 Alexander Korolev
|
c7d0eb5f7d
flare gb10: bench 2026-10-24_205024 resident-scales — 88/100, pp2048 2,892, tg 54.3/81.1, d8192 61.0/80.4
|
11 小時之前 |
|
Alexander Korolev
|
a8eea3d204
flare gb10: enable --embedding-table-resident-scales (engram scales resident, ~1.5 GB/rank)
|
12 小時之前 |
|
Alexander Korolev
|
8df94a8561
flare gb10: drop the stale 48 GiB line, describe resident-scales accurately
|
13 小時之前 |
|
Alexander Korolev
|
588d069921
flare gb10: comment matches the recipe (disk + 4 GiB) — resident-scales omitted on purpose
|
13 小時之前 |
|
Alexander Korolev
|
1c5ca3a73e
flare gb10: 4 GiB RAM row cache + local package (engram cannot be split by file)
|
21 小時之前 |
|
Alexander Korolev
|
9dab339aa5
flare gb10: port 30000, group name = recipe name
|
1 天之前 |
|
Alexander Korolev
|
bd819f030b
flare gb10: declare the served model id for the bench client
|
1 天之前 |
|
Alexander Korolev
|
e424dc50ab
1017: FLARE gb10 lane — the lab's own signed package for four Sparks
|
1 天之前 |
|
Alexander Korolev
|
db4c0d8e29
1017 vllm032: 512K context and the standard served-model-name
|
1 天之前 |
|
Alexander Korolev
|
833d3bde37
1017 vllm032: bench 91/100 (177/194) with locally staged engram — hard mode 48/56, pp 2,927 t/s c1
|
1 天之前 |
|
Alexander Korolev
|
f6e65187d2
1017 vllm032: engram staged locally via [[staging]] local-view — prefill 750 -> 3,957 t/s
|
1 天之前 |
|
Alexander Korolev
|
c48f50a6ba
dsv41-flash-1017-4x-vllm032: DSpark K5 on stock vLLM matches production decode (52.9 c1 / 66.4 c2)
|
1 天之前 |
|
Alexander Korolev
|
d92a1b9659
dsv41-flash-1017-4x-vllm032: stock vLLM baseline for 1017 — 29.7 t/s c1, 45.6 c2, prefill ~750
|
1 天之前 |
|
Alexander Korolev
|
2b7b4c711d
glm53-mxfp4-fp8-4x-r39: MTP on the 128K lane benches 90/100; ~120K prefill wedges the hosts
|
1 天之前 |
|
Alexander Korolev
|
8a88799019
glm53-mxfp4-fp8-4x-r39: bench with MTP on the 128K lane — 90/100; ~120K prefill hangs the hosts (repro)
|
1 天之前 |
|
Alexander Korolev
|
d74b8f37af
glm53-mxfp4-fp8-4x-r39: KV pin 7.5 GiB for the 131072 window (6.68 GiB needed, 6 GiB pinned -> refused)
|
2 天之前 |
|
Alexander Korolev
|
559e8d4db9
glm53-mxfp4-fp8-4x-r39: DSpark speculator attempt recorded — dies in the draft KV builder on r39 (needs vLLM 0.32)
|
2 天之前 |
|
Alexander Korolev
|
35de736d83
glm53-mxfp4-fp8-4x-r39: 196K window hung the host on a 150K prefill — back to 131072 / 6 GiB KV / 2048 chunks
|
2 天之前 |
|
Alexander Korolev
|
baabab7674
glm53-mxfp4-fp8-4x-r39: DCP4 rejected by every MLA backend ('DCP not supported'); window now 196608 at a 10 GiB KV pin
|
2 天之前 |
|
Alexander Korolev
|
866d05a06c
glm53-mxfp4-fp8-4x-r39: MTP k=5 measured — same throughput as k=3 (33-37 c1, 47.9 c2), acceptance 4.4/5
|
2 天之前 |
|
Alexander Korolev
|
8efe2f3d82
glm53-mxfp4-fp8-4x-r39: MTP k=3 breaks the decode ceiling — 19.0 -> 33-35 t/s c1, 47.6 t/s c2
|
2 天之前 |
|
Alexander Korolev
|
75276e8670
glm53-mxfp4-fp8-4x-karmic: A/B vs r39 — native b12x MoE experts buy +1% (19.2 vs 19.0 t/s)
|
2 天之前 |
|
Alexander Korolev
|
1c4d95db9d
glm53-mxfp4-fp8-4x-r39: measured on v2.8.0 — 88/100 (170/194), Hard Mode 46/56, 1 safety-critical, 8.8 s median turn
|
2 天之前 |
|
Alexander Korolev
|
76abd98c9c
glm53-mxfp4-fp8-4x-r39: GLM-5.3 (full) RedHat MXFP4xFP8 boots at TP4 on four Sparks
|
3 天之前 |
|
Alexander Korolev
|
fc210f50bd
glm53-kindling-2x: reproducibility run on v2.8.0 — 95/100 (185/194), Hard Mode 52/56 (93%), zero safety-critical, 3.6 s median turn
|
3 天之前 |
|
Alexander Korolev
|
817cb85d66
glm53-mxfp4-fp8-4x: GLM-5.3 (full, 78 layers) RedHatAI MXFP4xFP8 on 4 Sparks at TP4
|
3 天之前 |
|
Alexander Korolev
|
337f442e74
dsv41-flash-1017-2x-tensorfold: session cache A/B — same 89/100, median turn 6.7s -> 4.2s (responsiveness 23 -> 38)
|
3 天之前 |
|
Alexander Korolev
|
994a1ffa05
dsv41-flash-1017-2x-tensorfold: staging script moves into the recipe (scripts/prepare.sh) + TF_DSV41_SESSIONS=1 experiment
|
3 天之前 |
|
Alexander Korolev
|
7e9da804f3
glm53-kindling-2x: re-benchmarked on harness v2.8.0 — 94/100 (183/194), Hard Mode 50/56 (89%), zero safety-critical failures
|
3 天之前 |
|
Alexander Korolev
|
b1f054231b
beam lane: 384K context (13.6x concurrency) + needle-in-haystack at 131K and 327K tokens (both correct, 60s/175s TTFT)
|
3 天之前 |
