| 123456789101112131415161718 |
- {
- "competing_traffic": "none; four nodes dedicated",
- "context_limit": 524288,
- "hardware": "4 x ASUS Ascent GX10 (NVIDIA GB10, 128 GB unified memory each)",
- "kv_cache_dtype": "fp8_e4m3",
- "model": "nvidia/GLM-5.3-Flash-NVFP4 (ModelOpt)",
- "model_revision": "09b04e5e",
- "quantisation": "ModelOpt NVFP4 routed experts; dense linears FP8 after load, KDA in_proj, o_proj, shared experts and drafter NVFP4",
- "serving_engine": "vLLM nightly ddd6fbca, this repo's image, with every override in experimental/",
- "topology": "TP=4 over one 200G switch, both ConnectX ports per node",
- "selectors": {},
- "tensor_parallel_size": 4,
- "max_num_seqs": 50,
- "proposal_k": 7,
- "comparison_note": "Same comparison id and seed as ringside-redhat-rowsplit-20260926, for identical prompts; different checkpoint and stack",
- "checkpoint": "nvidia/GLM-5.3-Flash-NVFP4",
- "served_model_alias": "glm53"
- }
|
