feat: M1 benchmark suite — throughput, memory, perplexity comparison

Implements Issue #80: benchmark turboquant vs llama.cpp baseline on M1.

New files:
- benchmarks/run_m1_benchmark.py — comprehensive benchmark runner
- benchmarks/run_benchmark_m1.sh — shell wrapper for easy execution
- tests/test_m1_benchmark.py — unit tests for benchmark functions

Measures:
- Tokens/sec throughput (f16 vs turbo4, 3-run average)
- Memory usage (RSS monitoring during inference)
- Quality via perplexity (llama-perplexity on wikitext-2)

Generates:
- benchmarks/m1_benchmark_results.json — raw results
- benchmarks/m1_benchmark_report.md — markdown comparison table

Closes #80

benchmarks/run_benchmark_m1.sh

@@ -0,0 +1,91 @@
+#!/bin/bash
+# TurboQuant M1 Benchmark Runner (Issue #80)
+# Runs both f16 and turbo4 KV configs against same model, collects throughput + memory + perplexity.
+#
+# Prerequisites:
+#   - llama-server built from llama-cpp-turboquant fork (feature/turboquant-kv-cache)
+#   - Model GGUF file downloaded
+#   - wikitext-2 corpus in corpora/wiki.test.raw
+#
+# Usage:
+#   ./benchmarks/run_benchmark_m1.sh <model_name> <model_path> [llama_server_url]
+#
+# Example:
+#   ./benchmarks/run_benchmark_m1.sh qwen3.5:27b ~/models/qwen3.5-27b-q4_k_m.gguf
+
+set -euo pipefail
+
+MODEL_NAME="${1:?Usage: $0 <model_name> <model_path> [llama_server_url]}"
+MODEL_PATH="${2:?Model path required}"
+LLAMA_SERVER="${3:-http://localhost:8080}"
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+PROJECT_DIR="$(dirname "$SCRIPT_DIR")"
+LLAMA_BIN="${PROJECT_DIR}/llama.cpp-fork/build/bin"
+CORPUS="${PROJECT_DIR}/corpora/wiki.test.raw"
+OUTPUT_DIR="${PROJECT_DIR}/benchmarks"
+
+echo "=========================================="
+echo "TurboQuant M1 Benchmark"
+echo "=========================================="
+echo "Model:      ${MODEL_NAME}"
+echo "Model path: ${MODEL_PATH}"
+echo "Server:     ${LLAMA_SERVER}"
+echo "llama bin:  ${LLAMA_BIN}"
+echo "Corpus:     ${CORPUS}"
+echo ""
+
+# Check prerequisites
+if [ ! -f "${MODEL_PATH}" ]; then
+    echo "WARNING: Model file not found: ${MODEL_PATH}"
+    echo "  Perplexity tests will be skipped."
+fi
+
+if [ ! -f "${LLAMA_BIN}/llama-perplexity" ]; then
+    echo "WARNING: llama-perplexity not found at ${LLAMA_BIN}/llama-perplexity"
+    echo "  Perplexity tests will be skipped."
+fi
+
+if [ ! -f "${CORPUS}" ]; then
+    echo "WARNING: Corpus not found: ${CORPUS}"
+    echo "  Download with: curl -L https://raw.githubusercontent.com/pytorch/examples/main/word_language_model/data/wikitext-2/wiki.test.raw -o ${CORPUS}"
+fi
+
+# Check server is running
+echo "Checking llama-server at ${LLAMA_SERVER}..."
+if curl -sf "${LLAMA_SERVER}/health" > /dev/null 2>&1; then
+    echo "  Server is running ✓"
+else
+    echo "  Server not responding. Trying /v1/models..."
+    if curl -sf "${LLAMA_SERVER}/v1/models" > /dev/null 2>&1; then
+        echo "  Server is running (no /health endpoint) ✓"
+    else
+        echo "  ERROR: llama-server not reachable at ${LLAMA_SERVER}"
+        echo "  Start with: llama-server -m ${MODEL_PATH} --port 8080 -ctk turbo4 -ctv turbo4 -c 4096"
+        exit 1
+    fi
+fi
+
+# Run benchmark
+echo ""
+echo "Starting benchmark suite..."
