Updating toolbox and pushing GitHub Action

scripts/01-rocm-env-for-triton.sh

@@ -0,0 +1,3 @@
+# Required for Strix Halo / RDNA3.5 on vLLM
+export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+export VLLM_TARGET_DEVICE=rocm

scripts/vllm-env.sh

@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-# Auto-activate vLLM environment for toolbx
-
-# Activate PyTorch + vLLM environment
-source /torch-therock/.venv/bin/activate
-
-# ROCm and performance environment variables
-export PYTORCH_ROCM_ARCH=gfx1151
-export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
-export VLLM_USE_TRITON_FLASH_ATTN=0
-export TORCH_COMPILE_DEBUG=1
-export VLLM_COMPILE_LEVEL=3
-
-# Detect and export ROCm toolchain paths
-eval "$(
-python3 - <<'PY'
-try:
-    import pathlib, _rocm_sdk_core as r
-    base = pathlib.Path(r.__file__).parent / "lib" / "llvm" / "bin"
-    lib  = pathlib.Path(r.__file__).parent / "lib"
-    print(f'export TRITON_HIP_LLD_PATH="{base / "ld.lld"}"')
-    print(f'export TRITON_HIP_CLANG_PATH="{base / "clang++"}"')
-    print(f'export LD_LIBRARY_PATH="{lib}:$LD_LIBRARY_PATH"')
-except ImportError:
-    pass
-PY
-)" 2>/dev/null || true
-
-# Enable flash attention
-export FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE

scripts/zz-venv-last.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+# Ensure /opt/venv/bin is first even if ~/.local/bin or ~/.cargo/bin prepend themselves via user dotfiles.
+
+_venv_path_fix() {
+  # remove any existing /opt/venv/bin entries, then prepend one
+  local newpath
+  newpath="$(printf '%s' "$PATH" | awk -v RS=: -v ORS=: '$0!="/opt/venv/bin"{print}')"
+  PATH="/opt/venv/bin:${newpath%:}"
+}
+
+# run once after shell init; don't duplicate
+case "$PROMPT_COMMAND" in
+  *_venv_path_fix*) : ;;
+  *) PROMPT_COMMAND="_venv_path_fix${PROMPT_COMMAND:+;$PROMPT_COMMAND}" ;;
+esac
+