updated benchmarks including thunderbolt and configuratuion guides

benchmarks/benchmark_results/Qwen_Qwen3-14B-AWQ_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 485.412814248004,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.4120204373051785,
+    "tokens_per_second": 302.43330149293365
+}

benchmarks/benchmark_results/btbtyler09_Qwen3-Coder-30B-A3B-Instruct-gptq-4bit_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 503.28860085096676,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.397386310084984,
+    "tokens_per_second": 291.6914862601304
+}

benchmarks/benchmark_results/btbtyler09_Qwen3-Coder-30B-A3B-Instruct-gptq-8bit_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 571.4193902639672,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.35000562355367393,
+    "tokens_per_second": 256.91287782898553
+}

benchmarks/benchmark_results/dazipe_Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 824.4905019259895,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.2425740497104635,
+    "tokens_per_second": 178.05541683872298
+}

benchmarks/benchmark_results/google_gemma-3-12b-it_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 572.292031740013,
+    "num_requests": 200,
+    "total_num_tokens": 148857,
+    "requests_per_second": 0.3494719285046033,
+    "tokens_per_second": 260.10671430704866
+}

benchmarks/benchmark_results/meta-llama_Meta-Llama-3.1-8B-Instruct_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 284.23000320699066,
+    "num_requests": 200,
+    "total_num_tokens": 145877,
+    "requests_per_second": 0.7036554823325597,
+    "tokens_per_second": 513.235753981134
+}

benchmarks/benchmark_results/mratsim_MiniMax-M2.5-BF16-INT4-AWQ_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 1474.255295659008,
+    "num_requests": 200,
+    "total_num_tokens": 146523,
+    "requests_per_second": 0.13566171380825723,
+    "tokens_per_second": 99.38780646163637
+}

benchmarks/benchmark_results/openai_gpt-oss-120b_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 1338.8605944840237,
+    "num_requests": 200,
+    "total_num_tokens": 147036,
+    "requests_per_second": 0.1493807501871223,
+    "tokens_per_second": 109.82173992256857
+}

benchmarks/benchmark_results/openai_gpt-oss-20b_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 282.0738571010297,
+    "num_requests": 200,
+    "total_num_tokens": 147036,
+    "requests_per_second": 0.7090341588386423,
+    "tokens_per_second": 521.2677328949931
+}

benchmarks/benchmark_results/zai-org_GLM-4.7-Flash_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 455.7690629530116,
+    "num_requests": 200,
+    "total_num_tokens": 146278,
+    "requests_per_second": 0.4388187269758136,
+    "tokens_per_second": 320.9476287228403
+}

benchmarks/benchmark_results_rocm/Qwen_Qwen3-14B-AWQ_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 471.133652363962,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.4245079904534079,
+    "tokens_per_second": 311.59947769256274
+}

benchmarks/benchmark_results_rocm/btbtyler09_Qwen3-Coder-30B-A3B-Instruct-gptq-4bit_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 490.5911466999678,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.40767144157681784,
+    "tokens_per_second": 299.24102990342374
+}

benchmarks/benchmark_results_rocm/btbtyler09_Qwen3-Coder-30B-A3B-Instruct-gptq-8bit_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 548.4156070559984,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.3646869225214776,
+    "tokens_per_second": 267.6893183038276
+}

benchmarks/benchmark_results_rocm/dazipe_Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 839.8958681730437,
+    "num_requests": 200,
+    "total_num_tokens": 146805,
+    "requests_per_second": 0.23812475757863116,
+    "tokens_per_second": 174.78952518165474
+}

benchmarks/benchmark_results_rocm/google_gemma-3-12b-it_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 434.26602390100015,
+    "num_requests": 200,
+    "total_num_tokens": 148857,
+    "requests_per_second": 0.46054719686197254,
+    "tokens_per_second": 342.7783704164132
+}

benchmarks/benchmark_results_rocm/meta-llama_Meta-Llama-3.1-8B-Instruct_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 264.80451649799943,
+    "num_requests": 200,
+    "total_num_tokens": 145877,
+    "requests_per_second": 0.7552741269105618,
+    "tokens_per_second": 550.8856190566602
+}