+python3 "${SCRIPT_DIR}/run_m1_benchmark.py" \
+    --model "${MODEL_NAME}" \
+    --model-path "${MODEL_PATH}" \
+    --backend llama-server \
+    --llama-server "${LLAMA_SERVER}" \
+    --llama-bin "${LLAMA_BIN}" \
+    --corpus "${CORPUS}" \
+    --context 2048 \
+    --threads 4 \
+    --num-predict 256 \
+    --runs 3 \
+    --output-dir "${OUTPUT_DIR}" \
+    --ppl-threshold 0.5
+
+echo ""
+echo "=========================================="
+echo "Done. Results in:"
+echo "  ${OUTPUT_DIR}/m1_benchmark_results.json"
+echo "  ${OUTPUT_DIR}/m1_benchmark_report.md"
+echo "=========================================="

benchmarks/run_m1_benchmark.py

@@ -0,0 +1,681 @@
+#!/usr/bin/env python3
+"""
+TurboQuant M1 Benchmark Suite (Issue #80)
+
+Comprehensive benchmark comparing TurboQuant (turbo4 KV) vs baseline (f16 KV)
+on Apple M1 Mac. Measures: tokens/sec, memory usage, quality (perplexity).
+
+Usage:
+    python3 benchmarks/run_m1_benchmark.py \
+        --model qwen3.5:27b \
+        --llama-server http://localhost:8080 \
+        --llama-bin ~/llama-cpp-turboquant/build/bin
+
+    # Skip perplexity (quick throughput/memory only)
+    python3 benchmarks/run_m1_benchmark.py --model qwen3.5:27b --skip-perplexity
+
+Outputs:
+    - benchmarks/m1_benchmark_results.json
+    - benchmarks/m1_benchmark_report.md
+"""
+
+import argparse
+import json
+import os
+import re
+import subprocess
+import sys
+import time
+import threading
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+try:
+    import requests
+except ImportError:
+    print("ERROR: requests package required. Install with: pip install requests")
+    sys.exit(1)
+
+
+# ── Memory Monitoring ───────────────────────────────────────────────────────
+
+class MemoryMonitor:
+    """Monitor memory usage of a process in background."""
+
+    def __init__(self, pid: int, interval: float = 0.5):
+        self.pid = pid
+        self.interval = interval
+        self.samples = []
+        self._stop = threading.Event()
+        self._thread = None
+
+    def start(self):
+        self._stop.clear()
+        self._thread = threading.Thread(target=self._monitor_loop, daemon=True)
+        self._thread.start()
+
+    def stop(self):
+        self._stop.set()
+        if self._thread:
+            self._thread.join(timeout=2)
+        return self.get_stats()
+
+    def _monitor_loop(self):
+        while not self._stop.is_set():
+            try:
+                mem_mb = self._get_memory_mb()
+                if mem_mb > 0:
+                    self.samples.append(mem_mb)
+            except Exception:
+                pass
+            time.sleep(self.interval)
+
+    def _get_memory_mb(self) -> float:
+        if sys.platform == "darwin":
+            result = subprocess.run(
+                ["ps", "-o", "rss=", "-p", str(self.pid)],
+                capture_output=True, text=True
+            )
+            if result.returncode == 0 and result.stdout.strip():
+                return int(result.stdout.strip()) / 1024
+        else:
+            try:
+                with open(f"/proc/{self.pid}/status") as f:
+                    for line in f:
+                        if line.startswith("VmRSS:"):
+                            return int(line.split()[1]) / 1024
+            except FileNotFoundError:
+                pass
+        return 0.0
+
+    def get_stats(self) -> dict:
+        if not self.samples:
+            return {"avg_mb": 0, "peak_mb": 0, "min_mb": 0, "samples": 0}
+        return {
+            "avg_mb": round(sum(self.samples) / len(self.samples), 1),
+            "peak_mb": round(max(self.samples), 1),
+            "min_mb": round(min(self.samples), 1),
+            "samples": len(self.samples),
+        }
+
+
+# ── System Info ─────────────────────────────────────────────────────────────
+
+def get_system_info() -> dict:
+    info = {"platform": sys.platform, "python": sys.version.split()[0]}
+    try:
+        if sys.platform == "darwin":
+            info["chip"] = subprocess.run(
+                ["sysctl", "-n", "machdep.cpu.brand_string"],
+                capture_output=True, text=True
+            ).stdout.strip()
+            mem_bytes = int(subprocess.run(
+                ["sysctl", "-n", "hw.memsize"],
+                capture_output=True, text=True
+            ).stdout.strip())
+            info["memory_gb"] = round(mem_bytes / (1024**3), 1)
+            info["cpu_cores"] = os.cpu_count()
+        else:
+            info["cpu"] = subprocess.run(
+                ["uname", "-m"], capture_output=True, text=True
+            ).stdout.strip()
+            info["cpu_cores"] = os.cpu_count()
+    except Exception:
+        info["error"] = "Could not detect hardware"
+    return info
+
+
+# ── Benchmark Functions ─────────────────────────────────────────────────────
+
+def find_llama_server_pid() -> Optional[int]:
+    """Find PID of running llama-server process."""
+    try:
+        result = subprocess.run(
+            ["pgrep", "-f", "llama-server"],
+            capture_output=True, text=True
+        )
+        if result.stdout.strip():
+            return int(result.stdout.strip().split("\n")[0])
+    except Exception:
+        pass
+    return None
+
+
+def run_throughput_test(prompt: str, model: str, url: str, kv_type: str,
+                        num_predict: int = 256, timeout: int = 120) -> dict:
+    """Run a single throughput test against llama-server."""
