Page values for "GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2)"
From CMP Wiki
More actions
"Recipe" values
1 row is stored for this page| Field | Field type | Allowed values | Value |
|---|---|---|---|
| page | Page | GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) | |
| model | String | GLM-5.3-Flash (320B MoE, 18B active) | |
| quant | String | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw | |
| engine | String | ExLlamaV3 (patched, pinned) | |
| engine_version | String | upstream dev 0f685aa9fd plus 9 patches; repo db01c256 | |
| cards | String | 8 GB · 10 GB · both | 8 GB |
| card_count | Integer | 2 | |
| topology | String | TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting | |
| context | String | 392,960 per request; 655,360-token Q8 KV pool shared by two requests | |
| decode_tps | Float | 107.67 | |
| prefill_tps | Float | 1,533.0 | |
| vram_per_card | String | about 62.9 / 62.3 GiB idle, peaks about 64.4 / 63.0 GiB with vision | |
| power_cap | String | 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers | |
| driver | String | 610.57.04 | |
| unlock_build | String | cmpunlocker (64 GiB); P2P-capable driver for the NCCL backend | |
| flags | Text | scripts/run_tp.sh: -cq 8 -cs 655360 --context-cap 392960 -ndt 7 -chunk_size 2048 -ambs 2 -gs 62,63 -tp --experimental-dflash2-tp -tpb nccl --q8-staging --q8-head-gpu0 --concurrent-max-sequences 2; EXL3_MLA_KIDX_RING=32 | |
| status | String | Proven · Reported · Theory · Contested · Warning | Reported |
| tested_by | String | nestaa51 | |
| tested_on | Date | 2026-10-13 | |
| source | String | https://github.com/alexvnesta/glm53-flash-cmp170hx-tp | |
| files | Wikitext |