diff --git a/docs/hrx.md b/docs/hrx.md index 6078ec2e3..adddded4e 100644 --- a/docs/hrx.md +++ b/docs/hrx.md @@ -687,10 +687,14 @@ buffer — the loader resolves it through `select_weight_buft()` over `make_cpu_ repack path is allowed (`binding.weight` is that flag, engine#969). Measured, three arms in one ticket, Qwen3-4B-Q4_K_M, `-p 512 -n 32`, own peak `drm-total-gtt` from `bo_trace`: -| arm | BO bytes | own peak GTT | prefill | -|---|---:|---:|---:| -| everything on the device | 5,867.8 MiB | 5,495 MiB | 3,241 tok/s | -| host-placed `ffn_down` + repack | 7,743.8 MiB | 6,459 MiB | 3,879 tok/s | +| arm | load mode | BO bytes | own peak GTT | prefill | +|---|---|---:|---:|---:| +| everything on the device | `-lm mmap` (default) | 5,867.8 MiB | 5,495 MiB | 3,241 tok/s | +| host-placed `ffn_down` + repack | `-lm none` | 7,743.8 MiB | 6,459 MiB | 3,879 tok/s | + +The two rows differ in load mode as well as placement, so their prefill ratio (1.197×) is **not** a placement result: `-lm none` +alone is 1.1745× prefill on the same file and shape ([#1668](https://github.com/1bit-MONSTER/engine/issues/1668)). The third arm +holds the load mode fixed, and that is the placement number quoted below. The native range really does shrink — the whole-model buffer drops **2,376.0 → 1,786.0 MiB (−590)** — but the repacked and parked copies cost more than the range they displace: a new **896.0 MiB** class (`HostWeightCache`'s reserved-plus-parked slots), **37 ×