+    api_url = f"{url.rstrip('/')}/v1/chat/completions"
+    start = time.time()
+    ttft = None
+    tokens_per_sec = 0.0
+
+    try:
+        resp = requests.post(api_url, json={
+            "model": model,
+            "messages": [{"role": "user", "content": prompt}],
+            "max_tokens": num_predict,
+            "stream": False
+        }, timeout=timeout)
+        elapsed = time.time() - start
+        resp.raise_for_status()
+        data = resp.json()
+
+        response_text = data.get("choices", [{}])[0].get("message", {}).get("content", "")
+        usage = data.get("usage", {})
+        completion_tokens = usage.get("completion_tokens", 0)
+        prompt_tokens = usage.get("prompt_tokens", 0)
+
+        if elapsed > 0 and completion_tokens > 0:
+            tokens_per_sec = completion_tokens / max(elapsed - 0.1, 0.01)
+
+        return {
+            "response_len": len(response_text),
+            "latency_s": round(elapsed, 3),
+            "tokens_per_sec": round(tokens_per_sec, 2),
+            "completion_tokens": completion_tokens,
+            "prompt_tokens": prompt_tokens,
+            "kv_type": kv_type,
+            "status": "success"
+        }
+    except Exception as e:
+        return {"status": "failed", "error": str(e), "latency_s": round(time.time() - start, 3)}
+
+
+def run_ollama_test(prompt: str, model: str, url: str,
+                    num_predict: int = 256, timeout: int = 120) -> dict:
+    """Run a single throughput test against Ollama."""
+    api_url = f"{url.rstrip('/')}/api/generate"
+    start = time.time()
+
+    try:
+        resp = requests.post(api_url, json={
+            "model": model,
+            "prompt": prompt,
+            "stream": False,
+            "options": {"num_predict": num_predict}
+        }, timeout=timeout)
+        elapsed = time.time() - start
+        resp.raise_for_status()
+        data = resp.json()
+
+        response_text = data.get("response", "")
+        eval_count = data.get("eval_count", 0)
+        eval_duration_ns = data.get("eval_duration", 0)
+        prompt_eval_ns = data.get("prompt_eval_duration", 0)
+
+        tokens_per_sec = 0.0
+        if eval_duration_ns > 0:
+            tokens_per_sec = eval_count / (eval_duration_ns / 1e9)
+
+        ttft = None
+        if prompt_eval_ns > 0:
+            ttft = prompt_eval_ns / 1e9
+
+        return {
+            "response_len": len(response_text),
+            "latency_s": round(elapsed, 3),
+            "ttft_s": round(ttft, 3) if ttft else None,
+            "tokens_per_sec": round(tokens_per_sec, 2),
+            "completion_tokens": eval_count,
+            "prompt_tokens": data.get("prompt_eval_count", 0),
+            "status": "success"
+        }
+    except Exception as e:
+        return {"status": "failed", "error": str(e), "latency_s": round(time.time() - start, 3)}
+
+
+def run_perplexity_test(llama_bin: str, model_path: str, corpus: str,
+                        context: int, kv_type: str, threads: int = 4) -> dict:
+    """Run llama-perplexity and parse output."""
+    if not os.path.exists(llama_bin):
+        return {"error": f"Binary not found: {llama_bin}", "passed": False}
+    if not os.path.exists(model_path):
+        return {"error": f"Model not found: {model_path}", "passed": False}
+    if not os.path.exists(corpus):
+        return {"error": f"Corpus not found: {corpus}", "passed": False}
+
+    cmd = [
+        llama_bin,
+        "-m", model_path,
+        "-f", corpus,
+        "-c", str(context),
+        "-t", str(threads),
+        "--kv-type", kv_type,
+    ]
+    print(f"  Command: {' '.join(cmd)}")
+
+    start = time.time()
+    try:
+        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
+
+        return {
+            "kv_type": kv_type,
+            "perplexity": ppl,
+            "tokens": tokens,
+            "elapsed_seconds": round(elapsed, 1),
+            "exit_code": result.returncode,
+            "passed": result.returncode == 0 and ppl is not None,
+            "output_tail": output.strip()[-500:] if output else "",
+        }
+    except subprocess.TimeoutExpired:
+        return {"kv_type": kv_type, "perplexity": None, "error": "Timeout",
+                "passed": False, "elapsed_seconds": 3600}
+
+
+# ── Prompt Sets ─────────────────────────────────────────────────────────────
+
+THROUGHPUT_PROMPTS = [
+    "Explain the difference between TCP and UDP protocols. Include use cases for each.",
+    "Write a Python function that implements binary search on a sorted list.",
+    "What are the three laws of thermodynamics? Explain each in simple terms.",
+    "Describe the process of photosynthesis step by step.",
+    "Write a recursive function to calculate the Fibonacci sequence with memoization.",
+]
+
+
+# ── Report Generation ───────────────────────────────────────────────────────
+
+def generate_report(results: dict, output_path: str):
+    """Generate markdown report from benchmark results."""