benchmarks/benchmark_results_rocm/mratsim_MiniMax-M2.5-BF16-INT4-AWQ_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 1493.5249834020506,
+    "num_requests": 200,
+    "total_num_tokens": 146523,
+    "requests_per_second": 0.13391138562973798,
+    "tokens_per_second": 98.10548978313048
+}

benchmarks/benchmark_results_rocm/openai_gpt-oss-120b_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 1320.7432732739835,
+    "num_requests": 200,
+    "total_num_tokens": 147036,
+    "requests_per_second": 0.1514298834959962,
+    "tokens_per_second": 111.32822174858649
+}

benchmarks/benchmark_results_rocm/openai_gpt-oss-20b_cluster_tp2_eth_usb_throughput.json

@@ -0,0 +1,7 @@
+{
+    "elapsed_time": 282.87595599895576,
+    "num_requests": 200,
+    "total_num_tokens": 147036,
+    "requests_per_second": 0.707023682142636,
+    "tokens_per_second": 519.7896706376232
+}

benchmarks/run_vllm_bench.py

@@ -179,31 +179,59 @@ def run_throughput(model, tp_size, backend_name="Default", output_dir=RESULTS_DI
 
 
 def print_summary(tps):
-    print(f"\n{'MODEL':<40} | {'TP':<2} | {'Triton':<8} | {'ROCm':<8}")
-    print("-" * 75)
+    print(f"\n{'MODEL':<40} | {'TP':<2} | {'Tag':<15} | {'Triton':<8} | {'ROCm':<8}")
+    print("-" * 92)
     
     for m in MODELS_TO_RUN:
         msafe = m.replace("/", "_")
+        name_cell = m.split('/')[-1]
+        
         for tp in tps:
             if tp not in MODEL_TABLE[m]["valid_tp"]: continue
             
-            # Default
-            try: 
-                p1 = RESULTS_DIR / f"{msafe}_tp{tp}_throughput.json"
-                d1 = json.loads(p1.read_text())
-                val1 = f"{d1.get('tokens_per_second', 0):.1f}"
-            except: val1 = "N/A"
+            prefix = f"{msafe}_tp{tp}"
             
-            # ROCm
-            try:
-                p2 = Path("benchmark_results_rocm") / f"{msafe}_tp{tp}_throughput.json"
-                d2 = json.loads(p2.read_text())
-                val2 = f"{d2.get('tokens_per_second', 0):.1f}"
-            except: val2 = "N/A"
+            tags = set()
+            for p in RESULTS_DIR.glob(f"{prefix}*_throughput.json"):
+                name_part = p.name[len(prefix):-len("_throughput.json")]
+                tag = name_part.lstrip("_")
+                tags.add(tag)
+                
+            for p in Path("benchmark_results_rocm").glob(f"{prefix}*_throughput.json"):
+                name_part = p.name[len(prefix):-len("_throughput.json")]
+                tag = name_part.lstrip("_")
+                tags.add(tag)
+                
+            if not tags:
+                tags.add("") # Default empty tag if no files found
+                
+            for tag in sorted(list(tags)):
+                tag_suffix = f"_{tag}" if tag else ""
+                
+                # Default
+                try: 
+                    p1 = RESULTS_DIR / f"{prefix}{tag_suffix}_throughput.json"
+                    if p1.exists():
+                        d1 = json.loads(p1.read_text())
+                        val1 = f"{d1.get('tokens_per_second', 0):.1f}"
+                    else:
+                        val1 = "N/A"
+                except: val1 = "N/A"
+                
+                # ROCm
+                try:
+                    p2 = Path("benchmark_results_rocm") / f"{prefix}{tag_suffix}_throughput.json"
+                    if p2.exists():
+                        d2 = json.loads(p2.read_text())
+                        val2 = f"{d2.get('tokens_per_second', 0):.1f}"
+                    else:
+                        val2 = "N/A"
+                except: val2 = "N/A"
 
-            name_cell = m.split('/')[-1]
-            print(f"{name_cell:<40} | {tp:<2} | {val1:<8} | {val2:<8}")
-    print("-" * 75)
+                display_tag = tag if tag else "(Default)"
+                print(f"{name_cell:<40} | {tp:<2} | {display_tag:<15} | {val1:<8} | {val2:<8}")
+                
+    print("-" * 92)
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()

benchmarks/vllm_cluster_bench.py

@@ -213,7 +213,7 @@ def run_bench_set(model, backend_name, output_dir, extra_env=None, overrides=Non
 
     log(f"START {model} [TP={CLUSTER_TP} | {backend_name}]...")
     
