Audit Athena quantization inventory and runtime efficiency limits

2026-09-20 21:18:27 +02:00
parent 4007242709
commit 2425c999b0
11 changed files with 532 additions and 1 deletions
@@ -0,0 +1,71 @@
+[
+  {
+    "source": "benchmarks/qwen38-final-pre-move-20260822/qwen38-final-acceptance-20260822/cases/08-latest-nvfp4-72k-embedded-mtp3-p010/result.json",
+    "case": {
+      "id": "08-latest-nvfp4-72k-embedded-mtp3-p010",
+      "model": "nvfp4-latest",
+      "context": 72000,
+      "split": [
+        72,
+        28
+      ],
+      "projector": "off",
+      "mtp": true,
+      "mtp_max": 3,
+      "quality": false,
+      "long_fill": 0.0,
+      "vision": false
+    },
+    "model_path": "/models/qwen3.8-nvfp4-split/Qwen3.8-27B-NVFP4-Q4_K_M-mtp.gguf",
+    "status": "load_failed",
+    "loaded": false,
+    "mean_predicted_tps": null,
+    "warning": "Historical GGUF experiment, different runtime/prompts/settings. Not current native NVIDIA NVFP4 performance."
+  },
+  {
+    "source": "benchmarks/qwen38-final-pre-move-20260822/qwen38-final-acceptance-20260822/cases/09-latest-nvfp4-72k-split-mtp2-p010/result.json",
+    "case": {
+      "id": "09-latest-nvfp4-72k-split-mtp2-p010",
+      "model": "nvfp4-latest",
+      "context": 72000,
+      "split": [
+        85,
+        15
+      ],
+      "projector": "off",
+      "mtp": true,
+      "mtp_max": 2,
+      "quality": false,
+      "long_fill": 0.0,
+      "vision": false
+    },
+    "model_path": "/models/qwen3.8-nvfp4-split/Qwen3.8-27B-NVFP4-Q4_K_M-mtp.gguf",
+    "status": "complete",
+    "loaded": true,
+    "mean_predicted_tps": 42.27570687556832,
+    "warning": "Historical GGUF experiment, different runtime/prompts/settings. Not current native NVIDIA NVFP4 performance."
+  },
+  {
+    "source": "benchmarks/qwen38-final-pre-move-20260822/qwen38-final-acceptance-20260822/cases/10-latest-nvfp4-72k-split-mtp3-p010/result.json",
+    "case": {
+      "id": "10-latest-nvfp4-72k-split-mtp3-p010",
+      "model": "nvfp4-latest",
+      "context": 72000,
+      "split": [
+        85,
+        15
+      ],
+      "projector": "off",
+      "mtp": true,
+      "mtp_max": 3,
+      "quality": true,
+      "long_fill": 0.0,
+      "vision": false
+    },
+    "model_path": "/models/qwen3.8-nvfp4-split/Qwen3.8-27B-NVFP4-Q4_K_M-mtp.gguf",
+    "status": "complete",
+    "loaded": true,
+    "mean_predicted_tps": 35.94553417986984,
+    "warning": "Historical GGUF experiment, different runtime/prompts/settings. Not current native NVIDIA NVFP4 performance."
+  }
+]