From 67316774a00e747a11211ef5d4e326bfaf3ddeb8 Mon Sep 17 00:00:00 2001
From: claude-noether <claude@reauktion.de>
Date: Sat, 18 Jul 2026 15:21:55 +0200
Subject: [PATCH] ggml-rocket: make ROCKET_MIN_BATCH runtime-tunable via
 GGML_ROCKET_MIN_BATCH

The offload gate (M < 32 -> CPU) was compile-time only. Reading the env var
(default 32 unchanged) gives a clean NPU on/off toggle: GGML_ROCKET_MIN_BATCH=999999
forces CPU-only for benchmarking without rmmod or a CPU-only build. Used for the
2026-07-18 ampere NPU-vs-CPU gemma prefill benchmark.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EWpfhDgYNA21tETDP9ueBE
---
 ggml/src/ggml-rocket/ggml-rocket.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-rocket/ggml-rocket.cpp b/ggml/src/ggml-rocket/ggml-rocket.cpp
index 20e57f8e3..2a8665c81 100644
--- a/ggml/src/ggml-rocket/ggml-rocket.cpp
+++ b/ggml/src/ggml-rocket/ggml-rocket.cpp
@@ -384,7 +384,8 @@ static bool ggml_backend_rocket_device_supports_op(ggml_backend_dev_t dev, const
             const int64_t K = src1->ne[0]; // == src0->ne[0]
             const int64_t M = op->ne[1];   // batch rows (tokens)
             const char *why = NULL;
-            if (M < ROCKET_MIN_BATCH) why="M<min_batch";
+            static const int min_batch = (getenv("GGML_ROCKET_MIN_BATCH") ? atoi(getenv("GGML_ROCKET_MIN_BATCH")) : ROCKET_MIN_BATCH);
+            if (M < min_batch) why="M<min_batch";
             else if (src1->type != GGML_TYPE_F32) why="src1!=F32";
             else if (!ggml_is_contiguous(src0)) why="src0 noncontig";
             else if (!ggml_is_contiguous(src1)) why="src1 noncontig";