From 86d314050f41c30572e58d4ea50cc56f61086cf3 Mon Sep 17 00:00:00 2001 From: MOVZX Date: Sun, 4 Oct 2026 09:23:24 -0400 Subject: [PATCH] [Config] The MERLIN_* kernel toggles default to the shipped value when unset MERLIN_FULL_K, MERLIN_QSA_SORT, MERLIN_TIE_DET, MERLIN_TIE_RECENT and MERLIN_TOPK_SORTED_EMIT defaulted to off in code while every shipped launcher set them to 1, so a launch without them ran a different kernel configuration. vllm/env_override.py now sets each to 1 when unset; 0 still turns one off. tests/test_merlin_toggle_defaults.py covers it. Co-Authored-By: Hinata Signed-off-by: MOVZX --- csrc/libtorch_stable/moe/marlin_moe_wna16/ops.cu | 2 +- csrc/libtorch_stable/persistent_topk.cuh | 4 ++-- csrc/libtorch_stable/topk.cu | 8 ++++---- vllm/env_override.py | 16 ++++++++++++++++ vllm/envs.py | 2 +- vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py | 4 ++-- 6 files changed, 26 insertions(+), 10 deletions(-) diff --git a/csrc/libtorch_stable/moe/marlin_moe_wna16/ops.cu b/csrc/libtorch_stable/moe/marlin_moe_wna16/ops.cu index 0a6f5de67d..12a6a2cf78 100644 --- a/csrc/libtorch_stable/moe/marlin_moe_wna16/ops.cu +++ b/csrc/libtorch_stable/moe/marlin_moe_wna16/ops.cu @@ -49,7 +49,7 @@ namespace { // Resolved ONCE per process, not per launch, so it is a launch constant in // exactly the sense topk.cu's exact_medium_enabled() is: a cudagraph bakes in // the value that was live at capture time and every replay agrees with it. -// Default OFF -- enabling it by default is a separate release decision. +// Code default off; vllm/env_override.py sets 1 (the shipped value) when unset. // // This std::getenv is the SINGLE source of truth. vllm/envs.py carries a raw // passthrough entry for validate_environ and the compile-factor hash only; diff --git a/csrc/libtorch_stable/persistent_topk.cuh b/csrc/libtorch_stable/persistent_topk.cuh index e339e0ab99..68917f0ebe 100644 --- a/csrc/libtorch_stable/persistent_topk.cuh +++ b/csrc/libtorch_stable/persistent_topk.cuh @@ -58,7 +58,7 @@ __device__ __forceinline__ auto convert_to_uint32_v2(float x) -> uint32_t { // both the host (to report it) and the device (to obey it), so the reported // policy and the executed policy cannot drift. // -// `MERLIN_TIE_DET=0` is the shipped default and leaves the buffered paths on +// `MERLIN_TIE_DET=0` (the code default; launchers ship 1) leaves the buffered paths on // their original racy atomicAdd rule, which has no direction; the medium-path exact // restart is deterministic whatever happens, so it keeps the most-recent-wins // direction the medium-path restart shipped and was measured with. @@ -209,7 +209,7 @@ struct PersistentTopKParams { // large tail) re-emit their OWN selected set through sorted_emit_row(). // Selection is never re-derived, so the selected set cannot change. // Resolved once per process on the host, so it is a launch constant a - // cudagraph bakes in. Default OFF. + // cudagraph bakes in. Code default off; vllm/env_override.py sets 1 (the shipped value) when unset. bool sorted_emit; }; diff --git a/csrc/libtorch_stable/topk.cu b/csrc/libtorch_stable/topk.cu index ce0744bf11..51a4761a6b 100644 --- a/csrc/libtorch_stable/topk.cu +++ b/csrc/libtorch_stable/topk.cu @@ -39,7 +39,7 @@ namespace { // // Our own toggle rather than upstream's VLLM_BATCH_INVARIANT, which aborts // engine startup on GDN models (upstream #42960) and so can never reach this -// kernel on Flash-Next. Default OFF -- promotion is a release decision. +// kernel on Flash-Next. Code default off; vllm/env_override.py sets 1 (the shipped value) when unset. // // This std::getenv is the SINGLE source of truth; vllm/envs.py carries a raw // passthrough for validate_environ and the compile-factor hash only. @@ -73,7 +73,7 @@ bool publish_cached(std::atomic& mirror, bool value) { // most-recent-wins is +1.5 %. Both