#5803 [Engine] gfx1152: opt-in thread-per-block long-prompt scorer (STRATA_SCORES_THREAD=1)

open · @chadstthomas-svg · 0 comments · View on GitHub

BenchmarksMulti-GPUAMD / HIPNVIDIA / CUDAWindows

Description

**Problem and result**

The thread-per-block long-prompt QSA scorer is gated to gfx1100 (opt-in) and gfx12 (default) only. This adds gfx1152 to the opt-in class, following the same admission pattern as #5725.

The scorer is bitwise-identical to the warp scorer per the gfx1100 PR's own parity (`--full`, 35 cases; exact_pp residual dumps at 70K and 131K tokens), claimed 3.6-4.4x faster in the scorer stage with +8% (131K) to +15% (230K) on a prompt read.

**Approach**

One line in `scores_thread_arch()` (src/kernels/cuda/qsa_select.cu): gfx1152 joins gfx1100 as opt-in class 1. No default change for any other part.

**Validation**

Environment: CachyOS 7.3.1, ROCm 7.2.4 (Arch), Radeon 860M (gfx1152, 8 CUs), engine 0.1.55 + the gfx1152 WMMA enablement (#5725).

- Passed: the identical runtime-gate hunk is exercised in a local gfx1152 build across the full prefill ladder; the scorer engages with STRATA_SCORES_THREAD=1 and outputs stay greedy-identical.
- Failed: none observed.
- Not run: Windows, other gfx parts (the change is additive to the opt-in list only).

**Measured**

At 32768- and 65536-token prompts with 8192-token chunks: 185.0 and 186.8 tok/s against a 194.2 / 189.3 tok/s reference — within run variance, no gain at this chunking on this card. Kept opt-in: the claim zone is single-chunk prompts of 131-230K tokens, achievable on the gfx1152 64 GB memory budget for users who run those depths.

Related on this site

Quick links to related pages.