Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

docs/AMD_HIP.md: a hand-built engine has no hipBLASLt tuning table unless it is pointed at one (#4034)

Niko1221a3b7d9de3fc1
diff --git a/docs/AMD_HIP.md b/docs/AMD_HIP.md
index 2ff7549e4..d858002f4 100644
--- a/docs/AMD_HIP.md
+++ b/docs/AMD_HIP.md
@@ -193,6 +193,13 @@ the card's name, its architecture and the build's list.
- CMake fetches the pinned llama.cpp revision. For an offline build, point `STRATA_GGML_DIR` at a checkout of that exact
revision.
- Linux engine packages: `VARIANT=rx6000 packaging/build_hip_linux.sh`.
+- **A hand-built engine has no tuning table.** The packaged engines carry `tuning/<arch>-hipblaslt-<version>.txt` beside the binary
+ ([Ready-made packages](#ready-made-packages) lists which); a build made with plain `cmake` and started
+ through `serve/server.py` has none, and the prompt GEMMs fall back to plain hipBLAS. On Strix Halo (gfx1151) that read prompts
+ 20-25% slower in one report (1,043 against 826 tok/s at 8K, hipBLASLt 1.4.1, a single run per arm, not a paired A/B). Point
+ `STRATA_HIPBLASLT_TUNING` at the file in `tools/hip/` (`gfx1151-hipblaslt-100401.txt` for hipBLASLt 1.4.1), or copy it into a
+ `tuning/` folder next to the `strata` binary. The start log then has the line `prefill gemm: hipBLASLt tuning enabled (N rows,
+ gfx1151, version 100401)`; without it, no table was taken.
Run the registered checks with: