updated benchmarks, fix start-vllm

docs/index.html

@@ -469,6 +469,10 @@
                     style="font-size: 0.9rem; font-weight: 500; display: flex; align-items: center; gap: 4px; cursor: pointer;">
                     <input type="checkbox" id="toggleTP2" checked> TP2
                 </label>
+                <label
+                    style="font-size: 0.9rem; font-weight: 500; display: flex; align-items: center; gap: 4px; cursor: pointer;">
+                    <input type="checkbox" id="toggleTP2Eth" checked> TP2 (Eth)
+                </label>
             </div>
 
             <!-- Attention Group -->
@@ -544,6 +548,7 @@
             activeTab: "Throughput",
             showTP1: true,
             showTP2: true,
+            showTP2Eth: true,
             showTriton: true,
             showRocm: false
         };
@@ -615,6 +620,7 @@
             // Toggles
             $('toggleTP1').addEventListener('change', e => { state.showTP1 = e.target.checked; render(); });
             $('toggleTP2').addEventListener('change', e => { state.showTP2 = e.target.checked; render(); });
+            $('toggleTP2Eth').addEventListener('change', e => { state.showTP2Eth = e.target.checked; render(); });
             $('toggleTriton').addEventListener('change', e => { state.showTriton = e.target.checked; render(); });
             $('toggleRocm').addEventListener('change', e => { state.showRocm = e.target.checked; render(); });
         }
@@ -636,13 +642,17 @@
                         params: run.params_b || run.name_params_b,
                         results: {
                             1: { triton: null, rocm: null },
-                            2: { triton: null, rocm: null }
+                            2: { triton: null, rocm: null },
+                            "2_eth": { triton: null, rocm: null }
                         }
                     };
                 }
 
                 const m = testGroups[testName].models[modelName];
-                const tp = run.tp || 1;
+                let tp = run.tp || 1;
+                if (tp === 2 && run.network === "Ethernet") {
+                    tp = "2_eth";
+                }
 
                 if (!m.results[tp]) m.results[tp] = { triton: null, rocm: null };
 
@@ -749,8 +759,12 @@
                 if (state.showRocm) cols.push({ id: "tp1_rocm", label: "TP1 ROCm" });
             }
             if (state.showTP2) {
-                if (state.showTriton) cols.push({ id: "tp2_triton", label: "TP2 Triton" });
-                if (state.showRocm) cols.push({ id: "tp2_rocm", label: "TP2 ROCm" });
+                if (state.showTriton) cols.push({ id: "tp2_triton", label: "TP2 RoCE Triton" });
+                if (state.showRocm) cols.push({ id: "tp2_rocm", label: "TP2 RoCE ROCm" });
+            }
+            if (state.showTP2Eth) {
+                if (state.showTriton) cols.push({ id: "tp2_eth_triton", label: "TP2 Eth Triton" });
+                if (state.showRocm) cols.push({ id: "tp2_eth_rocm", label: "TP2 Eth ROCm" });
             }
 
             // Thead
@@ -790,11 +804,7 @@
 
                 // Data Cells
                 cols.forEach(c => {
-                    let val = null;
-                    if (c.id === "tp1_triton") val = m.results[1]?.triton;
-                    if (c.id === "tp1_rocm") val = m.results[1]?.rocm;
-                    if (c.id === "tp2_triton") val = m.results[2]?.triton;
-                    if (c.id === "tp2_rocm") val = m.results[2]?.rocm;
+                    let val = getVal(m, c.id);
 
                     const bg = c.id.startsWith("tp2") ? 'style="background:#fbfdff;"' : "";
                     rowHtml += `<td class="col-data" ${bg}>${formatVal(val, unit)}</td>`;
@@ -823,6 +833,8 @@
             if (colId === "tp1_rocm") return m.results[1]?.rocm;
             if (colId === "tp2_triton") return m.results[2]?.triton;
             if (colId === "tp2_rocm") return m.results[2]?.rocm;
+            if (colId === "tp2_eth_triton") return m.results["2_eth"]?.triton;
+            if (colId === "tp2_eth_rocm") return m.results["2_eth"]?.rocm;
             return null;
         }
 

docs/parse_results.py

@@ -66,6 +66,11 @@ def parse_logs():
             if not tp_match: continue
             tp = int(tp_match.group(1))
             
