...], <code><nowiki>patch/</nowiki></code> folder only || 38 tensors, 1.6 GB; Ivan's copy sha256 <code><nowiki>16e63cd3e79c3590855aad340b5b9dc37f27562bf2d0a40
# <code><nowiki>.env</nowiki></code> changes, all specific to Ivan's box:
...
15 KB (2,256 words) - 10:36, 30 October 2026
...en3.8 Flash Next on 2x 8 GB (mannydelrio PP2)]], matstatman's 4-card Qwen, Ivan's full GLM 5.3
* Ivan's lab: the full 753B GLM-5.3 (an abliterated NVFP4-expert / INT8 requant) r
...
34 KB (4,777 words) - 10:02, 30 October 2026
* Ivan (8-card box, two TP4 instances of this recipe, abliterated W4A16 + MTP chec
Ivan's measurements, same bench on all three (single stream code / structured /
...
31 KB (4,236 words) - 10:02, 30 October 2026
...> at 656 B per token per layer plus FP8 indexer keys (vllm-170hx-glm5). On Ivan's eight-card run the int8 KV tile matched plain FP8 in quality and was 10%
...FP8 rounding. See [[DeepSeek V4.1 Flash heretic on 4x 8 GB]]. Alternatives Ivan did not use: mrfakename's rank-2 edit (lowest KL, needs your own apply) and
...
26 KB (3,327 words) - 08:54, 30 October 2026
...oves it applied. {{Source|repo=https://github.com/cachenetics/170tune}} On Ivan's cards a miner's own core-offset call failed with "an internal driver erro
...r fuse map before tuning the memory clock.''' {{Status|Warning}} On two of Ivan's eight cards, which have FBPA0 fused off (one dead HBM stack each, read wi
...
27 KB (3,472 words) - 10:02, 30 October 2026
Field notes from Ivan's run:
* Ivan, 8 cards at Gen2 x16 with P2P, 1,024-token answers: at 180 W one user code
...
10 KB (1,509 words) - 08:14, 30 October 2026
== Ivan's eight-card server ==
{{Status|Reported}} Built by @darkmaniac7 (Ivan), notes from 2026-10-26 to 10-29. {{Source|handle=darkmaniac7|date=2026-10-
...
21 KB (2,841 words) - 10:01, 30 October 2026
* Ivan (TP4 on four cards of an 8-card box, Gen2 x16, P2P on, his own bench with 1
* '''Long-context copy errors (engine defect).''' On Ivan's test (75k to 92k-token roleplay turns that must copy image paths and name
...
14 KB (1,947 words) - 08:14, 30 October 2026
..., thinking off) and report three columns, '''structured / code / prose'''. Ivan's lab bench uses 1,024-token answers and lists '''code / structured / prose
* Ivan's lab rows are on his own 8-card box and are marked "lab notes". All other
...
38 KB (4,111 words) - 10:01, 30 October 2026
* Ivan's lab notes: the "ECC Enabled" readout comes from cmpunlocker's reporting p
* Ivan's notes on issue #9: the listing renames the card's device ID, which the pu
...
56 KB (7,395 words) - 10:12, 30 October 2026
listed on localairecipes and in Ivan's notes:
* Ivan's lab notes, 2026-10-27 (localairecipes listing).
...
6 KB (791 words) - 08:54, 30 October 2026
* Ivan (heretic patch applied, a miner sharing the cards so <code><nowiki>--gpu-me
On Ivan's long-context copy test (75k to 92k-token roleplay turns that must copy im
...
14 KB (1,914 words) - 08:14, 30 October 2026
...t (40 identifiers, shorter prompts) passed 300/300; it has not been run on Ivan's test.
...
10 KB (1,435 words) - 08:14, 30 October 2026
* Ivan ran release 0.2.0 on his box (Gen2 x16, P2P, 1,024-token answers): one user
* Keep vLLM's custom all-reduce on with this recipe: on Ivan's box turning it off cost 17 to 25% single-stream and 12 to 15% at 8 users.
...
14 KB (1,854 words) - 10:02, 30 October 2026
* Your intake air sets the floor. Ivan's eight-card server raises its power caps only while the inlet is below 24
...
17 KB (2,095 words) - 10:02, 30 October 2026
...com/channels/1518634379604136178/1518692084729511967/1562217046606090251}} Ivan's eight-card server and its UPS run on a dedicated 240 V circuit. {{Source|
...
27 KB (3,300 words) - 10:02, 30 October 2026
iron instead of an irreversible [[RIR fuse burn]]. Ivan's estimate for TP4 single user, 280-355 tok/s toward 440-480
...
28 KB (3,726 words) - 10:02, 30 October 2026
...ublished 153; PIECEWISE graphs are about 3x slower per step on an old CPU. Ivan's notes count it CPU-bound for the same reason. His W4A16 recipe keeps the
...
6 KB (782 words) - 08:54, 30 October 2026
...pactic's W4A16 on 8 cards, Mountains' EXL3 on 4 CMPs plus an RTX PRO 6000, Ivan's NVFP4 TP2 x PP4): [[Recipes]] and [[Benchmarks]].
...
11 KB (1,448 words) - 08:54, 30 October 2026
* The room matters too. Ivan's eight-card server put about 2 kW into the room; raising its fan wall from
...
8 KB (1,112 words) - 08:51, 30 October 2026
