fix(kyojin): serve wrappers get a C compiler, hipcc and the JIT sources

env.sh assumes a distro box: a system cc for Triton's runtime C builds, a hipcc
for the JIT HIP kernels and the repo root on PYTHONPATH. The wrappers now export
all three (, ,  for the amdgcn bitcode,
PYTHONPATH=${src}/gr for gr_mix_hip).

The wheel ships only the files pyproject lists, so the .hip kernel sources, the
qsa_proof marker gating the QSA prefill kernel, the dense-GEMM tuning seed and
gr/ never reached site-packages: kernels fell back to Triton and every start
re-tuned dense GEMM for 7 minutes.

Verified on the Strix Halo box: server READY, start-up 713 s -> 103 s, warm-up
418 s -> 4 s, 584-token prompt served over /v1/chat/completions with no
'unavailable' fallbacks.

packages/kyojin/package.nix

@@ -145,6 +145,14 @@ let
       sed -i '1i import sys as _sys, types as _types; _exl3_pkg = _types.ModuleType("exllamav3"); _exl3_pkg.__path__ = ["exllamav3"]; _sys.modules.setdefault("exllamav3", _exl3_pkg)' setup.py
       sed -i 's|^from exllamav3\.exllamav3_ext\.build_config import get_sources as _get_sources$|import sys as _exl3_sys; _exl3_sys.path.insert(0, "exllamav3/exllamav3_ext"); from build_config import get_sources as _get_sources; _exl3_sys.path.pop(0)|' setup.py
       grep -q 'from build_config import' setup.py
+
+      # Upstream serves these from the checkout (pip install -e .); the wheel only
+      # packages what pyproject lists, and three things are read out of the
+      # installed package at run time: the .hip sources the JIT kernels compile,
+      # the qsa_proof marker gating the QSA prefill kernel, and the dense-GEMM
+      # tuning seed (without it every start re-tunes for ~7 minutes).
+      sed -i 's|^    "exllamav3_ext/\*\*/\*",$|&\n    "**/*.hip",\n    "**/*.ok",\n    "model/dense_gemm_tune_seed.txt",|' pyproject.toml
+      grep -q '"\*\*/\*.hip"' pyproject.toml
     '';
 
     # tests/ needs a real gfx1151 GPU and the published model packs.
@@ -169,6 +177,19 @@ stdenv.mkDerivation {
 
   # Only the wrappers are installed here: the serve scripts are plain scripts,
   # they import their siblings by path, and the module itself is in `library`.
+  #
+  # Upstream `tools/strix_halo/env.sh` is sourced before serving, and on a
+  # distro box that brings a system `cc` (the AMD Triton backend compiles a
+  # CPython glue module at runtime, and again per kernel launcher) plus a hipcc
+  # for the JIT HIP kernels (qsa prefill, gdn, kda, ple). Nix gives the wrapper
+  # neither, so export them here instead: Triton takes the compiler from $CC,
+  # exllamav3 finds hipcc under $EXL3_ROCM_SDK, and hipclang only finds the
+  # amdgcn bitcode through $HIP_DEVICE_LIB_PATH (hipcc's own DEVICE_LIB_PATH is
+  # not enough for a --genco call). The ROCm paths stay out of LD_LIBRARY_PATH
+  # (upstream Trap 1/5: a second libhsa segfaults torch), and PYTHONPATH gets
+  # `gr/`: upstream env.sh puts the repo root there so `import gr_mix_hip` (the
+  # hand-written gated-residual WMMA kernel the server asks for with
+  # EXL3_GR_HIP=1) resolves, and it is not part of the wheel either.
   dontConfigure = true;
   dontBuild = true;
 
@@ -181,6 +202,11 @@ stdenv.mkDerivation {
     for model in glm mimo qwen; do
       makeWrapper ${python}/bin/python $out/bin/kyojin-serve-$model \
         --set PYTORCH_ROCM_ARCH ${rocmArch} \
+        --set CC ${stdenv.cc}/bin/cc \
+        --set EXL3_ROCM_SDK ${rocmToolkit} \
+        --set HIP_DEVICE_LIB_PATH ${rocmPackages."rocm-device-libs"}/amdgcn/bitcode \
+        --prefix PATH : ${lib.makeBinPath [ stdenv.cc ]} \
+        --prefix PYTHONPATH : ${src}/gr \
         --add-flags "${src}/tools/$model/serve.py"
     done