+    lines = []
+    lines.append("# TurboQuant M1 Benchmark Report")
+    lines.append("")
+    lines.append(f"**Date:** {results['timestamp']}")
+    lines.append(f"**Hardware:** {results['system'].get('chip', 'unknown')}, "
+                 f"{results['system'].get('memory_gb', '?')}GB RAM, "
+                 f"{results['system'].get('cpu_cores', '?')} cores")
+    lines.append(f"**Model:** {results['model']}")
+    lines.append("")
+
+    # Throughput comparison
+    lines.append("## Throughput Comparison")
+    lines.append("")
+
+    tp = results.get("throughput", {})
+    baseline = tp.get("f16", {})
+    turbo = tp.get("turbo4", {})
+
+    lines.append("| Metric | f16 (baseline) | turbo4 (TurboQuant) | Delta |")
+    lines.append("|:-------|:---------------|:--------------------|:------|")
+
+    def fmt_delta(baseline_val, turbo_val, suffix="", higher_is_better=True):
+        if baseline_val and turbo_val:
+            delta = turbo_val - baseline_val
+            pct = (delta / baseline_val) * 100 if baseline_val else 0
+            sign = "+" if delta >= 0 else ""
+            better = (delta >= 0) if higher_is_better else (delta <= 0)
+            marker = "✓" if better else "⚠"
+            return (f"{baseline_val}{suffix}", f"{turbo_val}{suffix}",
+                    f"{sign}{pct:.1f}% {marker}")
+        return ("N/A", "N/A", "N/A")
+
+    b_tok, t_tok, d_tok = fmt_delta(
+        baseline.get("avg_tok_per_sec"), turbo.get("avg_tok_per_sec"), " tok/s")
+    b_lat, t_lat, d_lat = fmt_delta(
+        baseline.get("avg_latency"), turbo.get("avg_latency"), "s", higher_is_better=False)
+    b_ttft, t_ttft, d_ttft = fmt_delta(
+        baseline.get("avg_ttft"), turbo.get("avg_ttft"), "s", higher_is_better=False)
+
+    lines.append(f"| Tokens/sec (avg) | {b_tok} | {t_tok} | {d_tok} |")
+    lines.append(f"| Latency (avg) | {b_lat} | {t_lat} | {d_lat} |")
+    lines.append(f"| TTFT (avg) | {b_ttft} | {t_ttft} | {d_ttft} |")
+    lines.append("")
+
+    # Per-prompt breakdown
+    lines.append("### Per-Prompt Results")
+    lines.append("")
+    lines.append("| Prompt # | f16 tok/s | turbo4 tok/s | Status |")
+    lines.append("|:---------|:----------|:-------------|:-------|")
+
+    baseline_results = baseline.get("results", [])
+    turbo_results = turbo.get("results", [])
+
+    for i, (b, t) in enumerate(zip(baseline_results, turbo_results), 1):
+        b_tps = b.get("tokens_per_sec", 0)
+        t_tps = t.get("tokens_per_sec", 0)
+        if b.get("status") == "success" and t.get("status") == "success":
+            delta_pct = ((t_tps - b_tps) / b_tps * 100) if b_tps else 0
+            status = "✓" if delta_pct > -20 else "⚠"
+            lines.append(f"| {i} | {b_tps:.1f} | {t_tps:.1f} | {status} ({delta_pct:+.1f}%) |")
+        else:
+            err_b = b.get("error", b.get("status", "?"))
+            err_t = t.get("error", t.get("status", "?"))
