        quality_corpus=args.quality_corpus,
+        llama_cpp_bin=args.llama_cpp,
+        context=args.context,
+        threads=args.threads,
+    )
 
 
 if __name__ == "__main__":

benchmarks/run_perplexity.py

@@ -1,8 +1,9 @@
 #!/usr/bin/env python3
 """
-TurboQuant Perplexity Quality Gate (Issue #21)
+TurboQuant Perplexity Quality Gate (Issues #21, #63)
 
-Compares text generation quality between f16 KV and turbo4 KV cache
+Measures true perplexity via llama-perplexity binary (logprob-based).
+Ollama cannot provide perplexity due to missing logprob API (issue #63).
 configurations using llama.cpp's perplexity tool on the wikitext-2 corpus.
 
 Usage: