cleanup(scripts): drop unread serve knobs and idle kill sleep

scripts/rdna_launcher_common.sh

@@ -80,7 +80,6 @@ rdna_init() {
     export GPU_MAX_HW_QUEUES=${GPU_MAX_HW_QUEUES:-2}
     export VLLM_ROCM_USE_AITER=${VLLM_ROCM_USE_AITER:-0}
     export VLLM_ROCM_USE_AITER_MOE=${VLLM_ROCM_USE_AITER_MOE:-0}
-    export VLLM_RDNA_FORCE_FP16=${VLLM_RDNA_FORCE_FP16:-1}
     export TORCH_BLAS_PREFER_HIPBLASLT=${TORCH_BLAS_PREFER_HIPBLASLT:-0}
     export VLLM_BATCH_INVARIANT=${VLLM_BATCH_INVARIANT:-0}
 
@@ -139,13 +138,17 @@ rdna_run_cwd() {
 }
 
 rdna_kill_stale() {
-    local _sig _p
+    local _sig _p _signaled=0
     for _sig in TERM KILL; do
         for _p in $(pgrep -f "entrypoints.cli.main serve|entrypoints.openai.api_server|VLLM::Worker|VLLM::EngineCore|PleOffloadWorker" 2>/dev/null); do
             [ -r "/proc/$_p/environ" ] || continue
-            tr '\0' '\n' < "/proc/$_p/environ" 2>/dev/null | grep -q "^VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT}$" && kill -"$_sig" "$_p" 2>/dev/null
+            tr '\0' '\n' < "/proc/$_p/environ" 2>/dev/null | grep -q "^VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT}$" && kill -"$_sig" "$_p" 2>/dev/null && _signaled=1
         done
-        [ "$_sig" = TERM ] && sleep 8
+        if [ "$_sig" = TERM ] && [ "$_signaled" = 1 ]; then
+            sleep 8
+        fi
     done
-    sleep 3
+    if [ "$_signaled" = 1 ]; then
+        sleep 3
+    fi
 }

scripts/recipes/flashnext-mtp0.env

@@ -33,7 +33,6 @@ VLLM_PLE_CPU_OFFLOAD=1
 
 # --- RDNA vLLM env block ---
 VLLM_USE_V2_MODEL_RUNNER=1
-VLLM_ROCM_MOE_PREFILL=0
 VLLM_GDN_HIP_PREFILL=0
 VLLM_RDNA_FUSED_SE=1
 VLLM_RDNA_DENSE_INT8=0

scripts/recipes/flashnext-mtp2.env

@@ -33,7 +33,6 @@ VLLM_PLE_CPU_OFFLOAD=1
 
 # --- RDNA vLLM env block ---
 VLLM_USE_V2_MODEL_RUNNER=1
-VLLM_ROCM_MOE_PREFILL=0
 VLLM_GDN_HIP_PREFILL=0
 VLLM_RDNA_FUSED_SE=1
 VLLM_RDNA_DENSE_INT8=0

scripts/recipes/full.env

@@ -32,8 +32,6 @@ ALLOC_CONF=auto
 # --- RDNA vLLM env block ---
 VLLM_USE_V2_MODEL_RUNNER=1
 VLLM_USE_RDNA2_FA=1
-# Do not hash the live last hybrid GDN+FA page while decode is appending.
-VLLM_ROCM_SKIP_LIVE_TAIL_HASH=1
 VLLM_USE_AOT_COMPILE=0
 VLLM_DISABLE_COMPILE_CACHE=1
 FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE

scripts/serve_rdna.sh

@@ -313,7 +313,7 @@ RDSE_ENV_KEYS=(
     LD_LIBRARY_PATH ROCM_HOME ROCM_PATH HIP_PATH HIP_ROOT_DIR
     PYTHONPATH HF_HUB_OFFLINE TRANSFORMERS_OFFLINE
     HIP_VISIBLE_DEVICES VLLM_WORKER_MULTIPROC_METHOD GPU_MAX_HW_QUEUES
-    VLLM_ROCM_USE_AITER VLLM_ROCM_USE_AITER_MOE VLLM_RDNA_FORCE_FP16
+    VLLM_ROCM_USE_AITER VLLM_ROCM_USE_AITER_MOE
     TORCH_BLAS_PREFER_HIPBLASLT VLLM_BATCH_INVARIANT
     VLLM_CACHE_ROOT TRITON_CACHE_DIR TORCHINDUCTOR_CACHE_DIR TORCH_EXTENSIONS_DIR
     PYTORCH_TUNABLEOP_ENABLED PYTORCH_TUNABLEOP_TUNING
@@ -323,14 +323,13 @@ RDSE_ENV_KEYS=(
     FLASH_ATTENTION_TRITON_AMD_ENABLE
     VLLM_USE_V2_MODEL_RUNNER VLLM_USE_AOT_COMPILE VLLM_DISABLE_COMPILE_CACHE
     VLLM_USE_BREAKABLE_CUDAGRAPH VLLM_RDNA_FUSED_HC
-    VLLM_ROCM_MOE_PREFILL VLLM_GDN_HIP_PREFILL VLLM_RDNA_FUSED_SE
+    VLLM_GDN_HIP_PREFILL VLLM_RDNA_FUSED_SE
     VLLM_RDNA_DENSE_INT8 VLLM_RDNA_DENSE_INT8_ONLY VLLM_RDNA_DENSE_GEMV
     VLLM_RDNA2_W4A8_SDOT4 VLLM_RDNA2_W4A8_DEBUG
     VLLM_FORCE_CUSTOM_ALL_REDUCE VLLM_RDNA_AR VLLM_RDNA_AR_MAX_KB
     VLLM_RDNA_AR_ONESHOT_KB VLLM_RDNA_AR_BLOCKS VLLM_RDNA_AR_PACE
     VLLM_CAUSAL_CONV1D_RDNA2_FWD VLLM_CAUSAL_CONV1D_RDNA2_UPDATE
     VLLM_ENABLE_STARTUP_PLAN VLLM_TUNED_CONFIG_FOLDER
-    VLLM_ROCM_SKIP_LIVE_TAIL_HASH
     VLLM_PLE_CPU_OFFLOAD VLLM_PLE_OFFLOAD_READY_TIMEOUT VLLM_PLE_QUANT_DIR
     VLLM_EXL3_DEBUG
     NCCL_P2P_LEVEL RCCL_P2P_NET_DISABLE RCCL_P2P_BATCH_ENABLE NCCL_PROTO

tools/rdna2/tunableop_env.sh

@@ -1,160 +1,5 @@
 #!/usr/bin/env bash
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Profile-aware TunableOp helper for the RDNA fork.
-#
-# This is the tools/rdna2 mirror of tools/rdna2_028/tunableop_env.sh; the two
-# files are kept byte-identical below this header. Prefer the rdna2_028 copy.
-#
-#   configure_tunableop <librocblas.so.5 path> <tunableop dir> [profile]
-#
-# Rows live in <tunableop dir>/<profile-dir>/tunableop_results{0..3}.csv, one
-# folder per rocBLAS build, registered in <tunableop dir>/profiles.json
-# (name -> {dir, lib_sha256, rocblas, rocm, torch, rows_per_rank, status}).
-# Solver IDs are build-specific, so rows are only ever used for the rocBLAS
-# build that produced them.
-#
-# Selection:
-#   * explicit (third argument or $TUNABLEOP_PROFILE): use that profile and
-#     validate its lib_sha256 against the loaded library. A mismatch makes the
-#     launch fail (return 1) unless TUNABLEOP_ALLOW_MISMATCH=1, which warns
-#     hard and proceeds anyway.
-#   * default: auto-select the profile whose lib_sha256[:12] equals the loaded
-#     library's sha256[:12].
-#   * no registered match: a legacy <tunableop dir>/rocblas-<sha12>/ folder is
-#     still honoured, then the per-user fallback
-#     $HOME/.cache/tunableop/tunableop_results.csv (created).
-#
-# Lookup only by default (PYTORCH_TUNABLEOP_TUNING=0); set TUNING=1 to re-tune
-# shapes missing from the table. Rows are never written to /tmp or the run CWD.
-#
-# configure_mtp_tunableop is an alias so existing callers do not break.
-# configure_v620_tunableop keeps the historical strict behaviour (lookup is
-# disabled unless this build's four rank files exist; no per-user fallback).
 
-# Print "<name>\t<dir>\t<lib_sha256>" for every profile in the registry.
-# Empty (and non-zero) when the registry is absent or unparseable.
-_tunableop_profiles() {
-    local registry=$1
-    [ -r "$registry" ] || return 1
-    if command -v python3 >/dev/null 2>&1; then
-        python3 -c 'import json, sys
-try:
-    prof = json.load(open(sys.argv[1])).get("profiles", {})
-except Exception:
-    sys.exit(0)
-for name in sorted(prof):
-    p = prof[name] or {}
-    d, s = p.get("dir"), p.get("lib_sha256")
-    if d and s:
-        print("%s\t%s\t%s" % (name, d, s))' "$registry" 2>/dev/null
-    elif command -v jq >/dev/null 2>&1; then
-        jq -r '.profiles // {} | to_entries[]
-               | select(.value.dir and .value.lib_sha256)
-               | "\(.key)\t\(.value.dir)\t\(.value.lib_sha256)"' \
-            "$registry" 2>/dev/null
-    fi
-}
-
-configure_tunableop() {
-    local library=${1:-} rows_root=${2:-} requested=${3:-${TUNABLEOP_PROFILE:-}}
-    local library_id selected_name="" rows_dir profile_line tune_state canonical
-    local profile_dir profile_sha
-    export PYTORCH_TUNABLEOP_HIPBLASLT_ENABLED=0
-    export PYTORCH_TUNABLEOP_TUNING=${PYTORCH_TUNABLEOP_TUNING:-0}
-    export PYTORCH_TUNABLEOP_ENABLED=0
-    unset PYTORCH_TUNABLEOP_FILENAME
-    canonical=${TUNEOP_PER_USER_DIR:-$HOME/.cache/tunableop}
-    mkdir -p -- "$canonical" 2>/dev/null || true
-
-    if [[ -z $library || ! -f $library ]]; then
-        printf 'WARNING: TunableOp lookup disabled: rocBLAS library "%s" is unavailable; using default FP16 algorithms.\n' "$library" >&2
-        return 0
-    fi
-
-    library_id=$(sha256sum -- "$library")
-    library_id=${library_id:0:12}
-
-    if [[ -n $requested ]]; then
-        profile_line=$(_tunableop_profiles "$rows_root/profiles.json" |
-            awk -F'\t' -v n="$requested" '$1 == n { print; exit }')
-        if [[ -z $profile_line ]]; then
-            printf 'ERROR: TunableOp profile "%s" is not registered in %s/profiles.json; refusing to start.\n' \
-                "$requested" "$rows_root" >&2
-            return 1
-        fi
-        selected_name=${profile_line%%$'\t'*}
-        profile_dir=$(printf '%s' "$profile_line" | cut -f2)
-        profile_sha=$(printf '%s' "$profile_line" | cut -f3)
-        if [[ ${profile_sha:0:12} != "$library_id" ]]; then
-            if [[ ${TUNABLEOP_ALLOW_MISMATCH:-0} == 1 ]]; then
-                printf 'WARNING: TunableOp profile "%s" was tuned for rocBLAS %s but the loaded library is %s; TUNABLEOP_ALLOW_MISMATCH=1 set, using it anyway (results may be wrong or slow).\n' \
-                    "$selected_name" "${profile_sha:0:12}" "$library_id" >&2
-            else
-                printf 'ERROR: TunableOp profile "%s" was tuned for rocBLAS %s, but the loaded librocblas.so.5 hashes to %s. Solver IDs are build-specific; refusing to start. Unset TUNABLEOP_PROFILE to auto-select, or set TUNABLEOP_ALLOW_MISMATCH=1 to override (experiments only).\n' \
-                    "$selected_name" "${profile_sha:0:12}" "$library_id" >&2
-                return 1
-            fi
-        fi
-        rows_dir=$rows_root/$profile_dir
-    else
-        profile_line=$(_tunableop_profiles "$rows_root/profiles.json" |
-            awk -F'\t' -v h="$library_id" 'substr($3, 1, 12) == h { print; exit }')
-        if [[ -n $profile_line ]]; then
-            selected_name=${profile_line%%$'\t'*}
-            profile_dir=$(printf '%s' "$profile_line" | cut -f2)