+            lines.append(f"| {i} | {err_b} | {err_t} | ✗ |")
+
+    lines.append("")
+
+    # Memory comparison
+    lines.append("## Memory Usage")
+    lines.append("")
+    mem = results.get("memory", {})
+    b_mem = mem.get("f16", {})
+    t_mem = mem.get("turbo4", {})
+
+    lines.append("| Metric | f16 (baseline) | turbo4 (TurboQuant) | Savings |")
+    lines.append("|:-------|:---------------|:--------------------|:--------|")
+
+    if b_mem.get("peak_mb") and t_mem.get("peak_mb"):
+        savings = b_mem["peak_mb"] - t_mem["peak_mb"]
+        savings_pct = (savings / b_mem["peak_mb"]) * 100
+        lines.append(f"| Peak RSS | {b_mem['peak_mb']:.0f} MB | {t_mem['peak_mb']:.0f} MB | "
+                     f"{savings:.0f} MB ({savings_pct:.1f}%) |")
+    if b_mem.get("avg_mb") and t_mem.get("avg_mb"):
+        lines.append(f"| Avg RSS | {b_mem['avg_mb']:.0f} MB | {t_mem['avg_mb']:.0f} MB | "
+                     f"{b_mem['avg_mb'] - t_mem['avg_mb']:.0f} MB |")
+
+    lines.append("")
+
+    # Perplexity
+    ppl = results.get("perplexity", {})
+    if ppl.get("f16") or ppl.get("turbo4"):
+        lines.append("## Quality (Perplexity)")
+        lines.append("")
+        lines.append("| KV Type | Perplexity | Tokens | Time |")
+        lines.append("|:--------|:-----------|:-------|:-----|")
+
+        for kv in ["f16", "turbo4"]:
+            r = ppl.get(kv, {})
+            ppl_val = r.get("perplexity")
+            tokens = r.get("tokens")
+            elapsed = r.get("elapsed_seconds")
+            lines.append(f"| {kv} | {ppl_val:.4f} if ppl_val else 'N/A' | "
+                         f"{tokens or 'N/A'} | {elapsed or 'N/A'}s |")
+
+        if ppl.get("delta") is not None:
+            lines.append("")
+            lines.append(f"**PPL Delta (turbo4 - f16):** {ppl['delta']:+.4f}")
+            lines.append(f"**Threshold:** ≤ {ppl.get('threshold', 0.5)}")
+            lines.append(f"**Result:** {'PASS ✓' if ppl.get('pass') else 'FAIL ✗'}")
+
+        lines.append("")
+
+    # Summary
+    lines.append("## Summary")
+    lines.append("")
+
+    # Compute overall verdict
+    throughput_ok = True
+    if turbo.get("avg_tok_per_sec") and baseline.get("avg_tok_per_sec"):
+        ratio = turbo["avg_tok_per_sec"] / baseline["avg_tok_per_sec"]
+        throughput_ok = ratio >= 0.80  # 80% of baseline is acceptable
+        lines.append(f"- **Throughput:** {ratio*100:.0f}% of baseline "
+                     f"({'PASS' if throughput_ok else 'BORDERLINE'})")
+
+    memory_ok = True
+    if t_mem.get("peak_mb") and b_mem.get("peak_mb"):
+        savings_pct = (b_mem["peak_mb"] - t_mem["peak_mb"]) / b_mem["peak_mb"] * 100
+        memory_ok = savings_pct > 50  # Expect >50% savings
+        lines.append(f"- **Memory savings:** {savings_pct:.1f}% "
+                     f"({'PASS' if memory_ok else 'CHECK'})")
+
+    if ppl.get("pass") is not None:
+        lines.append(f"- **Quality (PPL):** {'PASS' if ppl['pass'] else 'FAIL'} "
+                     f"(delta={ppl.get('delta', 'N/A')})")
+
+    lines.append("")
+
+    # Issues discovered
+    issues = results.get("issues_discovered", [])
+    if issues:
+        lines.append("## Issues Discovered")
+        lines.append("")
+        for issue in issues:
+            lines.append(f"- **{issue['title']}**")
+            lines.append(f"  {issue.get('description', '')}")
+        lines.append("")
+
+    lines.append("---")
+    lines.append(f"*Generated by run_m1_benchmark.py — Issue #80*")
+
+    report = "\n".join(lines)
+    with open(output_path, "w") as f:
+        f.write(report)
+    return report
+
+
+# ── Main ────────────────────────────────────────────────────────────────────
+
+def main():
+    parser = argparse.ArgumentParser(description="TurboQuant M1 Benchmark Suite")
+    parser.add_argument("--model", required=True, help="Model name (e.g. qwen3.5:27b)")
+    parser.add_argument("--model-path", default=None,
+                        help="Path to GGUF model file (for perplexity)")
+    parser.add_argument("--backend", choices=["llama-server", "ollama"],
+                        default="llama-server")
+    parser.add_argument("--llama-server", default="http://localhost:8080",
+                        help="llama-server URL")
+    parser.add_argument("--ollama-url", default="http://localhost:11434",
+                        help="Ollama URL")
+    parser.add_argument("--llama-bin", default=None,
+                        help="Path to llama.cpp build/bin directory")
+    parser.add_argument("--corpus", default="corpora/wiki.test.raw",
+                        help="Path to wikitext-2 corpus")
+    parser.add_argument("--context", type=int, default=2048,
+                        help="Context length for perplexity test")
+    parser.add_argument("--threads", type=int, default=4,
+                        help="Thread count for perplexity")
+    parser.add_argument("--num-predict", type=int, default=256,
+                        help="Max tokens to generate per prompt")
+    parser.add_argument("--runs", type=int, default=3,
+                        help="Number of runs per config for averaging")
+    parser.add_argument("--skip-perplexity", action="store_true",
+                        help="Skip perplexity measurement")
+    parser.add_argument("--output-dir", default="benchmarks",
+                        help="Output directory")
+    parser.add_argument("--ppl-threshold", type=float, default=0.5,
+                        help="Max acceptable PPL delta")
+    args = parser.parse_args()
+
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    # System info
+    print("Gathering system info...")