+            # Network
+            network = "RoCE"
+            if "_eth" in rest:
+                network = "Ethernet"
+            
             # Model Name
             if "_" in model_part:
                 model_display = model_part.replace("_", "/", 1)
@@ -87,6 +92,7 @@ def parse_logs():
                 "params_b": params_b,
                 "name_params_b": params_b,
                 "backend": backend_name, # "Triton" or "ROCm"
+                "network": network,
                 "error": False
             }
 

docs/results.json

@@ -1,131 +1,5 @@
 {
   "runs": [
-    {
-      "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-      "model_clean": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "BF16",
-      "params_b": 8.0,
-      "name_params_b": 8.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 383.3285005130725
-    },
-    {
-      "model": "google/gemma-3-12b-it",
-      "model_clean": "google/gemma-3-12b-it",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "BF16",
-      "params_b": 12.0,
-      "name_params_b": 12.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 169.75141153501525
-    },
-    {
-      "model": "Qwen/Qwen3-14B-AWQ",
-      "model_clean": "Qwen/Qwen3-14B-AWQ",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "AWQ",
-      "params_b": 14.0,
-      "name_params_b": 14.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 168.9724830093454
-    },
-    {
-      "model": "openai/gpt-oss-20b",
-      "model_clean": "openai/gpt-oss-20b",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "BF16",
-      "params_b": 20.0,
-      "name_params_b": 20.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 430.8537270233502
-    },
-    {
-      "model": "openai/gpt-oss-120b",
-      "model_clean": "openai/gpt-oss-120b",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "BF16",
-      "params_b": 120.0,
-      "name_params_b": 120.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 77.93077982357597
-    },
-    {
-      "model": "zai-org/GLM-4.7-Flash",
-      "model_clean": "zai-org/GLM-4.7-Flash",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "BF16",
-      "params_b": null,
-      "name_params_b": null,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 246.02994547299596
-    },
-    {
-      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
-      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "GPTQ",
-      "params_b": 30.0,
-      "name_params_b": 30.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 227.90361622402403
-    },
-    {
-      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
-      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "GPTQ",
-      "params_b": 30.0,
-      "name_params_b": 30.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 200.1426829468909
-    },
-    {
-      "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
-      "model_clean": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "GPTQ",
-      "params_b": 80.0,
-      "name_params_b": 80.0,
-      "backend": "Triton",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 132.2599488751683
-    },
     {
       "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
       "model_clean": "meta-llama/Meta-Llama-3.1-8B-Instruct",
@@ -135,10 +9,11 @@
       "params_b": 8.0,
       "name_params_b": 8.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 596.5891209659404
+      "tps_mean": 613.9321196754427
     },
     {
       "model": "google/gemma-3-12b-it",
@@ -149,10 +24,11 @@
       "params_b": 12.0,
       "name_params_b": 12.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 294.9472545848014
+      "tps_mean": 291.5155379231269
     },
     {
       "model": "Qwen/Qwen3-14B-AWQ",
@@ -163,10 +39,11 @@
       "params_b": 14.0,
       "name_params_b": 14.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 348.0799308087054
+      "tps_mean": 280.05330212131406
     },
     {
       "model": "openai/gpt-oss-20b",
@@ -177,10 +54,11 @@
       "params_b": 20.0,
       "name_params_b": 20.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 593.7843034224891
+      "tps_mean": 602.6345456319963
     },
     {
       "model": "openai/gpt-oss-120b",
@@ -191,10 +69,11 @@
       "params_b": 120.0,
       "name_params_b": 120.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 112.4781801162827
+      "tps_mean": 85.2809331488931
     },
     {
       "model": "zai-org/GLM-4.7-Flash",
@@ -205,10 +84,11 @@
       "params_b": null,
       "name_params_b": null,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 346.0061963818796
+      "tps_mean": 300.773560320048
     },
     {
       "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
@@ -219,10 +99,11 @@
       "params_b": 30.0,
       "name_params_b": 30.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 320.69249844623016
+      "tps_mean": 321.88057686801585
     },
     {
       "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
@@ -233,10 +114,11 @@
       "params_b": 30.0,
       "name_params_b": 30.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 274.7000183491961
+      "tps_mean": 274.46004720922855
     },
     {
       "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
@@ -247,10 +129,176 @@
       "params_b": 80.0,
       "name_params_b": 80.0,
       "backend": "Triton",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 167.00232766189475
+      "tps_mean": 182.16229690959702
+    },
+    {
+      "model": "mratsim/MiniMax-M2.5-BF16-INT4-AWQ",
+      "model_clean": "mratsim/MiniMax-M2.5-BF16-INT4-AWQ",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": null,
+      "name_params_b": null,
+      "backend": "Triton",
+      "network": "Ethernet",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 107.62460878889469
+    },
+    {
+      "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+      "model_clean": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 8.0,
+      "name_params_b": 8.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 590.0492703672895
+    },
+    {
+      "model": "google/gemma-3-12b-it",
+      "model_clean": "google/gemma-3-12b-it",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 12.0,
+      "name_params_b": 12.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 285.8275921888489
+    },
+    {
+      "model": "Qwen/Qwen3-14B-AWQ",
+      "model_clean": "Qwen/Qwen3-14B-AWQ",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "AWQ",
+      "params_b": 14.0,
+      "name_params_b": 14.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 346.2003835540928
+    },
+    {
+      "model": "openai/gpt-oss-20b",
+      "model_clean": "openai/gpt-oss-20b",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 20.0,
+      "name_params_b": 20.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 607.216674264294
+    },
+    {
+      "model": "openai/gpt-oss-120b",
+      "model_clean": "openai/gpt-oss-120b",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 120.0,
+      "name_params_b": 120.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 122.62029501860121
+    },
+    {
+      "model": "zai-org/GLM-4.7-Flash",
+      "model_clean": "zai-org/GLM-4.7-Flash",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": null,
+      "name_params_b": null,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 366.77052981888835
+    },
+    {
+      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
+      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "GPTQ",
+      "params_b": 30.0,
+      "name_params_b": 30.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 320.0197833991106
+    },
+    {
+      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
+      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "GPTQ",
+      "params_b": 30.0,
+      "name_params_b": 30.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 279.7240042842149
+    },
+    {
+      "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
+      "model_clean": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "GPTQ",
+      "params_b": 80.0,
+      "name_params_b": 80.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 196.2262690032198
+    },
+    {
+      "model": "mratsim/MiniMax-M2.5-BF16-INT4-AWQ",
+      "model_clean": "mratsim/MiniMax-M2.5-BF16-INT4-AWQ",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": null,
+      "name_params_b": null,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 98.85048345093716
     },
     {
       "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
@@ -260,25 +308,12 @@
       "quant": "BF16",
       "params_b": 8.0,
       "name_params_b": 8.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 452.7925628873698
-    },
-    {
-      "model": "google/gemma-3-12b-it",
-      "model_clean": "google/gemma-3-12b-it",
-      "env": "TP1",
-      "gpu_config": "single",
-      "quant": "BF16",
-      "params_b": 12.0,
-      "name_params_b": 12.0,
-      "backend": "ROCm",
-      "error": false,
-      "test": "Throughput",
-      "tp": 1,
-      "tps_mean": 292.4074663029466
+      "tps_mean": 369.23212230245684
     },
     {
       "model": "Qwen/Qwen3-14B-AWQ",
@@ -288,11 +323,12 @@
       "quant": "AWQ",
       "params_b": 14.0,
       "name_params_b": 14.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 190.86242019407229
+      "tps_mean": 159.8857312165796
     },
     {
       "model": "openai/gpt-oss-20b",
@@ -302,11 +338,12 @@
       "quant": "BF16",
       "params_b": 20.0,
       "name_params_b": 20.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 440.75738846836555
+      "tps_mean": 411.03557062490586
     },
     {
       "model": "openai/gpt-oss-120b",
@@ -316,11 +353,27 @@
       "quant": "BF16",
       "params_b": 120.0,
       "name_params_b": 120.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 76.44313314138553
+      "tps_mean": 75.0407548829671
+    },
+    {
+      "model": "zai-org/GLM-4.7-Flash",
+      "model_clean": "zai-org/GLM-4.7-Flash",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "BF16",
+      "params_b": null,
+      "name_params_b": null,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 239.57478116575834
     },
     {
       "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
@@ -330,11 +383,12 @@
       "quant": "GPTQ",
       "params_b": 30.0,
       "name_params_b": 30.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 229.9835374194385
+      "tps_mean": 213.74630950782364
     },
     {
       "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
@@ -344,11 +398,12 @@
       "quant": "GPTQ",
       "params_b": 30.0,
       "name_params_b": 30.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 203.38751203489863
+      "tps_mean": 186.03115379827653
     },
     {
       "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
@@ -358,11 +413,27 @@
       "quant": "GPTQ",
       "params_b": 80.0,
       "name_params_b": 80.0,
-      "backend": "ROCm",
+      "backend": "Triton",
+      "network": "RoCE",
       "error": false,
       "test": "Throughput",
       "tp": 1,
-      "tps_mean": 135.3839809398758
+      "tps_mean": 125.65027253668944
+    },
+    {
+      "model": "google/gemma-3-12b-it",
+      "model_clean": "google/gemma-3-12b-it",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "BF16",
+      "params_b": 12.0,
+      "name_params_b": 12.0,
+      "backend": "Triton",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 159.95620436815713
     },
     {
       "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
@@ -373,10 +444,11 @@
       "params_b": 8.0,
       "name_params_b": 8.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 649.02791593759
+      "tps_mean": 682.4360360233941
     },
     {
       "model": "google/gemma-3-12b-it",
@@ -387,10 +459,11 @@
       "params_b": 12.0,
       "name_params_b": 12.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 403.09652364564084
+      "tps_mean": 398.09474781142933
     },
     {
       "model": "Qwen/Qwen3-14B-AWQ",
@@ -401,10 +474,11 @@
       "params_b": 14.0,
       "name_params_b": 14.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 371.4058491591393
+      "tps_mean": 295.31575874126105
     },
     {
       "model": "openai/gpt-oss-20b",
@@ -415,10 +489,11 @@
       "params_b": 20.0,
       "name_params_b": 20.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 597.5787987620997
+      "tps_mean": 490.93757442090305
     },
     {
       "model": "openai/gpt-oss-120b",
@@ -429,10 +504,11 @@
       "params_b": 120.0,
       "name_params_b": 120.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 111.8113988472388
+      "tps_mean": 86.0910643999307
     },
     {
       "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
@@ -443,10 +519,11 @@
       "params_b": 30.0,
       "name_params_b": 30.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 315.906032423287
+      "tps_mean": 321.6166453306162
     },
     {
       "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
@@ -457,10 +534,11 @@
       "params_b": 30.0,
       "name_params_b": 30.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 292.0384117325289
+      "tps_mean": 283.6309502128471
     },
     {
       "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
@@ -471,10 +549,266 @@
       "params_b": 80.0,
       "name_params_b": 80.0,
       "backend": "ROCm",
+      "network": "Ethernet",
       "error": false,
       "test": "Throughput",
       "tp": 2,
-      "tps_mean": 165.5348293928834
+      "tps_mean": 182.9186467257061
+    },
+    {
+      "model": "mratsim/MiniMax-M2.5-BF16-INT4-AWQ",
+      "model_clean": "mratsim/MiniMax-M2.5-BF16-INT4-AWQ",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": null,
+      "name_params_b": null,
+      "backend": "ROCm",
+      "network": "Ethernet",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 98.67941666807306
+    },
+    {
+      "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+      "model_clean": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 8.0,
