kindling.json 2.4 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859
  1. {
  2. "version": 1,
  3. "title": "GLM-5.3-Flash",
  4. "variant": "NVFP4",
  5. "summary": "nvidia/GLM-5.3-Flash-NVFP4 on vLLM with mentat in place of Ray, over RoCE. DFlash2 speculative decoding, GB10 NVFP4 MoE kernels, RDMA collectives and processed-weight snapshots for ~3 minute restarts.",
  6. "model": "nvidia/GLM-5.3-Flash-NVFP4",
  7. "hardware": "ASUS GX10 (GB10)",
  8. "stack": ["vLLM", "mentat", "DFlash2", "RoCE"],
  9. "default_config": "tp4",
  10. "configs": [
  11. {
  12. "id": "tp2",
  13. "label": "TP=2",
  14. "nodes": 2,
  15. "decode": { "code": 66.2, "prose": 42.6, "structured": 94.7 },
  16. "prefill": [ { "context": 32768, "tps": 2931 }, { "context": 131072, "tps": 2871 } ],
  17. "concurrency": [ { "streams": 4, "aggregate_tps": 83 }, { "streams": 8, "aggregate_tps": 110 } ],
  18. "max_context": 160000,
  19. "kv_pool_tokens": 1100000,
  20. "max_concurrent": 16,
  21. "quality": [ { "name": "Needle recall", "value": "6/6 up to 128k" } ]
  22. },
  23. {
  24. "id": "tp3",
  25. "label": "TP=3",
  26. "nodes": 3,
  27. "decode": { "code": 88.0, "prose": 54.2, "structured": 121.4 },
  28. "prefill": [ { "context": 32768, "tps": 3864 }, { "context": 131072, "tps": 3736 } ],
  29. "concurrency": [ { "streams": 4, "aggregate_tps": 107 }, { "streams": 8, "aggregate_tps": 138 } ],
  30. "max_context": 524288,
  31. "kv_pool_tokens": 1910000,
  32. "max_concurrent": 64,
  33. "quality": [ { "name": "Needle recall", "value": "12/12 up to 507k" } ]
  34. },
  35. {
  36. "id": "tp4",
  37. "label": "TP=4",
  38. "nodes": 4,
  39. "decode": { "code": 117.3, "prose": 70.5, "structured": 164.3 },
  40. "prefill": [ { "context": 32768, "tps": 4934 }, { "context": 131072, "tps": 4776 } ],
  41. "concurrency": [ { "streams": 4, "aggregate_tps": 144 }, { "streams": 8, "aggregate_tps": 194 } ],
  42. "max_context": 524288,
  43. "kv_pool_tokens": 4400000,
  44. "max_concurrent": 64,
  45. "quality": [ { "name": "Needle recall", "value": "12/12 up to 507k" } ]
  46. },
  47. {
  48. "id": "tp6",
  49. "label": "TP=6",
  50. "nodes": 6,
  51. "prefill": [ { "context": 32768, "tps": 4907 }, { "context": 131072, "tps": 4731 } ],
  52. "max_context": 524288,
  53. "kv_pool_tokens": 4470000,
  54. "max_concurrent": 64,
  55. "quality": [ { "name": "Needle recall", "value": "12/12 up to 480k" } ],
  56. "notes": "Decode not measured at TP=6. Ran on other boxes and builds, with GPU clocks locked at 1989 MHz."
  57. }
  58. ]
  59. }