+    system_info = get_system_info()
+    print(f"  Platform: {system_info.get('chip', system_info.get('cpu', '?'))}")
+    print(f"  Memory: {system_info.get('memory_gb', '?')}GB")
+    print(f"  Cores: {system_info.get('cpu_cores', '?')}")
+
+    # URL
+    url = args.llama_server if args.backend == "llama-server" else args.ollama_url
+
+    # KV types to test
+    kv_types = ["f16", "turbo4"]
+
+    results = {
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "system": system_info,
+        "model": args.model,
+        "backend": args.backend,
+        "url": url,
+        "num_predict": args.num_predict,
+        "runs_per_config": args.runs,
+        "throughput": {},
+        "memory": {},
+        "perplexity": {},
+        "issues_discovered": [],
+    }
+
+    # ── Throughput + Memory Tests ────────────────────────────────────────
+    for kv_type in kv_types:
+        print(f"\n{'='*60}")
+        print(f"Testing: {kv_type} KV cache")
+        print(f"{'='*60}")
+
+        run_results = []
+
+        # Find server PID for memory monitoring
+        server_pid = find_llama_server_pid()
+        monitor = None
+        if server_pid:
+            print(f"  Monitoring PID {server_pid} for memory")
+            monitor = MemoryMonitor(server_pid)
+            monitor.start()
+
+        for i in range(args.runs):
+            prompt = THROUGHPUT_PROMPTS[i % len(THROUGHPUT_PROMPTS)]
+            print(f"  Run {i+1}/{args.runs}...", end=" ", flush=True)
+
+            if args.backend == "llama-server":
+                result = run_throughput_test(prompt, args.model, url, kv_type,
+                                             num_predict=args.num_predict)
+            else:
+                result = run_ollama_test(prompt, args.model, url,
+                                         num_predict=args.num_predict)
+                result["kv_type"] = "default"  # Ollama doesn't expose KV type
+
+            status = "✓" if result["status"] == "success" else "✗"
+            tps = result.get("tokens_per_sec", 0)
+            print(f"{status} {tps:.1f} tok/s, {result.get('latency_s', 0):.2f}s")
+            run_results.append(result)
+
+        # Stop memory monitor
+        mem_stats = {"avg_mb": 0, "peak_mb": 0, "min_mb": 0, "samples": 0}
+        if monitor:
+            mem_stats = monitor.stop()
+            print(f"  Memory: peak={mem_stats['peak_mb']:.0f}MB, "
+                  f"avg={mem_stats['avg_mb']:.0f}MB")
+
+        results["memory"][kv_type] = mem_stats
+
+        # Aggregate throughput
+        successful = [r for r in run_results if r["status"] == "success"]
+        if successful:
+            avg_tps = sum(r.get("tokens_per_sec", 0) for r in successful) / len(successful)
+            avg_lat = sum(r.get("latency_s", 0) for r in successful) / len(successful)
+            ttfts = [r.get("ttft_s") for r in successful if r.get("ttft_s")]
+            avg_ttft = sum(ttfts) / len(ttfts) if ttfts else None
+        else:
+            avg_tps = avg_lat = avg_ttft = 0
+
+        results["throughput"][kv_type] = {
+            "avg_tok_per_sec": round(avg_tps, 2),
+            "avg_latency": round(avg_lat, 3),
+            "avg_ttft": round(avg_ttft, 3) if avg_ttft else None,
+            "success_rate": f"{len(successful)}/{len(run_results)}",
+            "results": run_results,
+        }
+
+    # ── Perplexity Tests ─────────────────────────────────────────────────
+    if not args.skip_perplexity:
+        llama_bin = None
+        if args.llama_bin:
+            llama_bin = os.path.join(args.llama_bin, "llama-perplexity")
+            if not os.path.exists(llama_bin):
+                llama_bin = os.path.join(args.llama_bin, "bin", "llama-perplexity")
+
+        model_path = args.model_path
+
+        if llama_bin and os.path.exists(llama_bin) and model_path and os.path.exists(model_path) \
+                and os.path.exists(args.corpus):
+            print(f"\n{'='*60}")
+            print("Perplexity Tests")
+            print(f"{'='*60}")
+            print(f"  Model: {model_path}")
+            print(f"  Corpus: {args.corpus}")
+            print(f"  Context: {args.context}")
+
+            ppl_results = {"f16": {}, "turbo4": {}, "threshold": args.ppl_threshold}
+            for kv_type in kv_types:
+                print(f"\n  Running {kv_type} perplexity...")