-    nuke_vllm_cache(HEAD_IP)
+    nuke_vllm_cache()
 
     cmd = ["vllm", "bench", "throughput"] + get_model_args(model, overrides)
     cmd.extend([
@@ -279,29 +279,60 @@ def run_cluster_throughput(model, overrides=None):
 def print_summary():
     eth_suffix = "_eth" if FORCE_ETH else ""
     title_suffix = " (Ethernet ONLY)" if FORCE_ETH else ""
-    print(f"\n{f'MODEL (TP=2){title_suffix}':<50} | {'Triton':<8} | {'ROCm':<8}")
-    print("-" * 75)
+    print(f"\n{f'MODEL (TP={CLUSTER_TP}){title_suffix}':<50} | {'Tag':<15} | {'Triton':<8} | {'ROCm':<8}")
+    print("-" * 92)
     
     for m in MODELS_TO_RUN:
         msafe = m.replace("/", "_")
-        
-        # Default
-        try: 
-            p1 = RESULTS_DIR / f"{msafe}_cluster_tp{CLUSTER_TP}{eth_suffix}_throughput.json"
-            d1 = json.loads(p1.read_text())
-            val1 = f"{d1.get('tokens_per_second', 0):.1f}"
-        except: val1 = "N/A"
-        
-        # ROCm
-        try:
-            p2 = Path("benchmark_results_rocm") / f"{msafe}_cluster_tp{CLUSTER_TP}{eth_suffix}_throughput.json"
-            d2 = json.loads(p2.read_text())
-            val2 = f"{d2.get('tokens_per_second', 0):.1f}"
-        except: val2 = "N/A"
-
         name_cell = m.split('/')[-1]
-        print(f"{name_cell:<50} | {val1:<8} | {val2:<8}")
-    print("-" * 75)
+        
+        # Find all tags used for this model by looking at the files in RESULTS_DIR
+        prefix = f"{msafe}_cluster_tp{CLUSTER_TP}{eth_suffix}"
+        
+        # Gather all unique tags from both directories
+        tags = set()
+        for p in RESULTS_DIR.glob(f"{prefix}*_throughput.json"):
+            # Extract tag: {prefix}_{tag}_throughput.json or {prefix}_throughput.json
+            name_part = p.name[len(prefix):-len("_throughput.json")]
+            tag = name_part.lstrip("_")
+            tags.add(tag)
+            
+        for p in Path("benchmark_results_rocm").glob(f"{prefix}*_throughput.json"):
+            name_part = p.name[len(prefix):-len("_throughput.json")]
+            tag = name_part.lstrip("_")
+            tags.add(tag)
+            
+        if not tags:
+            tags.add("") # Default empty tag if no files found
+            
+        # Sort so empty tag (Default) comes first
+        for tag in sorted(list(tags)):
+            tag_suffix = f"_{tag}" if tag else ""
+            
+            # Default (Triton)
+            try: 
+                p1 = RESULTS_DIR / f"{prefix}{tag_suffix}_throughput.json"
+                if p1.exists():
+                    d1 = json.loads(p1.read_text())
+                    val1 = f"{d1.get('tokens_per_second', 0):.1f}"
+                else:
+                    val1 = "N/A"
+            except: val1 = "N/A"
+            
+            # ROCm
+            try:
+                p2 = Path("benchmark_results_rocm") / f"{prefix}{tag_suffix}_throughput.json"
+                if p2.exists():
+                    d2 = json.loads(p2.read_text())
+                    val2 = f"{d2.get('tokens_per_second', 0):.1f}"
+                else:
+                    val2 = "N/A"
+            except: val2 = "N/A"
+
+            display_tag = tag if tag else "(Default)"
+            print(f"{name_cell:<50} | {display_tag:<15} | {val1:<8} | {val2:<8}")
+            
+    print("-" * 92)
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="VLLM Cluster Benchmark")