feat: add interactive granular benchmark details to UI and update MTP summary data format

benchmark/mtp-bench.py

@@ -87,21 +87,33 @@ def run(args):
         # OpenAI-compatible endpoint: timings are in usage or top-level
         usage = r.get("usage", {}) or {}
         t = r.get("timings", {}) or {}
+        prompt_n = usage.get("prompt_tokens") or t.get("prompt_n")
+        prompt_ms = t.get("prompt_ms")
+        prompt_per_second = t.get("prompt_per_second")
         predicted_n = usage.get("completion_tokens") or t.get("predicted_n")
         predicted_per_second = t.get("predicted_per_second") or (predicted_n / wall if wall > 0 else 0)
         rec = {"name": p["name"], "wall_s": round(wall,3),
+               "prompt_n": prompt_n,
+               "prompt_ms": round(prompt_ms, 2) if prompt_ms is not None else None,
+               "prompt_per_second": round(prompt_per_second, 2) if prompt_per_second is not None else None,
                "predicted_n": predicted_n, "predicted_per_second": round(predicted_per_second, 2),
                "draft_n": t.get("draft_n",0), "draft_n_accepted": t.get("draft_n_accepted",0)}
         rec["accept_rate"] = round(rec["draft_n_accepted"]/rec["draft_n"],4) if rec["draft_n"] else None
         out["results"].append(rec)
         ar = f"{rec['accept_rate']:.3f}" if rec["accept_rate"] is not None else "n/a"
-        print(f"  {rec['name']:<18} pred={rec['predicted_n']:>4} draft={rec['draft_n']:>4} acc={rec['draft_n_accepted']:>4} rate={ar} tok/s={rec['predicted_per_second']:.1f}")
+        pps = f" pt/s={rec['prompt_per_second']:.1f}" if rec.get("prompt_per_second") else ""
+        print(f"  {rec['name']:<18} pred={rec['predicted_n']:>4} draft={rec['draft_n']:>4} acc={rec['draft_n_accepted']:>4} rate={ar} tok/s={rec['predicted_per_second']:.1f}{pps}")
     td  = sum(x["draft_n"] or 0 for x in out["results"])
     ta  = sum(x["draft_n_accepted"] or 0 for x in out["results"])
     tp  = sum(x["predicted_n"] or 0 for x in out["results"])
+    t_pn = sum(x["prompt_n"] or 0 for x in out["results"])
     tw  = sum(x["wall_s"] for x in out["results"])
+    pps_list = [x["prompt_per_second"] for x in out["results"] if x.get("prompt_per_second") is not None]
+    avg_pps = sum(pps_list)/len(pps_list) if pps_list else None
+    
     out["aggregate"] = {"n_requests": len(out["results"]), "total_predicted": tp, "total_draft": td, "total_draft_accepted": ta,
-                        "aggregate_accept_rate": round(ta/td,4) if td else None, "wall_s_total": round(tw,2)}
+                        "aggregate_accept_rate": round(ta/td,4) if td else None, "wall_s_total": round(tw,2),
+                        "total_prompt_tokens": t_pn, "avg_prompt_per_second": round(avg_pps, 2) if avg_pps is not None else None}
     print("\nAggregate:", json.dumps(out["aggregate"], indent=2))
     if args.out:
         json.dump(out, open(args.out,"w"), indent=2); print("Wrote", args.out)

benchmark/run_mtp_bench.py

@@ -29,8 +29,8 @@ from urllib.error import URLError
 # ── Toolbox definitions ──────────────────────────────────────────────────────
 
 TOOLBOXES = {
-    "rocm-7.2.3-mtp": {
-        "image": "docker.io/kyuz0/amd-strix-halo-toolboxes:rocm-7.2.3-mtp",
+    "rocm-7.2.3": {
+        "image": "docker.io/kyuz0/amd-strix-halo-toolboxes:rocm-7.2.3",
         "engine_args": [
             "--device", "/dev/dri",
             "--device", "/dev/kfd",
@@ -39,8 +39,8 @@ TOOLBOXES = {
             "--security-opt", "seccomp=unconfined",
         ],
     },
-    "vulkan-radv-mtp": {
-        "image": "docker.io/kyuz0/amd-strix-halo-toolboxes:vulkan-radv-mtp",
+    "vulkan-radv": {
+        "image": "docker.io/kyuz0/amd-strix-halo-toolboxes:vulkan-radv",
         "engine_args": [
             "--device", "/dev/dri",
             "--group-add", "video",
@@ -415,9 +415,9 @@ def print_summary(results_dir: Path):
             baselines[key] = r["_avg_toks"]
 
     # Print table
-    print("\n" + "=" * 100)
-    print(f"{'Model':<30} {'Toolbox':<20} {'Mode':<10} {'Avg tok/s':>10} {'Accept%':>9} {'Wall(s)':>8} {'Speedup':>8}")
-    print("-" * 100)
+    print("\n" + "=" * 115)
+    print(f"{'Model':<30} {'Toolbox':<20} {'Mode':<10} {'Prefill pt/s':>13} {'Avg tok/s':>10} {'Accept%':>9} {'Wall(s)':>8} {'Speedup':>8}")
+    print("-" * 115)
 
     for r in results:
         agg = r.get("aggregate", {})
@@ -426,6 +426,9 @@ def print_summary(results_dir: Path):
         accept_str = f"{accept * 100:.1f}%" if accept is not None else "—"
         avg_toks = r["_avg_toks"]
 
+        avg_prompt = agg.get("avg_prompt_per_second")
+        prefill_str = f"{avg_prompt:.1f}" if avg_prompt is not None else "—"
+        
         # Speedup relative to baseline
         baseline_key = (r["model"], r["toolbox"])
         baseline_toks = baselines.get(baseline_key)
@@ -434,9 +437,9 @@ def print_summary(results_dir: Path):
         else:
             speedup = "—"
 
-        print(f"{r['model']:<30} {r['toolbox']:<20} {r['mode']:<10} {avg_toks:>10.1f} {accept_str:>9} {wall:>8.1f} {speedup:>8}")
+        print(f"{r['model']:<30} {r['toolbox']:<20} {r['mode']:<10} {prefill_str:>13} {avg_toks:>10.1f} {accept_str:>9} {wall:>8.1f} {speedup:>8}")
 
-    print("=" * 100)
+    print("=" * 115)
 
     # Write summary.json
     summary_data = []
@@ -446,9 +449,11 @@ def print_summary(results_dir: Path):
             "model": r["model"],
             "toolbox": r["toolbox"],
             "mode": r["mode"],
+            "avg_prompt_tok_s": agg.get("avg_prompt_per_second"),
             "avg_tok_s": round(r["_avg_toks"], 1),
             "accept_rate": agg.get("aggregate_accept_rate"),
             "wall_s_total": agg.get("wall_s_total"),
+            "results": r.get("results", [])
         })
 
     summary_path = results_dir / "summary.json"