deterministic. Also the direction // medium_exact_restart() already implements, so setting this alongside // MERLIN_TIE_DET makes the medium path agree with itself on overflowing -// rows. Only read when MERLIN_TIE_DET is set. Default OFF. +// rows. Only read when MERLIN_TIE_DET is set; code default OFF; vllm/env_override.py sets 1 (the shipped value) when unset. bool tie_recent_enabled() { static const bool enabled = [] { const char* raw = std::getenv("MERLIN_TIE_RECENT"); @@ -149,8 +149,8 @@ bool exact_decode_enabled() { // (ascending index, padding last as -1), so qsa_indexer.py can drop that sort. // See sorted_emit_row() in persistent_topk.cuh for the contract; the selected // SET is never touched. Resolved once per process like every switch here, so -// it is a launch constant a cudagraph bakes in. Default OFF: the maintainer -// promotes after reference-host qualification. +// it is a launch constant a cudagraph bakes in. Code default OFF; vllm/env_override.py +// sets 1 (the shipped value) when unset. // // This std::getenv is the single source of truth. Python does not parse the // variable to decide anything: qsa_indexer.py asks THIS cache, through diff --git a/vllm/env_override.py b/vllm/env_override.py index bfcdfa7af2..d4080bd6ec 100644 --- a/vllm/env_override.py +++ b/vllm/env_override.py @@ -182,6 +182,22 @@ os.environ.setdefault("TRITON_CACHE_AUTOTUNING", "1") # debug layout (see https://github.com/vllm-project/vllm/issues/41410). os.environ.setdefault("TILELANG_CLEANUP_TEMP_FILES", "1") +# Fork (v2.5, knob-absent boot): the MERLIN_* kernel toggles are read by +# getenv in C++ (FULL_K, TIE_DET, TIE_RECENT, TOPK_SORTED_EMIT) and by +# vllm.envs (QSA_SORT) with OFF as the code default, while every shipped +# launcher sets them to 1. Unset now means the shipped value; an explicit "0" +# still turns a toggle off. Set here because this module runs on `import vllm`, +# before any reader, and worker processes inherit the environment. +_SHIPPED_KERNEL_TOGGLE_DEFAULTS = ( + "MERLIN_FULL_K", + "MERLIN_QSA_SORT", + "MERLIN_TIE_DET", + "MERLIN_TIE_RECENT", + "MERLIN_TOPK_SORTED_EMIT", +) +for _name in _SHIPPED_KERNEL_TOGGLE_DEFAULTS: + os.environ.setdefault(_name, "1") + # =================================================== # torch 2.9 Inductor PythonWrapperCodegen monkeypatch # =================================================== diff --git a/vllm/envs.py b/vllm/envs.py index 4780fcc97f..4f5faeb6e7 100755 --- a/vllm/envs.py +++ b/vllm/envs.py @@ -520,7 +520,7 @@ if TYPE_CHECKING: # --- determinism toggles ----------------------------------------- # The three nondeterminism mechanisms (FULL_K, QSA_SORT, TIE_DET/TIE_RECENT). # Every one of them is - # OFF by default: promotion is the maintainer's decision. + # OFF by default in this file; vllm/env_override.py sets each to 1 (the shipped value) when unset. # # Why these are registered at all: validate_environ() only inspects names # beginning with VLLM_, so a MERLIN_ name never trips the unknown-variable diff --git a/vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py b/vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py index a0dbe9e044..e156e1d144 100644 --- a/vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py +++ b/vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py @@ -45,8 +45,8 @@ _DECODE_BLOCK_N = 64 # on the reference host and is why this version has no Python-visible branch on data. # # Resolved once at import so it is a launch constant: a cudagraph captures the -# op sequence that was live at capture time. Default OFF -- promotion is a release -# decision, which is the maintainer's. +# op sequence that was live at capture time. Code default OFF; vllm/env_override.py +# sets 1 (the shipped value) when unset. # --------------------------------------------------------------------------- _QSA_SORT = envs.MERLIN_QSA_SORT