+      "name_params_b": 8.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 650.1471716939323
+    },
+    {
+      "model": "google/gemma-3-12b-it",
+      "model_clean": "google/gemma-3-12b-it",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 12.0,
+      "name_params_b": 12.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 397.9710386242193
+    },
+    {
+      "model": "Qwen/Qwen3-14B-AWQ",
+      "model_clean": "Qwen/Qwen3-14B-AWQ",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "AWQ",
+      "params_b": 14.0,
+      "name_params_b": 14.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 367.5704596781314
+    },
+    {
+      "model": "openai/gpt-oss-20b",
+      "model_clean": "openai/gpt-oss-20b",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 20.0,
+      "name_params_b": 20.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 601.2567608739705
+    },
+    {
+      "model": "openai/gpt-oss-120b",
+      "model_clean": "openai/gpt-oss-120b",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "BF16",
+      "params_b": 120.0,
+      "name_params_b": 120.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 118.34229353876268
+    },
+    {
+      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
+      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "GPTQ",
+      "params_b": 30.0,
+      "name_params_b": 30.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 333.147212194374
+    },
+    {
+      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
+      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "GPTQ",
+      "params_b": 30.0,
+      "name_params_b": 30.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 295.0301359026215
+    },
+    {
+      "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
+      "model_clean": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
+      "env": "TP2",
+      "gpu_config": "dual",
+      "quant": "GPTQ",
+      "params_b": 80.0,
+      "name_params_b": 80.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 2,
+      "tps_mean": 193.87438091607942
+    },
+    {
+      "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+      "model_clean": "meta-llama/Meta-Llama-3.1-8B-Instruct",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "BF16",
+      "params_b": 8.0,
+      "name_params_b": 8.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 433.5736973626181
+    },
+    {
+      "model": "Qwen/Qwen3-14B-AWQ",
+      "model_clean": "Qwen/Qwen3-14B-AWQ",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "AWQ",
+      "params_b": 14.0,
+      "name_params_b": 14.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 180.43566315423652
+    },
+    {
+      "model": "openai/gpt-oss-20b",
+      "model_clean": "openai/gpt-oss-20b",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "BF16",
+      "params_b": 20.0,
+      "name_params_b": 20.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 405.0974544317216
+    },
+    {
+      "model": "openai/gpt-oss-120b",
+      "model_clean": "openai/gpt-oss-120b",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "BF16",
+      "params_b": 120.0,
+      "name_params_b": 120.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 74.75385852312364
+    },
+    {
+      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
+      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-4bit",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "GPTQ",
+      "params_b": 30.0,
+      "name_params_b": 30.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 214.65152188564062
+    },
+    {
+      "model": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
+      "model_clean": "btbtyler09/Qwen3-Coder-30B-A3B-Instruct-gptq-8bit",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "GPTQ",
+      "params_b": 30.0,
+      "name_params_b": 30.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 188.17083503449163
+    },
+    {
+      "model": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
+      "model_clean": "dazipe/Qwen3-Next-80B-A3B-Instruct-GPTQ-Int4A16",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "GPTQ",
+      "params_b": 80.0,
+      "name_params_b": 80.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 128.30078036872973
+    },
+    {
+      "model": "google/gemma-3-12b-it",
+      "model_clean": "google/gemma-3-12b-it",
+      "env": "TP1",
+      "gpu_config": "single",
+      "quant": "BF16",
+      "params_b": 12.0,
+      "name_params_b": 12.0,
+      "backend": "ROCm",
+      "network": "RoCE",
+      "error": false,
+      "test": "Throughput",
+      "tp": 1,
+      "tps_mean": 267.99881204205957
     }
   ]
 }