-            rows_dir=$rows_root/$profile_dir
-        else
-            # Legacy hash-named folder, then the per-user fallback below.
-            rows_dir=$rows_root/rocblas-$library_id
-        fi
-    fi
-
-    for rank in 0 1 2 3; do
-        if [[ ! -s $rows_dir/tunableop_results$rank.csv ]]; then
-            if [[ ${TUNEOP_REQUIRE_HASH_ROWS:-0} == 1 ]]; then
-                printf 'WARNING: TunableOp disabled: missing rank %s rows for rocBLAS %s at %s; using default FP16 algorithms.\n' \
-                    "$rank" "$library_id" "$rows_dir" >&2
-                return 0
-            fi
-            tune_state=off
-            [[ $PYTORCH_TUNABLEOP_TUNING == 1 ]] && tune_state=enabled
-            printf 'WARNING: no fork rows for rocBLAS build %s (rank %s missing) at %s; falling back to per-user %s (tuning %s).\n' \
-                "$library_id" "$rank" "$rows_dir" "$canonical/tunableop_results.csv" "$tune_state" >&2
-            export PYTORCH_TUNABLEOP_FILENAME=$canonical/tunableop_results.csv
-            export PYTORCH_TUNABLEOP_ENABLED=1
-            return 0
-        fi
-    done
-
-    tune_state=off
-    [[ $PYTORCH_TUNABLEOP_TUNING == 1 ]] && tune_state=enabled
-    export PYTORCH_TUNABLEOP_FILENAME=$rows_dir/tunableop_results.csv
-    export PYTORCH_TUNABLEOP_ENABLED=1
-    if [[ -n $selected_name ]]; then
-        if [[ -n $requested ]]; then
-            printf 'TunableOp profile %s selected for rocBLAS build %s (rows: %s); tuning %s.\n' \
-                "$selected_name" "$library_id" "$rows_dir" "$tune_state" >&2
-        else
-            printf 'TunableOp profile %s auto-selected for rocBLAS build %s.\n' \
-                "$selected_name" "$library_id" >&2
-            printf 'TunableOp lookup enabled for profile %s (rocBLAS %s; rows: %s); tuning %s (set PYTORCH_TUNABLEOP_TUNING=1 to re-tune missing shapes).\n' \
-                "$selected_name" "$library_id" "$rows_dir" "$tune_state" >&2
-        fi
-    else
-        printf 'TunableOp lookup enabled for rocBLAS build %s (rows: %s); tuning %s.\n' \
-            "$library_id" "$rows_dir" "$tune_state" >&2
-    fi
-}
-
-# Alias kept for the MTP/Flash-Next launchers and other existing callers.
-configure_mtp_tunableop() {
-    configure_tunableop "$@"
-}
-
-# Historical V620 entry point: same as the generic helper but strict about the
-# hash-matched rank files (no per-user fallback). Callers unchanged.
-configure_v620_tunableop() {
-    TUNEOP_REQUIRE_HASH_ROWS=1 configure_tunableop "$@"
-}
+source "$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/../rdna2_028/tunableop_env.sh"

tools/rdna2_028/tunableop_env.sh

@@ -26,7 +26,6 @@
 # Lookup only by default (PYTORCH_TUNABLEOP_TUNING=0); set TUNING=1 to re-tune
 # shapes missing from the table. Rows are never written to /tmp or the run CWD.
 #
-# configure_mtp_tunableop is an alias so existing callers do not break.
 # configure_v620_tunableop keeps the historical strict behaviour (lookup is
 # disabled unless this build's four rank files exist; no per-user fallback).
 
@@ -145,11 +144,6 @@ configure_tunableop() {
     fi
 }
 
-# Alias kept for the MTP/Flash-Next launchers and other existing callers.
-configure_mtp_tunableop() {
-    configure_tunableop "$@"
-}
-
 # Historical V620 entry point: same as the generic helper but strict about the
 # hash-matched rank files (no per-user fallback). Callers unchanged.
 configure_v620_tunableop() {