From 8903e6d0655d42be5b4053b7739608cceb676018 Mon Sep 17 00:00:00 2001
From: Donato Capitella <donato.capitella@reversec.com>
Date: Mon, 1 Jun 2026 11:07:55 +0100
Subject: [PATCH] fix(rocm): allow chunked model copy without OOM

When DS4_CUDA_COPY_MODEL_CHUNKED is set, skip cudaHostRegister.
Registering a large memory map prevents posix_madvise(DONTNEED)
from freeing pages during the chunked copy, leading to catastrophic
system RAM exhaustion on APUs with unified memory.
---
 ds4_cuda.cu | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/ds4_cuda.cu b/ds4_cuda.cu
index 8b6241ca3..5db8b15b2 100644
--- a/ds4_cuda.cu
+++ b/ds4_cuda.cu
@@ -1430,6 +1430,11 @@ extern "C" int ds4_gpu_set_model_map(const void *model_map, uint64_t model_size)
         }
     }
 
+    const char *chunked_env = getenv("DS4_CUDA_COPY_MODEL_CHUNKED");
+    if (chunked_env && chunked_env[0]) {
+        return 1;
+    }
+
     cudaError_t err = cudaHostRegister((void *)model_map, (size_t)model_size,
                                        cudaHostRegisterMapped | cudaHostRegisterReadOnly);
     if (err == cudaSuccess) {