+                ppl_results[kv_type] = run_perplexity_test(
+                    llama_bin, model_path, args.corpus,
+                    args.context, kv_type, args.threads
+                )
+                ppl_val = ppl_results[kv_type].get("perplexity")
+                if ppl_val:
+                    print(f"    PPL = {ppl_val:.4f}")
+
+            # Calculate delta
+            b_ppl = ppl_results.get("f16", {}).get("perplexity")
+            t_ppl = ppl_results.get("turbo4", {}).get("perplexity")
+            if b_ppl and t_ppl:
+                delta = t_ppl - b_ppl
+                ppl_results["delta"] = round(delta, 4)
+                ppl_results["pass"] = delta <= args.ppl_threshold
+                print(f"\n  Delta: {delta:+.4f} (threshold: ≤{args.ppl_threshold})")
+                print(f"  Result: {'PASS ✓' if ppl_results['pass'] else 'FAIL ✗'}")
+
+            results["perplexity"] = ppl_results
+        else:
+            print("\nSkipping perplexity: need --llama-bin, --model-path, and corpus file")
+            if not llama_bin or not os.path.exists(llama_bin):
+                print(f"  llama-perplexity: {llama_bin or 'not specified'}")
+            if not model_path or not os.path.exists(model_path):
+                print(f"  model path: {model_path or 'not specified (use --model-path)'}")
+            if not os.path.exists(args.corpus):
+                print(f"  corpus: {args.corpus}")
+            results["perplexity"] = {"skipped": True, "reason": "missing binaries/model/corpus"}
+
+    # ── Issue Detection ──────────────────────────────────────────────────
+    tp = results["throughput"]
+    baseline_tps = tp.get("f16", {}).get("avg_tok_per_sec", 0)
+    turbo_tps = tp.get("turbo4", {}).get("avg_tok_per_sec", 0)
+
+    if baseline_tps > 0 and turbo_tps > 0:
+        ratio = turbo_tps / baseline_tps
+        if ratio < 0.75:
+            results["issues_discovered"].append({
+                "title": "turbo4 throughput below 75% of baseline",
+                "description": f"turbo4={turbo_tps:.1f} tok/s vs f16={baseline_tps:.1f} tok/s "
+                               f"({ratio*100:.0f}%). Investigate Metal kernel overhead.",
+            })
+
+    mem = results["memory"]
+    b_peak = mem.get("f16", {}).get("peak_mb", 0)
+    t_peak = mem.get("turbo4", {}).get("peak_mb", 0)
+    if b_peak > 0 and t_peak > 0:
+        savings_pct = (b_peak - t_peak) / b_peak * 100
+        if savings_pct < 50:
+            results["issues_discovered"].append({
+                "title": "turbo4 memory savings below expected 73%",
+                "description": f"Observed {savings_pct:.1f}% savings (expected ~73%). "
+                               f"Check if turbo4 KV is actually active.",
+            })
+
+    ppl = results.get("perplexity", {})
+    if ppl.get("delta") and ppl["delta"] > args.ppl_threshold:
+        results["issues_discovered"].append({
+            "title": f"PPL regression exceeds threshold ({ppl['delta']:.4f} > {args.ppl_threshold})",
+            "description": f"Quality degradation detected. Delta={ppl['delta']:.4f}. "
+                           f"Consider asymmetric K/V (q8_0/turbo4) or per-layer adaptive mode.",
+        })
+
+    # ── Save Results ─────────────────────────────────────────────────────
+    results_path = os.path.join(args.output_dir, "m1_benchmark_results.json")
+    with open(results_path, "w") as f:
+        json.dump(results, f, indent=2)
+    print(f"\nResults saved to {results_path}")
+
+    # ── Generate Report ──────────────────────────────────────────────────
+    report_path = os.path.join(args.output_dir, "m1_benchmark_report.md")
+    report = generate_report(results, report_path)
+    print(f"Report saved to {report_path}")
+
+    # Print summary
+    print(f"\n{'='*60}")
+    print("SUMMARY")
+    print(f"{'='*60}")
+    if baseline_tps and turbo_tps:
+        ratio = turbo_tps / baseline_tps
+        print(f"  Throughput: {turbo_tps:.1f} tok/s ({ratio*100:.0f}% of baseline {baseline_tps:.1f})")
+    if b_peak and t_peak:
+        savings = (b_peak - t_peak) / b_peak * 100
+        print(f"  Memory: {t_peak:.0f}MB peak ({savings:.0f}% savings)")
+    if ppl.get("delta") is not None:
+        print(f"  Quality: PPL delta={ppl['delta']:+.4f} ({'PASS' if ppl['pass'] else 'FAIL'})")
+    if results["issues_discovered"]:
+        print(f"  Issues: {len(results['issues_discovered'])} found")
+    print(f"{'='*60}")
+
+
+if __name__ == "__main__":
+    main()

tests/test_m1_benchmark.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python3
+"""
+Tests for run_m1_benchmark.py (Issue #80)
+
+Validates core benchmark functions without requiring a live server.
+"""
+
+import json
+import os
+import sys
+import tempfile
+import unittest
+from unittest.mock import patch, MagicMock
+
+# Add parent dir to path
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from benchmarks.run_m1_benchmark import (
+    MemoryMonitor,
+    get_system_info,
+    generate_report,
+)
+
+
+class TestMemoryMonitor(unittest.TestCase):
+    def test_init(self):
+        mon = MemoryMonitor(pid=1, interval=0.1)
+        self.assertEqual(mon.pid, 1)
+        self.assertEqual(mon.samples, [])
+
+    def test_get_stats_empty(self):
+        mon = MemoryMonitor(pid=1)
+        stats = mon.get_stats()
+        self.assertEqual(stats["avg_mb"], 0)
+        self.assertEqual(stats["peak_mb"], 0)
+        self.assertEqual(stats["samples"], 0)
+
+    def test_get_stats_with_samples(self):
+        mon = MemoryMonitor(pid=1)
+        mon.samples = [100.0, 150.0, 200.0, 120.0]
+        stats = mon.get_stats()
+        self.assertEqual(stats["peak_mb"], 200.0)
+        self.assertEqual(stats["min_mb"], 100.0)
+        self.assertEqual(stats["avg_mb"], 142.5)
+        self.assertEqual(stats["samples"], 4)
+
+
+class TestSystemInfo(unittest.TestCase):
+    def test_returns_dict(self):
+        info = get_system_info()
+        self.assertIsInstance(info, dict)
+        self.assertIn("platform", info)
+        self.assertIn("python", info)
+
+
+class TestReportGeneration(unittest.TestCase):
+    def test_basic_report(self):
+        results = {
+            "timestamp": "2026-04-15T12:00:00Z",
+            "system": {"chip": "Apple M1", "memory_gb": 16, "cpu_cores": 8},
+            "model": "test-model",
+            "throughput": {
+                "f16": {
+                    "avg_tok_per_sec": 100.0,
+                    "avg_latency": 2.5,
+                    "avg_ttft": 0.3,
+                    "results": [
+                        {"tokens_per_sec": 100, "latency_s": 2.5, "status": "success"},
+                    ],
+                },
+                "turbo4": {
+                    "avg_tok_per_sec": 90.0,
+                    "avg_latency": 2.8,
+                    "avg_ttft": 0.35,
+                    "results": [
+                        {"tokens_per_sec": 90, "latency_s": 2.8, "status": "success"},
+                    ],
+                },
+            },
+            "memory": {
+                "f16": {"peak_mb": 1000, "avg_mb": 900},
+                "turbo4": {"peak_mb": 300, "avg_mb": 250},
+            },
+            "perplexity": {
+                "f16": {"perplexity": 12.5, "tokens": 5000, "elapsed_seconds": 120},
+                "turbo4": {"perplexity": 12.8, "tokens": 5000, "elapsed_seconds": 130},
+                "delta": 0.3,
+                "pass": True,
+                "threshold": 0.5,
+            },
+            "issues_discovered": [],
+        }
+
+        with tempfile.NamedTemporaryFile(mode="w", suffix=".md", delete=False) as f:
+            report_path = f.name
+
+        try:
+            report = generate_report(results, report_path)
+            self.assertIn("TurboQuant M1 Benchmark Report", report)
+            self.assertIn("f16", report)
+            self.assertIn("turbo4", report)
+            self.assertIn("PASS", report)
+
+            # Verify file was written
+            with open(report_path) as f:
+                written = f.read()
+            self.assertEqual(written, report)
+        finally:
+            os.unlink(report_path)
+
+    def test_report_with_issues(self):
+        results = {
+            "timestamp": "2026-04-15T12:00:00Z",
+            "system": {"chip": "M1", "memory_gb": 16, "cpu_cores": 8},
+            "model": "test",
+            "throughput": {"f16": {"results": []}, "turbo4": {"results": []}},
+            "memory": {"f16": {}, "turbo4": {}},
+            "perplexity": {},
+            "issues_discovered": [
+                {"title": "Test issue", "description": "Something went wrong"}
+            ],
+        }
+
+        with tempfile.NamedTemporaryFile(mode="w", suffix=".md", delete=False) as f:
+            report_path = f.name
+
+        try:
+            report = generate_report(results, report_path)
+            self.assertIn("Issues Discovered", report)
+            self.assertIn("Test issue", report)
+        finally:
+            os.unlink(report_path)
+
+
+if __name__ == "__main__":
+    unittest.main()